diff --git a/it/1120ms/decoder.mlmodelc/analytics/coremldata.bin b/it/1120ms/decoder.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..85afd8d84c262c9e1ba71c6b460a5beb4d6b94c3 --- /dev/null +++ b/it/1120ms/decoder.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4cdca6bf678463f31354072f526088e5bdf5115ae94c04e387bb35b2c7a607d6 +size 243 diff --git a/it/1120ms/decoder.mlmodelc/coremldata.bin b/it/1120ms/decoder.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..8bb44f2c4c669bd785344007b33e6273bd87aa8c --- /dev/null +++ b/it/1120ms/decoder.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3c993f8b96ce22027cd2ed42d99b7e61f93a01197bb17cadada8eb989e946dec +size 433 diff --git a/it/1120ms/decoder.mlmodelc/model.mil b/it/1120ms/decoder.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..e034376bbf9a1dff11539e03ae80e7a65ea4f393 --- /dev/null +++ b/it/1120ms/decoder.mlmodelc/model.mil @@ -0,0 +1,64 @@ +program(1.3) +[buildInfo = dict({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.5.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.3.0"}})] +{ + func main(tensor c_in, tensor h_in, tensor token, tensor token_length) { + int32 y_axis_0 = const()[name = string("y_axis_0"), val = int32(0)]; + int32 y_batch_dims_0 = const()[name = string("y_batch_dims_0"), val = int32(0)]; + bool y_validate_indices_0 = const()[name = string("y_validate_indices_0"), val = bool(false)]; + tensor module_prediction_embed_weight_to_fp16 = const()[name = string("module_prediction_embed_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))]; + string token_to_int16_dtype_0 = const()[name = string("token_to_int16_dtype_0"), val = string("int16")]; + tensor token_to_int16 = cast(dtype = token_to_int16_dtype_0, x = token)[name = string("cast_8")]; + tensor y_cast_fp16_cast_uint16 = gather(axis = y_axis_0, batch_dims = y_batch_dims_0, indices = token_to_int16, validate_indices = y_validate_indices_0, x = module_prediction_embed_weight_to_fp16)[name = string("y_cast_fp16_cast_uint16")]; + tensor input_3_perm_0 = const()[name = string("input_3_perm_0"), val = tensor([1, 0, 2])]; + int32 split_0_num_splits_0 = const()[name = string("split_0_num_splits_0"), val = int32(2)]; + int32 split_0_axis_0 = const()[name = string("split_0_axis_0"), val = int32(0)]; + string h_in_to_fp16_dtype_0 = const()[name = string("h_in_to_fp16_dtype_0"), val = string("fp16")]; + tensor h_in_to_fp16 = cast(dtype = h_in_to_fp16_dtype_0, x = h_in)[name = string("cast_7")]; + tensor split_0_cast_fp16_0, tensor split_0_cast_fp16_1 = split(axis = split_0_axis_0, num_splits = split_0_num_splits_0, x = h_in_to_fp16)[name = string("split_0_cast_fp16")]; + int32 split_1_num_splits_0 = const()[name = string("split_1_num_splits_0"), val = int32(2)]; + int32 split_1_axis_0 = const()[name = string("split_1_axis_0"), val = int32(0)]; + string c_in_to_fp16_dtype_0 = const()[name = string("c_in_to_fp16_dtype_0"), val = string("fp16")]; + tensor c_in_to_fp16 = cast(dtype = c_in_to_fp16_dtype_0, x = c_in)[name = string("cast_6")]; + tensor split_1_cast_fp16_0, tensor split_1_cast_fp16_1 = split(axis = split_1_axis_0, num_splits = split_1_num_splits_0, x = c_in_to_fp16)[name = string("split_1_cast_fp16")]; + tensor input_lstm_layer_0_lstm_h0_squeeze_axes_0 = const()[name = string("input_lstm_layer_0_lstm_h0_squeeze_axes_0"), val = tensor([0])]; + tensor input_lstm_layer_0_lstm_h0_squeeze_cast_fp16 = squeeze(axes = input_lstm_layer_0_lstm_h0_squeeze_axes_0, x = split_0_cast_fp16_0)[name = string("input_lstm_layer_0_lstm_h0_squeeze_cast_fp16")]; + tensor input_lstm_layer_0_lstm_c0_squeeze_axes_0 = const()[name = string("input_lstm_layer_0_lstm_c0_squeeze_axes_0"), val = tensor([0])]; + tensor input_lstm_layer_0_lstm_c0_squeeze_cast_fp16 = squeeze(axes = input_lstm_layer_0_lstm_c0_squeeze_axes_0, x = split_1_cast_fp16_0)[name = string("input_lstm_layer_0_lstm_c0_squeeze_cast_fp16")]; + string input_lstm_layer_0_direction_0 = const()[name = string("input_lstm_layer_0_direction_0"), val = string("forward")]; + bool input_lstm_layer_0_output_sequence_0 = const()[name = string("input_lstm_layer_0_output_sequence_0"), val = bool(true)]; + string input_lstm_layer_0_recurrent_activation_0 = const()[name = string("input_lstm_layer_0_recurrent_activation_0"), val = string("sigmoid")]; + string input_lstm_layer_0_cell_activation_0 = const()[name = string("input_lstm_layer_0_cell_activation_0"), val = string("tanh")]; + string input_lstm_layer_0_activation_0 = const()[name = string("input_lstm_layer_0_activation_0"), val = string("tanh")]; + tensor concat_1_to_fp16 = const()[name = string("concat_1_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1031808)))]; + tensor concat_2_to_fp16 = const()[name = string("concat_2_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4308672)))]; + tensor concat_0_to_fp16 = const()[name = string("concat_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7585536)))]; + tensor input_3_cast_fp16 = transpose(perm = input_3_perm_0, x = y_cast_fp16_cast_uint16)[name = string("transpose_2")]; + tensor input_lstm_layer_0_cast_fp16_0, tensor input_lstm_layer_0_cast_fp16_1, tensor input_lstm_layer_0_cast_fp16_2 = lstm(activation = input_lstm_layer_0_activation_0, bias = concat_0_to_fp16, cell_activation = input_lstm_layer_0_cell_activation_0, direction = input_lstm_layer_0_direction_0, initial_c = input_lstm_layer_0_lstm_c0_squeeze_cast_fp16, initial_h = input_lstm_layer_0_lstm_h0_squeeze_cast_fp16, output_sequence = input_lstm_layer_0_output_sequence_0, recurrent_activation = input_lstm_layer_0_recurrent_activation_0, weight_hh = concat_2_to_fp16, weight_ih = concat_1_to_fp16, x = input_3_cast_fp16)[name = string("input_lstm_layer_0_cast_fp16")]; + tensor input_lstm_h0_squeeze_axes_0 = const()[name = string("input_lstm_h0_squeeze_axes_0"), val = tensor([0])]; + tensor input_lstm_h0_squeeze_cast_fp16 = squeeze(axes = input_lstm_h0_squeeze_axes_0, x = split_0_cast_fp16_1)[name = string("input_lstm_h0_squeeze_cast_fp16")]; + tensor input_lstm_c0_squeeze_axes_0 = const()[name = string("input_lstm_c0_squeeze_axes_0"), val = tensor([0])]; + tensor input_lstm_c0_squeeze_cast_fp16 = squeeze(axes = input_lstm_c0_squeeze_axes_0, x = split_1_cast_fp16_1)[name = string("input_lstm_c0_squeeze_cast_fp16")]; + string input_direction_0 = const()[name = string("input_direction_0"), val = string("forward")]; + bool input_output_sequence_0 = const()[name = string("input_output_sequence_0"), val = bool(true)]; + string input_recurrent_activation_0 = const()[name = string("input_recurrent_activation_0"), val = string("sigmoid")]; + string input_cell_activation_0 = const()[name = string("input_cell_activation_0"), val = string("tanh")]; + string input_activation_0 = const()[name = string("input_activation_0"), val = string("tanh")]; + tensor concat_4_to_fp16 = const()[name = string("concat_4_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7590720)))]; + tensor concat_5_to_fp16 = const()[name = string("concat_5_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10867584)))]; + tensor concat_3_to_fp16 = const()[name = string("concat_3_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14144448)))]; + tensor input_cast_fp16_0, tensor input_cast_fp16_1, tensor input_cast_fp16_2 = lstm(activation = input_activation_0, bias = concat_3_to_fp16, cell_activation = input_cell_activation_0, direction = input_direction_0, initial_c = input_lstm_c0_squeeze_cast_fp16, initial_h = input_lstm_h0_squeeze_cast_fp16, output_sequence = input_output_sequence_0, recurrent_activation = input_recurrent_activation_0, weight_hh = concat_5_to_fp16, weight_ih = concat_4_to_fp16, x = input_lstm_layer_0_cast_fp16_0)[name = string("input_cast_fp16")]; + int32 obj_3_axis_0 = const()[name = string("obj_3_axis_0"), val = int32(0)]; + tensor obj_3_cast_fp16 = stack(axis = obj_3_axis_0, values = (input_lstm_layer_0_cast_fp16_1, input_cast_fp16_1))[name = string("obj_3_cast_fp16")]; + string obj_3_cast_fp16_to_fp32_dtype_0 = const()[name = string("obj_3_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + int32 obj_axis_0 = const()[name = string("obj_axis_0"), val = int32(0)]; + tensor obj_cast_fp16 = stack(axis = obj_axis_0, values = (input_lstm_layer_0_cast_fp16_2, input_cast_fp16_2))[name = string("obj_cast_fp16")]; + string obj_cast_fp16_to_fp32_dtype_0 = const()[name = string("obj_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + tensor transpose_0_perm_0 = const()[name = string("transpose_0_perm_0"), val = tensor([1, 2, 0])]; + string transpose_0_cast_fp16_to_fp32_dtype_0 = const()[name = string("transpose_0_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + tensor transpose_0_cast_fp16 = transpose(perm = transpose_0_perm_0, x = input_cast_fp16_0)[name = string("transpose_1")]; + tensor decoder_out = cast(dtype = transpose_0_cast_fp16_to_fp32_dtype_0, x = transpose_0_cast_fp16)[name = string("cast_3")]; + tensor c_out = cast(dtype = obj_cast_fp16_to_fp32_dtype_0, x = obj_cast_fp16)[name = string("cast_4")]; + tensor h_out = cast(dtype = obj_3_cast_fp16_to_fp32_dtype_0, x = obj_3_cast_fp16)[name = string("cast_5")]; + tensor token_length_tmp = identity(x = token_length)[name = string("token_length_tmp")]; + } -> (decoder_out, h_out, c_out); +} \ No newline at end of file diff --git a/it/1120ms/decoder.mlmodelc/weights/weight.bin b/it/1120ms/decoder.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..2fac0f3a92f8a80de4d92a62819f6dba98aa4983 --- /dev/null +++ b/it/1120ms/decoder.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1360d93c68c3e9c54bda4adaec860753949f3b0dc93bc98f4edc9d6f8dd5595c +size 14149632 diff --git a/it/1120ms/decoder.mlpackage/Data/com.apple.CoreML/model.mlmodel b/it/1120ms/decoder.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..567c038e1e42f382639a9ececec8bb38c22cbde0 --- /dev/null +++ b/it/1120ms/decoder.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:73bb3afa62698bc822b6d32b3731d0bc40521e03737e3139e10a768542fca1fe +size 10359 diff --git a/it/1120ms/decoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/it/1120ms/decoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..2fac0f3a92f8a80de4d92a62819f6dba98aa4983 --- /dev/null +++ b/it/1120ms/decoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1360d93c68c3e9c54bda4adaec860753949f3b0dc93bc98f4edc9d6f8dd5595c +size 14149632 diff --git a/it/1120ms/decoder.mlpackage/Manifest.json b/it/1120ms/decoder.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..8fc74b00a3e9885d54546160ecae1f6da7736d01 --- /dev/null +++ b/it/1120ms/decoder.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "7CBCED8D-FA6A-45B0-BF60-30DB0A653074": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + }, + "AFD197FC-BECC-451A-961C-C0CA05D58065": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + } + }, + "rootModelIdentifier": "AFD197FC-BECC-451A-961C-C0CA05D58065" +} diff --git a/it/1120ms/decoder_joint.mlmodelc/analytics/coremldata.bin b/it/1120ms/decoder_joint.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..6063f90a9756de97c8450a89ef53ef04317ef653 --- /dev/null +++ b/it/1120ms/decoder_joint.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:12f4bcf5114baa2b3a37b8ebeab6c519109bd857e50ec345c458b7a6c4deb20e +size 243 diff --git a/it/1120ms/decoder_joint.mlmodelc/coremldata.bin b/it/1120ms/decoder_joint.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..6f49a6f6923a6d68c50bdf11730215b1db8a2d62 --- /dev/null +++ b/it/1120ms/decoder_joint.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:53754e2eaa7e0f7435220b47b621c5f3d8c5f2da83edd46efa5950fa723ef1d9 +size 454 diff --git a/it/1120ms/decoder_joint.mlmodelc/model.mil b/it/1120ms/decoder_joint.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..9e96b62349b7d1c4bd97fe8db2d7755704041510 --- /dev/null +++ b/it/1120ms/decoder_joint.mlmodelc/model.mil @@ -0,0 +1,83 @@ +program(1.3) +[buildInfo = dict({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.5.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.3.0"}})] +{ + func main(tensor c_in, tensor encoder, tensor h_in, tensor token, tensor token_length) { + int32 y_axis_0 = const()[name = string("y_axis_0"), val = int32(0)]; + int32 y_batch_dims_0 = const()[name = string("y_batch_dims_0"), val = int32(0)]; + bool y_validate_indices_0 = const()[name = string("y_validate_indices_0"), val = bool(false)]; + tensor decoder_module_prediction_embed_weight_to_fp16 = const()[name = string("decoder_module_prediction_embed_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))]; + string token_to_int16_dtype_0 = const()[name = string("token_to_int16_dtype_0"), val = string("int16")]; + tensor token_to_int16 = cast(dtype = token_to_int16_dtype_0, x = token)[name = string("cast_9")]; + tensor y_cast_fp16_cast_uint16 = gather(axis = y_axis_0, batch_dims = y_batch_dims_0, indices = token_to_int16, validate_indices = y_validate_indices_0, x = decoder_module_prediction_embed_weight_to_fp16)[name = string("y_cast_fp16_cast_uint16")]; + tensor input_3_perm_0 = const()[name = string("input_3_perm_0"), val = tensor([1, 0, 2])]; + int32 split_0_num_splits_0 = const()[name = string("split_0_num_splits_0"), val = int32(2)]; + int32 split_0_axis_0 = const()[name = string("split_0_axis_0"), val = int32(0)]; + string h_in_to_fp16_dtype_0 = const()[name = string("h_in_to_fp16_dtype_0"), val = string("fp16")]; + tensor h_in_to_fp16 = cast(dtype = h_in_to_fp16_dtype_0, x = h_in)[name = string("cast_8")]; + tensor split_0_cast_fp16_0, tensor split_0_cast_fp16_1 = split(axis = split_0_axis_0, num_splits = split_0_num_splits_0, x = h_in_to_fp16)[name = string("split_0_cast_fp16")]; + int32 split_1_num_splits_0 = const()[name = string("split_1_num_splits_0"), val = int32(2)]; + int32 split_1_axis_0 = const()[name = string("split_1_axis_0"), val = int32(0)]; + string c_in_to_fp16_dtype_0 = const()[name = string("c_in_to_fp16_dtype_0"), val = string("fp16")]; + tensor c_in_to_fp16 = cast(dtype = c_in_to_fp16_dtype_0, x = c_in)[name = string("cast_7")]; + tensor split_1_cast_fp16_0, tensor split_1_cast_fp16_1 = split(axis = split_1_axis_0, num_splits = split_1_num_splits_0, x = c_in_to_fp16)[name = string("split_1_cast_fp16")]; + tensor input_5_lstm_layer_0_lstm_h0_squeeze_axes_0 = const()[name = string("input_5_lstm_layer_0_lstm_h0_squeeze_axes_0"), val = tensor([0])]; + tensor input_5_lstm_layer_0_lstm_h0_squeeze_cast_fp16 = squeeze(axes = input_5_lstm_layer_0_lstm_h0_squeeze_axes_0, x = split_0_cast_fp16_0)[name = string("input_5_lstm_layer_0_lstm_h0_squeeze_cast_fp16")]; + tensor input_5_lstm_layer_0_lstm_c0_squeeze_axes_0 = const()[name = string("input_5_lstm_layer_0_lstm_c0_squeeze_axes_0"), val = tensor([0])]; + tensor input_5_lstm_layer_0_lstm_c0_squeeze_cast_fp16 = squeeze(axes = input_5_lstm_layer_0_lstm_c0_squeeze_axes_0, x = split_1_cast_fp16_0)[name = string("input_5_lstm_layer_0_lstm_c0_squeeze_cast_fp16")]; + string input_5_lstm_layer_0_direction_0 = const()[name = string("input_5_lstm_layer_0_direction_0"), val = string("forward")]; + bool input_5_lstm_layer_0_output_sequence_0 = const()[name = string("input_5_lstm_layer_0_output_sequence_0"), val = bool(true)]; + string input_5_lstm_layer_0_recurrent_activation_0 = const()[name = string("input_5_lstm_layer_0_recurrent_activation_0"), val = string("sigmoid")]; + string input_5_lstm_layer_0_cell_activation_0 = const()[name = string("input_5_lstm_layer_0_cell_activation_0"), val = string("tanh")]; + string input_5_lstm_layer_0_activation_0 = const()[name = string("input_5_lstm_layer_0_activation_0"), val = string("tanh")]; + tensor concat_1_to_fp16 = const()[name = string("concat_1_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1031808)))]; + tensor concat_2_to_fp16 = const()[name = string("concat_2_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4308672)))]; + tensor concat_0_to_fp16 = const()[name = string("concat_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7585536)))]; + tensor input_3_cast_fp16 = transpose(perm = input_3_perm_0, x = y_cast_fp16_cast_uint16)[name = string("transpose_4")]; + tensor input_5_lstm_layer_0_cast_fp16_0, tensor input_5_lstm_layer_0_cast_fp16_1, tensor input_5_lstm_layer_0_cast_fp16_2 = lstm(activation = input_5_lstm_layer_0_activation_0, bias = concat_0_to_fp16, cell_activation = input_5_lstm_layer_0_cell_activation_0, direction = input_5_lstm_layer_0_direction_0, initial_c = input_5_lstm_layer_0_lstm_c0_squeeze_cast_fp16, initial_h = input_5_lstm_layer_0_lstm_h0_squeeze_cast_fp16, output_sequence = input_5_lstm_layer_0_output_sequence_0, recurrent_activation = input_5_lstm_layer_0_recurrent_activation_0, weight_hh = concat_2_to_fp16, weight_ih = concat_1_to_fp16, x = input_3_cast_fp16)[name = string("input_5_lstm_layer_0_cast_fp16")]; + tensor input_5_lstm_h0_squeeze_axes_0 = const()[name = string("input_5_lstm_h0_squeeze_axes_0"), val = tensor([0])]; + tensor input_5_lstm_h0_squeeze_cast_fp16 = squeeze(axes = input_5_lstm_h0_squeeze_axes_0, x = split_0_cast_fp16_1)[name = string("input_5_lstm_h0_squeeze_cast_fp16")]; + tensor input_5_lstm_c0_squeeze_axes_0 = const()[name = string("input_5_lstm_c0_squeeze_axes_0"), val = tensor([0])]; + tensor input_5_lstm_c0_squeeze_cast_fp16 = squeeze(axes = input_5_lstm_c0_squeeze_axes_0, x = split_1_cast_fp16_1)[name = string("input_5_lstm_c0_squeeze_cast_fp16")]; + string input_5_direction_0 = const()[name = string("input_5_direction_0"), val = string("forward")]; + bool input_5_output_sequence_0 = const()[name = string("input_5_output_sequence_0"), val = bool(true)]; + string input_5_recurrent_activation_0 = const()[name = string("input_5_recurrent_activation_0"), val = string("sigmoid")]; + string input_5_cell_activation_0 = const()[name = string("input_5_cell_activation_0"), val = string("tanh")]; + string input_5_activation_0 = const()[name = string("input_5_activation_0"), val = string("tanh")]; + tensor concat_4_to_fp16 = const()[name = string("concat_4_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7590720)))]; + tensor concat_5_to_fp16 = const()[name = string("concat_5_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10867584)))]; + tensor concat_3_to_fp16 = const()[name = string("concat_3_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14144448)))]; + tensor input_5_cast_fp16_0, tensor input_5_cast_fp16_1, tensor input_5_cast_fp16_2 = lstm(activation = input_5_activation_0, bias = concat_3_to_fp16, cell_activation = input_5_cell_activation_0, direction = input_5_direction_0, initial_c = input_5_lstm_c0_squeeze_cast_fp16, initial_h = input_5_lstm_h0_squeeze_cast_fp16, output_sequence = input_5_output_sequence_0, recurrent_activation = input_5_recurrent_activation_0, weight_hh = concat_5_to_fp16, weight_ih = concat_4_to_fp16, x = input_5_lstm_layer_0_cast_fp16_0)[name = string("input_5_cast_fp16")]; + int32 obj_3_axis_0 = const()[name = string("obj_3_axis_0"), val = int32(0)]; + tensor obj_3_cast_fp16 = stack(axis = obj_3_axis_0, values = (input_5_lstm_layer_0_cast_fp16_1, input_5_cast_fp16_1))[name = string("obj_3_cast_fp16")]; + string obj_3_cast_fp16_to_fp32_dtype_0 = const()[name = string("obj_3_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + int32 obj_axis_0 = const()[name = string("obj_axis_0"), val = int32(0)]; + tensor obj_cast_fp16 = stack(axis = obj_axis_0, values = (input_5_lstm_layer_0_cast_fp16_2, input_5_cast_fp16_2))[name = string("obj_cast_fp16")]; + string obj_cast_fp16_to_fp32_dtype_0 = const()[name = string("obj_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + tensor transpose_1_perm_0 = const()[name = string("transpose_1_perm_0"), val = tensor([1, 0, 2])]; + tensor input_7_perm_0 = const()[name = string("input_7_perm_0"), val = tensor([0, 2, 1])]; + string encoder_to_fp16_dtype_0 = const()[name = string("encoder_to_fp16_dtype_0"), val = string("fp16")]; + tensor joint_module_enc_weight_to_fp16 = const()[name = string("joint_module_enc_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14149632)))]; + tensor joint_module_enc_bias_to_fp16 = const()[name = string("joint_module_enc_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15460416)))]; + tensor encoder_to_fp16 = cast(dtype = encoder_to_fp16_dtype_0, x = encoder)[name = string("cast_4")]; + tensor input_7_cast_fp16 = transpose(perm = input_7_perm_0, x = encoder_to_fp16)[name = string("transpose_2")]; + tensor linear_0_cast_fp16 = linear(bias = joint_module_enc_bias_to_fp16, weight = joint_module_enc_weight_to_fp16, x = input_7_cast_fp16)[name = string("linear_0_cast_fp16")]; + tensor joint_module_pred_weight_to_fp16 = const()[name = string("joint_module_pred_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15461760)))]; + tensor joint_module_pred_bias_to_fp16 = const()[name = string("joint_module_pred_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16281024)))]; + tensor transpose_1_cast_fp16 = transpose(perm = transpose_1_perm_0, x = input_5_cast_fp16_0)[name = string("transpose_3")]; + tensor linear_1_cast_fp16 = linear(bias = joint_module_pred_bias_to_fp16, weight = joint_module_pred_weight_to_fp16, x = transpose_1_cast_fp16)[name = string("linear_1_cast_fp16")]; + tensor var_79_axes_0 = const()[name = string("op_79_axes_0"), val = tensor([2])]; + tensor var_79_cast_fp16 = expand_dims(axes = var_79_axes_0, x = linear_0_cast_fp16)[name = string("op_79_cast_fp16")]; + tensor var_80_axes_0 = const()[name = string("op_80_axes_0"), val = tensor([1])]; + tensor var_80_cast_fp16 = expand_dims(axes = var_80_axes_0, x = linear_1_cast_fp16)[name = string("op_80_cast_fp16")]; + tensor input_11_cast_fp16 = add(x = var_79_cast_fp16, y = var_80_cast_fp16)[name = string("input_11_cast_fp16")]; + tensor input_13_cast_fp16 = relu(x = input_11_cast_fp16)[name = string("input_13_cast_fp16")]; + tensor joint_module_joint_net_2_weight_to_fp16 = const()[name = string("joint_module_joint_net_2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16282368)))]; + tensor joint_module_joint_net_2_bias_to_fp16 = const()[name = string("joint_module_joint_net_2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17314112)))]; + tensor linear_2_cast_fp16 = linear(bias = joint_module_joint_net_2_bias_to_fp16, weight = joint_module_joint_net_2_weight_to_fp16, x = input_13_cast_fp16)[name = string("linear_2_cast_fp16")]; + string linear_2_cast_fp16_to_fp32_dtype_0 = const()[name = string("linear_2_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + tensor logits = cast(dtype = linear_2_cast_fp16_to_fp32_dtype_0, x = linear_2_cast_fp16)[name = string("cast_3")]; + tensor c_out = cast(dtype = obj_cast_fp16_to_fp32_dtype_0, x = obj_cast_fp16)[name = string("cast_5")]; + tensor h_out = cast(dtype = obj_3_cast_fp16_to_fp32_dtype_0, x = obj_3_cast_fp16)[name = string("cast_6")]; + tensor token_length_tmp = identity(x = token_length)[name = string("token_length_tmp")]; + } -> (logits, h_out, c_out); +} \ No newline at end of file diff --git a/it/1120ms/decoder_joint.mlmodelc/weights/weight.bin b/it/1120ms/decoder_joint.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..33ea2ec2f210db3873bae9b152a8fa5b13171f2e --- /dev/null +++ b/it/1120ms/decoder_joint.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e507a69196a04e30adfafc302b6a5f5f527e45c1965c65dd81d63a621cae2064 +size 17315788 diff --git a/it/1120ms/decoder_joint.mlpackage/Data/com.apple.CoreML/model.mlmodel b/it/1120ms/decoder_joint.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..cc3525ebd701acd827b72a5d6a05caf5ddff80e9 --- /dev/null +++ b/it/1120ms/decoder_joint.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5fa6a6d89c1c07ae16f162f5a3b6809b12aafe57a663f5cdb270be3dec7b1427 +size 13745 diff --git a/it/1120ms/decoder_joint.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/it/1120ms/decoder_joint.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..33ea2ec2f210db3873bae9b152a8fa5b13171f2e --- /dev/null +++ b/it/1120ms/decoder_joint.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e507a69196a04e30adfafc302b6a5f5f527e45c1965c65dd81d63a621cae2064 +size 17315788 diff --git a/it/1120ms/decoder_joint.mlpackage/Manifest.json b/it/1120ms/decoder_joint.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..48ab93415f34542ace6e74a66d563506a10f114a --- /dev/null +++ b/it/1120ms/decoder_joint.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "2B19A50C-1D16-4D97-BE3C-D9BCF35884CF": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + }, + "9CA734BC-CFD2-4F39-B068-BE69ABCAAD1F": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + } + }, + "rootModelIdentifier": "2B19A50C-1D16-4D97-BE3C-D9BCF35884CF" +} diff --git a/it/1120ms/decoder_joint_noencproj.mlmodelc/analytics/coremldata.bin b/it/1120ms/decoder_joint_noencproj.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..3573ed8dea8350501693449f8d9e59b9543d1e3b --- /dev/null +++ b/it/1120ms/decoder_joint_noencproj.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:40ec657603479e7dbf8cdb3d6368349eb8b766a52439a26a735d1fadf1b4281d +size 243 diff --git a/it/1120ms/decoder_joint_noencproj.mlmodelc/coremldata.bin b/it/1120ms/decoder_joint_noencproj.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..94d83d6a74b8a602fbbc8c932d43ab754ba51b88 --- /dev/null +++ b/it/1120ms/decoder_joint_noencproj.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c682ba0e028fb7ab6557f8ac1006febc8ec8dd81e4ef8d3a2c05d876e2dbcc8e +size 519 diff --git a/it/1120ms/decoder_joint_noencproj.mlmodelc/model.mil b/it/1120ms/decoder_joint_noencproj.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..5d5cbd528590956dead59657945f5dab997a7da9 --- /dev/null +++ b/it/1120ms/decoder_joint_noencproj.mlmodelc/model.mil @@ -0,0 +1,91 @@ +program(1.3) +[buildInfo = dict({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.10.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor c_in, tensor encoder_proj, tensor h_in, tensor token, tensor token_length) { + int32 y_batch_dims_0 = const()[name = string("y_batch_dims_0"), val = int32(0)]; + bool y_validate_indices_0 = const()[name = string("y_validate_indices_0"), val = bool(false)]; + tensor decoder_module_prediction_embed_weight_to_fp16 = const()[name = string("decoder_module_prediction_embed_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))]; + string token_to_int16_dtype_0 = const()[name = string("token_to_int16_dtype_0"), val = string("int16")]; + string cast_1_dtype_0 = const()[name = string("cast_1_dtype_0"), val = string("int32")]; + int32 greater_equal_0_y_0 = const()[name = string("greater_equal_0_y_0"), val = int32(0)]; + tensor token_to_int16 = cast(dtype = token_to_int16_dtype_0, x = token)[name = string("cast_10")]; + tensor cast_1 = cast(dtype = cast_1_dtype_0, x = token_to_int16)[name = string("cast_9")]; + tensor greater_equal_0 = greater_equal(x = cast_1, y = greater_equal_0_y_0)[name = string("greater_equal_0")]; + int32 slice_by_index_0 = const()[name = string("slice_by_index_0"), val = int32(806)]; + tensor add_2 = add(x = cast_1, y = slice_by_index_0)[name = string("add_2")]; + tensor select_0 = select(a = cast_1, b = add_2, cond = greater_equal_0)[name = string("select_0")]; + int32 y_cast_fp16_cast_uint16_axis_0 = const()[name = string("y_cast_fp16_cast_uint16_axis_0"), val = int32(0)]; + string select_0_to_int16_dtype_0 = const()[name = string("select_0_to_int16_dtype_0"), val = string("int16")]; + tensor select_0_to_int16 = cast(dtype = select_0_to_int16_dtype_0, x = select_0)[name = string("cast_8")]; + tensor y_cast_fp16_cast_uint16_cast_uint16 = gather(axis = y_cast_fp16_cast_uint16_axis_0, batch_dims = y_batch_dims_0, indices = select_0_to_int16, validate_indices = y_validate_indices_0, x = decoder_module_prediction_embed_weight_to_fp16)[name = string("y_cast_fp16_cast_uint16_cast_uint16")]; + tensor input_3_perm_0 = const()[name = string("input_3_perm_0"), val = tensor([1, 0, 2])]; + int32 split_0_num_splits_0 = const()[name = string("split_0_num_splits_0"), val = int32(2)]; + int32 split_0_axis_0 = const()[name = string("split_0_axis_0"), val = int32(0)]; + string h_in_to_fp16_dtype_0 = const()[name = string("h_in_to_fp16_dtype_0"), val = string("fp16")]; + tensor h_in_to_fp16 = cast(dtype = h_in_to_fp16_dtype_0, x = h_in)[name = string("cast_7")]; + tensor split_0_cast_fp16_0, tensor split_0_cast_fp16_1 = split(axis = split_0_axis_0, num_splits = split_0_num_splits_0, x = h_in_to_fp16)[name = string("split_0_cast_fp16")]; + int32 split_1_num_splits_0 = const()[name = string("split_1_num_splits_0"), val = int32(2)]; + int32 split_1_axis_0 = const()[name = string("split_1_axis_0"), val = int32(0)]; + string c_in_to_fp16_dtype_0 = const()[name = string("c_in_to_fp16_dtype_0"), val = string("fp16")]; + tensor c_in_to_fp16 = cast(dtype = c_in_to_fp16_dtype_0, x = c_in)[name = string("cast_6")]; + tensor split_1_cast_fp16_0, tensor split_1_cast_fp16_1 = split(axis = split_1_axis_0, num_splits = split_1_num_splits_0, x = c_in_to_fp16)[name = string("split_1_cast_fp16")]; + tensor input_5_lstm_layer_0_lstm_h0_squeeze_axes_0 = const()[name = string("input_5_lstm_layer_0_lstm_h0_squeeze_axes_0"), val = tensor([0])]; + tensor input_5_lstm_layer_0_lstm_h0_squeeze_cast_fp16 = squeeze(axes = input_5_lstm_layer_0_lstm_h0_squeeze_axes_0, x = split_0_cast_fp16_0)[name = string("input_5_lstm_layer_0_lstm_h0_squeeze_cast_fp16")]; + tensor input_5_lstm_layer_0_lstm_c0_squeeze_axes_0 = const()[name = string("input_5_lstm_layer_0_lstm_c0_squeeze_axes_0"), val = tensor([0])]; + tensor input_5_lstm_layer_0_lstm_c0_squeeze_cast_fp16 = squeeze(axes = input_5_lstm_layer_0_lstm_c0_squeeze_axes_0, x = split_1_cast_fp16_0)[name = string("input_5_lstm_layer_0_lstm_c0_squeeze_cast_fp16")]; + string input_5_lstm_layer_0_direction_0 = const()[name = string("input_5_lstm_layer_0_direction_0"), val = string("forward")]; + bool input_5_lstm_layer_0_output_sequence_0 = const()[name = string("input_5_lstm_layer_0_output_sequence_0"), val = bool(true)]; + string input_5_lstm_layer_0_recurrent_activation_0 = const()[name = string("input_5_lstm_layer_0_recurrent_activation_0"), val = string("sigmoid")]; + string input_5_lstm_layer_0_cell_activation_0 = const()[name = string("input_5_lstm_layer_0_cell_activation_0"), val = string("tanh")]; + string input_5_lstm_layer_0_activation_0 = const()[name = string("input_5_lstm_layer_0_activation_0"), val = string("tanh")]; + tensor concat_1_to_fp16 = const()[name = string("concat_1_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1031808)))]; + tensor concat_2_to_fp16 = const()[name = string("concat_2_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4308672)))]; + tensor concat_0_to_fp16 = const()[name = string("concat_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7585536)))]; + tensor input_3_cast_fp16 = transpose(perm = input_3_perm_0, x = y_cast_fp16_cast_uint16_cast_uint16)[name = string("transpose_3")]; + tensor input_5_lstm_layer_0_cast_fp16_0, tensor input_5_lstm_layer_0_cast_fp16_1, tensor input_5_lstm_layer_0_cast_fp16_2 = lstm(activation = input_5_lstm_layer_0_activation_0, bias = concat_0_to_fp16, cell_activation = input_5_lstm_layer_0_cell_activation_0, direction = input_5_lstm_layer_0_direction_0, initial_c = input_5_lstm_layer_0_lstm_c0_squeeze_cast_fp16, initial_h = input_5_lstm_layer_0_lstm_h0_squeeze_cast_fp16, output_sequence = input_5_lstm_layer_0_output_sequence_0, recurrent_activation = input_5_lstm_layer_0_recurrent_activation_0, weight_hh = concat_2_to_fp16, weight_ih = concat_1_to_fp16, x = input_3_cast_fp16)[name = string("input_5_lstm_layer_0_cast_fp16")]; + tensor input_5_lstm_h0_squeeze_axes_0 = const()[name = string("input_5_lstm_h0_squeeze_axes_0"), val = tensor([0])]; + tensor input_5_lstm_h0_squeeze_cast_fp16 = squeeze(axes = input_5_lstm_h0_squeeze_axes_0, x = split_0_cast_fp16_1)[name = string("input_5_lstm_h0_squeeze_cast_fp16")]; + tensor input_5_lstm_c0_squeeze_axes_0 = const()[name = string("input_5_lstm_c0_squeeze_axes_0"), val = tensor([0])]; + tensor input_5_lstm_c0_squeeze_cast_fp16 = squeeze(axes = input_5_lstm_c0_squeeze_axes_0, x = split_1_cast_fp16_1)[name = string("input_5_lstm_c0_squeeze_cast_fp16")]; + string input_5_direction_0 = const()[name = string("input_5_direction_0"), val = string("forward")]; + bool input_5_output_sequence_0 = const()[name = string("input_5_output_sequence_0"), val = bool(true)]; + string input_5_recurrent_activation_0 = const()[name = string("input_5_recurrent_activation_0"), val = string("sigmoid")]; + string input_5_cell_activation_0 = const()[name = string("input_5_cell_activation_0"), val = string("tanh")]; + string input_5_activation_0 = const()[name = string("input_5_activation_0"), val = string("tanh")]; + tensor concat_4_to_fp16 = const()[name = string("concat_4_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7590720)))]; + tensor concat_5_to_fp16 = const()[name = string("concat_5_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10867584)))]; + tensor concat_3_to_fp16 = const()[name = string("concat_3_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14144448)))]; + tensor input_5_cast_fp16_0, tensor input_5_cast_fp16_1, tensor input_5_cast_fp16_2 = lstm(activation = input_5_activation_0, bias = concat_3_to_fp16, cell_activation = input_5_cell_activation_0, direction = input_5_direction_0, initial_c = input_5_lstm_c0_squeeze_cast_fp16, initial_h = input_5_lstm_h0_squeeze_cast_fp16, output_sequence = input_5_output_sequence_0, recurrent_activation = input_5_recurrent_activation_0, weight_hh = concat_5_to_fp16, weight_ih = concat_4_to_fp16, x = input_5_lstm_layer_0_cast_fp16_0)[name = string("input_5_cast_fp16")]; + int32 obj_3_axis_0 = const()[name = string("obj_3_axis_0"), val = int32(0)]; + tensor obj_3_cast_fp16 = stack(axis = obj_3_axis_0, values = (input_5_lstm_layer_0_cast_fp16_1, input_5_cast_fp16_1))[name = string("obj_3_cast_fp16")]; + string obj_3_cast_fp16_to_fp32_dtype_0 = const()[name = string("obj_3_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + int32 obj_axis_0 = const()[name = string("obj_axis_0"), val = int32(0)]; + tensor obj_cast_fp16 = stack(axis = obj_axis_0, values = (input_5_lstm_layer_0_cast_fp16_2, input_5_cast_fp16_2))[name = string("obj_cast_fp16")]; + string obj_cast_fp16_to_fp32_dtype_0 = const()[name = string("obj_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + tensor transpose_1_perm_0 = const()[name = string("transpose_1_perm_0"), val = tensor([1, 0, 2])]; + tensor joint_module_pred_weight_to_fp16 = const()[name = string("joint_module_pred_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14149632)))]; + tensor joint_module_pred_bias_to_fp16 = const()[name = string("joint_module_pred_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14968896)))]; + tensor transpose_1_cast_fp16 = transpose(perm = transpose_1_perm_0, x = input_5_cast_fp16_0)[name = string("transpose_2")]; + tensor linear_0_cast_fp16 = linear(bias = joint_module_pred_bias_to_fp16, weight = joint_module_pred_weight_to_fp16, x = transpose_1_cast_fp16)[name = string("linear_0_cast_fp16")]; + tensor f_axes_0 = const()[name = string("f_axes_0"), val = tensor([2])]; + string encoder_proj_to_fp16_dtype_0 = const()[name = string("encoder_proj_to_fp16_dtype_0"), val = string("fp16")]; + tensor encoder_proj_to_fp16 = cast(dtype = encoder_proj_to_fp16_dtype_0, x = encoder_proj)[name = string("cast_3")]; + tensor f_cast_fp16 = expand_dims(axes = f_axes_0, x = encoder_proj_to_fp16)[name = string("f_cast_fp16")]; + tensor g_axes_0 = const()[name = string("g_axes_0"), val = tensor([1])]; + tensor g_cast_fp16 = expand_dims(axes = g_axes_0, x = linear_0_cast_fp16)[name = string("g_cast_fp16")]; + tensor input_9_cast_fp16 = add(x = f_cast_fp16, y = g_cast_fp16)[name = string("input_9_cast_fp16")]; + tensor input_11_cast_fp16 = relu(x = input_9_cast_fp16)[name = string("input_11_cast_fp16")]; + tensor joint_module_joint_net_2_weight_to_fp16 = const()[name = string("joint_module_joint_net_2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14970240)))]; + tensor joint_module_joint_net_2_bias_to_fp16 = const()[name = string("joint_module_joint_net_2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16001984)))]; + tensor linear_1_cast_fp16 = linear(bias = joint_module_joint_net_2_bias_to_fp16, weight = joint_module_joint_net_2_weight_to_fp16, x = input_11_cast_fp16)[name = string("linear_1_cast_fp16")]; + int32 var_83 = const()[name = string("op_83"), val = int32(-1)]; + tensor var_85_softmax_cast_fp16 = softmax(axis = var_83, x = linear_1_cast_fp16)[name = string("op_85_softmax_cast_fp16")]; + fp32 var_85_epsilon_0 = const()[name = string("op_85_epsilon_0"), val = fp32(0x1p-149)]; + tensor var_85_cast_fp16 = log(epsilon = var_85_epsilon_0, x = var_85_softmax_cast_fp16)[name = string("op_85_cast_fp16")]; + string var_85_cast_fp16_to_fp32_dtype_0 = const()[name = string("op_85_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + tensor logits = cast(dtype = var_85_cast_fp16_to_fp32_dtype_0, x = var_85_cast_fp16)[name = string("cast_2")]; + tensor c_out = cast(dtype = obj_cast_fp16_to_fp32_dtype_0, x = obj_cast_fp16)[name = string("cast_4")]; + tensor h_out = cast(dtype = obj_3_cast_fp16_to_fp32_dtype_0, x = obj_3_cast_fp16)[name = string("cast_5")]; + tensor token_length_tmp = identity(x = token_length)[name = string("token_length_tmp")]; + } -> (logits, h_out, c_out); +} \ No newline at end of file diff --git a/it/1120ms/decoder_joint_noencproj.mlmodelc/weights/weight.bin b/it/1120ms/decoder_joint_noencproj.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..80af5cec724e7e6b117fcd6a7bc8046c27c26e75 --- /dev/null +++ b/it/1120ms/decoder_joint_noencproj.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5ca6a467ebf44612a8032c6c4ddf323e35a9ffaa15c01822528fb97144ec9439 +size 16003660 diff --git a/it/1120ms/decoder_joint_noencproj.mlpackage/Data/com.apple.CoreML/model.mlmodel b/it/1120ms/decoder_joint_noencproj.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..492f62656a116fa907f67e5963520bcaa37729f8 --- /dev/null +++ b/it/1120ms/decoder_joint_noencproj.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1a46ef6436e8c1e5af6f379658d871f2a40dff3a13c946d4f8df736b19cdc45e +size 14630 diff --git a/it/1120ms/decoder_joint_noencproj.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/it/1120ms/decoder_joint_noencproj.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..80af5cec724e7e6b117fcd6a7bc8046c27c26e75 --- /dev/null +++ b/it/1120ms/decoder_joint_noencproj.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5ca6a467ebf44612a8032c6c4ddf323e35a9ffaa15c01822528fb97144ec9439 +size 16003660 diff --git a/it/1120ms/decoder_joint_noencproj.mlpackage/Manifest.json b/it/1120ms/decoder_joint_noencproj.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..2135f9be1a1ec8555bef5683841b910c941cbd7c --- /dev/null +++ b/it/1120ms/decoder_joint_noencproj.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "E92AE661-EE59-447C-B9BE-52D50BC4FC37": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + }, + "F7138F3F-8633-4C54-A3F4-D0F6E8220E39": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + } + }, + "rootModelIdentifier": "F7138F3F-8633-4C54-A3F4-D0F6E8220E39" +} diff --git a/it/1120ms/encoder.mlmodelc/analytics/coremldata.bin b/it/1120ms/encoder.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..cdc6c00cf855a2b827e4760c156a421699a9f39e --- /dev/null +++ b/it/1120ms/encoder.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0638780288ed25026170632a153dbed5d798954deac843fa58ddaae3190914d4 +size 243 diff --git a/it/1120ms/encoder.mlmodelc/coremldata.bin b/it/1120ms/encoder.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..c384ec54b6b9ec8a6492debd235a2551bf1e9d22 --- /dev/null +++ b/it/1120ms/encoder.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0f76d62779c43fa5272f8977cfc6d3e368abd5cc38841834d05834021633dd2f +size 662 diff --git a/it/1120ms/encoder.mlmodelc/model.mil b/it/1120ms/encoder.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..8243eb3dcf4d99803bf1e78f45f94ceda08ed8ee --- /dev/null +++ b/it/1120ms/encoder.mlmodelc/model.mil @@ -0,0 +1,4439 @@ +program(1.3) +[buildInfo = dict({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}})] +{ + func main(tensor cache_channel, tensor cache_len, tensor cache_time, tensor mel, tensor mel_length, tensor prompt_id) { + tensor value_3_perm_0 = const()[name = string("value_3_perm_0"), val = tensor([1, 0, 2, 3])]; + string cache_channel_to_fp16_dtype_0 = const()[name = string("cache_channel_to_fp16_dtype_0"), val = string("fp16")]; + tensor value_5_perm_0 = const()[name = string("value_5_perm_0"), val = tensor([1, 0, 2, 3])]; + string cache_time_to_fp16_dtype_0 = const()[name = string("cache_time_to_fp16_dtype_0"), val = string("fp16")]; + int32 var_60 = const()[name = string("op_60"), val = int32(-1)]; + int32 var_69 = const()[name = string("op_69"), val = int32(1)]; + tensor x_1_perm_0 = const()[name = string("x_1_perm_0"), val = tensor([0, 2, 1])]; + string mel_to_fp16_dtype_0 = const()[name = string("mel_to_fp16_dtype_0"), val = string("fp16")]; + tensor tensor_1_axes_0 = const()[name = string("tensor_1_axes_0"), val = tensor([1])]; + tensor mel_to_fp16 = cast(dtype = mel_to_fp16_dtype_0, x = mel)[name = string("cast_22")]; + tensor x_1_cast_fp16 = transpose(perm = x_1_perm_0, x = mel_to_fp16)[name = string("transpose_367")]; + tensor tensor_1_cast_fp16 = expand_dims(axes = tensor_1_axes_0, x = x_1_cast_fp16)[name = string("tensor_1_cast_fp16")]; + tensor expand_dims_0 = const()[name = string("expand_dims_0"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))]; + tensor var_138_axes_0 = const()[name = string("op_138_axes_0"), val = tensor([1])]; + tensor var_138 = expand_dims(axes = var_138_axes_0, x = mel_length)[name = string("op_138")]; + tensor time_mask_1 = less(x = expand_dims_0, y = var_138)[name = string("time_mask_1")]; + tensor var_140_axes_0 = const()[name = string("op_140_axes_0"), val = tensor([-1])]; + tensor var_140 = expand_dims(axes = var_140_axes_0, x = time_mask_1)[name = string("op_140")]; + tensor var_142_reps_0 = const()[name = string("op_142_reps_0"), val = tensor([1, 1, 128])]; + tensor var_142 = tile(reps = var_142_reps_0, x = var_140)[name = string("op_142")]; + tensor var_148_axes_0 = const()[name = string("op_148_axes_0"), val = tensor([1])]; + string mask_1_to_fp16_dtype_0 = const()[name = string("mask_1_to_fp16_dtype_0"), val = string("fp16")]; + tensor var_142_to_fp16 = cast(dtype = mask_1_to_fp16_dtype_0, x = var_142)[name = string("cast_21")]; + tensor var_148_cast_fp16 = expand_dims(axes = var_148_axes_0, x = var_142_to_fp16)[name = string("op_148_cast_fp16")]; + tensor input_1_cast_fp16 = mul(x = tensor_1_cast_fp16, y = var_148_cast_fp16)[name = string("input_1_cast_fp16")]; + tensor input_3_pad_0 = const()[name = string("input_3_pad_0"), val = tensor([0, 0, 0, 0, 2, 1, 2, 1])]; + string input_3_mode_0 = const()[name = string("input_3_mode_0"), val = string("constant")]; + fp16 const_9_to_fp16 = const()[name = string("const_9_to_fp16"), val = fp16(0x0p+0)]; + tensor input_3_cast_fp16 = pad(constant_val = const_9_to_fp16, mode = input_3_mode_0, pad = input_3_pad_0, x = input_1_cast_fp16)[name = string("input_3_cast_fp16")]; + string tensor_3_pad_type_0 = const()[name = string("tensor_3_pad_type_0"), val = string("valid")]; + tensor tensor_3_strides_0 = const()[name = string("tensor_3_strides_0"), val = tensor([2, 2])]; + tensor tensor_3_pad_0 = const()[name = string("tensor_3_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor tensor_3_dilations_0 = const()[name = string("tensor_3_dilations_0"), val = tensor([1, 1])]; + int32 tensor_3_groups_0 = const()[name = string("tensor_3_groups_0"), val = int32(1)]; + tensor encoder_pre_encode_conv_0_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(640))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3008))))[name = string("encoder_pre_encode_conv_0_weight_to_fp16_quantized")]; + tensor encoder_pre_encode_conv_0_bias_to_fp16 = const()[name = string("encoder_pre_encode_conv_0_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3584)))]; + tensor tensor_3_cast_fp16 = conv(bias = encoder_pre_encode_conv_0_bias_to_fp16, dilations = tensor_3_dilations_0, groups = tensor_3_groups_0, pad = tensor_3_pad_0, pad_type = tensor_3_pad_type_0, strides = tensor_3_strides_0, weight = encoder_pre_encode_conv_0_weight_to_fp16_quantized, x = input_3_cast_fp16)[name = string("tensor_3_cast_fp16")]; + string current_lengths_1_to_fp16_dtype_0 = const()[name = string("current_lengths_1_to_fp16_dtype_0"), val = string("fp16")]; + fp16 var_161_promoted_to_fp16 = const()[name = string("op_161_promoted_to_fp16"), val = fp16(0x1p+1)]; + tensor mel_length_to_fp16 = cast(dtype = current_lengths_1_to_fp16_dtype_0, x = mel_length)[name = string("cast_20")]; + tensor var_162_cast_fp16 = add(x = mel_length_to_fp16, y = var_161_promoted_to_fp16)[name = string("op_162_cast_fp16")]; + fp16 var_163_promoted_to_fp16 = const()[name = string("op_163_promoted_to_fp16"), val = fp16(0x1p+0)]; + tensor var_164_cast_fp16 = add(x = var_162_cast_fp16, y = var_163_promoted_to_fp16)[name = string("op_164_cast_fp16")]; + fp16 var_165_promoted_to_fp16 = const()[name = string("op_165_promoted_to_fp16"), val = fp16(0x1.8p+1)]; + tensor var_166_cast_fp16 = sub(x = var_164_cast_fp16, y = var_165_promoted_to_fp16)[name = string("op_166_cast_fp16")]; + fp16 var_57_promoted_to_fp16 = const()[name = string("op_57_promoted_to_fp16"), val = fp16(0x1p+1)]; + tensor floor_div_0_cast_fp16 = floor_div(x = var_166_cast_fp16, y = var_57_promoted_to_fp16)[name = string("floor_div_0_cast_fp16")]; + fp16 var_168_promoted_to_fp16 = const()[name = string("op_168_promoted_to_fp16"), val = fp16(0x1p+0)]; + tensor current_lengths_3_cast_fp16 = add(x = floor_div_0_cast_fp16, y = var_168_promoted_to_fp16)[name = string("current_lengths_3_cast_fp16")]; + string lengths_19_dtype_0 = const()[name = string("lengths_19_dtype_0"), val = string("int32")]; + tensor expand_dims_1 = const()[name = string("expand_dims_1"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4160)))]; + tensor var_177_axes_0 = const()[name = string("op_177_axes_0"), val = tensor([1])]; + tensor current_lengths_3_cast_fp16_to_int32 = cast(dtype = lengths_19_dtype_0, x = current_lengths_3_cast_fp16)[name = string("cast_19")]; + tensor var_177 = expand_dims(axes = var_177_axes_0, x = current_lengths_3_cast_fp16_to_int32)[name = string("op_177")]; + tensor time_mask_3 = less(x = expand_dims_1, y = var_177)[name = string("time_mask_3")]; + tensor var_179_axes_0 = const()[name = string("op_179_axes_0"), val = tensor([-1])]; + tensor var_179 = expand_dims(axes = var_179_axes_0, x = time_mask_3)[name = string("op_179")]; + tensor var_181_reps_0 = const()[name = string("op_181_reps_0"), val = tensor([1, 1, 65])]; + tensor var_181 = tile(reps = var_181_reps_0, x = var_179)[name = string("op_181")]; + tensor var_187_axes_0 = const()[name = string("op_187_axes_0"), val = tensor([1])]; + string mask_3_to_fp16_dtype_0 = const()[name = string("mask_3_to_fp16_dtype_0"), val = string("fp16")]; + tensor var_181_to_fp16 = cast(dtype = mask_3_to_fp16_dtype_0, x = var_181)[name = string("cast_18")]; + tensor var_187_cast_fp16 = expand_dims(axes = var_187_axes_0, x = var_181_to_fp16)[name = string("op_187_cast_fp16")]; + tensor expanded_mask_3_reps_0 = const()[name = string("expanded_mask_3_reps_0"), val = tensor([1, 256, 1, 1])]; + tensor expanded_mask_3_cast_fp16 = tile(reps = expanded_mask_3_reps_0, x = var_187_cast_fp16)[name = string("expanded_mask_3_cast_fp16")]; + tensor input_5_cast_fp16 = mul(x = tensor_3_cast_fp16, y = expanded_mask_3_cast_fp16)[name = string("input_5_cast_fp16")]; + tensor tensor_5_cast_fp16 = relu(x = input_5_cast_fp16)[name = string("tensor_5_cast_fp16")]; + tensor input_7_cast_fp16 = mul(x = tensor_5_cast_fp16, y = expanded_mask_3_cast_fp16)[name = string("input_7_cast_fp16")]; + tensor input_9_pad_0 = const()[name = string("input_9_pad_0"), val = tensor([0, 0, 0, 0, 2, 1, 2, 1])]; + string input_9_mode_0 = const()[name = string("input_9_mode_0"), val = string("constant")]; + fp16 const_23_to_fp16 = const()[name = string("const_23_to_fp16"), val = fp16(0x0p+0)]; + tensor input_9_cast_fp16 = pad(constant_val = const_23_to_fp16, mode = input_9_mode_0, pad = input_9_pad_0, x = input_7_cast_fp16)[name = string("input_9_cast_fp16")]; + string tensor_7_pad_type_0 = const()[name = string("tensor_7_pad_type_0"), val = string("valid")]; + tensor tensor_7_strides_0 = const()[name = string("tensor_7_strides_0"), val = tensor([2, 2])]; + int32 tensor_7_groups_0 = const()[name = string("tensor_7_groups_0"), val = int32(256)]; + tensor tensor_7_pad_0 = const()[name = string("tensor_7_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor tensor_7_dilations_0 = const()[name = string("tensor_7_dilations_0"), val = tensor([1, 1])]; + tensor encoder_pre_encode_conv_2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4480))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6848))))[name = string("encoder_pre_encode_conv_2_weight_to_fp16_quantized")]; + tensor encoder_pre_encode_conv_2_bias_to_fp16 = const()[name = string("encoder_pre_encode_conv_2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7424)))]; + tensor tensor_7_cast_fp16 = conv(bias = encoder_pre_encode_conv_2_bias_to_fp16, dilations = tensor_7_dilations_0, groups = tensor_7_groups_0, pad = tensor_7_pad_0, pad_type = tensor_7_pad_type_0, strides = tensor_7_strides_0, weight = encoder_pre_encode_conv_2_weight_to_fp16_quantized, x = input_9_cast_fp16)[name = string("tensor_7_cast_fp16")]; + fp16 var_209_promoted_to_fp16 = const()[name = string("op_209_promoted_to_fp16"), val = fp16(0x1p+1)]; + tensor var_210_cast_fp16 = add(x = current_lengths_3_cast_fp16, y = var_209_promoted_to_fp16)[name = string("op_210_cast_fp16")]; + fp16 var_211_promoted_to_fp16 = const()[name = string("op_211_promoted_to_fp16"), val = fp16(0x1p+0)]; + tensor var_212_cast_fp16 = add(x = var_210_cast_fp16, y = var_211_promoted_to_fp16)[name = string("op_212_cast_fp16")]; + fp16 var_213_promoted_to_fp16 = const()[name = string("op_213_promoted_to_fp16"), val = fp16(0x1.8p+1)]; + tensor var_214_cast_fp16 = sub(x = var_212_cast_fp16, y = var_213_promoted_to_fp16)[name = string("op_214_cast_fp16")]; + fp16 var_57_promoted_1_to_fp16 = const()[name = string("op_57_promoted_1_to_fp16"), val = fp16(0x1p+1)]; + tensor floor_div_1_cast_fp16 = floor_div(x = var_214_cast_fp16, y = var_57_promoted_1_to_fp16)[name = string("floor_div_1_cast_fp16")]; + fp16 var_216_promoted_to_fp16 = const()[name = string("op_216_promoted_to_fp16"), val = fp16(0x1p+0)]; + tensor current_lengths_5_cast_fp16 = add(x = floor_div_1_cast_fp16, y = var_216_promoted_to_fp16)[name = string("current_lengths_5_cast_fp16")]; + string lengths_21_dtype_0 = const()[name = string("lengths_21_dtype_0"), val = string("int32")]; + tensor expand_dims_2 = const()[name = string("expand_dims_2"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8000)))]; + tensor var_225_axes_0 = const()[name = string("op_225_axes_0"), val = tensor([1])]; + tensor current_lengths_5_cast_fp16_to_int32 = cast(dtype = lengths_21_dtype_0, x = current_lengths_5_cast_fp16)[name = string("cast_17")]; + tensor var_225 = expand_dims(axes = var_225_axes_0, x = current_lengths_5_cast_fp16_to_int32)[name = string("op_225")]; + tensor time_mask_5 = less(x = expand_dims_2, y = var_225)[name = string("time_mask_5")]; + tensor var_227_axes_0 = const()[name = string("op_227_axes_0"), val = tensor([-1])]; + tensor var_227 = expand_dims(axes = var_227_axes_0, x = time_mask_5)[name = string("op_227")]; + tensor var_229_reps_0 = const()[name = string("op_229_reps_0"), val = tensor([1, 1, 33])]; + tensor var_229 = tile(reps = var_229_reps_0, x = var_227)[name = string("op_229")]; + tensor var_235_axes_0 = const()[name = string("op_235_axes_0"), val = tensor([1])]; + string mask_5_to_fp16_dtype_0 = const()[name = string("mask_5_to_fp16_dtype_0"), val = string("fp16")]; + tensor var_229_to_fp16 = cast(dtype = mask_5_to_fp16_dtype_0, x = var_229)[name = string("cast_16")]; + tensor var_235_cast_fp16 = expand_dims(axes = var_235_axes_0, x = var_229_to_fp16)[name = string("op_235_cast_fp16")]; + tensor expanded_mask_7_reps_0 = const()[name = string("expanded_mask_7_reps_0"), val = tensor([1, 256, 1, 1])]; + tensor expanded_mask_7_cast_fp16 = tile(reps = expanded_mask_7_reps_0, x = var_235_cast_fp16)[name = string("expanded_mask_7_cast_fp16")]; + tensor input_11_cast_fp16 = mul(x = tensor_7_cast_fp16, y = expanded_mask_7_cast_fp16)[name = string("input_11_cast_fp16")]; + string tensor_9_pad_type_0 = const()[name = string("tensor_9_pad_type_0"), val = string("valid")]; + tensor tensor_9_strides_0 = const()[name = string("tensor_9_strides_0"), val = tensor([1, 1])]; + tensor tensor_9_pad_0 = const()[name = string("tensor_9_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor tensor_9_dilations_0 = const()[name = string("tensor_9_dilations_0"), val = tensor([1, 1])]; + int32 tensor_9_groups_0 = const()[name = string("tensor_9_groups_0"), val = int32(1)]; + tensor encoder_pre_encode_conv_3_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8192))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(73792))))[name = string("encoder_pre_encode_conv_3_weight_to_fp16_quantized")]; + tensor encoder_pre_encode_conv_3_bias_to_fp16 = const()[name = string("encoder_pre_encode_conv_3_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(74368)))]; + tensor tensor_9_cast_fp16 = conv(bias = encoder_pre_encode_conv_3_bias_to_fp16, dilations = tensor_9_dilations_0, groups = tensor_9_groups_0, pad = tensor_9_pad_0, pad_type = tensor_9_pad_type_0, strides = tensor_9_strides_0, weight = encoder_pre_encode_conv_3_weight_to_fp16_quantized, x = input_11_cast_fp16)[name = string("tensor_9_cast_fp16")]; + tensor input_13_cast_fp16 = mul(x = tensor_9_cast_fp16, y = expanded_mask_7_cast_fp16)[name = string("input_13_cast_fp16")]; + tensor tensor_11_cast_fp16 = relu(x = input_13_cast_fp16)[name = string("tensor_11_cast_fp16")]; + tensor input_15_cast_fp16 = mul(x = tensor_11_cast_fp16, y = expanded_mask_7_cast_fp16)[name = string("input_15_cast_fp16")]; + tensor input_17_pad_0 = const()[name = string("input_17_pad_0"), val = tensor([0, 0, 0, 0, 2, 1, 2, 1])]; + string input_17_mode_0 = const()[name = string("input_17_mode_0"), val = string("constant")]; + fp16 const_41_to_fp16 = const()[name = string("const_41_to_fp16"), val = fp16(0x0p+0)]; + tensor input_17_cast_fp16 = pad(constant_val = const_41_to_fp16, mode = input_17_mode_0, pad = input_17_pad_0, x = input_15_cast_fp16)[name = string("input_17_cast_fp16")]; + string tensor_13_pad_type_0 = const()[name = string("tensor_13_pad_type_0"), val = string("valid")]; + tensor tensor_13_strides_0 = const()[name = string("tensor_13_strides_0"), val = tensor([2, 2])]; + int32 tensor_13_groups_0 = const()[name = string("tensor_13_groups_0"), val = int32(256)]; + tensor tensor_13_pad_0 = const()[name = string("tensor_13_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor tensor_13_dilations_0 = const()[name = string("tensor_13_dilations_0"), val = tensor([1, 1])]; + tensor encoder_pre_encode_conv_5_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(74944))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77312))))[name = string("encoder_pre_encode_conv_5_weight_to_fp16_quantized")]; + tensor encoder_pre_encode_conv_5_bias_to_fp16 = const()[name = string("encoder_pre_encode_conv_5_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77888)))]; + tensor tensor_13_cast_fp16 = conv(bias = encoder_pre_encode_conv_5_bias_to_fp16, dilations = tensor_13_dilations_0, groups = tensor_13_groups_0, pad = tensor_13_pad_0, pad_type = tensor_13_pad_type_0, strides = tensor_13_strides_0, weight = encoder_pre_encode_conv_5_weight_to_fp16_quantized, x = input_17_cast_fp16)[name = string("tensor_13_cast_fp16")]; + fp16 var_272_promoted_to_fp16 = const()[name = string("op_272_promoted_to_fp16"), val = fp16(0x1p+1)]; + tensor var_273_cast_fp16 = add(x = current_lengths_5_cast_fp16, y = var_272_promoted_to_fp16)[name = string("op_273_cast_fp16")]; + fp16 var_274_promoted_to_fp16 = const()[name = string("op_274_promoted_to_fp16"), val = fp16(0x1p+0)]; + tensor var_275_cast_fp16 = add(x = var_273_cast_fp16, y = var_274_promoted_to_fp16)[name = string("op_275_cast_fp16")]; + fp16 var_276_promoted_to_fp16 = const()[name = string("op_276_promoted_to_fp16"), val = fp16(0x1.8p+1)]; + tensor var_277_cast_fp16 = sub(x = var_275_cast_fp16, y = var_276_promoted_to_fp16)[name = string("op_277_cast_fp16")]; + fp16 var_57_promoted_2_to_fp16 = const()[name = string("op_57_promoted_2_to_fp16"), val = fp16(0x1p+1)]; + tensor floor_div_2_cast_fp16 = floor_div(x = var_277_cast_fp16, y = var_57_promoted_2_to_fp16)[name = string("floor_div_2_cast_fp16")]; + fp16 var_279_promoted_to_fp16 = const()[name = string("op_279_promoted_to_fp16"), val = fp16(0x1p+0)]; + tensor current_lengths_cast_fp16 = add(x = floor_div_2_cast_fp16, y = var_279_promoted_to_fp16)[name = string("current_lengths_cast_fp16")]; + string lengths_dtype_0 = const()[name = string("lengths_dtype_0"), val = string("int32")]; + tensor expand_dims_3 = const()[name = string("expand_dims_3"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78464)))]; + tensor var_288_axes_0 = const()[name = string("op_288_axes_0"), val = tensor([1])]; + tensor current_lengths_cast_fp16_to_int32 = cast(dtype = lengths_dtype_0, x = current_lengths_cast_fp16)[name = string("cast_15")]; + tensor var_288 = expand_dims(axes = var_288_axes_0, x = current_lengths_cast_fp16_to_int32)[name = string("op_288")]; + tensor time_mask = less(x = expand_dims_3, y = var_288)[name = string("time_mask")]; + tensor var_290_axes_0 = const()[name = string("op_290_axes_0"), val = tensor([-1])]; + tensor var_290 = expand_dims(axes = var_290_axes_0, x = time_mask)[name = string("op_290")]; + tensor var_292_reps_0 = const()[name = string("op_292_reps_0"), val = tensor([1, 1, 17])]; + tensor var_292 = tile(reps = var_292_reps_0, x = var_290)[name = string("op_292")]; + tensor var_298_axes_0 = const()[name = string("op_298_axes_0"), val = tensor([1])]; + string mask_7_to_fp16_dtype_0 = const()[name = string("mask_7_to_fp16_dtype_0"), val = string("fp16")]; + tensor var_292_to_fp16 = cast(dtype = mask_7_to_fp16_dtype_0, x = var_292)[name = string("cast_14")]; + tensor var_298_cast_fp16 = expand_dims(axes = var_298_axes_0, x = var_292_to_fp16)[name = string("op_298_cast_fp16")]; + tensor expanded_mask_13_reps_0 = const()[name = string("expanded_mask_13_reps_0"), val = tensor([1, 256, 1, 1])]; + tensor expanded_mask_13_cast_fp16 = tile(reps = expanded_mask_13_reps_0, x = var_298_cast_fp16)[name = string("expanded_mask_13_cast_fp16")]; + tensor input_19_cast_fp16 = mul(x = tensor_13_cast_fp16, y = expanded_mask_13_cast_fp16)[name = string("input_19_cast_fp16")]; + string tensor_15_pad_type_0 = const()[name = string("tensor_15_pad_type_0"), val = string("valid")]; + tensor tensor_15_strides_0 = const()[name = string("tensor_15_strides_0"), val = tensor([1, 1])]; + tensor tensor_15_pad_0 = const()[name = string("tensor_15_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor tensor_15_dilations_0 = const()[name = string("tensor_15_dilations_0"), val = tensor([1, 1])]; + int32 tensor_15_groups_0 = const()[name = string("tensor_15_groups_0"), val = int32(1)]; + tensor encoder_pre_encode_conv_6_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78592))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144192))))[name = string("encoder_pre_encode_conv_6_weight_to_fp16_quantized")]; + tensor encoder_pre_encode_conv_6_bias_to_fp16 = const()[name = string("encoder_pre_encode_conv_6_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144768)))]; + tensor tensor_15_cast_fp16 = conv(bias = encoder_pre_encode_conv_6_bias_to_fp16, dilations = tensor_15_dilations_0, groups = tensor_15_groups_0, pad = tensor_15_pad_0, pad_type = tensor_15_pad_type_0, strides = tensor_15_strides_0, weight = encoder_pre_encode_conv_6_weight_to_fp16_quantized, x = input_19_cast_fp16)[name = string("tensor_15_cast_fp16")]; + tensor input_21_cast_fp16 = mul(x = tensor_15_cast_fp16, y = expanded_mask_13_cast_fp16)[name = string("input_21_cast_fp16")]; + tensor tensor_cast_fp16 = relu(x = input_21_cast_fp16)[name = string("tensor_cast_fp16")]; + tensor x_3_cast_fp16 = mul(x = tensor_cast_fp16, y = expanded_mask_13_cast_fp16)[name = string("x_3_cast_fp16")]; + tensor var_332_perm_0 = const()[name = string("op_332_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_333 = const()[name = string("op_333"), val = tensor([1, 16, -1])]; + tensor var_332_cast_fp16 = transpose(perm = var_332_perm_0, x = x_3_cast_fp16)[name = string("transpose_366")]; + tensor input_23_cast_fp16 = reshape(shape = var_333, x = var_332_cast_fp16)[name = string("input_23_cast_fp16")]; + tensor encoder_pre_encode_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(145344))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4601856))))[name = string("encoder_pre_encode_out_weight_to_fp16_quantized")]; + tensor encoder_pre_encode_out_bias_to_fp16 = const()[name = string("encoder_pre_encode_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4603968)))]; + tensor linear_0_cast_fp16 = linear(bias = encoder_pre_encode_out_bias_to_fp16, weight = encoder_pre_encode_out_weight_to_fp16_quantized, x = input_23_cast_fp16)[name = string("linear_0_cast_fp16")]; + tensor var_343_begin_0 = const()[name = string("op_343_begin_0"), val = tensor([0, 2, 0])]; + tensor var_343_end_0 = const()[name = string("op_343_end_0"), val = tensor([1, 16, 1024])]; + tensor var_343_end_mask_0 = const()[name = string("op_343_end_mask_0"), val = tensor([true, true, true])]; + tensor var_343_cast_fp16 = slice_by_index(begin = var_343_begin_0, end = var_343_end_0, end_mask = var_343_end_mask_0, x = linear_0_cast_fp16)[name = string("op_343_cast_fp16")]; + int32 var_345 = const()[name = string("op_345"), val = int32(2)]; + tensor var_346 = sub(x = current_lengths_cast_fp16_to_int32, y = var_345)[name = string("op_346")]; + string var_346_promoted_to_fp16_dtype_0 = const()[name = string("op_346_promoted_to_fp16_dtype_0"), val = string("fp16")]; + fp16 var_63_promoted_to_fp16 = const()[name = string("op_63_promoted_to_fp16"), val = fp16(0x0p+0)]; + fp16 const_61_to_fp16 = const()[name = string("const_61_to_fp16"), val = fp16(inf)]; + tensor var_346_to_fp16 = cast(dtype = var_346_promoted_to_fp16_dtype_0, x = var_346)[name = string("cast_13")]; + tensor clip_0_cast_fp16 = clip(alpha = var_63_promoted_to_fp16, beta = const_61_to_fp16, x = var_346_to_fp16)[name = string("clip_0_cast_fp16")]; + tensor max_audio_length_1 = const()[name = string("max_audio_length_1"), val = tensor([14])]; + fp16 var_362_promoted_to_fp16 = const()[name = string("op_362_promoted_to_fp16"), val = fp16(0x1.5p+5)]; + tensor padding_length_cast_fp16 = add(x = clip_0_cast_fp16, y = var_362_promoted_to_fp16)[name = string("padding_length_cast_fp16")]; + int32 const_63 = const()[name = string("const_63"), val = int32(-1)]; + tensor var_364 = mul(x = cache_len, y = const_63)[name = string("op_364")]; + int32 var_365 = const()[name = string("op_365"), val = int32(42)]; + tensor offset = add(x = var_364, y = var_365)[name = string("offset")]; + tensor var_405_axes_0 = const()[name = string("op_405_axes_0"), val = tensor([-1])]; + tensor var_405_cast_fp16 = expand_dims(axes = var_405_axes_0, x = padding_length_cast_fp16)[name = string("op_405_cast_fp16")]; + tensor var_404_promoted_to_fp16 = const()[name = string("op_404_promoted_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4606080)))]; + tensor pad_mask_1_cast_fp16 = less(x = var_404_promoted_to_fp16, y = var_405_cast_fp16)[name = string("pad_mask_1_cast_fp16")]; + tensor expand_dims_5 = const()[name = string("expand_dims_5"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4606272)))]; + tensor var_411_axes_0 = const()[name = string("op_411_axes_0"), val = tensor([-1])]; + tensor var_411 = expand_dims(axes = var_411_axes_0, x = offset)[name = string("op_411")]; + tensor pad_mask_off = greater_equal(x = expand_dims_5, y = var_411)[name = string("pad_mask_off")]; + tensor pad_mask_3 = logical_and(x = pad_mask_off, y = pad_mask_1_cast_fp16)[name = string("pad_mask_3")]; + tensor var_414_axes_0 = const()[name = string("op_414_axes_0"), val = tensor([1])]; + tensor var_414 = expand_dims(axes = var_414_axes_0, x = pad_mask_3)[name = string("op_414")]; + tensor var_415 = const()[name = string("op_415"), val = tensor([1, 56, 1])]; + tensor pad_mask_for_att_mask_1 = tile(reps = var_415, x = var_414)[name = string("pad_mask_for_att_mask_1")]; + tensor var_417_perm_0 = const()[name = string("op_417_perm_0"), val = tensor([0, 2, 1])]; + tensor var_417 = transpose(perm = var_417_perm_0, x = pad_mask_for_att_mask_1)[name = string("transpose_365")]; + tensor pad_mask_for_att_mask = logical_and(x = pad_mask_for_att_mask_1, y = var_417)[name = string("pad_mask_for_att_mask")]; + tensor const_71 = const()[name = string("const_71"), val = tensor([[[true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true]]])]; + tensor att_mask_9 = logical_and(x = pad_mask_for_att_mask, y = const_71)[name = string("att_mask_9")]; + tensor att_mask = logical_not(x = att_mask_9)[name = string("att_mask")]; + tensor pad_mask_5 = logical_not(x = pad_mask_3)[name = string("pad_mask_5")]; + tensor pad_mask_begin_0 = const()[name = string("pad_mask_begin_0"), val = tensor([0, 42])]; + tensor pad_mask_end_0 = const()[name = string("pad_mask_end_0"), val = tensor([1, 56])]; + tensor pad_mask_end_mask_0 = const()[name = string("pad_mask_end_mask_0"), val = tensor([true, true])]; + tensor pad_mask = slice_by_index(begin = pad_mask_begin_0, end = pad_mask_end_0, end_mask = pad_mask_end_mask_0, x = pad_mask_5)[name = string("pad_mask")]; + tensor mask_9_begin_0 = const()[name = string("mask_9_begin_0"), val = tensor([0, 42, 0])]; + tensor mask_9_end_0 = const()[name = string("mask_9_end_0"), val = tensor([1, 56, 56])]; + tensor mask_9_end_mask_0 = const()[name = string("mask_9_end_mask_0"), val = tensor([true, true, true])]; + tensor mask_9 = slice_by_index(begin = mask_9_begin_0, end = mask_9_end_0, end_mask = mask_9_end_mask_0, x = att_mask)[name = string("mask_9")]; + tensor cache_1_begin_0 = const()[name = string("cache_1_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor cache_1_end_0 = const()[name = string("cache_1_end_0"), val = tensor([1, 1, 42, 1024])]; + tensor cache_1_end_mask_0 = const()[name = string("cache_1_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_1_squeeze_mask_0 = const()[name = string("cache_1_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_channel_to_fp16 = cast(dtype = cache_channel_to_fp16_dtype_0, x = cache_channel)[name = string("cast_12")]; + tensor value_3_cast_fp16 = transpose(perm = value_3_perm_0, x = cache_channel_to_fp16)[name = string("transpose_364")]; + tensor cache_1_cast_fp16 = slice_by_index(begin = cache_1_begin_0, end = cache_1_end_0, end_mask = cache_1_end_mask_0, squeeze_mask = cache_1_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_1_cast_fp16")]; + tensor cache_3_begin_0 = const()[name = string("cache_3_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor cache_3_end_0 = const()[name = string("cache_3_end_0"), val = tensor([1, 1, 1024, 8])]; + tensor cache_3_end_mask_0 = const()[name = string("cache_3_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_3_squeeze_mask_0 = const()[name = string("cache_3_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_time_to_fp16 = cast(dtype = cache_time_to_fp16_dtype_0, x = cache_time)[name = string("cast_11")]; + tensor value_5_cast_fp16 = transpose(perm = value_5_perm_0, x = cache_time_to_fp16)[name = string("transpose_363")]; + tensor cache_3_cast_fp16 = slice_by_index(begin = cache_3_begin_0, end = cache_3_end_0, end_mask = cache_3_end_mask_0, squeeze_mask = cache_3_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_3_cast_fp16")]; + tensor input_27_axes_0 = const()[name = string("input_27_axes_0"), val = tensor([-1])]; + tensor encoder_layers_0_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_0_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4606592)))]; + tensor encoder_layers_0_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_0_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4608704)))]; + fp16 var_43_to_fp16 = const()[name = string("op_43_to_fp16"), val = fp16(0x1.5p-17)]; + tensor input_27_cast_fp16 = layer_norm(axes = input_27_axes_0, beta = encoder_layers_0_norm_feed_forward1_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_0_norm_feed_forward1_weight_to_fp16, x = var_343_cast_fp16)[name = string("input_27_cast_fp16")]; + tensor encoder_layers_0_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4610816))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8805184))))[name = string("encoder_layers_0_feed_forward1_linear1_weight_to_fp16_quantized")]; + tensor encoder_layers_0_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_0_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8813440)))]; + tensor linear_1_cast_fp16 = linear(bias = encoder_layers_0_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_0_feed_forward1_linear1_weight_to_fp16_quantized, x = input_27_cast_fp16)[name = string("linear_1_cast_fp16")]; + tensor input_31_cast_fp16 = silu(x = linear_1_cast_fp16)[name = string("input_31_cast_fp16")]; + tensor encoder_layers_0_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8821696))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13016064))))[name = string("encoder_layers_0_feed_forward1_linear2_weight_to_fp16_quantized")]; + tensor encoder_layers_0_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_0_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13018176)))]; + tensor linear_2_cast_fp16 = linear(bias = encoder_layers_0_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_0_feed_forward1_linear2_weight_to_fp16_quantized, x = input_31_cast_fp16)[name = string("linear_2_cast_fp16")]; + fp16 var_456_to_fp16 = const()[name = string("op_456_to_fp16"), val = fp16(0x1p-1)]; + tensor var_457_cast_fp16 = mul(x = linear_2_cast_fp16, y = var_456_to_fp16)[name = string("op_457_cast_fp16")]; + tensor input_37_cast_fp16 = add(x = var_343_cast_fp16, y = var_457_cast_fp16)[name = string("input_37_cast_fp16")]; + tensor key_1_axes_0 = const()[name = string("key_1_axes_0"), val = tensor([-1])]; + tensor encoder_layers_0_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_0_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13020288)))]; + tensor encoder_layers_0_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_0_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13022400)))]; + tensor key_1_cast_fp16 = layer_norm(axes = key_1_axes_0, beta = encoder_layers_0_norm_self_att_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_0_norm_self_att_weight_to_fp16, x = input_37_cast_fp16)[name = string("key_1_cast_fp16")]; + bool input_39_interleave_0 = const()[name = string("input_39_interleave_0"), val = bool(false)]; + tensor input_39_cast_fp16 = concat(axis = var_69, interleave = input_39_interleave_0, values = (cache_1_cast_fp16, key_1_cast_fp16))[name = string("input_39_cast_fp16")]; + tensor var_479_begin_0 = const()[name = string("op_479_begin_0"), val = tensor([0, 14, 0])]; + tensor var_479_end_0 = const()[name = string("op_479_end_0"), val = tensor([1, 42, 1024])]; + tensor var_479_end_mask_0 = const()[name = string("op_479_end_mask_0"), val = tensor([true, true, true])]; + tensor var_479_cast_fp16 = slice_by_index(begin = var_479_begin_0, end = var_479_end_0, end_mask = var_479_end_mask_0, x = cache_1_cast_fp16)[name = string("op_479_cast_fp16")]; + bool var_485_interleave_0 = const()[name = string("op_485_interleave_0"), val = bool(false)]; + tensor var_485_cast_fp16 = concat(axis = var_69, interleave = var_485_interleave_0, values = (var_479_cast_fp16, key_1_cast_fp16))[name = string("op_485_cast_fp16")]; + tensor encoder_layers_0_self_attn_linear_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13024512))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14073152))))[name = string("encoder_layers_0_self_attn_linear_q_weight_to_fp16_quantized")]; + tensor encoder_layers_0_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_0_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14075264)))]; + tensor linear_3_cast_fp16 = linear(bias = encoder_layers_0_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_0_self_attn_linear_q_weight_to_fp16_quantized, x = key_1_cast_fp16)[name = string("linear_3_cast_fp16")]; + tensor var_490 = const()[name = string("op_490"), val = tensor([1, -1, 8, 128])]; + tensor q_1_cast_fp16 = reshape(shape = var_490, x = linear_3_cast_fp16)[name = string("q_1_cast_fp16")]; + tensor encoder_layers_0_self_attn_linear_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14077376))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15126016))))[name = string("encoder_layers_0_self_attn_linear_k_weight_to_fp16_quantized")]; + tensor encoder_layers_0_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_0_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15128128)))]; + tensor linear_4_cast_fp16 = linear(bias = encoder_layers_0_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_0_self_attn_linear_k_weight_to_fp16_quantized, x = input_39_cast_fp16)[name = string("linear_4_cast_fp16")]; + tensor var_495 = const()[name = string("op_495"), val = tensor([1, -1, 8, 128])]; + tensor k_1_cast_fp16 = reshape(shape = var_495, x = linear_4_cast_fp16)[name = string("k_1_cast_fp16")]; + tensor encoder_layers_0_self_attn_linear_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15130240))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16178880))))[name = string("encoder_layers_0_self_attn_linear_v_weight_to_fp16_quantized")]; + tensor encoder_layers_0_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_0_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16180992)))]; + tensor linear_5_cast_fp16 = linear(bias = encoder_layers_0_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_0_self_attn_linear_v_weight_to_fp16_quantized, x = input_39_cast_fp16)[name = string("linear_5_cast_fp16")]; + tensor var_500 = const()[name = string("op_500"), val = tensor([1, -1, 8, 128])]; + tensor v_1_cast_fp16 = reshape(shape = var_500, x = linear_5_cast_fp16)[name = string("v_1_cast_fp16")]; + tensor value_9_perm_0 = const()[name = string("value_9_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_0_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_0_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16183104)))]; + tensor var_513_cast_fp16 = add(x = q_1_cast_fp16, y = encoder_layers_0_self_attn_pos_bias_u_to_fp16)[name = string("op_513_cast_fp16")]; + tensor encoder_layers_0_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_0_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16185216)))]; + tensor var_515_cast_fp16 = add(x = q_1_cast_fp16, y = encoder_layers_0_self_attn_pos_bias_v_to_fp16)[name = string("op_515_cast_fp16")]; + tensor q_with_bias_v_1_perm_0 = const()[name = string("q_with_bias_v_1_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_7_transpose_x_0 = const()[name = string("x_7_transpose_x_0"), val = bool(false)]; + bool x_7_transpose_y_0 = const()[name = string("x_7_transpose_y_0"), val = bool(false)]; + tensor op_517_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16187328))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16301056))))[name = string("op_517_to_fp16_quantized")]; + tensor q_with_bias_v_1_cast_fp16 = transpose(perm = q_with_bias_v_1_perm_0, x = var_515_cast_fp16)[name = string("transpose_362")]; + tensor x_7_cast_fp16 = matmul(transpose_x = x_7_transpose_x_0, transpose_y = x_7_transpose_y_0, x = q_with_bias_v_1_cast_fp16, y = op_517_to_fp16_quantized)[name = string("x_7_cast_fp16")]; + tensor x_9_pad_0 = const()[name = string("x_9_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_9_mode_0 = const()[name = string("x_9_mode_0"), val = string("constant")]; + fp16 const_79_to_fp16 = const()[name = string("const_79_to_fp16"), val = fp16(0x0p+0)]; + tensor x_9_cast_fp16 = pad(constant_val = const_79_to_fp16, mode = x_9_mode_0, pad = x_9_pad_0, x = x_7_cast_fp16)[name = string("x_9_cast_fp16")]; + tensor var_525 = const()[name = string("op_525"), val = tensor([1, 8, -1, 14])]; + tensor x_11_cast_fp16 = reshape(shape = var_525, x = x_9_cast_fp16)[name = string("x_11_cast_fp16")]; + tensor var_529_begin_0 = const()[name = string("op_529_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_529_end_0 = const()[name = string("op_529_end_0"), val = tensor([1, 8, 112, 14])]; + tensor var_529_end_mask_0 = const()[name = string("op_529_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_529_cast_fp16 = slice_by_index(begin = var_529_begin_0, end = var_529_end_0, end_mask = var_529_end_mask_0, x = x_11_cast_fp16)[name = string("op_529_cast_fp16")]; + tensor var_530 = const()[name = string("op_530"), val = tensor([1, 8, 14, 111])]; + tensor matrix_bd_1_cast_fp16 = reshape(shape = var_530, x = var_529_cast_fp16)[name = string("matrix_bd_1_cast_fp16")]; + bool matrix_ac_1_transpose_x_0 = const()[name = string("matrix_ac_1_transpose_x_0"), val = bool(false)]; + bool matrix_ac_1_transpose_y_0 = const()[name = string("matrix_ac_1_transpose_y_0"), val = bool(false)]; + tensor transpose_96_perm_0 = const()[name = string("transpose_96_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_97_perm_0 = const()[name = string("transpose_97_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_97 = transpose(perm = transpose_97_perm_0, x = k_1_cast_fp16)[name = string("transpose_360")]; + tensor transpose_96 = transpose(perm = transpose_96_perm_0, x = var_513_cast_fp16)[name = string("transpose_361")]; + tensor matrix_ac_1_cast_fp16 = matmul(transpose_x = matrix_ac_1_transpose_x_0, transpose_y = matrix_ac_1_transpose_y_0, x = transpose_96, y = transpose_97)[name = string("matrix_ac_1_cast_fp16")]; + tensor matrix_bd_3_begin_0 = const()[name = string("matrix_bd_3_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_3_end_0 = const()[name = string("matrix_bd_3_end_0"), val = tensor([1, 8, 14, 56])]; + tensor matrix_bd_3_end_mask_0 = const()[name = string("matrix_bd_3_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_3_cast_fp16 = slice_by_index(begin = matrix_bd_3_begin_0, end = matrix_bd_3_end_0, end_mask = matrix_bd_3_end_mask_0, x = matrix_bd_1_cast_fp16)[name = string("matrix_bd_3_cast_fp16")]; + tensor var_539_cast_fp16 = add(x = matrix_ac_1_cast_fp16, y = matrix_bd_3_cast_fp16)[name = string("op_539_cast_fp16")]; + fp16 _inversed_scores_1_y_0_to_fp16 = const()[name = string("_inversed_scores_1_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_1_cast_fp16 = mul(x = var_539_cast_fp16, y = _inversed_scores_1_y_0_to_fp16)[name = string("_inversed_scores_1_cast_fp16")]; + tensor mask_11_axes_0 = const()[name = string("mask_11_axes_0"), val = tensor([1])]; + tensor mask_11 = expand_dims(axes = mask_11_axes_0, x = mask_9)[name = string("mask_11")]; + fp16 var_46_to_fp16 = const()[name = string("op_46_to_fp16"), val = fp16(-0x1.388p+13)]; + tensor scores_3_cast_fp16 = select(a = var_46_to_fp16, b = _inversed_scores_1_cast_fp16, cond = mask_11)[name = string("scores_3_cast_fp16")]; + tensor var_545_cast_fp16 = softmax(axis = var_60, x = scores_3_cast_fp16)[name = string("op_545_cast_fp16")]; + fp16 var_45_to_fp16 = const()[name = string("op_45_to_fp16"), val = fp16(0x0p+0)]; + tensor input_41_cast_fp16 = select(a = var_45_to_fp16, b = var_545_cast_fp16, cond = mask_11)[name = string("input_41_cast_fp16")]; + bool x_13_transpose_x_0 = const()[name = string("x_13_transpose_x_0"), val = bool(false)]; + bool x_13_transpose_y_0 = const()[name = string("x_13_transpose_y_0"), val = bool(false)]; + tensor value_9_cast_fp16 = transpose(perm = value_9_perm_0, x = v_1_cast_fp16)[name = string("transpose_359")]; + tensor x_13_cast_fp16 = matmul(transpose_x = x_13_transpose_x_0, transpose_y = x_13_transpose_y_0, x = input_41_cast_fp16, y = value_9_cast_fp16)[name = string("x_13_cast_fp16")]; + tensor var_549_perm_0 = const()[name = string("op_549_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_550 = const()[name = string("op_550"), val = tensor([1, -1, 1024])]; + tensor var_549_cast_fp16 = transpose(perm = var_549_perm_0, x = x_13_cast_fp16)[name = string("transpose_358")]; + tensor input_43_cast_fp16 = reshape(shape = var_550, x = var_549_cast_fp16)[name = string("input_43_cast_fp16")]; + tensor encoder_layers_0_self_attn_linear_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16301376))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17350016))))[name = string("encoder_layers_0_self_attn_linear_out_weight_to_fp16_quantized")]; + tensor encoder_layers_0_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_0_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17352128)))]; + tensor linear_7_cast_fp16 = linear(bias = encoder_layers_0_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_0_self_attn_linear_out_weight_to_fp16_quantized, x = input_43_cast_fp16)[name = string("linear_7_cast_fp16")]; + tensor input_47_cast_fp16 = add(x = input_37_cast_fp16, y = linear_7_cast_fp16)[name = string("input_47_cast_fp16")]; + tensor x_17_axes_0 = const()[name = string("x_17_axes_0"), val = tensor([-1])]; + tensor encoder_layers_0_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_0_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17354240)))]; + tensor encoder_layers_0_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_0_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17356352)))]; + tensor x_17_cast_fp16 = layer_norm(axes = x_17_axes_0, beta = encoder_layers_0_norm_conv_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_0_norm_conv_weight_to_fp16, x = input_47_cast_fp16)[name = string("x_17_cast_fp16")]; + tensor input_49_perm_0 = const()[name = string("input_49_perm_0"), val = tensor([0, 2, 1])]; + string input_51_pad_type_0 = const()[name = string("input_51_pad_type_0"), val = string("valid")]; + tensor input_51_strides_0 = const()[name = string("input_51_strides_0"), val = tensor([1])]; + tensor input_51_pad_0 = const()[name = string("input_51_pad_0"), val = tensor([0, 0])]; + tensor input_51_dilations_0 = const()[name = string("input_51_dilations_0"), val = tensor([1])]; + int32 input_51_groups_0 = const()[name = string("input_51_groups_0"), val = int32(1)]; + tensor encoder_layers_0_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17358464))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19455680))))[name = string("encoder_layers_0_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_49_cast_fp16 = transpose(perm = input_49_perm_0, x = x_17_cast_fp16)[name = string("transpose_357")]; + tensor input_51_cast_fp16 = conv(dilations = input_51_dilations_0, groups = input_51_groups_0, pad = input_51_pad_0, pad_type = input_51_pad_type_0, strides = input_51_strides_0, weight = encoder_layers_0_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_49_cast_fp16)[name = string("input_51_cast_fp16")]; + int32 x_19_split_num_splits_0 = const()[name = string("x_19_split_num_splits_0"), val = int32(2)]; + int32 x_19_split_axis_0 = const()[name = string("x_19_split_axis_0"), val = int32(1)]; + tensor x_19_split_cast_fp16_0, tensor x_19_split_cast_fp16_1 = split(axis = x_19_split_axis_0, num_splits = x_19_split_num_splits_0, x = input_51_cast_fp16)[name = string("x_19_split_cast_fp16")]; + tensor x_19_split_1_sigmoid_cast_fp16 = sigmoid(x = x_19_split_cast_fp16_1)[name = string("x_19_split_1_sigmoid_cast_fp16")]; + tensor x_19_cast_fp16 = mul(x = x_19_split_cast_fp16_0, y = x_19_split_1_sigmoid_cast_fp16)[name = string("x_19_cast_fp16")]; + tensor var_576_axes_0 = const()[name = string("op_576_axes_0"), val = tensor([1])]; + tensor var_576 = expand_dims(axes = var_576_axes_0, x = pad_mask)[name = string("op_576")]; + tensor input_53_cast_fp16 = select(a = var_45_to_fp16, b = x_19_cast_fp16, cond = var_576)[name = string("input_53_cast_fp16")]; + bool new_x_3_interleave_0 = const()[name = string("new_x_3_interleave_0"), val = bool(false)]; + tensor new_x_3_cast_fp16 = concat(axis = var_60, interleave = new_x_3_interleave_0, values = (cache_3_cast_fp16, input_53_cast_fp16))[name = string("new_x_3_cast_fp16")]; + tensor var_589_begin_0 = const()[name = string("op_589_begin_0"), val = tensor([0, 0, 14])]; + tensor var_589_end_0 = const()[name = string("op_589_end_0"), val = tensor([1, 1024, 22])]; + tensor var_589_end_mask_0 = const()[name = string("op_589_end_mask_0"), val = tensor([true, true, true])]; + tensor var_589_cast_fp16 = slice_by_index(begin = var_589_begin_0, end = var_589_end_0, end_mask = var_589_end_mask_0, x = new_x_3_cast_fp16)[name = string("op_589_cast_fp16")]; + string x_21_pad_type_0 = const()[name = string("x_21_pad_type_0"), val = string("valid")]; + int32 x_21_groups_0 = const()[name = string("x_21_groups_0"), val = int32(1024)]; + tensor x_21_strides_0 = const()[name = string("x_21_strides_0"), val = tensor([1])]; + tensor x_21_pad_0 = const()[name = string("x_21_pad_0"), val = tensor([0, 0])]; + tensor x_21_dilations_0 = const()[name = string("x_21_dilations_0"), val = tensor([1])]; + tensor encoder_layers_0_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19459840))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19469120))))[name = string("encoder_layers_0_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_21_cast_fp16 = conv(dilations = x_21_dilations_0, groups = x_21_groups_0, pad = x_21_pad_0, pad_type = x_21_pad_type_0, strides = x_21_strides_0, weight = encoder_layers_0_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_3_cast_fp16)[name = string("x_21_cast_fp16")]; + tensor input_55_perm_0 = const()[name = string("input_55_perm_0"), val = tensor([0, 2, 1])]; + tensor x_23_axes_0 = const()[name = string("x_23_axes_0"), val = tensor([-1])]; + tensor encoder_layers_0_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_0_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19471232)))]; + tensor encoder_layers_0_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_0_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19473344)))]; + tensor input_55_cast_fp16 = transpose(perm = input_55_perm_0, x = x_21_cast_fp16)[name = string("transpose_356")]; + tensor x_23_cast_fp16 = layer_norm(axes = x_23_axes_0, beta = encoder_layers_0_conv_batch_norm_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_0_conv_batch_norm_weight_to_fp16, x = input_55_cast_fp16)[name = string("x_23_cast_fp16")]; + tensor input_57_perm_0 = const()[name = string("input_57_perm_0"), val = tensor([0, 2, 1])]; + tensor input_57_cast_fp16 = transpose(perm = input_57_perm_0, x = x_23_cast_fp16)[name = string("transpose_355")]; + tensor input_59_cast_fp16 = silu(x = input_57_cast_fp16)[name = string("input_59_cast_fp16")]; + string x_25_pad_type_0 = const()[name = string("x_25_pad_type_0"), val = string("valid")]; + tensor x_25_strides_0 = const()[name = string("x_25_strides_0"), val = tensor([1])]; + tensor x_25_pad_0 = const()[name = string("x_25_pad_0"), val = tensor([0, 0])]; + tensor x_25_dilations_0 = const()[name = string("x_25_dilations_0"), val = tensor([1])]; + int32 x_25_groups_0 = const()[name = string("x_25_groups_0"), val = int32(1)]; + tensor encoder_layers_0_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19475456))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20524096))))[name = string("encoder_layers_0_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_25_cast_fp16 = conv(dilations = x_25_dilations_0, groups = x_25_groups_0, pad = x_25_pad_0, pad_type = x_25_pad_type_0, strides = x_25_strides_0, weight = encoder_layers_0_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_59_cast_fp16)[name = string("x_25_cast_fp16")]; + tensor input_61_perm_0 = const()[name = string("input_61_perm_0"), val = tensor([0, 2, 1])]; + tensor input_61_cast_fp16 = transpose(perm = input_61_perm_0, x = x_25_cast_fp16)[name = string("transpose_354")]; + tensor input_63_cast_fp16 = add(x = input_47_cast_fp16, y = input_61_cast_fp16)[name = string("input_63_cast_fp16")]; + tensor input_65_axes_0 = const()[name = string("input_65_axes_0"), val = tensor([-1])]; + tensor encoder_layers_0_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_0_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20526208)))]; + tensor encoder_layers_0_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_0_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20528320)))]; + tensor input_65_cast_fp16 = layer_norm(axes = input_65_axes_0, beta = encoder_layers_0_norm_feed_forward2_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_0_norm_feed_forward2_weight_to_fp16, x = input_63_cast_fp16)[name = string("input_65_cast_fp16")]; + tensor encoder_layers_0_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20530432))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24724800))))[name = string("encoder_layers_0_feed_forward2_linear1_weight_to_fp16_quantized")]; + tensor encoder_layers_0_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_0_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24733056)))]; + tensor linear_8_cast_fp16 = linear(bias = encoder_layers_0_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_0_feed_forward2_linear1_weight_to_fp16_quantized, x = input_65_cast_fp16)[name = string("linear_8_cast_fp16")]; + tensor input_69_cast_fp16 = silu(x = linear_8_cast_fp16)[name = string("input_69_cast_fp16")]; + tensor encoder_layers_0_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24741312))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28935680))))[name = string("encoder_layers_0_feed_forward2_linear2_weight_to_fp16_quantized")]; + tensor encoder_layers_0_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_0_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28937792)))]; + tensor linear_9_cast_fp16 = linear(bias = encoder_layers_0_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_0_feed_forward2_linear2_weight_to_fp16_quantized, x = input_69_cast_fp16)[name = string("linear_9_cast_fp16")]; + fp16 var_632_to_fp16 = const()[name = string("op_632_to_fp16"), val = fp16(0x1p-1)]; + tensor var_633_cast_fp16 = mul(x = linear_9_cast_fp16, y = var_632_to_fp16)[name = string("op_633_cast_fp16")]; + tensor input_75_cast_fp16 = add(x = input_63_cast_fp16, y = var_633_cast_fp16)[name = string("input_75_cast_fp16")]; + tensor input_77_axes_0 = const()[name = string("input_77_axes_0"), val = tensor([-1])]; + tensor encoder_layers_0_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_0_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28939904)))]; + tensor encoder_layers_0_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_0_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28942016)))]; + tensor input_77_cast_fp16 = layer_norm(axes = input_77_axes_0, beta = encoder_layers_0_norm_out_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_0_norm_out_weight_to_fp16, x = input_75_cast_fp16)[name = string("input_77_cast_fp16")]; + tensor cache_5_begin_0 = const()[name = string("cache_5_begin_0"), val = tensor([1, 0, 0, 0])]; + tensor cache_5_end_0 = const()[name = string("cache_5_end_0"), val = tensor([2, 1, 42, 1024])]; + tensor cache_5_end_mask_0 = const()[name = string("cache_5_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_5_squeeze_mask_0 = const()[name = string("cache_5_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_5_cast_fp16 = slice_by_index(begin = cache_5_begin_0, end = cache_5_end_0, end_mask = cache_5_end_mask_0, squeeze_mask = cache_5_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_5_cast_fp16")]; + tensor cache_7_begin_0 = const()[name = string("cache_7_begin_0"), val = tensor([1, 0, 0, 0])]; + tensor cache_7_end_0 = const()[name = string("cache_7_end_0"), val = tensor([2, 1, 1024, 8])]; + tensor cache_7_end_mask_0 = const()[name = string("cache_7_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_7_squeeze_mask_0 = const()[name = string("cache_7_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_7_cast_fp16 = slice_by_index(begin = cache_7_begin_0, end = cache_7_end_0, end_mask = cache_7_end_mask_0, squeeze_mask = cache_7_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_7_cast_fp16")]; + tensor input_79_axes_0 = const()[name = string("input_79_axes_0"), val = tensor([-1])]; + tensor encoder_layers_1_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_1_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28944128)))]; + tensor encoder_layers_1_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_1_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28946240)))]; + tensor input_79_cast_fp16 = layer_norm(axes = input_79_axes_0, beta = encoder_layers_1_norm_feed_forward1_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_1_norm_feed_forward1_weight_to_fp16, x = input_77_cast_fp16)[name = string("input_79_cast_fp16")]; + tensor encoder_layers_1_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28948352))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33142720))))[name = string("encoder_layers_1_feed_forward1_linear1_weight_to_fp16_quantized")]; + tensor encoder_layers_1_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_1_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33150976)))]; + tensor linear_10_cast_fp16 = linear(bias = encoder_layers_1_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_1_feed_forward1_linear1_weight_to_fp16_quantized, x = input_79_cast_fp16)[name = string("linear_10_cast_fp16")]; + tensor input_83_cast_fp16 = silu(x = linear_10_cast_fp16)[name = string("input_83_cast_fp16")]; + tensor encoder_layers_1_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33159232))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37353600))))[name = string("encoder_layers_1_feed_forward1_linear2_weight_to_fp16_quantized")]; + tensor encoder_layers_1_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_1_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37355712)))]; + tensor linear_11_cast_fp16 = linear(bias = encoder_layers_1_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_1_feed_forward1_linear2_weight_to_fp16_quantized, x = input_83_cast_fp16)[name = string("linear_11_cast_fp16")]; + fp16 var_669_to_fp16 = const()[name = string("op_669_to_fp16"), val = fp16(0x1p-1)]; + tensor var_670_cast_fp16 = mul(x = linear_11_cast_fp16, y = var_669_to_fp16)[name = string("op_670_cast_fp16")]; + tensor input_89_cast_fp16 = add(x = input_77_cast_fp16, y = var_670_cast_fp16)[name = string("input_89_cast_fp16")]; + tensor key_3_axes_0 = const()[name = string("key_3_axes_0"), val = tensor([-1])]; + tensor encoder_layers_1_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_1_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37357824)))]; + tensor encoder_layers_1_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_1_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37359936)))]; + tensor key_3_cast_fp16 = layer_norm(axes = key_3_axes_0, beta = encoder_layers_1_norm_self_att_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_1_norm_self_att_weight_to_fp16, x = input_89_cast_fp16)[name = string("key_3_cast_fp16")]; + bool input_91_interleave_0 = const()[name = string("input_91_interleave_0"), val = bool(false)]; + tensor input_91_cast_fp16 = concat(axis = var_69, interleave = input_91_interleave_0, values = (cache_5_cast_fp16, key_3_cast_fp16))[name = string("input_91_cast_fp16")]; + tensor var_692_begin_0 = const()[name = string("op_692_begin_0"), val = tensor([0, 14, 0])]; + tensor var_692_end_0 = const()[name = string("op_692_end_0"), val = tensor([1, 42, 1024])]; + tensor var_692_end_mask_0 = const()[name = string("op_692_end_mask_0"), val = tensor([true, true, true])]; + tensor var_692_cast_fp16 = slice_by_index(begin = var_692_begin_0, end = var_692_end_0, end_mask = var_692_end_mask_0, x = cache_5_cast_fp16)[name = string("op_692_cast_fp16")]; + bool var_698_interleave_0 = const()[name = string("op_698_interleave_0"), val = bool(false)]; + tensor var_698_cast_fp16 = concat(axis = var_69, interleave = var_698_interleave_0, values = (var_692_cast_fp16, key_3_cast_fp16))[name = string("op_698_cast_fp16")]; + tensor encoder_layers_1_self_attn_linear_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37362048))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38410688))))[name = string("encoder_layers_1_self_attn_linear_q_weight_to_fp16_quantized")]; + tensor encoder_layers_1_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_1_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38412800)))]; + tensor linear_12_cast_fp16 = linear(bias = encoder_layers_1_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_1_self_attn_linear_q_weight_to_fp16_quantized, x = key_3_cast_fp16)[name = string("linear_12_cast_fp16")]; + tensor var_703 = const()[name = string("op_703"), val = tensor([1, -1, 8, 128])]; + tensor q_7_cast_fp16 = reshape(shape = var_703, x = linear_12_cast_fp16)[name = string("q_7_cast_fp16")]; + tensor encoder_layers_1_self_attn_linear_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38414912))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39463552))))[name = string("encoder_layers_1_self_attn_linear_k_weight_to_fp16_quantized")]; + tensor encoder_layers_1_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_1_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39465664)))]; + tensor linear_13_cast_fp16 = linear(bias = encoder_layers_1_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_1_self_attn_linear_k_weight_to_fp16_quantized, x = input_91_cast_fp16)[name = string("linear_13_cast_fp16")]; + tensor var_708 = const()[name = string("op_708"), val = tensor([1, -1, 8, 128])]; + tensor k_5_cast_fp16 = reshape(shape = var_708, x = linear_13_cast_fp16)[name = string("k_5_cast_fp16")]; + tensor encoder_layers_1_self_attn_linear_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39467776))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40516416))))[name = string("encoder_layers_1_self_attn_linear_v_weight_to_fp16_quantized")]; + tensor encoder_layers_1_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_1_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40518528)))]; + tensor linear_14_cast_fp16 = linear(bias = encoder_layers_1_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_1_self_attn_linear_v_weight_to_fp16_quantized, x = input_91_cast_fp16)[name = string("linear_14_cast_fp16")]; + tensor var_713 = const()[name = string("op_713"), val = tensor([1, -1, 8, 128])]; + tensor v_3_cast_fp16 = reshape(shape = var_713, x = linear_14_cast_fp16)[name = string("v_3_cast_fp16")]; + tensor value_11_perm_0 = const()[name = string("value_11_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_1_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_1_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40520640)))]; + tensor var_726_cast_fp16 = add(x = q_7_cast_fp16, y = encoder_layers_1_self_attn_pos_bias_u_to_fp16)[name = string("op_726_cast_fp16")]; + tensor encoder_layers_1_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_1_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40522752)))]; + tensor var_728_cast_fp16 = add(x = q_7_cast_fp16, y = encoder_layers_1_self_attn_pos_bias_v_to_fp16)[name = string("op_728_cast_fp16")]; + tensor q_with_bias_v_3_perm_0 = const()[name = string("q_with_bias_v_3_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_33_transpose_x_0 = const()[name = string("x_33_transpose_x_0"), val = bool(false)]; + bool x_33_transpose_y_0 = const()[name = string("x_33_transpose_y_0"), val = bool(false)]; + tensor op_730_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40524864))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40638592))))[name = string("op_730_to_fp16_quantized")]; + tensor q_with_bias_v_3_cast_fp16 = transpose(perm = q_with_bias_v_3_perm_0, x = var_728_cast_fp16)[name = string("transpose_353")]; + tensor x_33_cast_fp16 = matmul(transpose_x = x_33_transpose_x_0, transpose_y = x_33_transpose_y_0, x = q_with_bias_v_3_cast_fp16, y = op_730_to_fp16_quantized)[name = string("x_33_cast_fp16")]; + tensor x_35_pad_0 = const()[name = string("x_35_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_35_mode_0 = const()[name = string("x_35_mode_0"), val = string("constant")]; + fp16 const_92_to_fp16 = const()[name = string("const_92_to_fp16"), val = fp16(0x0p+0)]; + tensor x_35_cast_fp16 = pad(constant_val = const_92_to_fp16, mode = x_35_mode_0, pad = x_35_pad_0, x = x_33_cast_fp16)[name = string("x_35_cast_fp16")]; + tensor var_738 = const()[name = string("op_738"), val = tensor([1, 8, -1, 14])]; + tensor x_37_cast_fp16 = reshape(shape = var_738, x = x_35_cast_fp16)[name = string("x_37_cast_fp16")]; + tensor var_742_begin_0 = const()[name = string("op_742_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_742_end_0 = const()[name = string("op_742_end_0"), val = tensor([1, 8, 112, 14])]; + tensor var_742_end_mask_0 = const()[name = string("op_742_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_742_cast_fp16 = slice_by_index(begin = var_742_begin_0, end = var_742_end_0, end_mask = var_742_end_mask_0, x = x_37_cast_fp16)[name = string("op_742_cast_fp16")]; + tensor var_743 = const()[name = string("op_743"), val = tensor([1, 8, 14, 111])]; + tensor matrix_bd_5_cast_fp16 = reshape(shape = var_743, x = var_742_cast_fp16)[name = string("matrix_bd_5_cast_fp16")]; + bool matrix_ac_3_transpose_x_0 = const()[name = string("matrix_ac_3_transpose_x_0"), val = bool(false)]; + bool matrix_ac_3_transpose_y_0 = const()[name = string("matrix_ac_3_transpose_y_0"), val = bool(false)]; + tensor transpose_98_perm_0 = const()[name = string("transpose_98_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_99_perm_0 = const()[name = string("transpose_99_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_99 = transpose(perm = transpose_99_perm_0, x = k_5_cast_fp16)[name = string("transpose_351")]; + tensor transpose_98 = transpose(perm = transpose_98_perm_0, x = var_726_cast_fp16)[name = string("transpose_352")]; + tensor matrix_ac_3_cast_fp16 = matmul(transpose_x = matrix_ac_3_transpose_x_0, transpose_y = matrix_ac_3_transpose_y_0, x = transpose_98, y = transpose_99)[name = string("matrix_ac_3_cast_fp16")]; + tensor matrix_bd_7_begin_0 = const()[name = string("matrix_bd_7_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_7_end_0 = const()[name = string("matrix_bd_7_end_0"), val = tensor([1, 8, 14, 56])]; + tensor matrix_bd_7_end_mask_0 = const()[name = string("matrix_bd_7_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_7_cast_fp16 = slice_by_index(begin = matrix_bd_7_begin_0, end = matrix_bd_7_end_0, end_mask = matrix_bd_7_end_mask_0, x = matrix_bd_5_cast_fp16)[name = string("matrix_bd_7_cast_fp16")]; + tensor var_752_cast_fp16 = add(x = matrix_ac_3_cast_fp16, y = matrix_bd_7_cast_fp16)[name = string("op_752_cast_fp16")]; + fp16 _inversed_scores_5_y_0_to_fp16 = const()[name = string("_inversed_scores_5_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_5_cast_fp16 = mul(x = var_752_cast_fp16, y = _inversed_scores_5_y_0_to_fp16)[name = string("_inversed_scores_5_cast_fp16")]; + tensor scores_7_cast_fp16 = select(a = var_46_to_fp16, b = _inversed_scores_5_cast_fp16, cond = mask_11)[name = string("scores_7_cast_fp16")]; + tensor var_758_cast_fp16 = softmax(axis = var_60, x = scores_7_cast_fp16)[name = string("op_758_cast_fp16")]; + tensor input_93_cast_fp16 = select(a = var_45_to_fp16, b = var_758_cast_fp16, cond = mask_11)[name = string("input_93_cast_fp16")]; + bool x_39_transpose_x_0 = const()[name = string("x_39_transpose_x_0"), val = bool(false)]; + bool x_39_transpose_y_0 = const()[name = string("x_39_transpose_y_0"), val = bool(false)]; + tensor value_11_cast_fp16 = transpose(perm = value_11_perm_0, x = v_3_cast_fp16)[name = string("transpose_350")]; + tensor x_39_cast_fp16 = matmul(transpose_x = x_39_transpose_x_0, transpose_y = x_39_transpose_y_0, x = input_93_cast_fp16, y = value_11_cast_fp16)[name = string("x_39_cast_fp16")]; + tensor var_762_perm_0 = const()[name = string("op_762_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_763 = const()[name = string("op_763"), val = tensor([1, -1, 1024])]; + tensor var_762_cast_fp16 = transpose(perm = var_762_perm_0, x = x_39_cast_fp16)[name = string("transpose_349")]; + tensor input_95_cast_fp16 = reshape(shape = var_763, x = var_762_cast_fp16)[name = string("input_95_cast_fp16")]; + tensor encoder_layers_1_self_attn_linear_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40638912))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41687552))))[name = string("encoder_layers_1_self_attn_linear_out_weight_to_fp16_quantized")]; + tensor encoder_layers_1_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_1_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41689664)))]; + tensor linear_16_cast_fp16 = linear(bias = encoder_layers_1_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_1_self_attn_linear_out_weight_to_fp16_quantized, x = input_95_cast_fp16)[name = string("linear_16_cast_fp16")]; + tensor input_99_cast_fp16 = add(x = input_89_cast_fp16, y = linear_16_cast_fp16)[name = string("input_99_cast_fp16")]; + tensor x_43_axes_0 = const()[name = string("x_43_axes_0"), val = tensor([-1])]; + tensor encoder_layers_1_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_1_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41691776)))]; + tensor encoder_layers_1_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_1_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41693888)))]; + tensor x_43_cast_fp16 = layer_norm(axes = x_43_axes_0, beta = encoder_layers_1_norm_conv_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_1_norm_conv_weight_to_fp16, x = input_99_cast_fp16)[name = string("x_43_cast_fp16")]; + tensor input_101_perm_0 = const()[name = string("input_101_perm_0"), val = tensor([0, 2, 1])]; + string input_103_pad_type_0 = const()[name = string("input_103_pad_type_0"), val = string("valid")]; + tensor input_103_strides_0 = const()[name = string("input_103_strides_0"), val = tensor([1])]; + tensor input_103_pad_0 = const()[name = string("input_103_pad_0"), val = tensor([0, 0])]; + tensor input_103_dilations_0 = const()[name = string("input_103_dilations_0"), val = tensor([1])]; + int32 input_103_groups_0 = const()[name = string("input_103_groups_0"), val = int32(1)]; + tensor encoder_layers_1_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41696000))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43793216))))[name = string("encoder_layers_1_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_101_cast_fp16 = transpose(perm = input_101_perm_0, x = x_43_cast_fp16)[name = string("transpose_348")]; + tensor input_103_cast_fp16 = conv(dilations = input_103_dilations_0, groups = input_103_groups_0, pad = input_103_pad_0, pad_type = input_103_pad_type_0, strides = input_103_strides_0, weight = encoder_layers_1_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_101_cast_fp16)[name = string("input_103_cast_fp16")]; + int32 x_45_split_num_splits_0 = const()[name = string("x_45_split_num_splits_0"), val = int32(2)]; + int32 x_45_split_axis_0 = const()[name = string("x_45_split_axis_0"), val = int32(1)]; + tensor x_45_split_cast_fp16_0, tensor x_45_split_cast_fp16_1 = split(axis = x_45_split_axis_0, num_splits = x_45_split_num_splits_0, x = input_103_cast_fp16)[name = string("x_45_split_cast_fp16")]; + tensor x_45_split_1_sigmoid_cast_fp16 = sigmoid(x = x_45_split_cast_fp16_1)[name = string("x_45_split_1_sigmoid_cast_fp16")]; + tensor x_45_cast_fp16 = mul(x = x_45_split_cast_fp16_0, y = x_45_split_1_sigmoid_cast_fp16)[name = string("x_45_cast_fp16")]; + tensor input_105_cast_fp16 = select(a = var_45_to_fp16, b = x_45_cast_fp16, cond = var_576)[name = string("input_105_cast_fp16")]; + bool new_x_7_interleave_0 = const()[name = string("new_x_7_interleave_0"), val = bool(false)]; + tensor new_x_7_cast_fp16 = concat(axis = var_60, interleave = new_x_7_interleave_0, values = (cache_7_cast_fp16, input_105_cast_fp16))[name = string("new_x_7_cast_fp16")]; + tensor var_802_begin_0 = const()[name = string("op_802_begin_0"), val = tensor([0, 0, 14])]; + tensor var_802_end_0 = const()[name = string("op_802_end_0"), val = tensor([1, 1024, 22])]; + tensor var_802_end_mask_0 = const()[name = string("op_802_end_mask_0"), val = tensor([true, true, true])]; + tensor var_802_cast_fp16 = slice_by_index(begin = var_802_begin_0, end = var_802_end_0, end_mask = var_802_end_mask_0, x = new_x_7_cast_fp16)[name = string("op_802_cast_fp16")]; + string x_47_pad_type_0 = const()[name = string("x_47_pad_type_0"), val = string("valid")]; + int32 x_47_groups_0 = const()[name = string("x_47_groups_0"), val = int32(1024)]; + tensor x_47_strides_0 = const()[name = string("x_47_strides_0"), val = tensor([1])]; + tensor x_47_pad_0 = const()[name = string("x_47_pad_0"), val = tensor([0, 0])]; + tensor x_47_dilations_0 = const()[name = string("x_47_dilations_0"), val = tensor([1])]; + tensor encoder_layers_1_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43797376))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43806656))))[name = string("encoder_layers_1_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_47_cast_fp16 = conv(dilations = x_47_dilations_0, groups = x_47_groups_0, pad = x_47_pad_0, pad_type = x_47_pad_type_0, strides = x_47_strides_0, weight = encoder_layers_1_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_7_cast_fp16)[name = string("x_47_cast_fp16")]; + tensor input_107_perm_0 = const()[name = string("input_107_perm_0"), val = tensor([0, 2, 1])]; + tensor x_49_axes_0 = const()[name = string("x_49_axes_0"), val = tensor([-1])]; + tensor encoder_layers_1_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_1_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43808768)))]; + tensor encoder_layers_1_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_1_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43810880)))]; + tensor input_107_cast_fp16 = transpose(perm = input_107_perm_0, x = x_47_cast_fp16)[name = string("transpose_347")]; + tensor x_49_cast_fp16 = layer_norm(axes = x_49_axes_0, beta = encoder_layers_1_conv_batch_norm_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_1_conv_batch_norm_weight_to_fp16, x = input_107_cast_fp16)[name = string("x_49_cast_fp16")]; + tensor input_109_perm_0 = const()[name = string("input_109_perm_0"), val = tensor([0, 2, 1])]; + tensor input_109_cast_fp16 = transpose(perm = input_109_perm_0, x = x_49_cast_fp16)[name = string("transpose_346")]; + tensor input_111_cast_fp16 = silu(x = input_109_cast_fp16)[name = string("input_111_cast_fp16")]; + string x_51_pad_type_0 = const()[name = string("x_51_pad_type_0"), val = string("valid")]; + tensor x_51_strides_0 = const()[name = string("x_51_strides_0"), val = tensor([1])]; + tensor x_51_pad_0 = const()[name = string("x_51_pad_0"), val = tensor([0, 0])]; + tensor x_51_dilations_0 = const()[name = string("x_51_dilations_0"), val = tensor([1])]; + int32 x_51_groups_0 = const()[name = string("x_51_groups_0"), val = int32(1)]; + tensor encoder_layers_1_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43812992))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44861632))))[name = string("encoder_layers_1_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_51_cast_fp16 = conv(dilations = x_51_dilations_0, groups = x_51_groups_0, pad = x_51_pad_0, pad_type = x_51_pad_type_0, strides = x_51_strides_0, weight = encoder_layers_1_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_111_cast_fp16)[name = string("x_51_cast_fp16")]; + tensor input_113_perm_0 = const()[name = string("input_113_perm_0"), val = tensor([0, 2, 1])]; + tensor input_113_cast_fp16 = transpose(perm = input_113_perm_0, x = x_51_cast_fp16)[name = string("transpose_345")]; + tensor input_115_cast_fp16 = add(x = input_99_cast_fp16, y = input_113_cast_fp16)[name = string("input_115_cast_fp16")]; + tensor input_117_axes_0 = const()[name = string("input_117_axes_0"), val = tensor([-1])]; + tensor encoder_layers_1_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_1_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44863744)))]; + tensor encoder_layers_1_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_1_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44865856)))]; + tensor input_117_cast_fp16 = layer_norm(axes = input_117_axes_0, beta = encoder_layers_1_norm_feed_forward2_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_1_norm_feed_forward2_weight_to_fp16, x = input_115_cast_fp16)[name = string("input_117_cast_fp16")]; + tensor encoder_layers_1_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44867968))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49062336))))[name = string("encoder_layers_1_feed_forward2_linear1_weight_to_fp16_quantized")]; + tensor encoder_layers_1_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_1_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49070592)))]; + tensor linear_17_cast_fp16 = linear(bias = encoder_layers_1_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_1_feed_forward2_linear1_weight_to_fp16_quantized, x = input_117_cast_fp16)[name = string("linear_17_cast_fp16")]; + tensor input_121_cast_fp16 = silu(x = linear_17_cast_fp16)[name = string("input_121_cast_fp16")]; + tensor encoder_layers_1_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49078848))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53273216))))[name = string("encoder_layers_1_feed_forward2_linear2_weight_to_fp16_quantized")]; + tensor encoder_layers_1_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_1_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53275328)))]; + tensor linear_18_cast_fp16 = linear(bias = encoder_layers_1_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_1_feed_forward2_linear2_weight_to_fp16_quantized, x = input_121_cast_fp16)[name = string("linear_18_cast_fp16")]; + fp16 var_845_to_fp16 = const()[name = string("op_845_to_fp16"), val = fp16(0x1p-1)]; + tensor var_846_cast_fp16 = mul(x = linear_18_cast_fp16, y = var_845_to_fp16)[name = string("op_846_cast_fp16")]; + tensor input_127_cast_fp16 = add(x = input_115_cast_fp16, y = var_846_cast_fp16)[name = string("input_127_cast_fp16")]; + tensor input_129_axes_0 = const()[name = string("input_129_axes_0"), val = tensor([-1])]; + tensor encoder_layers_1_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_1_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53277440)))]; + tensor encoder_layers_1_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_1_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53279552)))]; + tensor input_129_cast_fp16 = layer_norm(axes = input_129_axes_0, beta = encoder_layers_1_norm_out_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_1_norm_out_weight_to_fp16, x = input_127_cast_fp16)[name = string("input_129_cast_fp16")]; + tensor cache_9_begin_0 = const()[name = string("cache_9_begin_0"), val = tensor([2, 0, 0, 0])]; + tensor cache_9_end_0 = const()[name = string("cache_9_end_0"), val = tensor([3, 1, 42, 1024])]; + tensor cache_9_end_mask_0 = const()[name = string("cache_9_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_9_squeeze_mask_0 = const()[name = string("cache_9_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_9_cast_fp16 = slice_by_index(begin = cache_9_begin_0, end = cache_9_end_0, end_mask = cache_9_end_mask_0, squeeze_mask = cache_9_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_9_cast_fp16")]; + tensor cache_11_begin_0 = const()[name = string("cache_11_begin_0"), val = tensor([2, 0, 0, 0])]; + tensor cache_11_end_0 = const()[name = string("cache_11_end_0"), val = tensor([3, 1, 1024, 8])]; + tensor cache_11_end_mask_0 = const()[name = string("cache_11_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_11_squeeze_mask_0 = const()[name = string("cache_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_11_cast_fp16 = slice_by_index(begin = cache_11_begin_0, end = cache_11_end_0, end_mask = cache_11_end_mask_0, squeeze_mask = cache_11_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_11_cast_fp16")]; + tensor input_131_axes_0 = const()[name = string("input_131_axes_0"), val = tensor([-1])]; + tensor encoder_layers_2_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_2_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53281664)))]; + tensor encoder_layers_2_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_2_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53283776)))]; + tensor input_131_cast_fp16 = layer_norm(axes = input_131_axes_0, beta = encoder_layers_2_norm_feed_forward1_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_2_norm_feed_forward1_weight_to_fp16, x = input_129_cast_fp16)[name = string("input_131_cast_fp16")]; + tensor encoder_layers_2_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53285888))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57480256))))[name = string("encoder_layers_2_feed_forward1_linear1_weight_to_fp16_quantized")]; + tensor encoder_layers_2_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_2_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57488512)))]; + tensor linear_19_cast_fp16 = linear(bias = encoder_layers_2_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_2_feed_forward1_linear1_weight_to_fp16_quantized, x = input_131_cast_fp16)[name = string("linear_19_cast_fp16")]; + tensor input_135_cast_fp16 = silu(x = linear_19_cast_fp16)[name = string("input_135_cast_fp16")]; + tensor encoder_layers_2_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57496768))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61691136))))[name = string("encoder_layers_2_feed_forward1_linear2_weight_to_fp16_quantized")]; + tensor encoder_layers_2_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_2_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61693248)))]; + tensor linear_20_cast_fp16 = linear(bias = encoder_layers_2_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_2_feed_forward1_linear2_weight_to_fp16_quantized, x = input_135_cast_fp16)[name = string("linear_20_cast_fp16")]; + fp16 var_882_to_fp16 = const()[name = string("op_882_to_fp16"), val = fp16(0x1p-1)]; + tensor var_883_cast_fp16 = mul(x = linear_20_cast_fp16, y = var_882_to_fp16)[name = string("op_883_cast_fp16")]; + tensor input_141_cast_fp16 = add(x = input_129_cast_fp16, y = var_883_cast_fp16)[name = string("input_141_cast_fp16")]; + tensor key_5_axes_0 = const()[name = string("key_5_axes_0"), val = tensor([-1])]; + tensor encoder_layers_2_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_2_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61695360)))]; + tensor encoder_layers_2_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_2_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61697472)))]; + tensor key_5_cast_fp16 = layer_norm(axes = key_5_axes_0, beta = encoder_layers_2_norm_self_att_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_2_norm_self_att_weight_to_fp16, x = input_141_cast_fp16)[name = string("key_5_cast_fp16")]; + bool input_143_interleave_0 = const()[name = string("input_143_interleave_0"), val = bool(false)]; + tensor input_143_cast_fp16 = concat(axis = var_69, interleave = input_143_interleave_0, values = (cache_9_cast_fp16, key_5_cast_fp16))[name = string("input_143_cast_fp16")]; + tensor var_905_begin_0 = const()[name = string("op_905_begin_0"), val = tensor([0, 14, 0])]; + tensor var_905_end_0 = const()[name = string("op_905_end_0"), val = tensor([1, 42, 1024])]; + tensor var_905_end_mask_0 = const()[name = string("op_905_end_mask_0"), val = tensor([true, true, true])]; + tensor var_905_cast_fp16 = slice_by_index(begin = var_905_begin_0, end = var_905_end_0, end_mask = var_905_end_mask_0, x = cache_9_cast_fp16)[name = string("op_905_cast_fp16")]; + bool var_911_interleave_0 = const()[name = string("op_911_interleave_0"), val = bool(false)]; + tensor var_911_cast_fp16 = concat(axis = var_69, interleave = var_911_interleave_0, values = (var_905_cast_fp16, key_5_cast_fp16))[name = string("op_911_cast_fp16")]; + tensor encoder_layers_2_self_attn_linear_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61699584))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62748224))))[name = string("encoder_layers_2_self_attn_linear_q_weight_to_fp16_quantized")]; + tensor encoder_layers_2_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_2_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62750336)))]; + tensor linear_21_cast_fp16 = linear(bias = encoder_layers_2_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_2_self_attn_linear_q_weight_to_fp16_quantized, x = key_5_cast_fp16)[name = string("linear_21_cast_fp16")]; + tensor var_916 = const()[name = string("op_916"), val = tensor([1, -1, 8, 128])]; + tensor q_13_cast_fp16 = reshape(shape = var_916, x = linear_21_cast_fp16)[name = string("q_13_cast_fp16")]; + tensor encoder_layers_2_self_attn_linear_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62752448))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(63801088))))[name = string("encoder_layers_2_self_attn_linear_k_weight_to_fp16_quantized")]; + tensor encoder_layers_2_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_2_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(63803200)))]; + tensor linear_22_cast_fp16 = linear(bias = encoder_layers_2_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_2_self_attn_linear_k_weight_to_fp16_quantized, x = input_143_cast_fp16)[name = string("linear_22_cast_fp16")]; + tensor var_921 = const()[name = string("op_921"), val = tensor([1, -1, 8, 128])]; + tensor k_9_cast_fp16 = reshape(shape = var_921, x = linear_22_cast_fp16)[name = string("k_9_cast_fp16")]; + tensor encoder_layers_2_self_attn_linear_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(63805312))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64853952))))[name = string("encoder_layers_2_self_attn_linear_v_weight_to_fp16_quantized")]; + tensor encoder_layers_2_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_2_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64856064)))]; + tensor linear_23_cast_fp16 = linear(bias = encoder_layers_2_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_2_self_attn_linear_v_weight_to_fp16_quantized, x = input_143_cast_fp16)[name = string("linear_23_cast_fp16")]; + tensor var_926 = const()[name = string("op_926"), val = tensor([1, -1, 8, 128])]; + tensor v_5_cast_fp16 = reshape(shape = var_926, x = linear_23_cast_fp16)[name = string("v_5_cast_fp16")]; + tensor value_13_perm_0 = const()[name = string("value_13_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_2_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_2_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64858176)))]; + tensor var_939_cast_fp16 = add(x = q_13_cast_fp16, y = encoder_layers_2_self_attn_pos_bias_u_to_fp16)[name = string("op_939_cast_fp16")]; + tensor encoder_layers_2_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_2_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64860288)))]; + tensor var_941_cast_fp16 = add(x = q_13_cast_fp16, y = encoder_layers_2_self_attn_pos_bias_v_to_fp16)[name = string("op_941_cast_fp16")]; + tensor q_with_bias_v_5_perm_0 = const()[name = string("q_with_bias_v_5_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_59_transpose_x_0 = const()[name = string("x_59_transpose_x_0"), val = bool(false)]; + bool x_59_transpose_y_0 = const()[name = string("x_59_transpose_y_0"), val = bool(false)]; + tensor op_943_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64862400))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64976128))))[name = string("op_943_to_fp16_quantized")]; + tensor q_with_bias_v_5_cast_fp16 = transpose(perm = q_with_bias_v_5_perm_0, x = var_941_cast_fp16)[name = string("transpose_344")]; + tensor x_59_cast_fp16 = matmul(transpose_x = x_59_transpose_x_0, transpose_y = x_59_transpose_y_0, x = q_with_bias_v_5_cast_fp16, y = op_943_to_fp16_quantized)[name = string("x_59_cast_fp16")]; + tensor x_61_pad_0 = const()[name = string("x_61_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_61_mode_0 = const()[name = string("x_61_mode_0"), val = string("constant")]; + fp16 const_105_to_fp16 = const()[name = string("const_105_to_fp16"), val = fp16(0x0p+0)]; + tensor x_61_cast_fp16 = pad(constant_val = const_105_to_fp16, mode = x_61_mode_0, pad = x_61_pad_0, x = x_59_cast_fp16)[name = string("x_61_cast_fp16")]; + tensor var_951 = const()[name = string("op_951"), val = tensor([1, 8, -1, 14])]; + tensor x_63_cast_fp16 = reshape(shape = var_951, x = x_61_cast_fp16)[name = string("x_63_cast_fp16")]; + tensor var_955_begin_0 = const()[name = string("op_955_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_955_end_0 = const()[name = string("op_955_end_0"), val = tensor([1, 8, 112, 14])]; + tensor var_955_end_mask_0 = const()[name = string("op_955_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_955_cast_fp16 = slice_by_index(begin = var_955_begin_0, end = var_955_end_0, end_mask = var_955_end_mask_0, x = x_63_cast_fp16)[name = string("op_955_cast_fp16")]; + tensor var_956 = const()[name = string("op_956"), val = tensor([1, 8, 14, 111])]; + tensor matrix_bd_9_cast_fp16 = reshape(shape = var_956, x = var_955_cast_fp16)[name = string("matrix_bd_9_cast_fp16")]; + bool matrix_ac_5_transpose_x_0 = const()[name = string("matrix_ac_5_transpose_x_0"), val = bool(false)]; + bool matrix_ac_5_transpose_y_0 = const()[name = string("matrix_ac_5_transpose_y_0"), val = bool(false)]; + tensor transpose_100_perm_0 = const()[name = string("transpose_100_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_101_perm_0 = const()[name = string("transpose_101_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_101 = transpose(perm = transpose_101_perm_0, x = k_9_cast_fp16)[name = string("transpose_342")]; + tensor transpose_100 = transpose(perm = transpose_100_perm_0, x = var_939_cast_fp16)[name = string("transpose_343")]; + tensor matrix_ac_5_cast_fp16 = matmul(transpose_x = matrix_ac_5_transpose_x_0, transpose_y = matrix_ac_5_transpose_y_0, x = transpose_100, y = transpose_101)[name = string("matrix_ac_5_cast_fp16")]; + tensor matrix_bd_11_begin_0 = const()[name = string("matrix_bd_11_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_11_end_0 = const()[name = string("matrix_bd_11_end_0"), val = tensor([1, 8, 14, 56])]; + tensor matrix_bd_11_end_mask_0 = const()[name = string("matrix_bd_11_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_11_cast_fp16 = slice_by_index(begin = matrix_bd_11_begin_0, end = matrix_bd_11_end_0, end_mask = matrix_bd_11_end_mask_0, x = matrix_bd_9_cast_fp16)[name = string("matrix_bd_11_cast_fp16")]; + tensor var_965_cast_fp16 = add(x = matrix_ac_5_cast_fp16, y = matrix_bd_11_cast_fp16)[name = string("op_965_cast_fp16")]; + fp16 _inversed_scores_9_y_0_to_fp16 = const()[name = string("_inversed_scores_9_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_9_cast_fp16 = mul(x = var_965_cast_fp16, y = _inversed_scores_9_y_0_to_fp16)[name = string("_inversed_scores_9_cast_fp16")]; + tensor scores_11_cast_fp16 = select(a = var_46_to_fp16, b = _inversed_scores_9_cast_fp16, cond = mask_11)[name = string("scores_11_cast_fp16")]; + tensor var_971_cast_fp16 = softmax(axis = var_60, x = scores_11_cast_fp16)[name = string("op_971_cast_fp16")]; + tensor input_145_cast_fp16 = select(a = var_45_to_fp16, b = var_971_cast_fp16, cond = mask_11)[name = string("input_145_cast_fp16")]; + bool x_65_transpose_x_0 = const()[name = string("x_65_transpose_x_0"), val = bool(false)]; + bool x_65_transpose_y_0 = const()[name = string("x_65_transpose_y_0"), val = bool(false)]; + tensor value_13_cast_fp16 = transpose(perm = value_13_perm_0, x = v_5_cast_fp16)[name = string("transpose_341")]; + tensor x_65_cast_fp16 = matmul(transpose_x = x_65_transpose_x_0, transpose_y = x_65_transpose_y_0, x = input_145_cast_fp16, y = value_13_cast_fp16)[name = string("x_65_cast_fp16")]; + tensor var_975_perm_0 = const()[name = string("op_975_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_976 = const()[name = string("op_976"), val = tensor([1, -1, 1024])]; + tensor var_975_cast_fp16 = transpose(perm = var_975_perm_0, x = x_65_cast_fp16)[name = string("transpose_340")]; + tensor input_147_cast_fp16 = reshape(shape = var_976, x = var_975_cast_fp16)[name = string("input_147_cast_fp16")]; + tensor encoder_layers_2_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64976448))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65762944))))[name = string("encoder_layers_2_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_2_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_2_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65763136)))]; + tensor linear_25_cast_fp16 = linear(bias = encoder_layers_2_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_2_self_attn_linear_out_weight_to_fp16_palettized, x = input_147_cast_fp16)[name = string("linear_25_cast_fp16")]; + tensor input_151_cast_fp16 = add(x = input_141_cast_fp16, y = linear_25_cast_fp16)[name = string("input_151_cast_fp16")]; + tensor x_69_axes_0 = const()[name = string("x_69_axes_0"), val = tensor([-1])]; + tensor encoder_layers_2_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_2_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65765248)))]; + tensor encoder_layers_2_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_2_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65767360)))]; + tensor x_69_cast_fp16 = layer_norm(axes = x_69_axes_0, beta = encoder_layers_2_norm_conv_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_2_norm_conv_weight_to_fp16, x = input_151_cast_fp16)[name = string("x_69_cast_fp16")]; + tensor input_153_perm_0 = const()[name = string("input_153_perm_0"), val = tensor([0, 2, 1])]; + string input_155_pad_type_0 = const()[name = string("input_155_pad_type_0"), val = string("valid")]; + tensor input_155_strides_0 = const()[name = string("input_155_strides_0"), val = tensor([1])]; + tensor input_155_pad_0 = const()[name = string("input_155_pad_0"), val = tensor([0, 0])]; + tensor input_155_dilations_0 = const()[name = string("input_155_dilations_0"), val = tensor([1])]; + int32 input_155_groups_0 = const()[name = string("input_155_groups_0"), val = int32(1)]; + tensor encoder_layers_2_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65769472))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67866688))))[name = string("encoder_layers_2_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_153_cast_fp16 = transpose(perm = input_153_perm_0, x = x_69_cast_fp16)[name = string("transpose_339")]; + tensor input_155_cast_fp16 = conv(dilations = input_155_dilations_0, groups = input_155_groups_0, pad = input_155_pad_0, pad_type = input_155_pad_type_0, strides = input_155_strides_0, weight = encoder_layers_2_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_153_cast_fp16)[name = string("input_155_cast_fp16")]; + int32 x_71_split_num_splits_0 = const()[name = string("x_71_split_num_splits_0"), val = int32(2)]; + int32 x_71_split_axis_0 = const()[name = string("x_71_split_axis_0"), val = int32(1)]; + tensor x_71_split_cast_fp16_0, tensor x_71_split_cast_fp16_1 = split(axis = x_71_split_axis_0, num_splits = x_71_split_num_splits_0, x = input_155_cast_fp16)[name = string("x_71_split_cast_fp16")]; + tensor x_71_split_1_sigmoid_cast_fp16 = sigmoid(x = x_71_split_cast_fp16_1)[name = string("x_71_split_1_sigmoid_cast_fp16")]; + tensor x_71_cast_fp16 = mul(x = x_71_split_cast_fp16_0, y = x_71_split_1_sigmoid_cast_fp16)[name = string("x_71_cast_fp16")]; + tensor input_157_cast_fp16 = select(a = var_45_to_fp16, b = x_71_cast_fp16, cond = var_576)[name = string("input_157_cast_fp16")]; + bool new_x_11_interleave_0 = const()[name = string("new_x_11_interleave_0"), val = bool(false)]; + tensor new_x_11_cast_fp16 = concat(axis = var_60, interleave = new_x_11_interleave_0, values = (cache_11_cast_fp16, input_157_cast_fp16))[name = string("new_x_11_cast_fp16")]; + tensor var_1015_begin_0 = const()[name = string("op_1015_begin_0"), val = tensor([0, 0, 14])]; + tensor var_1015_end_0 = const()[name = string("op_1015_end_0"), val = tensor([1, 1024, 22])]; + tensor var_1015_end_mask_0 = const()[name = string("op_1015_end_mask_0"), val = tensor([true, true, true])]; + tensor var_1015_cast_fp16 = slice_by_index(begin = var_1015_begin_0, end = var_1015_end_0, end_mask = var_1015_end_mask_0, x = new_x_11_cast_fp16)[name = string("op_1015_cast_fp16")]; + string x_73_pad_type_0 = const()[name = string("x_73_pad_type_0"), val = string("valid")]; + int32 x_73_groups_0 = const()[name = string("x_73_groups_0"), val = int32(1024)]; + tensor x_73_strides_0 = const()[name = string("x_73_strides_0"), val = tensor([1])]; + tensor x_73_pad_0 = const()[name = string("x_73_pad_0"), val = tensor([0, 0])]; + tensor x_73_dilations_0 = const()[name = string("x_73_dilations_0"), val = tensor([1])]; + tensor encoder_layers_2_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67870848))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67880128))))[name = string("encoder_layers_2_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_73_cast_fp16 = conv(dilations = x_73_dilations_0, groups = x_73_groups_0, pad = x_73_pad_0, pad_type = x_73_pad_type_0, strides = x_73_strides_0, weight = encoder_layers_2_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_11_cast_fp16)[name = string("x_73_cast_fp16")]; + tensor input_159_perm_0 = const()[name = string("input_159_perm_0"), val = tensor([0, 2, 1])]; + tensor x_75_axes_0 = const()[name = string("x_75_axes_0"), val = tensor([-1])]; + tensor encoder_layers_2_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_2_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67882240)))]; + tensor encoder_layers_2_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_2_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67884352)))]; + tensor input_159_cast_fp16 = transpose(perm = input_159_perm_0, x = x_73_cast_fp16)[name = string("transpose_338")]; + tensor x_75_cast_fp16 = layer_norm(axes = x_75_axes_0, beta = encoder_layers_2_conv_batch_norm_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_2_conv_batch_norm_weight_to_fp16, x = input_159_cast_fp16)[name = string("x_75_cast_fp16")]; + tensor input_161_perm_0 = const()[name = string("input_161_perm_0"), val = tensor([0, 2, 1])]; + tensor input_161_cast_fp16 = transpose(perm = input_161_perm_0, x = x_75_cast_fp16)[name = string("transpose_337")]; + tensor input_163_cast_fp16 = silu(x = input_161_cast_fp16)[name = string("input_163_cast_fp16")]; + string x_77_pad_type_0 = const()[name = string("x_77_pad_type_0"), val = string("valid")]; + tensor x_77_strides_0 = const()[name = string("x_77_strides_0"), val = tensor([1])]; + tensor x_77_pad_0 = const()[name = string("x_77_pad_0"), val = tensor([0, 0])]; + tensor x_77_dilations_0 = const()[name = string("x_77_dilations_0"), val = tensor([1])]; + int32 x_77_groups_0 = const()[name = string("x_77_groups_0"), val = int32(1)]; + tensor encoder_layers_2_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67886464))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68935104))))[name = string("encoder_layers_2_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_77_cast_fp16 = conv(dilations = x_77_dilations_0, groups = x_77_groups_0, pad = x_77_pad_0, pad_type = x_77_pad_type_0, strides = x_77_strides_0, weight = encoder_layers_2_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_163_cast_fp16)[name = string("x_77_cast_fp16")]; + tensor input_165_perm_0 = const()[name = string("input_165_perm_0"), val = tensor([0, 2, 1])]; + tensor input_165_cast_fp16 = transpose(perm = input_165_perm_0, x = x_77_cast_fp16)[name = string("transpose_336")]; + tensor input_167_cast_fp16 = add(x = input_151_cast_fp16, y = input_165_cast_fp16)[name = string("input_167_cast_fp16")]; + tensor input_169_axes_0 = const()[name = string("input_169_axes_0"), val = tensor([-1])]; + tensor encoder_layers_2_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_2_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68937216)))]; + tensor encoder_layers_2_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_2_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68939328)))]; + tensor input_169_cast_fp16 = layer_norm(axes = input_169_axes_0, beta = encoder_layers_2_norm_feed_forward2_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_2_norm_feed_forward2_weight_to_fp16, x = input_167_cast_fp16)[name = string("input_169_cast_fp16")]; + tensor encoder_layers_2_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68941440))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72087232))))[name = string("encoder_layers_2_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_2_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_2_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72087424)))]; + tensor linear_26_cast_fp16 = linear(bias = encoder_layers_2_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_2_feed_forward2_linear1_weight_to_fp16_palettized, x = input_169_cast_fp16)[name = string("linear_26_cast_fp16")]; + tensor input_173_cast_fp16 = silu(x = linear_26_cast_fp16)[name = string("input_173_cast_fp16")]; + tensor encoder_layers_2_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72095680))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75241472))))[name = string("encoder_layers_2_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_2_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_2_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75241664)))]; + tensor linear_27_cast_fp16 = linear(bias = encoder_layers_2_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_2_feed_forward2_linear2_weight_to_fp16_palettized, x = input_173_cast_fp16)[name = string("linear_27_cast_fp16")]; + fp16 var_1058_to_fp16 = const()[name = string("op_1058_to_fp16"), val = fp16(0x1p-1)]; + tensor var_1059_cast_fp16 = mul(x = linear_27_cast_fp16, y = var_1058_to_fp16)[name = string("op_1059_cast_fp16")]; + tensor input_179_cast_fp16 = add(x = input_167_cast_fp16, y = var_1059_cast_fp16)[name = string("input_179_cast_fp16")]; + tensor input_181_axes_0 = const()[name = string("input_181_axes_0"), val = tensor([-1])]; + tensor encoder_layers_2_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_2_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75243776)))]; + tensor encoder_layers_2_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_2_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75245888)))]; + tensor input_181_cast_fp16 = layer_norm(axes = input_181_axes_0, beta = encoder_layers_2_norm_out_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_2_norm_out_weight_to_fp16, x = input_179_cast_fp16)[name = string("input_181_cast_fp16")]; + tensor cache_13_begin_0 = const()[name = string("cache_13_begin_0"), val = tensor([3, 0, 0, 0])]; + tensor cache_13_end_0 = const()[name = string("cache_13_end_0"), val = tensor([4, 1, 42, 1024])]; + tensor cache_13_end_mask_0 = const()[name = string("cache_13_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_13_squeeze_mask_0 = const()[name = string("cache_13_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_13_cast_fp16 = slice_by_index(begin = cache_13_begin_0, end = cache_13_end_0, end_mask = cache_13_end_mask_0, squeeze_mask = cache_13_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_13_cast_fp16")]; + tensor cache_15_begin_0 = const()[name = string("cache_15_begin_0"), val = tensor([3, 0, 0, 0])]; + tensor cache_15_end_0 = const()[name = string("cache_15_end_0"), val = tensor([4, 1, 1024, 8])]; + tensor cache_15_end_mask_0 = const()[name = string("cache_15_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_15_squeeze_mask_0 = const()[name = string("cache_15_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_15_cast_fp16 = slice_by_index(begin = cache_15_begin_0, end = cache_15_end_0, end_mask = cache_15_end_mask_0, squeeze_mask = cache_15_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_15_cast_fp16")]; + tensor input_183_axes_0 = const()[name = string("input_183_axes_0"), val = tensor([-1])]; + tensor encoder_layers_3_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_3_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75248000)))]; + tensor encoder_layers_3_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_3_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75250112)))]; + tensor input_183_cast_fp16 = layer_norm(axes = input_183_axes_0, beta = encoder_layers_3_norm_feed_forward1_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_3_norm_feed_forward1_weight_to_fp16, x = input_181_cast_fp16)[name = string("input_183_cast_fp16")]; + tensor encoder_layers_3_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75252224))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78398016))))[name = string("encoder_layers_3_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_3_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_3_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78398208)))]; + tensor linear_28_cast_fp16 = linear(bias = encoder_layers_3_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_3_feed_forward1_linear1_weight_to_fp16_palettized, x = input_183_cast_fp16)[name = string("linear_28_cast_fp16")]; + tensor input_187_cast_fp16 = silu(x = linear_28_cast_fp16)[name = string("input_187_cast_fp16")]; + tensor encoder_layers_3_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78406464))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81552256))))[name = string("encoder_layers_3_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_3_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_3_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81552448)))]; + tensor linear_29_cast_fp16 = linear(bias = encoder_layers_3_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_3_feed_forward1_linear2_weight_to_fp16_palettized, x = input_187_cast_fp16)[name = string("linear_29_cast_fp16")]; + fp16 var_1095_to_fp16 = const()[name = string("op_1095_to_fp16"), val = fp16(0x1p-1)]; + tensor var_1096_cast_fp16 = mul(x = linear_29_cast_fp16, y = var_1095_to_fp16)[name = string("op_1096_cast_fp16")]; + tensor input_193_cast_fp16 = add(x = input_181_cast_fp16, y = var_1096_cast_fp16)[name = string("input_193_cast_fp16")]; + tensor key_7_axes_0 = const()[name = string("key_7_axes_0"), val = tensor([-1])]; + tensor encoder_layers_3_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_3_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81554560)))]; + tensor encoder_layers_3_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_3_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81556672)))]; + tensor key_7_cast_fp16 = layer_norm(axes = key_7_axes_0, beta = encoder_layers_3_norm_self_att_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_3_norm_self_att_weight_to_fp16, x = input_193_cast_fp16)[name = string("key_7_cast_fp16")]; + bool input_195_interleave_0 = const()[name = string("input_195_interleave_0"), val = bool(false)]; + tensor input_195_cast_fp16 = concat(axis = var_69, interleave = input_195_interleave_0, values = (cache_13_cast_fp16, key_7_cast_fp16))[name = string("input_195_cast_fp16")]; + tensor var_1118_begin_0 = const()[name = string("op_1118_begin_0"), val = tensor([0, 14, 0])]; + tensor var_1118_end_0 = const()[name = string("op_1118_end_0"), val = tensor([1, 42, 1024])]; + tensor var_1118_end_mask_0 = const()[name = string("op_1118_end_mask_0"), val = tensor([true, true, true])]; + tensor var_1118_cast_fp16 = slice_by_index(begin = var_1118_begin_0, end = var_1118_end_0, end_mask = var_1118_end_mask_0, x = cache_13_cast_fp16)[name = string("op_1118_cast_fp16")]; + bool var_1124_interleave_0 = const()[name = string("op_1124_interleave_0"), val = bool(false)]; + tensor var_1124_cast_fp16 = concat(axis = var_69, interleave = var_1124_interleave_0, values = (var_1118_cast_fp16, key_7_cast_fp16))[name = string("op_1124_cast_fp16")]; + tensor encoder_layers_3_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81558784))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82345280))))[name = string("encoder_layers_3_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_3_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_3_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82345472)))]; + tensor linear_30_cast_fp16 = linear(bias = encoder_layers_3_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_3_self_attn_linear_q_weight_to_fp16_palettized, x = key_7_cast_fp16)[name = string("linear_30_cast_fp16")]; + tensor var_1129 = const()[name = string("op_1129"), val = tensor([1, -1, 8, 128])]; + tensor q_19_cast_fp16 = reshape(shape = var_1129, x = linear_30_cast_fp16)[name = string("q_19_cast_fp16")]; + tensor encoder_layers_3_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82347584))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83134080))))[name = string("encoder_layers_3_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_3_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_3_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83134272)))]; + tensor linear_31_cast_fp16 = linear(bias = encoder_layers_3_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_3_self_attn_linear_k_weight_to_fp16_palettized, x = input_195_cast_fp16)[name = string("linear_31_cast_fp16")]; + tensor var_1134 = const()[name = string("op_1134"), val = tensor([1, -1, 8, 128])]; + tensor k_13_cast_fp16 = reshape(shape = var_1134, x = linear_31_cast_fp16)[name = string("k_13_cast_fp16")]; + tensor encoder_layers_3_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83136384))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83922880))))[name = string("encoder_layers_3_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_3_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_3_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83923072)))]; + tensor linear_32_cast_fp16 = linear(bias = encoder_layers_3_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_3_self_attn_linear_v_weight_to_fp16_palettized, x = input_195_cast_fp16)[name = string("linear_32_cast_fp16")]; + tensor var_1139 = const()[name = string("op_1139"), val = tensor([1, -1, 8, 128])]; + tensor v_7_cast_fp16 = reshape(shape = var_1139, x = linear_32_cast_fp16)[name = string("v_7_cast_fp16")]; + tensor value_15_perm_0 = const()[name = string("value_15_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_3_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_3_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83925184)))]; + tensor var_1152_cast_fp16 = add(x = q_19_cast_fp16, y = encoder_layers_3_self_attn_pos_bias_u_to_fp16)[name = string("op_1152_cast_fp16")]; + tensor encoder_layers_3_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_3_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83927296)))]; + tensor var_1154_cast_fp16 = add(x = q_19_cast_fp16, y = encoder_layers_3_self_attn_pos_bias_v_to_fp16)[name = string("op_1154_cast_fp16")]; + tensor q_with_bias_v_7_perm_0 = const()[name = string("q_with_bias_v_7_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_85_transpose_x_0 = const()[name = string("x_85_transpose_x_0"), val = bool(false)]; + bool x_85_transpose_y_0 = const()[name = string("x_85_transpose_y_0"), val = bool(false)]; + tensor op_1156_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83929408))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84043136))))[name = string("op_1156_to_fp16_quantized")]; + tensor q_with_bias_v_7_cast_fp16 = transpose(perm = q_with_bias_v_7_perm_0, x = var_1154_cast_fp16)[name = string("transpose_335")]; + tensor x_85_cast_fp16 = matmul(transpose_x = x_85_transpose_x_0, transpose_y = x_85_transpose_y_0, x = q_with_bias_v_7_cast_fp16, y = op_1156_to_fp16_quantized)[name = string("x_85_cast_fp16")]; + tensor x_87_pad_0 = const()[name = string("x_87_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_87_mode_0 = const()[name = string("x_87_mode_0"), val = string("constant")]; + fp16 const_118_to_fp16 = const()[name = string("const_118_to_fp16"), val = fp16(0x0p+0)]; + tensor x_87_cast_fp16 = pad(constant_val = const_118_to_fp16, mode = x_87_mode_0, pad = x_87_pad_0, x = x_85_cast_fp16)[name = string("x_87_cast_fp16")]; + tensor var_1164 = const()[name = string("op_1164"), val = tensor([1, 8, -1, 14])]; + tensor x_89_cast_fp16 = reshape(shape = var_1164, x = x_87_cast_fp16)[name = string("x_89_cast_fp16")]; + tensor var_1168_begin_0 = const()[name = string("op_1168_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_1168_end_0 = const()[name = string("op_1168_end_0"), val = tensor([1, 8, 112, 14])]; + tensor var_1168_end_mask_0 = const()[name = string("op_1168_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_1168_cast_fp16 = slice_by_index(begin = var_1168_begin_0, end = var_1168_end_0, end_mask = var_1168_end_mask_0, x = x_89_cast_fp16)[name = string("op_1168_cast_fp16")]; + tensor var_1169 = const()[name = string("op_1169"), val = tensor([1, 8, 14, 111])]; + tensor matrix_bd_13_cast_fp16 = reshape(shape = var_1169, x = var_1168_cast_fp16)[name = string("matrix_bd_13_cast_fp16")]; + bool matrix_ac_7_transpose_x_0 = const()[name = string("matrix_ac_7_transpose_x_0"), val = bool(false)]; + bool matrix_ac_7_transpose_y_0 = const()[name = string("matrix_ac_7_transpose_y_0"), val = bool(false)]; + tensor transpose_102_perm_0 = const()[name = string("transpose_102_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_103_perm_0 = const()[name = string("transpose_103_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_103 = transpose(perm = transpose_103_perm_0, x = k_13_cast_fp16)[name = string("transpose_333")]; + tensor transpose_102 = transpose(perm = transpose_102_perm_0, x = var_1152_cast_fp16)[name = string("transpose_334")]; + tensor matrix_ac_7_cast_fp16 = matmul(transpose_x = matrix_ac_7_transpose_x_0, transpose_y = matrix_ac_7_transpose_y_0, x = transpose_102, y = transpose_103)[name = string("matrix_ac_7_cast_fp16")]; + tensor matrix_bd_15_begin_0 = const()[name = string("matrix_bd_15_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_15_end_0 = const()[name = string("matrix_bd_15_end_0"), val = tensor([1, 8, 14, 56])]; + tensor matrix_bd_15_end_mask_0 = const()[name = string("matrix_bd_15_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_15_cast_fp16 = slice_by_index(begin = matrix_bd_15_begin_0, end = matrix_bd_15_end_0, end_mask = matrix_bd_15_end_mask_0, x = matrix_bd_13_cast_fp16)[name = string("matrix_bd_15_cast_fp16")]; + tensor var_1178_cast_fp16 = add(x = matrix_ac_7_cast_fp16, y = matrix_bd_15_cast_fp16)[name = string("op_1178_cast_fp16")]; + fp16 _inversed_scores_13_y_0_to_fp16 = const()[name = string("_inversed_scores_13_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_13_cast_fp16 = mul(x = var_1178_cast_fp16, y = _inversed_scores_13_y_0_to_fp16)[name = string("_inversed_scores_13_cast_fp16")]; + tensor scores_15_cast_fp16 = select(a = var_46_to_fp16, b = _inversed_scores_13_cast_fp16, cond = mask_11)[name = string("scores_15_cast_fp16")]; + tensor var_1184_cast_fp16 = softmax(axis = var_60, x = scores_15_cast_fp16)[name = string("op_1184_cast_fp16")]; + tensor input_197_cast_fp16 = select(a = var_45_to_fp16, b = var_1184_cast_fp16, cond = mask_11)[name = string("input_197_cast_fp16")]; + bool x_91_transpose_x_0 = const()[name = string("x_91_transpose_x_0"), val = bool(false)]; + bool x_91_transpose_y_0 = const()[name = string("x_91_transpose_y_0"), val = bool(false)]; + tensor value_15_cast_fp16 = transpose(perm = value_15_perm_0, x = v_7_cast_fp16)[name = string("transpose_332")]; + tensor x_91_cast_fp16 = matmul(transpose_x = x_91_transpose_x_0, transpose_y = x_91_transpose_y_0, x = input_197_cast_fp16, y = value_15_cast_fp16)[name = string("x_91_cast_fp16")]; + tensor var_1188_perm_0 = const()[name = string("op_1188_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1189 = const()[name = string("op_1189"), val = tensor([1, -1, 1024])]; + tensor var_1188_cast_fp16 = transpose(perm = var_1188_perm_0, x = x_91_cast_fp16)[name = string("transpose_331")]; + tensor input_199_cast_fp16 = reshape(shape = var_1189, x = var_1188_cast_fp16)[name = string("input_199_cast_fp16")]; + tensor encoder_layers_3_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84043456))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84829952))))[name = string("encoder_layers_3_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_3_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_3_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84830144)))]; + tensor linear_34_cast_fp16 = linear(bias = encoder_layers_3_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_3_self_attn_linear_out_weight_to_fp16_palettized, x = input_199_cast_fp16)[name = string("linear_34_cast_fp16")]; + tensor input_203_cast_fp16 = add(x = input_193_cast_fp16, y = linear_34_cast_fp16)[name = string("input_203_cast_fp16")]; + tensor x_95_axes_0 = const()[name = string("x_95_axes_0"), val = tensor([-1])]; + tensor encoder_layers_3_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_3_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84832256)))]; + tensor encoder_layers_3_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_3_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84834368)))]; + tensor x_95_cast_fp16 = layer_norm(axes = x_95_axes_0, beta = encoder_layers_3_norm_conv_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_3_norm_conv_weight_to_fp16, x = input_203_cast_fp16)[name = string("x_95_cast_fp16")]; + tensor input_205_perm_0 = const()[name = string("input_205_perm_0"), val = tensor([0, 2, 1])]; + string input_207_pad_type_0 = const()[name = string("input_207_pad_type_0"), val = string("valid")]; + tensor input_207_strides_0 = const()[name = string("input_207_strides_0"), val = tensor([1])]; + tensor input_207_pad_0 = const()[name = string("input_207_pad_0"), val = tensor([0, 0])]; + tensor input_207_dilations_0 = const()[name = string("input_207_dilations_0"), val = tensor([1])]; + int32 input_207_groups_0 = const()[name = string("input_207_groups_0"), val = int32(1)]; + tensor encoder_layers_3_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84836480))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86933696))))[name = string("encoder_layers_3_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_205_cast_fp16 = transpose(perm = input_205_perm_0, x = x_95_cast_fp16)[name = string("transpose_330")]; + tensor input_207_cast_fp16 = conv(dilations = input_207_dilations_0, groups = input_207_groups_0, pad = input_207_pad_0, pad_type = input_207_pad_type_0, strides = input_207_strides_0, weight = encoder_layers_3_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_205_cast_fp16)[name = string("input_207_cast_fp16")]; + int32 x_97_split_num_splits_0 = const()[name = string("x_97_split_num_splits_0"), val = int32(2)]; + int32 x_97_split_axis_0 = const()[name = string("x_97_split_axis_0"), val = int32(1)]; + tensor x_97_split_cast_fp16_0, tensor x_97_split_cast_fp16_1 = split(axis = x_97_split_axis_0, num_splits = x_97_split_num_splits_0, x = input_207_cast_fp16)[name = string("x_97_split_cast_fp16")]; + tensor x_97_split_1_sigmoid_cast_fp16 = sigmoid(x = x_97_split_cast_fp16_1)[name = string("x_97_split_1_sigmoid_cast_fp16")]; + tensor x_97_cast_fp16 = mul(x = x_97_split_cast_fp16_0, y = x_97_split_1_sigmoid_cast_fp16)[name = string("x_97_cast_fp16")]; + tensor input_209_cast_fp16 = select(a = var_45_to_fp16, b = x_97_cast_fp16, cond = var_576)[name = string("input_209_cast_fp16")]; + bool new_x_15_interleave_0 = const()[name = string("new_x_15_interleave_0"), val = bool(false)]; + tensor new_x_15_cast_fp16 = concat(axis = var_60, interleave = new_x_15_interleave_0, values = (cache_15_cast_fp16, input_209_cast_fp16))[name = string("new_x_15_cast_fp16")]; + tensor var_1228_begin_0 = const()[name = string("op_1228_begin_0"), val = tensor([0, 0, 14])]; + tensor var_1228_end_0 = const()[name = string("op_1228_end_0"), val = tensor([1, 1024, 22])]; + tensor var_1228_end_mask_0 = const()[name = string("op_1228_end_mask_0"), val = tensor([true, true, true])]; + tensor var_1228_cast_fp16 = slice_by_index(begin = var_1228_begin_0, end = var_1228_end_0, end_mask = var_1228_end_mask_0, x = new_x_15_cast_fp16)[name = string("op_1228_cast_fp16")]; + string x_99_pad_type_0 = const()[name = string("x_99_pad_type_0"), val = string("valid")]; + int32 x_99_groups_0 = const()[name = string("x_99_groups_0"), val = int32(1024)]; + tensor x_99_strides_0 = const()[name = string("x_99_strides_0"), val = tensor([1])]; + tensor x_99_pad_0 = const()[name = string("x_99_pad_0"), val = tensor([0, 0])]; + tensor x_99_dilations_0 = const()[name = string("x_99_dilations_0"), val = tensor([1])]; + tensor encoder_layers_3_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86937856))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86947136))))[name = string("encoder_layers_3_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_99_cast_fp16 = conv(dilations = x_99_dilations_0, groups = x_99_groups_0, pad = x_99_pad_0, pad_type = x_99_pad_type_0, strides = x_99_strides_0, weight = encoder_layers_3_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_15_cast_fp16)[name = string("x_99_cast_fp16")]; + tensor input_211_perm_0 = const()[name = string("input_211_perm_0"), val = tensor([0, 2, 1])]; + tensor x_101_axes_0 = const()[name = string("x_101_axes_0"), val = tensor([-1])]; + tensor encoder_layers_3_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_3_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86949248)))]; + tensor encoder_layers_3_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_3_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86951360)))]; + tensor input_211_cast_fp16 = transpose(perm = input_211_perm_0, x = x_99_cast_fp16)[name = string("transpose_329")]; + tensor x_101_cast_fp16 = layer_norm(axes = x_101_axes_0, beta = encoder_layers_3_conv_batch_norm_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_3_conv_batch_norm_weight_to_fp16, x = input_211_cast_fp16)[name = string("x_101_cast_fp16")]; + tensor input_213_perm_0 = const()[name = string("input_213_perm_0"), val = tensor([0, 2, 1])]; + tensor input_213_cast_fp16 = transpose(perm = input_213_perm_0, x = x_101_cast_fp16)[name = string("transpose_328")]; + tensor input_215_cast_fp16 = silu(x = input_213_cast_fp16)[name = string("input_215_cast_fp16")]; + string x_103_pad_type_0 = const()[name = string("x_103_pad_type_0"), val = string("valid")]; + tensor x_103_strides_0 = const()[name = string("x_103_strides_0"), val = tensor([1])]; + tensor x_103_pad_0 = const()[name = string("x_103_pad_0"), val = tensor([0, 0])]; + tensor x_103_dilations_0 = const()[name = string("x_103_dilations_0"), val = tensor([1])]; + int32 x_103_groups_0 = const()[name = string("x_103_groups_0"), val = int32(1)]; + tensor encoder_layers_3_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86953472))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88002112))))[name = string("encoder_layers_3_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_103_cast_fp16 = conv(dilations = x_103_dilations_0, groups = x_103_groups_0, pad = x_103_pad_0, pad_type = x_103_pad_type_0, strides = x_103_strides_0, weight = encoder_layers_3_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_215_cast_fp16)[name = string("x_103_cast_fp16")]; + tensor input_217_perm_0 = const()[name = string("input_217_perm_0"), val = tensor([0, 2, 1])]; + tensor input_217_cast_fp16 = transpose(perm = input_217_perm_0, x = x_103_cast_fp16)[name = string("transpose_327")]; + tensor input_219_cast_fp16 = add(x = input_203_cast_fp16, y = input_217_cast_fp16)[name = string("input_219_cast_fp16")]; + tensor input_221_axes_0 = const()[name = string("input_221_axes_0"), val = tensor([-1])]; + tensor encoder_layers_3_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_3_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88004224)))]; + tensor encoder_layers_3_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_3_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88006336)))]; + tensor input_221_cast_fp16 = layer_norm(axes = input_221_axes_0, beta = encoder_layers_3_norm_feed_forward2_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_3_norm_feed_forward2_weight_to_fp16, x = input_219_cast_fp16)[name = string("input_221_cast_fp16")]; + tensor encoder_layers_3_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88008448))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(91154240))))[name = string("encoder_layers_3_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_3_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_3_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(91154432)))]; + tensor linear_35_cast_fp16 = linear(bias = encoder_layers_3_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_3_feed_forward2_linear1_weight_to_fp16_palettized, x = input_221_cast_fp16)[name = string("linear_35_cast_fp16")]; + tensor input_225_cast_fp16 = silu(x = linear_35_cast_fp16)[name = string("input_225_cast_fp16")]; + tensor encoder_layers_3_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(91162688))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94308480))))[name = string("encoder_layers_3_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_3_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_3_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94308672)))]; + tensor linear_36_cast_fp16 = linear(bias = encoder_layers_3_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_3_feed_forward2_linear2_weight_to_fp16_palettized, x = input_225_cast_fp16)[name = string("linear_36_cast_fp16")]; + fp16 var_1271_to_fp16 = const()[name = string("op_1271_to_fp16"), val = fp16(0x1p-1)]; + tensor var_1272_cast_fp16 = mul(x = linear_36_cast_fp16, y = var_1271_to_fp16)[name = string("op_1272_cast_fp16")]; + tensor input_231_cast_fp16 = add(x = input_219_cast_fp16, y = var_1272_cast_fp16)[name = string("input_231_cast_fp16")]; + tensor input_233_axes_0 = const()[name = string("input_233_axes_0"), val = tensor([-1])]; + tensor encoder_layers_3_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_3_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94310784)))]; + tensor encoder_layers_3_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_3_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94312896)))]; + tensor input_233_cast_fp16 = layer_norm(axes = input_233_axes_0, beta = encoder_layers_3_norm_out_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_3_norm_out_weight_to_fp16, x = input_231_cast_fp16)[name = string("input_233_cast_fp16")]; + tensor cache_17_begin_0 = const()[name = string("cache_17_begin_0"), val = tensor([4, 0, 0, 0])]; + tensor cache_17_end_0 = const()[name = string("cache_17_end_0"), val = tensor([5, 1, 42, 1024])]; + tensor cache_17_end_mask_0 = const()[name = string("cache_17_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_17_squeeze_mask_0 = const()[name = string("cache_17_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_17_cast_fp16 = slice_by_index(begin = cache_17_begin_0, end = cache_17_end_0, end_mask = cache_17_end_mask_0, squeeze_mask = cache_17_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_17_cast_fp16")]; + tensor cache_19_begin_0 = const()[name = string("cache_19_begin_0"), val = tensor([4, 0, 0, 0])]; + tensor cache_19_end_0 = const()[name = string("cache_19_end_0"), val = tensor([5, 1, 1024, 8])]; + tensor cache_19_end_mask_0 = const()[name = string("cache_19_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_19_squeeze_mask_0 = const()[name = string("cache_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_19_cast_fp16 = slice_by_index(begin = cache_19_begin_0, end = cache_19_end_0, end_mask = cache_19_end_mask_0, squeeze_mask = cache_19_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_19_cast_fp16")]; + tensor input_235_axes_0 = const()[name = string("input_235_axes_0"), val = tensor([-1])]; + tensor encoder_layers_4_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_4_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94315008)))]; + tensor encoder_layers_4_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_4_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94317120)))]; + tensor input_235_cast_fp16 = layer_norm(axes = input_235_axes_0, beta = encoder_layers_4_norm_feed_forward1_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_4_norm_feed_forward1_weight_to_fp16, x = input_233_cast_fp16)[name = string("input_235_cast_fp16")]; + tensor encoder_layers_4_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94319232))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97465024))))[name = string("encoder_layers_4_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_4_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_4_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97465216)))]; + tensor linear_37_cast_fp16 = linear(bias = encoder_layers_4_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_4_feed_forward1_linear1_weight_to_fp16_palettized, x = input_235_cast_fp16)[name = string("linear_37_cast_fp16")]; + tensor input_239_cast_fp16 = silu(x = linear_37_cast_fp16)[name = string("input_239_cast_fp16")]; + tensor encoder_layers_4_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97473472))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100619264))))[name = string("encoder_layers_4_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_4_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_4_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100619456)))]; + tensor linear_38_cast_fp16 = linear(bias = encoder_layers_4_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_4_feed_forward1_linear2_weight_to_fp16_palettized, x = input_239_cast_fp16)[name = string("linear_38_cast_fp16")]; + fp16 var_1308_to_fp16 = const()[name = string("op_1308_to_fp16"), val = fp16(0x1p-1)]; + tensor var_1309_cast_fp16 = mul(x = linear_38_cast_fp16, y = var_1308_to_fp16)[name = string("op_1309_cast_fp16")]; + tensor input_245_cast_fp16 = add(x = input_233_cast_fp16, y = var_1309_cast_fp16)[name = string("input_245_cast_fp16")]; + tensor key_9_axes_0 = const()[name = string("key_9_axes_0"), val = tensor([-1])]; + tensor encoder_layers_4_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_4_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100621568)))]; + tensor encoder_layers_4_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_4_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100623680)))]; + tensor key_9_cast_fp16 = layer_norm(axes = key_9_axes_0, beta = encoder_layers_4_norm_self_att_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_4_norm_self_att_weight_to_fp16, x = input_245_cast_fp16)[name = string("key_9_cast_fp16")]; + bool input_247_interleave_0 = const()[name = string("input_247_interleave_0"), val = bool(false)]; + tensor input_247_cast_fp16 = concat(axis = var_69, interleave = input_247_interleave_0, values = (cache_17_cast_fp16, key_9_cast_fp16))[name = string("input_247_cast_fp16")]; + tensor var_1331_begin_0 = const()[name = string("op_1331_begin_0"), val = tensor([0, 14, 0])]; + tensor var_1331_end_0 = const()[name = string("op_1331_end_0"), val = tensor([1, 42, 1024])]; + tensor var_1331_end_mask_0 = const()[name = string("op_1331_end_mask_0"), val = tensor([true, true, true])]; + tensor var_1331_cast_fp16 = slice_by_index(begin = var_1331_begin_0, end = var_1331_end_0, end_mask = var_1331_end_mask_0, x = cache_17_cast_fp16)[name = string("op_1331_cast_fp16")]; + bool var_1337_interleave_0 = const()[name = string("op_1337_interleave_0"), val = bool(false)]; + tensor var_1337_cast_fp16 = concat(axis = var_69, interleave = var_1337_interleave_0, values = (var_1331_cast_fp16, key_9_cast_fp16))[name = string("op_1337_cast_fp16")]; + tensor encoder_layers_4_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100625792))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(101412288))))[name = string("encoder_layers_4_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_4_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_4_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(101412480)))]; + tensor linear_39_cast_fp16 = linear(bias = encoder_layers_4_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_4_self_attn_linear_q_weight_to_fp16_palettized, x = key_9_cast_fp16)[name = string("linear_39_cast_fp16")]; + tensor var_1342 = const()[name = string("op_1342"), val = tensor([1, -1, 8, 128])]; + tensor q_25_cast_fp16 = reshape(shape = var_1342, x = linear_39_cast_fp16)[name = string("q_25_cast_fp16")]; + tensor encoder_layers_4_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(101414592))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102201088))))[name = string("encoder_layers_4_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_4_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_4_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102201280)))]; + tensor linear_40_cast_fp16 = linear(bias = encoder_layers_4_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_4_self_attn_linear_k_weight_to_fp16_palettized, x = input_247_cast_fp16)[name = string("linear_40_cast_fp16")]; + tensor var_1347 = const()[name = string("op_1347"), val = tensor([1, -1, 8, 128])]; + tensor k_17_cast_fp16 = reshape(shape = var_1347, x = linear_40_cast_fp16)[name = string("k_17_cast_fp16")]; + tensor encoder_layers_4_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102203392))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102989888))))[name = string("encoder_layers_4_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_4_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_4_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102990080)))]; + tensor linear_41_cast_fp16 = linear(bias = encoder_layers_4_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_4_self_attn_linear_v_weight_to_fp16_palettized, x = input_247_cast_fp16)[name = string("linear_41_cast_fp16")]; + tensor var_1352 = const()[name = string("op_1352"), val = tensor([1, -1, 8, 128])]; + tensor v_9_cast_fp16 = reshape(shape = var_1352, x = linear_41_cast_fp16)[name = string("v_9_cast_fp16")]; + tensor value_17_perm_0 = const()[name = string("value_17_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_4_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_4_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102992192)))]; + tensor var_1365_cast_fp16 = add(x = q_25_cast_fp16, y = encoder_layers_4_self_attn_pos_bias_u_to_fp16)[name = string("op_1365_cast_fp16")]; + tensor encoder_layers_4_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_4_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102994304)))]; + tensor var_1367_cast_fp16 = add(x = q_25_cast_fp16, y = encoder_layers_4_self_attn_pos_bias_v_to_fp16)[name = string("op_1367_cast_fp16")]; + tensor q_with_bias_v_9_perm_0 = const()[name = string("q_with_bias_v_9_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_111_transpose_x_0 = const()[name = string("x_111_transpose_x_0"), val = bool(false)]; + bool x_111_transpose_y_0 = const()[name = string("x_111_transpose_y_0"), val = bool(false)]; + tensor op_1369_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102996416))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103110144))))[name = string("op_1369_to_fp16_quantized")]; + tensor q_with_bias_v_9_cast_fp16 = transpose(perm = q_with_bias_v_9_perm_0, x = var_1367_cast_fp16)[name = string("transpose_326")]; + tensor x_111_cast_fp16 = matmul(transpose_x = x_111_transpose_x_0, transpose_y = x_111_transpose_y_0, x = q_with_bias_v_9_cast_fp16, y = op_1369_to_fp16_quantized)[name = string("x_111_cast_fp16")]; + tensor x_113_pad_0 = const()[name = string("x_113_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_113_mode_0 = const()[name = string("x_113_mode_0"), val = string("constant")]; + fp16 const_131_to_fp16 = const()[name = string("const_131_to_fp16"), val = fp16(0x0p+0)]; + tensor x_113_cast_fp16 = pad(constant_val = const_131_to_fp16, mode = x_113_mode_0, pad = x_113_pad_0, x = x_111_cast_fp16)[name = string("x_113_cast_fp16")]; + tensor var_1377 = const()[name = string("op_1377"), val = tensor([1, 8, -1, 14])]; + tensor x_115_cast_fp16 = reshape(shape = var_1377, x = x_113_cast_fp16)[name = string("x_115_cast_fp16")]; + tensor var_1381_begin_0 = const()[name = string("op_1381_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_1381_end_0 = const()[name = string("op_1381_end_0"), val = tensor([1, 8, 112, 14])]; + tensor var_1381_end_mask_0 = const()[name = string("op_1381_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_1381_cast_fp16 = slice_by_index(begin = var_1381_begin_0, end = var_1381_end_0, end_mask = var_1381_end_mask_0, x = x_115_cast_fp16)[name = string("op_1381_cast_fp16")]; + tensor var_1382 = const()[name = string("op_1382"), val = tensor([1, 8, 14, 111])]; + tensor matrix_bd_17_cast_fp16 = reshape(shape = var_1382, x = var_1381_cast_fp16)[name = string("matrix_bd_17_cast_fp16")]; + bool matrix_ac_9_transpose_x_0 = const()[name = string("matrix_ac_9_transpose_x_0"), val = bool(false)]; + bool matrix_ac_9_transpose_y_0 = const()[name = string("matrix_ac_9_transpose_y_0"), val = bool(false)]; + tensor transpose_104_perm_0 = const()[name = string("transpose_104_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_105_perm_0 = const()[name = string("transpose_105_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_105 = transpose(perm = transpose_105_perm_0, x = k_17_cast_fp16)[name = string("transpose_324")]; + tensor transpose_104 = transpose(perm = transpose_104_perm_0, x = var_1365_cast_fp16)[name = string("transpose_325")]; + tensor matrix_ac_9_cast_fp16 = matmul(transpose_x = matrix_ac_9_transpose_x_0, transpose_y = matrix_ac_9_transpose_y_0, x = transpose_104, y = transpose_105)[name = string("matrix_ac_9_cast_fp16")]; + tensor matrix_bd_19_begin_0 = const()[name = string("matrix_bd_19_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_19_end_0 = const()[name = string("matrix_bd_19_end_0"), val = tensor([1, 8, 14, 56])]; + tensor matrix_bd_19_end_mask_0 = const()[name = string("matrix_bd_19_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_19_cast_fp16 = slice_by_index(begin = matrix_bd_19_begin_0, end = matrix_bd_19_end_0, end_mask = matrix_bd_19_end_mask_0, x = matrix_bd_17_cast_fp16)[name = string("matrix_bd_19_cast_fp16")]; + tensor var_1391_cast_fp16 = add(x = matrix_ac_9_cast_fp16, y = matrix_bd_19_cast_fp16)[name = string("op_1391_cast_fp16")]; + fp16 _inversed_scores_17_y_0_to_fp16 = const()[name = string("_inversed_scores_17_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_17_cast_fp16 = mul(x = var_1391_cast_fp16, y = _inversed_scores_17_y_0_to_fp16)[name = string("_inversed_scores_17_cast_fp16")]; + tensor scores_19_cast_fp16 = select(a = var_46_to_fp16, b = _inversed_scores_17_cast_fp16, cond = mask_11)[name = string("scores_19_cast_fp16")]; + tensor var_1397_cast_fp16 = softmax(axis = var_60, x = scores_19_cast_fp16)[name = string("op_1397_cast_fp16")]; + tensor input_249_cast_fp16 = select(a = var_45_to_fp16, b = var_1397_cast_fp16, cond = mask_11)[name = string("input_249_cast_fp16")]; + bool x_117_transpose_x_0 = const()[name = string("x_117_transpose_x_0"), val = bool(false)]; + bool x_117_transpose_y_0 = const()[name = string("x_117_transpose_y_0"), val = bool(false)]; + tensor value_17_cast_fp16 = transpose(perm = value_17_perm_0, x = v_9_cast_fp16)[name = string("transpose_323")]; + tensor x_117_cast_fp16 = matmul(transpose_x = x_117_transpose_x_0, transpose_y = x_117_transpose_y_0, x = input_249_cast_fp16, y = value_17_cast_fp16)[name = string("x_117_cast_fp16")]; + tensor var_1401_perm_0 = const()[name = string("op_1401_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1402 = const()[name = string("op_1402"), val = tensor([1, -1, 1024])]; + tensor var_1401_cast_fp16 = transpose(perm = var_1401_perm_0, x = x_117_cast_fp16)[name = string("transpose_322")]; + tensor input_251_cast_fp16 = reshape(shape = var_1402, x = var_1401_cast_fp16)[name = string("input_251_cast_fp16")]; + tensor encoder_layers_4_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103110464))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103896960))))[name = string("encoder_layers_4_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_4_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_4_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103897152)))]; + tensor linear_43_cast_fp16 = linear(bias = encoder_layers_4_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_4_self_attn_linear_out_weight_to_fp16_palettized, x = input_251_cast_fp16)[name = string("linear_43_cast_fp16")]; + tensor input_255_cast_fp16 = add(x = input_245_cast_fp16, y = linear_43_cast_fp16)[name = string("input_255_cast_fp16")]; + tensor x_121_axes_0 = const()[name = string("x_121_axes_0"), val = tensor([-1])]; + tensor encoder_layers_4_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_4_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103899264)))]; + tensor encoder_layers_4_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_4_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103901376)))]; + tensor x_121_cast_fp16 = layer_norm(axes = x_121_axes_0, beta = encoder_layers_4_norm_conv_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_4_norm_conv_weight_to_fp16, x = input_255_cast_fp16)[name = string("x_121_cast_fp16")]; + tensor input_257_perm_0 = const()[name = string("input_257_perm_0"), val = tensor([0, 2, 1])]; + string input_259_pad_type_0 = const()[name = string("input_259_pad_type_0"), val = string("valid")]; + tensor input_259_strides_0 = const()[name = string("input_259_strides_0"), val = tensor([1])]; + tensor input_259_pad_0 = const()[name = string("input_259_pad_0"), val = tensor([0, 0])]; + tensor input_259_dilations_0 = const()[name = string("input_259_dilations_0"), val = tensor([1])]; + int32 input_259_groups_0 = const()[name = string("input_259_groups_0"), val = int32(1)]; + tensor encoder_layers_4_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103903488))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106000704))))[name = string("encoder_layers_4_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_257_cast_fp16 = transpose(perm = input_257_perm_0, x = x_121_cast_fp16)[name = string("transpose_321")]; + tensor input_259_cast_fp16 = conv(dilations = input_259_dilations_0, groups = input_259_groups_0, pad = input_259_pad_0, pad_type = input_259_pad_type_0, strides = input_259_strides_0, weight = encoder_layers_4_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_257_cast_fp16)[name = string("input_259_cast_fp16")]; + int32 x_123_split_num_splits_0 = const()[name = string("x_123_split_num_splits_0"), val = int32(2)]; + int32 x_123_split_axis_0 = const()[name = string("x_123_split_axis_0"), val = int32(1)]; + tensor x_123_split_cast_fp16_0, tensor x_123_split_cast_fp16_1 = split(axis = x_123_split_axis_0, num_splits = x_123_split_num_splits_0, x = input_259_cast_fp16)[name = string("x_123_split_cast_fp16")]; + tensor x_123_split_1_sigmoid_cast_fp16 = sigmoid(x = x_123_split_cast_fp16_1)[name = string("x_123_split_1_sigmoid_cast_fp16")]; + tensor x_123_cast_fp16 = mul(x = x_123_split_cast_fp16_0, y = x_123_split_1_sigmoid_cast_fp16)[name = string("x_123_cast_fp16")]; + tensor input_261_cast_fp16 = select(a = var_45_to_fp16, b = x_123_cast_fp16, cond = var_576)[name = string("input_261_cast_fp16")]; + bool new_x_19_interleave_0 = const()[name = string("new_x_19_interleave_0"), val = bool(false)]; + tensor new_x_19_cast_fp16 = concat(axis = var_60, interleave = new_x_19_interleave_0, values = (cache_19_cast_fp16, input_261_cast_fp16))[name = string("new_x_19_cast_fp16")]; + tensor var_1441_begin_0 = const()[name = string("op_1441_begin_0"), val = tensor([0, 0, 14])]; + tensor var_1441_end_0 = const()[name = string("op_1441_end_0"), val = tensor([1, 1024, 22])]; + tensor var_1441_end_mask_0 = const()[name = string("op_1441_end_mask_0"), val = tensor([true, true, true])]; + tensor var_1441_cast_fp16 = slice_by_index(begin = var_1441_begin_0, end = var_1441_end_0, end_mask = var_1441_end_mask_0, x = new_x_19_cast_fp16)[name = string("op_1441_cast_fp16")]; + string x_125_pad_type_0 = const()[name = string("x_125_pad_type_0"), val = string("valid")]; + int32 x_125_groups_0 = const()[name = string("x_125_groups_0"), val = int32(1024)]; + tensor x_125_strides_0 = const()[name = string("x_125_strides_0"), val = tensor([1])]; + tensor x_125_pad_0 = const()[name = string("x_125_pad_0"), val = tensor([0, 0])]; + tensor x_125_dilations_0 = const()[name = string("x_125_dilations_0"), val = tensor([1])]; + tensor encoder_layers_4_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106004864))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106014144))))[name = string("encoder_layers_4_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_125_cast_fp16 = conv(dilations = x_125_dilations_0, groups = x_125_groups_0, pad = x_125_pad_0, pad_type = x_125_pad_type_0, strides = x_125_strides_0, weight = encoder_layers_4_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_19_cast_fp16)[name = string("x_125_cast_fp16")]; + tensor input_263_perm_0 = const()[name = string("input_263_perm_0"), val = tensor([0, 2, 1])]; + tensor x_127_axes_0 = const()[name = string("x_127_axes_0"), val = tensor([-1])]; + tensor encoder_layers_4_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_4_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106016256)))]; + tensor encoder_layers_4_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_4_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106018368)))]; + tensor input_263_cast_fp16 = transpose(perm = input_263_perm_0, x = x_125_cast_fp16)[name = string("transpose_320")]; + tensor x_127_cast_fp16 = layer_norm(axes = x_127_axes_0, beta = encoder_layers_4_conv_batch_norm_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_4_conv_batch_norm_weight_to_fp16, x = input_263_cast_fp16)[name = string("x_127_cast_fp16")]; + tensor input_265_perm_0 = const()[name = string("input_265_perm_0"), val = tensor([0, 2, 1])]; + tensor input_265_cast_fp16 = transpose(perm = input_265_perm_0, x = x_127_cast_fp16)[name = string("transpose_319")]; + tensor input_267_cast_fp16 = silu(x = input_265_cast_fp16)[name = string("input_267_cast_fp16")]; + string x_129_pad_type_0 = const()[name = string("x_129_pad_type_0"), val = string("valid")]; + tensor x_129_strides_0 = const()[name = string("x_129_strides_0"), val = tensor([1])]; + tensor x_129_pad_0 = const()[name = string("x_129_pad_0"), val = tensor([0, 0])]; + tensor x_129_dilations_0 = const()[name = string("x_129_dilations_0"), val = tensor([1])]; + int32 x_129_groups_0 = const()[name = string("x_129_groups_0"), val = int32(1)]; + tensor encoder_layers_4_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106020480))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107069120))))[name = string("encoder_layers_4_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_129_cast_fp16 = conv(dilations = x_129_dilations_0, groups = x_129_groups_0, pad = x_129_pad_0, pad_type = x_129_pad_type_0, strides = x_129_strides_0, weight = encoder_layers_4_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_267_cast_fp16)[name = string("x_129_cast_fp16")]; + tensor input_269_perm_0 = const()[name = string("input_269_perm_0"), val = tensor([0, 2, 1])]; + tensor input_269_cast_fp16 = transpose(perm = input_269_perm_0, x = x_129_cast_fp16)[name = string("transpose_318")]; + tensor input_271_cast_fp16 = add(x = input_255_cast_fp16, y = input_269_cast_fp16)[name = string("input_271_cast_fp16")]; + tensor input_273_axes_0 = const()[name = string("input_273_axes_0"), val = tensor([-1])]; + tensor encoder_layers_4_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_4_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107071232)))]; + tensor encoder_layers_4_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_4_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107073344)))]; + tensor input_273_cast_fp16 = layer_norm(axes = input_273_axes_0, beta = encoder_layers_4_norm_feed_forward2_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_4_norm_feed_forward2_weight_to_fp16, x = input_271_cast_fp16)[name = string("input_273_cast_fp16")]; + tensor encoder_layers_4_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107075456))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110221248))))[name = string("encoder_layers_4_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_4_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_4_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110221440)))]; + tensor linear_44_cast_fp16 = linear(bias = encoder_layers_4_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_4_feed_forward2_linear1_weight_to_fp16_palettized, x = input_273_cast_fp16)[name = string("linear_44_cast_fp16")]; + tensor input_277_cast_fp16 = silu(x = linear_44_cast_fp16)[name = string("input_277_cast_fp16")]; + tensor encoder_layers_4_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110229696))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113375488))))[name = string("encoder_layers_4_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_4_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_4_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113375680)))]; + tensor linear_45_cast_fp16 = linear(bias = encoder_layers_4_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_4_feed_forward2_linear2_weight_to_fp16_palettized, x = input_277_cast_fp16)[name = string("linear_45_cast_fp16")]; + fp16 var_1484_to_fp16 = const()[name = string("op_1484_to_fp16"), val = fp16(0x1p-1)]; + tensor var_1485_cast_fp16 = mul(x = linear_45_cast_fp16, y = var_1484_to_fp16)[name = string("op_1485_cast_fp16")]; + tensor input_283_cast_fp16 = add(x = input_271_cast_fp16, y = var_1485_cast_fp16)[name = string("input_283_cast_fp16")]; + tensor input_285_axes_0 = const()[name = string("input_285_axes_0"), val = tensor([-1])]; + tensor encoder_layers_4_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_4_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113377792)))]; + tensor encoder_layers_4_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_4_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113379904)))]; + tensor input_285_cast_fp16 = layer_norm(axes = input_285_axes_0, beta = encoder_layers_4_norm_out_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_4_norm_out_weight_to_fp16, x = input_283_cast_fp16)[name = string("input_285_cast_fp16")]; + tensor cache_21_begin_0 = const()[name = string("cache_21_begin_0"), val = tensor([5, 0, 0, 0])]; + tensor cache_21_end_0 = const()[name = string("cache_21_end_0"), val = tensor([6, 1, 42, 1024])]; + tensor cache_21_end_mask_0 = const()[name = string("cache_21_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_21_squeeze_mask_0 = const()[name = string("cache_21_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_21_cast_fp16 = slice_by_index(begin = cache_21_begin_0, end = cache_21_end_0, end_mask = cache_21_end_mask_0, squeeze_mask = cache_21_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_21_cast_fp16")]; + tensor cache_23_begin_0 = const()[name = string("cache_23_begin_0"), val = tensor([5, 0, 0, 0])]; + tensor cache_23_end_0 = const()[name = string("cache_23_end_0"), val = tensor([6, 1, 1024, 8])]; + tensor cache_23_end_mask_0 = const()[name = string("cache_23_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_23_squeeze_mask_0 = const()[name = string("cache_23_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_23_cast_fp16 = slice_by_index(begin = cache_23_begin_0, end = cache_23_end_0, end_mask = cache_23_end_mask_0, squeeze_mask = cache_23_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_23_cast_fp16")]; + tensor input_287_axes_0 = const()[name = string("input_287_axes_0"), val = tensor([-1])]; + tensor encoder_layers_5_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_5_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113382016)))]; + tensor encoder_layers_5_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_5_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113384128)))]; + tensor input_287_cast_fp16 = layer_norm(axes = input_287_axes_0, beta = encoder_layers_5_norm_feed_forward1_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_5_norm_feed_forward1_weight_to_fp16, x = input_285_cast_fp16)[name = string("input_287_cast_fp16")]; + tensor encoder_layers_5_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113386240))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116532032))))[name = string("encoder_layers_5_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_5_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_5_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116532224)))]; + tensor linear_46_cast_fp16 = linear(bias = encoder_layers_5_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_5_feed_forward1_linear1_weight_to_fp16_palettized, x = input_287_cast_fp16)[name = string("linear_46_cast_fp16")]; + tensor input_291_cast_fp16 = silu(x = linear_46_cast_fp16)[name = string("input_291_cast_fp16")]; + tensor encoder_layers_5_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116540480))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119686272))))[name = string("encoder_layers_5_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_5_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_5_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119686464)))]; + tensor linear_47_cast_fp16 = linear(bias = encoder_layers_5_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_5_feed_forward1_linear2_weight_to_fp16_palettized, x = input_291_cast_fp16)[name = string("linear_47_cast_fp16")]; + fp16 var_1521_to_fp16 = const()[name = string("op_1521_to_fp16"), val = fp16(0x1p-1)]; + tensor var_1522_cast_fp16 = mul(x = linear_47_cast_fp16, y = var_1521_to_fp16)[name = string("op_1522_cast_fp16")]; + tensor input_297_cast_fp16 = add(x = input_285_cast_fp16, y = var_1522_cast_fp16)[name = string("input_297_cast_fp16")]; + tensor key_11_axes_0 = const()[name = string("key_11_axes_0"), val = tensor([-1])]; + tensor encoder_layers_5_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_5_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119688576)))]; + tensor encoder_layers_5_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_5_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119690688)))]; + tensor key_11_cast_fp16 = layer_norm(axes = key_11_axes_0, beta = encoder_layers_5_norm_self_att_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_5_norm_self_att_weight_to_fp16, x = input_297_cast_fp16)[name = string("key_11_cast_fp16")]; + bool input_299_interleave_0 = const()[name = string("input_299_interleave_0"), val = bool(false)]; + tensor input_299_cast_fp16 = concat(axis = var_69, interleave = input_299_interleave_0, values = (cache_21_cast_fp16, key_11_cast_fp16))[name = string("input_299_cast_fp16")]; + tensor var_1544_begin_0 = const()[name = string("op_1544_begin_0"), val = tensor([0, 14, 0])]; + tensor var_1544_end_0 = const()[name = string("op_1544_end_0"), val = tensor([1, 42, 1024])]; + tensor var_1544_end_mask_0 = const()[name = string("op_1544_end_mask_0"), val = tensor([true, true, true])]; + tensor var_1544_cast_fp16 = slice_by_index(begin = var_1544_begin_0, end = var_1544_end_0, end_mask = var_1544_end_mask_0, x = cache_21_cast_fp16)[name = string("op_1544_cast_fp16")]; + bool var_1550_interleave_0 = const()[name = string("op_1550_interleave_0"), val = bool(false)]; + tensor var_1550_cast_fp16 = concat(axis = var_69, interleave = var_1550_interleave_0, values = (var_1544_cast_fp16, key_11_cast_fp16))[name = string("op_1550_cast_fp16")]; + tensor encoder_layers_5_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119692800))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120479296))))[name = string("encoder_layers_5_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_5_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_5_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120479488)))]; + tensor linear_48_cast_fp16 = linear(bias = encoder_layers_5_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_5_self_attn_linear_q_weight_to_fp16_palettized, x = key_11_cast_fp16)[name = string("linear_48_cast_fp16")]; + tensor var_1555 = const()[name = string("op_1555"), val = tensor([1, -1, 8, 128])]; + tensor q_31_cast_fp16 = reshape(shape = var_1555, x = linear_48_cast_fp16)[name = string("q_31_cast_fp16")]; + tensor encoder_layers_5_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120481600))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121268096))))[name = string("encoder_layers_5_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_5_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_5_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121268288)))]; + tensor linear_49_cast_fp16 = linear(bias = encoder_layers_5_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_5_self_attn_linear_k_weight_to_fp16_palettized, x = input_299_cast_fp16)[name = string("linear_49_cast_fp16")]; + tensor var_1560 = const()[name = string("op_1560"), val = tensor([1, -1, 8, 128])]; + tensor k_21_cast_fp16 = reshape(shape = var_1560, x = linear_49_cast_fp16)[name = string("k_21_cast_fp16")]; + tensor encoder_layers_5_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121270400))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122056896))))[name = string("encoder_layers_5_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_5_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_5_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122057088)))]; + tensor linear_50_cast_fp16 = linear(bias = encoder_layers_5_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_5_self_attn_linear_v_weight_to_fp16_palettized, x = input_299_cast_fp16)[name = string("linear_50_cast_fp16")]; + tensor var_1565 = const()[name = string("op_1565"), val = tensor([1, -1, 8, 128])]; + tensor v_11_cast_fp16 = reshape(shape = var_1565, x = linear_50_cast_fp16)[name = string("v_11_cast_fp16")]; + tensor value_19_perm_0 = const()[name = string("value_19_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_5_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_5_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122059200)))]; + tensor var_1578_cast_fp16 = add(x = q_31_cast_fp16, y = encoder_layers_5_self_attn_pos_bias_u_to_fp16)[name = string("op_1578_cast_fp16")]; + tensor encoder_layers_5_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_5_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122061312)))]; + tensor var_1580_cast_fp16 = add(x = q_31_cast_fp16, y = encoder_layers_5_self_attn_pos_bias_v_to_fp16)[name = string("op_1580_cast_fp16")]; + tensor q_with_bias_v_11_perm_0 = const()[name = string("q_with_bias_v_11_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_137_transpose_x_0 = const()[name = string("x_137_transpose_x_0"), val = bool(false)]; + bool x_137_transpose_y_0 = const()[name = string("x_137_transpose_y_0"), val = bool(false)]; + tensor op_1582_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122063424))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122177152))))[name = string("op_1582_to_fp16_quantized")]; + tensor q_with_bias_v_11_cast_fp16 = transpose(perm = q_with_bias_v_11_perm_0, x = var_1580_cast_fp16)[name = string("transpose_317")]; + tensor x_137_cast_fp16 = matmul(transpose_x = x_137_transpose_x_0, transpose_y = x_137_transpose_y_0, x = q_with_bias_v_11_cast_fp16, y = op_1582_to_fp16_quantized)[name = string("x_137_cast_fp16")]; + tensor x_139_pad_0 = const()[name = string("x_139_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_139_mode_0 = const()[name = string("x_139_mode_0"), val = string("constant")]; + fp16 const_144_to_fp16 = const()[name = string("const_144_to_fp16"), val = fp16(0x0p+0)]; + tensor x_139_cast_fp16 = pad(constant_val = const_144_to_fp16, mode = x_139_mode_0, pad = x_139_pad_0, x = x_137_cast_fp16)[name = string("x_139_cast_fp16")]; + tensor var_1590 = const()[name = string("op_1590"), val = tensor([1, 8, -1, 14])]; + tensor x_141_cast_fp16 = reshape(shape = var_1590, x = x_139_cast_fp16)[name = string("x_141_cast_fp16")]; + tensor var_1594_begin_0 = const()[name = string("op_1594_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_1594_end_0 = const()[name = string("op_1594_end_0"), val = tensor([1, 8, 112, 14])]; + tensor var_1594_end_mask_0 = const()[name = string("op_1594_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_1594_cast_fp16 = slice_by_index(begin = var_1594_begin_0, end = var_1594_end_0, end_mask = var_1594_end_mask_0, x = x_141_cast_fp16)[name = string("op_1594_cast_fp16")]; + tensor var_1595 = const()[name = string("op_1595"), val = tensor([1, 8, 14, 111])]; + tensor matrix_bd_21_cast_fp16 = reshape(shape = var_1595, x = var_1594_cast_fp16)[name = string("matrix_bd_21_cast_fp16")]; + bool matrix_ac_11_transpose_x_0 = const()[name = string("matrix_ac_11_transpose_x_0"), val = bool(false)]; + bool matrix_ac_11_transpose_y_0 = const()[name = string("matrix_ac_11_transpose_y_0"), val = bool(false)]; + tensor transpose_106_perm_0 = const()[name = string("transpose_106_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_107_perm_0 = const()[name = string("transpose_107_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_107 = transpose(perm = transpose_107_perm_0, x = k_21_cast_fp16)[name = string("transpose_315")]; + tensor transpose_106 = transpose(perm = transpose_106_perm_0, x = var_1578_cast_fp16)[name = string("transpose_316")]; + tensor matrix_ac_11_cast_fp16 = matmul(transpose_x = matrix_ac_11_transpose_x_0, transpose_y = matrix_ac_11_transpose_y_0, x = transpose_106, y = transpose_107)[name = string("matrix_ac_11_cast_fp16")]; + tensor matrix_bd_23_begin_0 = const()[name = string("matrix_bd_23_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_23_end_0 = const()[name = string("matrix_bd_23_end_0"), val = tensor([1, 8, 14, 56])]; + tensor matrix_bd_23_end_mask_0 = const()[name = string("matrix_bd_23_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_23_cast_fp16 = slice_by_index(begin = matrix_bd_23_begin_0, end = matrix_bd_23_end_0, end_mask = matrix_bd_23_end_mask_0, x = matrix_bd_21_cast_fp16)[name = string("matrix_bd_23_cast_fp16")]; + tensor var_1604_cast_fp16 = add(x = matrix_ac_11_cast_fp16, y = matrix_bd_23_cast_fp16)[name = string("op_1604_cast_fp16")]; + fp16 _inversed_scores_21_y_0_to_fp16 = const()[name = string("_inversed_scores_21_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_21_cast_fp16 = mul(x = var_1604_cast_fp16, y = _inversed_scores_21_y_0_to_fp16)[name = string("_inversed_scores_21_cast_fp16")]; + tensor scores_23_cast_fp16 = select(a = var_46_to_fp16, b = _inversed_scores_21_cast_fp16, cond = mask_11)[name = string("scores_23_cast_fp16")]; + tensor var_1610_cast_fp16 = softmax(axis = var_60, x = scores_23_cast_fp16)[name = string("op_1610_cast_fp16")]; + tensor input_301_cast_fp16 = select(a = var_45_to_fp16, b = var_1610_cast_fp16, cond = mask_11)[name = string("input_301_cast_fp16")]; + bool x_143_transpose_x_0 = const()[name = string("x_143_transpose_x_0"), val = bool(false)]; + bool x_143_transpose_y_0 = const()[name = string("x_143_transpose_y_0"), val = bool(false)]; + tensor value_19_cast_fp16 = transpose(perm = value_19_perm_0, x = v_11_cast_fp16)[name = string("transpose_314")]; + tensor x_143_cast_fp16 = matmul(transpose_x = x_143_transpose_x_0, transpose_y = x_143_transpose_y_0, x = input_301_cast_fp16, y = value_19_cast_fp16)[name = string("x_143_cast_fp16")]; + tensor var_1614_perm_0 = const()[name = string("op_1614_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1615 = const()[name = string("op_1615"), val = tensor([1, -1, 1024])]; + tensor var_1614_cast_fp16 = transpose(perm = var_1614_perm_0, x = x_143_cast_fp16)[name = string("transpose_313")]; + tensor input_303_cast_fp16 = reshape(shape = var_1615, x = var_1614_cast_fp16)[name = string("input_303_cast_fp16")]; + tensor encoder_layers_5_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122177472))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122963968))))[name = string("encoder_layers_5_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_5_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_5_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122964160)))]; + tensor linear_52_cast_fp16 = linear(bias = encoder_layers_5_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_5_self_attn_linear_out_weight_to_fp16_palettized, x = input_303_cast_fp16)[name = string("linear_52_cast_fp16")]; + tensor input_307_cast_fp16 = add(x = input_297_cast_fp16, y = linear_52_cast_fp16)[name = string("input_307_cast_fp16")]; + tensor x_147_axes_0 = const()[name = string("x_147_axes_0"), val = tensor([-1])]; + tensor encoder_layers_5_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_5_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122966272)))]; + tensor encoder_layers_5_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_5_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122968384)))]; + tensor x_147_cast_fp16 = layer_norm(axes = x_147_axes_0, beta = encoder_layers_5_norm_conv_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_5_norm_conv_weight_to_fp16, x = input_307_cast_fp16)[name = string("x_147_cast_fp16")]; + tensor input_309_perm_0 = const()[name = string("input_309_perm_0"), val = tensor([0, 2, 1])]; + string input_311_pad_type_0 = const()[name = string("input_311_pad_type_0"), val = string("valid")]; + tensor input_311_strides_0 = const()[name = string("input_311_strides_0"), val = tensor([1])]; + tensor input_311_pad_0 = const()[name = string("input_311_pad_0"), val = tensor([0, 0])]; + tensor input_311_dilations_0 = const()[name = string("input_311_dilations_0"), val = tensor([1])]; + int32 input_311_groups_0 = const()[name = string("input_311_groups_0"), val = int32(1)]; + tensor encoder_layers_5_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122970496))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125067712))))[name = string("encoder_layers_5_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_309_cast_fp16 = transpose(perm = input_309_perm_0, x = x_147_cast_fp16)[name = string("transpose_312")]; + tensor input_311_cast_fp16 = conv(dilations = input_311_dilations_0, groups = input_311_groups_0, pad = input_311_pad_0, pad_type = input_311_pad_type_0, strides = input_311_strides_0, weight = encoder_layers_5_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_309_cast_fp16)[name = string("input_311_cast_fp16")]; + int32 x_149_split_num_splits_0 = const()[name = string("x_149_split_num_splits_0"), val = int32(2)]; + int32 x_149_split_axis_0 = const()[name = string("x_149_split_axis_0"), val = int32(1)]; + tensor x_149_split_cast_fp16_0, tensor x_149_split_cast_fp16_1 = split(axis = x_149_split_axis_0, num_splits = x_149_split_num_splits_0, x = input_311_cast_fp16)[name = string("x_149_split_cast_fp16")]; + tensor x_149_split_1_sigmoid_cast_fp16 = sigmoid(x = x_149_split_cast_fp16_1)[name = string("x_149_split_1_sigmoid_cast_fp16")]; + tensor x_149_cast_fp16 = mul(x = x_149_split_cast_fp16_0, y = x_149_split_1_sigmoid_cast_fp16)[name = string("x_149_cast_fp16")]; + tensor input_313_cast_fp16 = select(a = var_45_to_fp16, b = x_149_cast_fp16, cond = var_576)[name = string("input_313_cast_fp16")]; + bool new_x_23_interleave_0 = const()[name = string("new_x_23_interleave_0"), val = bool(false)]; + tensor new_x_23_cast_fp16 = concat(axis = var_60, interleave = new_x_23_interleave_0, values = (cache_23_cast_fp16, input_313_cast_fp16))[name = string("new_x_23_cast_fp16")]; + tensor var_1654_begin_0 = const()[name = string("op_1654_begin_0"), val = tensor([0, 0, 14])]; + tensor var_1654_end_0 = const()[name = string("op_1654_end_0"), val = tensor([1, 1024, 22])]; + tensor var_1654_end_mask_0 = const()[name = string("op_1654_end_mask_0"), val = tensor([true, true, true])]; + tensor var_1654_cast_fp16 = slice_by_index(begin = var_1654_begin_0, end = var_1654_end_0, end_mask = var_1654_end_mask_0, x = new_x_23_cast_fp16)[name = string("op_1654_cast_fp16")]; + string x_151_pad_type_0 = const()[name = string("x_151_pad_type_0"), val = string("valid")]; + int32 x_151_groups_0 = const()[name = string("x_151_groups_0"), val = int32(1024)]; + tensor x_151_strides_0 = const()[name = string("x_151_strides_0"), val = tensor([1])]; + tensor x_151_pad_0 = const()[name = string("x_151_pad_0"), val = tensor([0, 0])]; + tensor x_151_dilations_0 = const()[name = string("x_151_dilations_0"), val = tensor([1])]; + tensor encoder_layers_5_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125071872))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125081152))))[name = string("encoder_layers_5_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_151_cast_fp16 = conv(dilations = x_151_dilations_0, groups = x_151_groups_0, pad = x_151_pad_0, pad_type = x_151_pad_type_0, strides = x_151_strides_0, weight = encoder_layers_5_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_23_cast_fp16)[name = string("x_151_cast_fp16")]; + tensor input_315_perm_0 = const()[name = string("input_315_perm_0"), val = tensor([0, 2, 1])]; + tensor x_153_axes_0 = const()[name = string("x_153_axes_0"), val = tensor([-1])]; + tensor encoder_layers_5_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_5_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125083264)))]; + tensor encoder_layers_5_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_5_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125085376)))]; + tensor input_315_cast_fp16 = transpose(perm = input_315_perm_0, x = x_151_cast_fp16)[name = string("transpose_311")]; + tensor x_153_cast_fp16 = layer_norm(axes = x_153_axes_0, beta = encoder_layers_5_conv_batch_norm_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_5_conv_batch_norm_weight_to_fp16, x = input_315_cast_fp16)[name = string("x_153_cast_fp16")]; + tensor input_317_perm_0 = const()[name = string("input_317_perm_0"), val = tensor([0, 2, 1])]; + tensor input_317_cast_fp16 = transpose(perm = input_317_perm_0, x = x_153_cast_fp16)[name = string("transpose_310")]; + tensor input_319_cast_fp16 = silu(x = input_317_cast_fp16)[name = string("input_319_cast_fp16")]; + string x_155_pad_type_0 = const()[name = string("x_155_pad_type_0"), val = string("valid")]; + tensor x_155_strides_0 = const()[name = string("x_155_strides_0"), val = tensor([1])]; + tensor x_155_pad_0 = const()[name = string("x_155_pad_0"), val = tensor([0, 0])]; + tensor x_155_dilations_0 = const()[name = string("x_155_dilations_0"), val = tensor([1])]; + int32 x_155_groups_0 = const()[name = string("x_155_groups_0"), val = int32(1)]; + tensor encoder_layers_5_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125087488))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126136128))))[name = string("encoder_layers_5_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_155_cast_fp16 = conv(dilations = x_155_dilations_0, groups = x_155_groups_0, pad = x_155_pad_0, pad_type = x_155_pad_type_0, strides = x_155_strides_0, weight = encoder_layers_5_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_319_cast_fp16)[name = string("x_155_cast_fp16")]; + tensor input_321_perm_0 = const()[name = string("input_321_perm_0"), val = tensor([0, 2, 1])]; + tensor input_321_cast_fp16 = transpose(perm = input_321_perm_0, x = x_155_cast_fp16)[name = string("transpose_309")]; + tensor input_323_cast_fp16 = add(x = input_307_cast_fp16, y = input_321_cast_fp16)[name = string("input_323_cast_fp16")]; + tensor input_325_axes_0 = const()[name = string("input_325_axes_0"), val = tensor([-1])]; + tensor encoder_layers_5_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_5_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126138240)))]; + tensor encoder_layers_5_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_5_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126140352)))]; + tensor input_325_cast_fp16 = layer_norm(axes = input_325_axes_0, beta = encoder_layers_5_norm_feed_forward2_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_5_norm_feed_forward2_weight_to_fp16, x = input_323_cast_fp16)[name = string("input_325_cast_fp16")]; + tensor encoder_layers_5_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126142464))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(129288256))))[name = string("encoder_layers_5_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_5_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_5_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(129288448)))]; + tensor linear_53_cast_fp16 = linear(bias = encoder_layers_5_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_5_feed_forward2_linear1_weight_to_fp16_palettized, x = input_325_cast_fp16)[name = string("linear_53_cast_fp16")]; + tensor input_329_cast_fp16 = silu(x = linear_53_cast_fp16)[name = string("input_329_cast_fp16")]; + tensor encoder_layers_5_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(129296704))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132442496))))[name = string("encoder_layers_5_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_5_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_5_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132442688)))]; + tensor linear_54_cast_fp16 = linear(bias = encoder_layers_5_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_5_feed_forward2_linear2_weight_to_fp16_palettized, x = input_329_cast_fp16)[name = string("linear_54_cast_fp16")]; + fp16 var_1697_to_fp16 = const()[name = string("op_1697_to_fp16"), val = fp16(0x1p-1)]; + tensor var_1698_cast_fp16 = mul(x = linear_54_cast_fp16, y = var_1697_to_fp16)[name = string("op_1698_cast_fp16")]; + tensor input_335_cast_fp16 = add(x = input_323_cast_fp16, y = var_1698_cast_fp16)[name = string("input_335_cast_fp16")]; + tensor input_337_axes_0 = const()[name = string("input_337_axes_0"), val = tensor([-1])]; + tensor encoder_layers_5_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_5_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132444800)))]; + tensor encoder_layers_5_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_5_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132446912)))]; + tensor input_337_cast_fp16 = layer_norm(axes = input_337_axes_0, beta = encoder_layers_5_norm_out_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_5_norm_out_weight_to_fp16, x = input_335_cast_fp16)[name = string("input_337_cast_fp16")]; + tensor cache_25_begin_0 = const()[name = string("cache_25_begin_0"), val = tensor([6, 0, 0, 0])]; + tensor cache_25_end_0 = const()[name = string("cache_25_end_0"), val = tensor([7, 1, 42, 1024])]; + tensor cache_25_end_mask_0 = const()[name = string("cache_25_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_25_squeeze_mask_0 = const()[name = string("cache_25_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_25_cast_fp16 = slice_by_index(begin = cache_25_begin_0, end = cache_25_end_0, end_mask = cache_25_end_mask_0, squeeze_mask = cache_25_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_25_cast_fp16")]; + tensor cache_27_begin_0 = const()[name = string("cache_27_begin_0"), val = tensor([6, 0, 0, 0])]; + tensor cache_27_end_0 = const()[name = string("cache_27_end_0"), val = tensor([7, 1, 1024, 8])]; + tensor cache_27_end_mask_0 = const()[name = string("cache_27_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_27_squeeze_mask_0 = const()[name = string("cache_27_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_27_cast_fp16 = slice_by_index(begin = cache_27_begin_0, end = cache_27_end_0, end_mask = cache_27_end_mask_0, squeeze_mask = cache_27_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_27_cast_fp16")]; + tensor input_339_axes_0 = const()[name = string("input_339_axes_0"), val = tensor([-1])]; + tensor encoder_layers_6_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_6_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132449024)))]; + tensor encoder_layers_6_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_6_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132451136)))]; + tensor input_339_cast_fp16 = layer_norm(axes = input_339_axes_0, beta = encoder_layers_6_norm_feed_forward1_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_6_norm_feed_forward1_weight_to_fp16, x = input_337_cast_fp16)[name = string("input_339_cast_fp16")]; + tensor encoder_layers_6_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132453248))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135599040))))[name = string("encoder_layers_6_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_6_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_6_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135599232)))]; + tensor linear_55_cast_fp16 = linear(bias = encoder_layers_6_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_6_feed_forward1_linear1_weight_to_fp16_palettized, x = input_339_cast_fp16)[name = string("linear_55_cast_fp16")]; + tensor input_343_cast_fp16 = silu(x = linear_55_cast_fp16)[name = string("input_343_cast_fp16")]; + tensor encoder_layers_6_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135607488))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138753280))))[name = string("encoder_layers_6_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_6_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_6_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138753472)))]; + tensor linear_56_cast_fp16 = linear(bias = encoder_layers_6_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_6_feed_forward1_linear2_weight_to_fp16_palettized, x = input_343_cast_fp16)[name = string("linear_56_cast_fp16")]; + fp16 var_1734_to_fp16 = const()[name = string("op_1734_to_fp16"), val = fp16(0x1p-1)]; + tensor var_1735_cast_fp16 = mul(x = linear_56_cast_fp16, y = var_1734_to_fp16)[name = string("op_1735_cast_fp16")]; + tensor input_349_cast_fp16 = add(x = input_337_cast_fp16, y = var_1735_cast_fp16)[name = string("input_349_cast_fp16")]; + tensor key_13_axes_0 = const()[name = string("key_13_axes_0"), val = tensor([-1])]; + tensor encoder_layers_6_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_6_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138755584)))]; + tensor encoder_layers_6_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_6_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138757696)))]; + tensor key_13_cast_fp16 = layer_norm(axes = key_13_axes_0, beta = encoder_layers_6_norm_self_att_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_6_norm_self_att_weight_to_fp16, x = input_349_cast_fp16)[name = string("key_13_cast_fp16")]; + bool input_351_interleave_0 = const()[name = string("input_351_interleave_0"), val = bool(false)]; + tensor input_351_cast_fp16 = concat(axis = var_69, interleave = input_351_interleave_0, values = (cache_25_cast_fp16, key_13_cast_fp16))[name = string("input_351_cast_fp16")]; + tensor var_1757_begin_0 = const()[name = string("op_1757_begin_0"), val = tensor([0, 14, 0])]; + tensor var_1757_end_0 = const()[name = string("op_1757_end_0"), val = tensor([1, 42, 1024])]; + tensor var_1757_end_mask_0 = const()[name = string("op_1757_end_mask_0"), val = tensor([true, true, true])]; + tensor var_1757_cast_fp16 = slice_by_index(begin = var_1757_begin_0, end = var_1757_end_0, end_mask = var_1757_end_mask_0, x = cache_25_cast_fp16)[name = string("op_1757_cast_fp16")]; + bool var_1763_interleave_0 = const()[name = string("op_1763_interleave_0"), val = bool(false)]; + tensor var_1763_cast_fp16 = concat(axis = var_69, interleave = var_1763_interleave_0, values = (var_1757_cast_fp16, key_13_cast_fp16))[name = string("op_1763_cast_fp16")]; + tensor encoder_layers_6_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138759808))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(139546304))))[name = string("encoder_layers_6_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_6_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_6_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(139546496)))]; + tensor linear_57_cast_fp16 = linear(bias = encoder_layers_6_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_6_self_attn_linear_q_weight_to_fp16_palettized, x = key_13_cast_fp16)[name = string("linear_57_cast_fp16")]; + tensor var_1768 = const()[name = string("op_1768"), val = tensor([1, -1, 8, 128])]; + tensor q_37_cast_fp16 = reshape(shape = var_1768, x = linear_57_cast_fp16)[name = string("q_37_cast_fp16")]; + tensor encoder_layers_6_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(139548608))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140335104))))[name = string("encoder_layers_6_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_6_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_6_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140335296)))]; + tensor linear_58_cast_fp16 = linear(bias = encoder_layers_6_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_6_self_attn_linear_k_weight_to_fp16_palettized, x = input_351_cast_fp16)[name = string("linear_58_cast_fp16")]; + tensor var_1773 = const()[name = string("op_1773"), val = tensor([1, -1, 8, 128])]; + tensor k_25_cast_fp16 = reshape(shape = var_1773, x = linear_58_cast_fp16)[name = string("k_25_cast_fp16")]; + tensor encoder_layers_6_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140337408))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141123904))))[name = string("encoder_layers_6_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_6_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_6_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141124096)))]; + tensor linear_59_cast_fp16 = linear(bias = encoder_layers_6_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_6_self_attn_linear_v_weight_to_fp16_palettized, x = input_351_cast_fp16)[name = string("linear_59_cast_fp16")]; + tensor var_1778 = const()[name = string("op_1778"), val = tensor([1, -1, 8, 128])]; + tensor v_13_cast_fp16 = reshape(shape = var_1778, x = linear_59_cast_fp16)[name = string("v_13_cast_fp16")]; + tensor value_21_perm_0 = const()[name = string("value_21_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_6_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_6_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141126208)))]; + tensor var_1791_cast_fp16 = add(x = q_37_cast_fp16, y = encoder_layers_6_self_attn_pos_bias_u_to_fp16)[name = string("op_1791_cast_fp16")]; + tensor encoder_layers_6_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_6_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141128320)))]; + tensor var_1793_cast_fp16 = add(x = q_37_cast_fp16, y = encoder_layers_6_self_attn_pos_bias_v_to_fp16)[name = string("op_1793_cast_fp16")]; + tensor q_with_bias_v_13_perm_0 = const()[name = string("q_with_bias_v_13_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_163_transpose_x_0 = const()[name = string("x_163_transpose_x_0"), val = bool(false)]; + bool x_163_transpose_y_0 = const()[name = string("x_163_transpose_y_0"), val = bool(false)]; + tensor op_1795_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141130432))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141244160))))[name = string("op_1795_to_fp16_quantized")]; + tensor q_with_bias_v_13_cast_fp16 = transpose(perm = q_with_bias_v_13_perm_0, x = var_1793_cast_fp16)[name = string("transpose_308")]; + tensor x_163_cast_fp16 = matmul(transpose_x = x_163_transpose_x_0, transpose_y = x_163_transpose_y_0, x = q_with_bias_v_13_cast_fp16, y = op_1795_to_fp16_quantized)[name = string("x_163_cast_fp16")]; + tensor x_165_pad_0 = const()[name = string("x_165_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_165_mode_0 = const()[name = string("x_165_mode_0"), val = string("constant")]; + fp16 const_157_to_fp16 = const()[name = string("const_157_to_fp16"), val = fp16(0x0p+0)]; + tensor x_165_cast_fp16 = pad(constant_val = const_157_to_fp16, mode = x_165_mode_0, pad = x_165_pad_0, x = x_163_cast_fp16)[name = string("x_165_cast_fp16")]; + tensor var_1803 = const()[name = string("op_1803"), val = tensor([1, 8, -1, 14])]; + tensor x_167_cast_fp16 = reshape(shape = var_1803, x = x_165_cast_fp16)[name = string("x_167_cast_fp16")]; + tensor var_1807_begin_0 = const()[name = string("op_1807_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_1807_end_0 = const()[name = string("op_1807_end_0"), val = tensor([1, 8, 112, 14])]; + tensor var_1807_end_mask_0 = const()[name = string("op_1807_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_1807_cast_fp16 = slice_by_index(begin = var_1807_begin_0, end = var_1807_end_0, end_mask = var_1807_end_mask_0, x = x_167_cast_fp16)[name = string("op_1807_cast_fp16")]; + tensor var_1808 = const()[name = string("op_1808"), val = tensor([1, 8, 14, 111])]; + tensor matrix_bd_25_cast_fp16 = reshape(shape = var_1808, x = var_1807_cast_fp16)[name = string("matrix_bd_25_cast_fp16")]; + bool matrix_ac_13_transpose_x_0 = const()[name = string("matrix_ac_13_transpose_x_0"), val = bool(false)]; + bool matrix_ac_13_transpose_y_0 = const()[name = string("matrix_ac_13_transpose_y_0"), val = bool(false)]; + tensor transpose_108_perm_0 = const()[name = string("transpose_108_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_109_perm_0 = const()[name = string("transpose_109_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_109 = transpose(perm = transpose_109_perm_0, x = k_25_cast_fp16)[name = string("transpose_306")]; + tensor transpose_108 = transpose(perm = transpose_108_perm_0, x = var_1791_cast_fp16)[name = string("transpose_307")]; + tensor matrix_ac_13_cast_fp16 = matmul(transpose_x = matrix_ac_13_transpose_x_0, transpose_y = matrix_ac_13_transpose_y_0, x = transpose_108, y = transpose_109)[name = string("matrix_ac_13_cast_fp16")]; + tensor matrix_bd_27_begin_0 = const()[name = string("matrix_bd_27_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_27_end_0 = const()[name = string("matrix_bd_27_end_0"), val = tensor([1, 8, 14, 56])]; + tensor matrix_bd_27_end_mask_0 = const()[name = string("matrix_bd_27_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_27_cast_fp16 = slice_by_index(begin = matrix_bd_27_begin_0, end = matrix_bd_27_end_0, end_mask = matrix_bd_27_end_mask_0, x = matrix_bd_25_cast_fp16)[name = string("matrix_bd_27_cast_fp16")]; + tensor var_1817_cast_fp16 = add(x = matrix_ac_13_cast_fp16, y = matrix_bd_27_cast_fp16)[name = string("op_1817_cast_fp16")]; + fp16 _inversed_scores_25_y_0_to_fp16 = const()[name = string("_inversed_scores_25_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_25_cast_fp16 = mul(x = var_1817_cast_fp16, y = _inversed_scores_25_y_0_to_fp16)[name = string("_inversed_scores_25_cast_fp16")]; + tensor scores_27_cast_fp16 = select(a = var_46_to_fp16, b = _inversed_scores_25_cast_fp16, cond = mask_11)[name = string("scores_27_cast_fp16")]; + tensor var_1823_cast_fp16 = softmax(axis = var_60, x = scores_27_cast_fp16)[name = string("op_1823_cast_fp16")]; + tensor input_353_cast_fp16 = select(a = var_45_to_fp16, b = var_1823_cast_fp16, cond = mask_11)[name = string("input_353_cast_fp16")]; + bool x_169_transpose_x_0 = const()[name = string("x_169_transpose_x_0"), val = bool(false)]; + bool x_169_transpose_y_0 = const()[name = string("x_169_transpose_y_0"), val = bool(false)]; + tensor value_21_cast_fp16 = transpose(perm = value_21_perm_0, x = v_13_cast_fp16)[name = string("transpose_305")]; + tensor x_169_cast_fp16 = matmul(transpose_x = x_169_transpose_x_0, transpose_y = x_169_transpose_y_0, x = input_353_cast_fp16, y = value_21_cast_fp16)[name = string("x_169_cast_fp16")]; + tensor var_1827_perm_0 = const()[name = string("op_1827_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1828 = const()[name = string("op_1828"), val = tensor([1, -1, 1024])]; + tensor var_1827_cast_fp16 = transpose(perm = var_1827_perm_0, x = x_169_cast_fp16)[name = string("transpose_304")]; + tensor input_355_cast_fp16 = reshape(shape = var_1828, x = var_1827_cast_fp16)[name = string("input_355_cast_fp16")]; + tensor encoder_layers_6_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141244480))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142030976))))[name = string("encoder_layers_6_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_6_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_6_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142031168)))]; + tensor linear_61_cast_fp16 = linear(bias = encoder_layers_6_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_6_self_attn_linear_out_weight_to_fp16_palettized, x = input_355_cast_fp16)[name = string("linear_61_cast_fp16")]; + tensor input_359_cast_fp16 = add(x = input_349_cast_fp16, y = linear_61_cast_fp16)[name = string("input_359_cast_fp16")]; + tensor x_173_axes_0 = const()[name = string("x_173_axes_0"), val = tensor([-1])]; + tensor encoder_layers_6_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_6_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142033280)))]; + tensor encoder_layers_6_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_6_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142035392)))]; + tensor x_173_cast_fp16 = layer_norm(axes = x_173_axes_0, beta = encoder_layers_6_norm_conv_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_6_norm_conv_weight_to_fp16, x = input_359_cast_fp16)[name = string("x_173_cast_fp16")]; + tensor input_361_perm_0 = const()[name = string("input_361_perm_0"), val = tensor([0, 2, 1])]; + string input_363_pad_type_0 = const()[name = string("input_363_pad_type_0"), val = string("valid")]; + tensor input_363_strides_0 = const()[name = string("input_363_strides_0"), val = tensor([1])]; + tensor input_363_pad_0 = const()[name = string("input_363_pad_0"), val = tensor([0, 0])]; + tensor input_363_dilations_0 = const()[name = string("input_363_dilations_0"), val = tensor([1])]; + int32 input_363_groups_0 = const()[name = string("input_363_groups_0"), val = int32(1)]; + tensor encoder_layers_6_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142037504))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144134720))))[name = string("encoder_layers_6_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_361_cast_fp16 = transpose(perm = input_361_perm_0, x = x_173_cast_fp16)[name = string("transpose_303")]; + tensor input_363_cast_fp16 = conv(dilations = input_363_dilations_0, groups = input_363_groups_0, pad = input_363_pad_0, pad_type = input_363_pad_type_0, strides = input_363_strides_0, weight = encoder_layers_6_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_361_cast_fp16)[name = string("input_363_cast_fp16")]; + int32 x_175_split_num_splits_0 = const()[name = string("x_175_split_num_splits_0"), val = int32(2)]; + int32 x_175_split_axis_0 = const()[name = string("x_175_split_axis_0"), val = int32(1)]; + tensor x_175_split_cast_fp16_0, tensor x_175_split_cast_fp16_1 = split(axis = x_175_split_axis_0, num_splits = x_175_split_num_splits_0, x = input_363_cast_fp16)[name = string("x_175_split_cast_fp16")]; + tensor x_175_split_1_sigmoid_cast_fp16 = sigmoid(x = x_175_split_cast_fp16_1)[name = string("x_175_split_1_sigmoid_cast_fp16")]; + tensor x_175_cast_fp16 = mul(x = x_175_split_cast_fp16_0, y = x_175_split_1_sigmoid_cast_fp16)[name = string("x_175_cast_fp16")]; + tensor input_365_cast_fp16 = select(a = var_45_to_fp16, b = x_175_cast_fp16, cond = var_576)[name = string("input_365_cast_fp16")]; + bool new_x_27_interleave_0 = const()[name = string("new_x_27_interleave_0"), val = bool(false)]; + tensor new_x_27_cast_fp16 = concat(axis = var_60, interleave = new_x_27_interleave_0, values = (cache_27_cast_fp16, input_365_cast_fp16))[name = string("new_x_27_cast_fp16")]; + tensor var_1867_begin_0 = const()[name = string("op_1867_begin_0"), val = tensor([0, 0, 14])]; + tensor var_1867_end_0 = const()[name = string("op_1867_end_0"), val = tensor([1, 1024, 22])]; + tensor var_1867_end_mask_0 = const()[name = string("op_1867_end_mask_0"), val = tensor([true, true, true])]; + tensor var_1867_cast_fp16 = slice_by_index(begin = var_1867_begin_0, end = var_1867_end_0, end_mask = var_1867_end_mask_0, x = new_x_27_cast_fp16)[name = string("op_1867_cast_fp16")]; + string x_177_pad_type_0 = const()[name = string("x_177_pad_type_0"), val = string("valid")]; + int32 x_177_groups_0 = const()[name = string("x_177_groups_0"), val = int32(1024)]; + tensor x_177_strides_0 = const()[name = string("x_177_strides_0"), val = tensor([1])]; + tensor x_177_pad_0 = const()[name = string("x_177_pad_0"), val = tensor([0, 0])]; + tensor x_177_dilations_0 = const()[name = string("x_177_dilations_0"), val = tensor([1])]; + tensor encoder_layers_6_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144138880))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144148160))))[name = string("encoder_layers_6_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_177_cast_fp16 = conv(dilations = x_177_dilations_0, groups = x_177_groups_0, pad = x_177_pad_0, pad_type = x_177_pad_type_0, strides = x_177_strides_0, weight = encoder_layers_6_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_27_cast_fp16)[name = string("x_177_cast_fp16")]; + tensor input_367_perm_0 = const()[name = string("input_367_perm_0"), val = tensor([0, 2, 1])]; + tensor x_179_axes_0 = const()[name = string("x_179_axes_0"), val = tensor([-1])]; + tensor encoder_layers_6_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_6_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144150272)))]; + tensor encoder_layers_6_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_6_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144152384)))]; + tensor input_367_cast_fp16 = transpose(perm = input_367_perm_0, x = x_177_cast_fp16)[name = string("transpose_302")]; + tensor x_179_cast_fp16 = layer_norm(axes = x_179_axes_0, beta = encoder_layers_6_conv_batch_norm_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_6_conv_batch_norm_weight_to_fp16, x = input_367_cast_fp16)[name = string("x_179_cast_fp16")]; + tensor input_369_perm_0 = const()[name = string("input_369_perm_0"), val = tensor([0, 2, 1])]; + tensor input_369_cast_fp16 = transpose(perm = input_369_perm_0, x = x_179_cast_fp16)[name = string("transpose_301")]; + tensor input_371_cast_fp16 = silu(x = input_369_cast_fp16)[name = string("input_371_cast_fp16")]; + string x_181_pad_type_0 = const()[name = string("x_181_pad_type_0"), val = string("valid")]; + tensor x_181_strides_0 = const()[name = string("x_181_strides_0"), val = tensor([1])]; + tensor x_181_pad_0 = const()[name = string("x_181_pad_0"), val = tensor([0, 0])]; + tensor x_181_dilations_0 = const()[name = string("x_181_dilations_0"), val = tensor([1])]; + int32 x_181_groups_0 = const()[name = string("x_181_groups_0"), val = int32(1)]; + tensor encoder_layers_6_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144154496))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(145203136))))[name = string("encoder_layers_6_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_181_cast_fp16 = conv(dilations = x_181_dilations_0, groups = x_181_groups_0, pad = x_181_pad_0, pad_type = x_181_pad_type_0, strides = x_181_strides_0, weight = encoder_layers_6_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_371_cast_fp16)[name = string("x_181_cast_fp16")]; + tensor input_373_perm_0 = const()[name = string("input_373_perm_0"), val = tensor([0, 2, 1])]; + tensor input_373_cast_fp16 = transpose(perm = input_373_perm_0, x = x_181_cast_fp16)[name = string("transpose_300")]; + tensor input_375_cast_fp16 = add(x = input_359_cast_fp16, y = input_373_cast_fp16)[name = string("input_375_cast_fp16")]; + tensor input_377_axes_0 = const()[name = string("input_377_axes_0"), val = tensor([-1])]; + tensor encoder_layers_6_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_6_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(145205248)))]; + tensor encoder_layers_6_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_6_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(145207360)))]; + tensor input_377_cast_fp16 = layer_norm(axes = input_377_axes_0, beta = encoder_layers_6_norm_feed_forward2_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_6_norm_feed_forward2_weight_to_fp16, x = input_375_cast_fp16)[name = string("input_377_cast_fp16")]; + tensor encoder_layers_6_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(145209472))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(148355264))))[name = string("encoder_layers_6_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_6_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_6_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(148355456)))]; + tensor linear_62_cast_fp16 = linear(bias = encoder_layers_6_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_6_feed_forward2_linear1_weight_to_fp16_palettized, x = input_377_cast_fp16)[name = string("linear_62_cast_fp16")]; + tensor input_381_cast_fp16 = silu(x = linear_62_cast_fp16)[name = string("input_381_cast_fp16")]; + tensor encoder_layers_6_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(148363712))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151509504))))[name = string("encoder_layers_6_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_6_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_6_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151509696)))]; + tensor linear_63_cast_fp16 = linear(bias = encoder_layers_6_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_6_feed_forward2_linear2_weight_to_fp16_palettized, x = input_381_cast_fp16)[name = string("linear_63_cast_fp16")]; + fp16 var_1910_to_fp16 = const()[name = string("op_1910_to_fp16"), val = fp16(0x1p-1)]; + tensor var_1911_cast_fp16 = mul(x = linear_63_cast_fp16, y = var_1910_to_fp16)[name = string("op_1911_cast_fp16")]; + tensor input_387_cast_fp16 = add(x = input_375_cast_fp16, y = var_1911_cast_fp16)[name = string("input_387_cast_fp16")]; + tensor input_389_axes_0 = const()[name = string("input_389_axes_0"), val = tensor([-1])]; + tensor encoder_layers_6_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_6_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151511808)))]; + tensor encoder_layers_6_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_6_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151513920)))]; + tensor input_389_cast_fp16 = layer_norm(axes = input_389_axes_0, beta = encoder_layers_6_norm_out_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_6_norm_out_weight_to_fp16, x = input_387_cast_fp16)[name = string("input_389_cast_fp16")]; + tensor cache_29_begin_0 = const()[name = string("cache_29_begin_0"), val = tensor([7, 0, 0, 0])]; + tensor cache_29_end_0 = const()[name = string("cache_29_end_0"), val = tensor([8, 1, 42, 1024])]; + tensor cache_29_end_mask_0 = const()[name = string("cache_29_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_29_squeeze_mask_0 = const()[name = string("cache_29_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_29_cast_fp16 = slice_by_index(begin = cache_29_begin_0, end = cache_29_end_0, end_mask = cache_29_end_mask_0, squeeze_mask = cache_29_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_29_cast_fp16")]; + tensor cache_31_begin_0 = const()[name = string("cache_31_begin_0"), val = tensor([7, 0, 0, 0])]; + tensor cache_31_end_0 = const()[name = string("cache_31_end_0"), val = tensor([8, 1, 1024, 8])]; + tensor cache_31_end_mask_0 = const()[name = string("cache_31_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_31_squeeze_mask_0 = const()[name = string("cache_31_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_31_cast_fp16 = slice_by_index(begin = cache_31_begin_0, end = cache_31_end_0, end_mask = cache_31_end_mask_0, squeeze_mask = cache_31_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_31_cast_fp16")]; + tensor input_391_axes_0 = const()[name = string("input_391_axes_0"), val = tensor([-1])]; + tensor encoder_layers_7_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_7_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151516032)))]; + tensor encoder_layers_7_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_7_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151518144)))]; + tensor input_391_cast_fp16 = layer_norm(axes = input_391_axes_0, beta = encoder_layers_7_norm_feed_forward1_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_7_norm_feed_forward1_weight_to_fp16, x = input_389_cast_fp16)[name = string("input_391_cast_fp16")]; + tensor encoder_layers_7_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151520256))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(154666048))))[name = string("encoder_layers_7_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_7_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_7_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(154666240)))]; + tensor linear_64_cast_fp16 = linear(bias = encoder_layers_7_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_7_feed_forward1_linear1_weight_to_fp16_palettized, x = input_391_cast_fp16)[name = string("linear_64_cast_fp16")]; + tensor input_395_cast_fp16 = silu(x = linear_64_cast_fp16)[name = string("input_395_cast_fp16")]; + tensor encoder_layers_7_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(154674496))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(157820288))))[name = string("encoder_layers_7_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_7_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_7_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(157820480)))]; + tensor linear_65_cast_fp16 = linear(bias = encoder_layers_7_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_7_feed_forward1_linear2_weight_to_fp16_palettized, x = input_395_cast_fp16)[name = string("linear_65_cast_fp16")]; + fp16 var_1947_to_fp16 = const()[name = string("op_1947_to_fp16"), val = fp16(0x1p-1)]; + tensor var_1948_cast_fp16 = mul(x = linear_65_cast_fp16, y = var_1947_to_fp16)[name = string("op_1948_cast_fp16")]; + tensor input_401_cast_fp16 = add(x = input_389_cast_fp16, y = var_1948_cast_fp16)[name = string("input_401_cast_fp16")]; + tensor key_15_axes_0 = const()[name = string("key_15_axes_0"), val = tensor([-1])]; + tensor encoder_layers_7_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_7_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(157822592)))]; + tensor encoder_layers_7_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_7_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(157824704)))]; + tensor key_15_cast_fp16 = layer_norm(axes = key_15_axes_0, beta = encoder_layers_7_norm_self_att_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_7_norm_self_att_weight_to_fp16, x = input_401_cast_fp16)[name = string("key_15_cast_fp16")]; + bool input_403_interleave_0 = const()[name = string("input_403_interleave_0"), val = bool(false)]; + tensor input_403_cast_fp16 = concat(axis = var_69, interleave = input_403_interleave_0, values = (cache_29_cast_fp16, key_15_cast_fp16))[name = string("input_403_cast_fp16")]; + tensor var_1970_begin_0 = const()[name = string("op_1970_begin_0"), val = tensor([0, 14, 0])]; + tensor var_1970_end_0 = const()[name = string("op_1970_end_0"), val = tensor([1, 42, 1024])]; + tensor var_1970_end_mask_0 = const()[name = string("op_1970_end_mask_0"), val = tensor([true, true, true])]; + tensor var_1970_cast_fp16 = slice_by_index(begin = var_1970_begin_0, end = var_1970_end_0, end_mask = var_1970_end_mask_0, x = cache_29_cast_fp16)[name = string("op_1970_cast_fp16")]; + bool var_1976_interleave_0 = const()[name = string("op_1976_interleave_0"), val = bool(false)]; + tensor var_1976_cast_fp16 = concat(axis = var_69, interleave = var_1976_interleave_0, values = (var_1970_cast_fp16, key_15_cast_fp16))[name = string("op_1976_cast_fp16")]; + tensor encoder_layers_7_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(157826816))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158613312))))[name = string("encoder_layers_7_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_7_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_7_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158613504)))]; + tensor linear_66_cast_fp16 = linear(bias = encoder_layers_7_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_7_self_attn_linear_q_weight_to_fp16_palettized, x = key_15_cast_fp16)[name = string("linear_66_cast_fp16")]; + tensor var_1981 = const()[name = string("op_1981"), val = tensor([1, -1, 8, 128])]; + tensor q_43_cast_fp16 = reshape(shape = var_1981, x = linear_66_cast_fp16)[name = string("q_43_cast_fp16")]; + tensor encoder_layers_7_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158615616))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(159402112))))[name = string("encoder_layers_7_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_7_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_7_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(159402304)))]; + tensor linear_67_cast_fp16 = linear(bias = encoder_layers_7_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_7_self_attn_linear_k_weight_to_fp16_palettized, x = input_403_cast_fp16)[name = string("linear_67_cast_fp16")]; + tensor var_1986 = const()[name = string("op_1986"), val = tensor([1, -1, 8, 128])]; + tensor k_29_cast_fp16 = reshape(shape = var_1986, x = linear_67_cast_fp16)[name = string("k_29_cast_fp16")]; + tensor encoder_layers_7_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(159404416))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160190912))))[name = string("encoder_layers_7_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_7_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_7_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160191104)))]; + tensor linear_68_cast_fp16 = linear(bias = encoder_layers_7_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_7_self_attn_linear_v_weight_to_fp16_palettized, x = input_403_cast_fp16)[name = string("linear_68_cast_fp16")]; + tensor var_1991 = const()[name = string("op_1991"), val = tensor([1, -1, 8, 128])]; + tensor v_15_cast_fp16 = reshape(shape = var_1991, x = linear_68_cast_fp16)[name = string("v_15_cast_fp16")]; + tensor value_23_perm_0 = const()[name = string("value_23_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_7_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_7_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160193216)))]; + tensor var_2004_cast_fp16 = add(x = q_43_cast_fp16, y = encoder_layers_7_self_attn_pos_bias_u_to_fp16)[name = string("op_2004_cast_fp16")]; + tensor encoder_layers_7_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_7_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160195328)))]; + tensor var_2006_cast_fp16 = add(x = q_43_cast_fp16, y = encoder_layers_7_self_attn_pos_bias_v_to_fp16)[name = string("op_2006_cast_fp16")]; + tensor q_with_bias_v_15_perm_0 = const()[name = string("q_with_bias_v_15_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_189_transpose_x_0 = const()[name = string("x_189_transpose_x_0"), val = bool(false)]; + bool x_189_transpose_y_0 = const()[name = string("x_189_transpose_y_0"), val = bool(false)]; + tensor op_2008_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160197440))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160311168))))[name = string("op_2008_to_fp16_quantized")]; + tensor q_with_bias_v_15_cast_fp16 = transpose(perm = q_with_bias_v_15_perm_0, x = var_2006_cast_fp16)[name = string("transpose_299")]; + tensor x_189_cast_fp16 = matmul(transpose_x = x_189_transpose_x_0, transpose_y = x_189_transpose_y_0, x = q_with_bias_v_15_cast_fp16, y = op_2008_to_fp16_quantized)[name = string("x_189_cast_fp16")]; + tensor x_191_pad_0 = const()[name = string("x_191_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_191_mode_0 = const()[name = string("x_191_mode_0"), val = string("constant")]; + fp16 const_170_to_fp16 = const()[name = string("const_170_to_fp16"), val = fp16(0x0p+0)]; + tensor x_191_cast_fp16 = pad(constant_val = const_170_to_fp16, mode = x_191_mode_0, pad = x_191_pad_0, x = x_189_cast_fp16)[name = string("x_191_cast_fp16")]; + tensor var_2016 = const()[name = string("op_2016"), val = tensor([1, 8, -1, 14])]; + tensor x_193_cast_fp16 = reshape(shape = var_2016, x = x_191_cast_fp16)[name = string("x_193_cast_fp16")]; + tensor var_2020_begin_0 = const()[name = string("op_2020_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_2020_end_0 = const()[name = string("op_2020_end_0"), val = tensor([1, 8, 112, 14])]; + tensor var_2020_end_mask_0 = const()[name = string("op_2020_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_2020_cast_fp16 = slice_by_index(begin = var_2020_begin_0, end = var_2020_end_0, end_mask = var_2020_end_mask_0, x = x_193_cast_fp16)[name = string("op_2020_cast_fp16")]; + tensor var_2021 = const()[name = string("op_2021"), val = tensor([1, 8, 14, 111])]; + tensor matrix_bd_29_cast_fp16 = reshape(shape = var_2021, x = var_2020_cast_fp16)[name = string("matrix_bd_29_cast_fp16")]; + bool matrix_ac_15_transpose_x_0 = const()[name = string("matrix_ac_15_transpose_x_0"), val = bool(false)]; + bool matrix_ac_15_transpose_y_0 = const()[name = string("matrix_ac_15_transpose_y_0"), val = bool(false)]; + tensor transpose_110_perm_0 = const()[name = string("transpose_110_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_111_perm_0 = const()[name = string("transpose_111_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_111 = transpose(perm = transpose_111_perm_0, x = k_29_cast_fp16)[name = string("transpose_297")]; + tensor transpose_110 = transpose(perm = transpose_110_perm_0, x = var_2004_cast_fp16)[name = string("transpose_298")]; + tensor matrix_ac_15_cast_fp16 = matmul(transpose_x = matrix_ac_15_transpose_x_0, transpose_y = matrix_ac_15_transpose_y_0, x = transpose_110, y = transpose_111)[name = string("matrix_ac_15_cast_fp16")]; + tensor matrix_bd_31_begin_0 = const()[name = string("matrix_bd_31_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_31_end_0 = const()[name = string("matrix_bd_31_end_0"), val = tensor([1, 8, 14, 56])]; + tensor matrix_bd_31_end_mask_0 = const()[name = string("matrix_bd_31_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_31_cast_fp16 = slice_by_index(begin = matrix_bd_31_begin_0, end = matrix_bd_31_end_0, end_mask = matrix_bd_31_end_mask_0, x = matrix_bd_29_cast_fp16)[name = string("matrix_bd_31_cast_fp16")]; + tensor var_2030_cast_fp16 = add(x = matrix_ac_15_cast_fp16, y = matrix_bd_31_cast_fp16)[name = string("op_2030_cast_fp16")]; + fp16 _inversed_scores_29_y_0_to_fp16 = const()[name = string("_inversed_scores_29_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_29_cast_fp16 = mul(x = var_2030_cast_fp16, y = _inversed_scores_29_y_0_to_fp16)[name = string("_inversed_scores_29_cast_fp16")]; + tensor scores_31_cast_fp16 = select(a = var_46_to_fp16, b = _inversed_scores_29_cast_fp16, cond = mask_11)[name = string("scores_31_cast_fp16")]; + tensor var_2036_cast_fp16 = softmax(axis = var_60, x = scores_31_cast_fp16)[name = string("op_2036_cast_fp16")]; + tensor input_405_cast_fp16 = select(a = var_45_to_fp16, b = var_2036_cast_fp16, cond = mask_11)[name = string("input_405_cast_fp16")]; + bool x_195_transpose_x_0 = const()[name = string("x_195_transpose_x_0"), val = bool(false)]; + bool x_195_transpose_y_0 = const()[name = string("x_195_transpose_y_0"), val = bool(false)]; + tensor value_23_cast_fp16 = transpose(perm = value_23_perm_0, x = v_15_cast_fp16)[name = string("transpose_296")]; + tensor x_195_cast_fp16 = matmul(transpose_x = x_195_transpose_x_0, transpose_y = x_195_transpose_y_0, x = input_405_cast_fp16, y = value_23_cast_fp16)[name = string("x_195_cast_fp16")]; + tensor var_2040_perm_0 = const()[name = string("op_2040_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_2041 = const()[name = string("op_2041"), val = tensor([1, -1, 1024])]; + tensor var_2040_cast_fp16 = transpose(perm = var_2040_perm_0, x = x_195_cast_fp16)[name = string("transpose_295")]; + tensor input_407_cast_fp16 = reshape(shape = var_2041, x = var_2040_cast_fp16)[name = string("input_407_cast_fp16")]; + tensor encoder_layers_7_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160311488))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161097984))))[name = string("encoder_layers_7_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_7_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_7_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161098176)))]; + tensor linear_70_cast_fp16 = linear(bias = encoder_layers_7_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_7_self_attn_linear_out_weight_to_fp16_palettized, x = input_407_cast_fp16)[name = string("linear_70_cast_fp16")]; + tensor input_411_cast_fp16 = add(x = input_401_cast_fp16, y = linear_70_cast_fp16)[name = string("input_411_cast_fp16")]; + tensor x_199_axes_0 = const()[name = string("x_199_axes_0"), val = tensor([-1])]; + tensor encoder_layers_7_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_7_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161100288)))]; + tensor encoder_layers_7_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_7_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161102400)))]; + tensor x_199_cast_fp16 = layer_norm(axes = x_199_axes_0, beta = encoder_layers_7_norm_conv_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_7_norm_conv_weight_to_fp16, x = input_411_cast_fp16)[name = string("x_199_cast_fp16")]; + tensor input_413_perm_0 = const()[name = string("input_413_perm_0"), val = tensor([0, 2, 1])]; + string input_415_pad_type_0 = const()[name = string("input_415_pad_type_0"), val = string("valid")]; + tensor input_415_strides_0 = const()[name = string("input_415_strides_0"), val = tensor([1])]; + tensor input_415_pad_0 = const()[name = string("input_415_pad_0"), val = tensor([0, 0])]; + tensor input_415_dilations_0 = const()[name = string("input_415_dilations_0"), val = tensor([1])]; + int32 input_415_groups_0 = const()[name = string("input_415_groups_0"), val = int32(1)]; + tensor encoder_layers_7_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161104512))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163201728))))[name = string("encoder_layers_7_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_413_cast_fp16 = transpose(perm = input_413_perm_0, x = x_199_cast_fp16)[name = string("transpose_294")]; + tensor input_415_cast_fp16 = conv(dilations = input_415_dilations_0, groups = input_415_groups_0, pad = input_415_pad_0, pad_type = input_415_pad_type_0, strides = input_415_strides_0, weight = encoder_layers_7_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_413_cast_fp16)[name = string("input_415_cast_fp16")]; + int32 x_201_split_num_splits_0 = const()[name = string("x_201_split_num_splits_0"), val = int32(2)]; + int32 x_201_split_axis_0 = const()[name = string("x_201_split_axis_0"), val = int32(1)]; + tensor x_201_split_cast_fp16_0, tensor x_201_split_cast_fp16_1 = split(axis = x_201_split_axis_0, num_splits = x_201_split_num_splits_0, x = input_415_cast_fp16)[name = string("x_201_split_cast_fp16")]; + tensor x_201_split_1_sigmoid_cast_fp16 = sigmoid(x = x_201_split_cast_fp16_1)[name = string("x_201_split_1_sigmoid_cast_fp16")]; + tensor x_201_cast_fp16 = mul(x = x_201_split_cast_fp16_0, y = x_201_split_1_sigmoid_cast_fp16)[name = string("x_201_cast_fp16")]; + tensor input_417_cast_fp16 = select(a = var_45_to_fp16, b = x_201_cast_fp16, cond = var_576)[name = string("input_417_cast_fp16")]; + bool new_x_31_interleave_0 = const()[name = string("new_x_31_interleave_0"), val = bool(false)]; + tensor new_x_31_cast_fp16 = concat(axis = var_60, interleave = new_x_31_interleave_0, values = (cache_31_cast_fp16, input_417_cast_fp16))[name = string("new_x_31_cast_fp16")]; + tensor var_2080_begin_0 = const()[name = string("op_2080_begin_0"), val = tensor([0, 0, 14])]; + tensor var_2080_end_0 = const()[name = string("op_2080_end_0"), val = tensor([1, 1024, 22])]; + tensor var_2080_end_mask_0 = const()[name = string("op_2080_end_mask_0"), val = tensor([true, true, true])]; + tensor var_2080_cast_fp16 = slice_by_index(begin = var_2080_begin_0, end = var_2080_end_0, end_mask = var_2080_end_mask_0, x = new_x_31_cast_fp16)[name = string("op_2080_cast_fp16")]; + string x_203_pad_type_0 = const()[name = string("x_203_pad_type_0"), val = string("valid")]; + int32 x_203_groups_0 = const()[name = string("x_203_groups_0"), val = int32(1024)]; + tensor x_203_strides_0 = const()[name = string("x_203_strides_0"), val = tensor([1])]; + tensor x_203_pad_0 = const()[name = string("x_203_pad_0"), val = tensor([0, 0])]; + tensor x_203_dilations_0 = const()[name = string("x_203_dilations_0"), val = tensor([1])]; + tensor encoder_layers_7_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163205888))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163215168))))[name = string("encoder_layers_7_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_203_cast_fp16 = conv(dilations = x_203_dilations_0, groups = x_203_groups_0, pad = x_203_pad_0, pad_type = x_203_pad_type_0, strides = x_203_strides_0, weight = encoder_layers_7_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_31_cast_fp16)[name = string("x_203_cast_fp16")]; + tensor input_419_perm_0 = const()[name = string("input_419_perm_0"), val = tensor([0, 2, 1])]; + tensor x_205_axes_0 = const()[name = string("x_205_axes_0"), val = tensor([-1])]; + tensor encoder_layers_7_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_7_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163217280)))]; + tensor encoder_layers_7_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_7_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163219392)))]; + tensor input_419_cast_fp16 = transpose(perm = input_419_perm_0, x = x_203_cast_fp16)[name = string("transpose_293")]; + tensor x_205_cast_fp16 = layer_norm(axes = x_205_axes_0, beta = encoder_layers_7_conv_batch_norm_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_7_conv_batch_norm_weight_to_fp16, x = input_419_cast_fp16)[name = string("x_205_cast_fp16")]; + tensor input_421_perm_0 = const()[name = string("input_421_perm_0"), val = tensor([0, 2, 1])]; + tensor input_421_cast_fp16 = transpose(perm = input_421_perm_0, x = x_205_cast_fp16)[name = string("transpose_292")]; + tensor input_423_cast_fp16 = silu(x = input_421_cast_fp16)[name = string("input_423_cast_fp16")]; + string x_207_pad_type_0 = const()[name = string("x_207_pad_type_0"), val = string("valid")]; + tensor x_207_strides_0 = const()[name = string("x_207_strides_0"), val = tensor([1])]; + tensor x_207_pad_0 = const()[name = string("x_207_pad_0"), val = tensor([0, 0])]; + tensor x_207_dilations_0 = const()[name = string("x_207_dilations_0"), val = tensor([1])]; + int32 x_207_groups_0 = const()[name = string("x_207_groups_0"), val = int32(1)]; + tensor encoder_layers_7_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163221504))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(164270144))))[name = string("encoder_layers_7_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_207_cast_fp16 = conv(dilations = x_207_dilations_0, groups = x_207_groups_0, pad = x_207_pad_0, pad_type = x_207_pad_type_0, strides = x_207_strides_0, weight = encoder_layers_7_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_423_cast_fp16)[name = string("x_207_cast_fp16")]; + tensor input_425_perm_0 = const()[name = string("input_425_perm_0"), val = tensor([0, 2, 1])]; + tensor input_425_cast_fp16 = transpose(perm = input_425_perm_0, x = x_207_cast_fp16)[name = string("transpose_291")]; + tensor input_427_cast_fp16 = add(x = input_411_cast_fp16, y = input_425_cast_fp16)[name = string("input_427_cast_fp16")]; + tensor input_429_axes_0 = const()[name = string("input_429_axes_0"), val = tensor([-1])]; + tensor encoder_layers_7_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_7_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(164272256)))]; + tensor encoder_layers_7_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_7_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(164274368)))]; + tensor input_429_cast_fp16 = layer_norm(axes = input_429_axes_0, beta = encoder_layers_7_norm_feed_forward2_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_7_norm_feed_forward2_weight_to_fp16, x = input_427_cast_fp16)[name = string("input_429_cast_fp16")]; + tensor encoder_layers_7_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(164276480))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167422272))))[name = string("encoder_layers_7_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_7_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_7_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167422464)))]; + tensor linear_71_cast_fp16 = linear(bias = encoder_layers_7_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_7_feed_forward2_linear1_weight_to_fp16_palettized, x = input_429_cast_fp16)[name = string("linear_71_cast_fp16")]; + tensor input_433_cast_fp16 = silu(x = linear_71_cast_fp16)[name = string("input_433_cast_fp16")]; + tensor encoder_layers_7_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167430720))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170576512))))[name = string("encoder_layers_7_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_7_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_7_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170576704)))]; + tensor linear_72_cast_fp16 = linear(bias = encoder_layers_7_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_7_feed_forward2_linear2_weight_to_fp16_palettized, x = input_433_cast_fp16)[name = string("linear_72_cast_fp16")]; + fp16 var_2123_to_fp16 = const()[name = string("op_2123_to_fp16"), val = fp16(0x1p-1)]; + tensor var_2124_cast_fp16 = mul(x = linear_72_cast_fp16, y = var_2123_to_fp16)[name = string("op_2124_cast_fp16")]; + tensor input_439_cast_fp16 = add(x = input_427_cast_fp16, y = var_2124_cast_fp16)[name = string("input_439_cast_fp16")]; + tensor input_441_axes_0 = const()[name = string("input_441_axes_0"), val = tensor([-1])]; + tensor encoder_layers_7_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_7_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170578816)))]; + tensor encoder_layers_7_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_7_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170580928)))]; + tensor input_441_cast_fp16 = layer_norm(axes = input_441_axes_0, beta = encoder_layers_7_norm_out_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_7_norm_out_weight_to_fp16, x = input_439_cast_fp16)[name = string("input_441_cast_fp16")]; + tensor cache_33_begin_0 = const()[name = string("cache_33_begin_0"), val = tensor([8, 0, 0, 0])]; + tensor cache_33_end_0 = const()[name = string("cache_33_end_0"), val = tensor([9, 1, 42, 1024])]; + tensor cache_33_end_mask_0 = const()[name = string("cache_33_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_33_squeeze_mask_0 = const()[name = string("cache_33_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_33_cast_fp16 = slice_by_index(begin = cache_33_begin_0, end = cache_33_end_0, end_mask = cache_33_end_mask_0, squeeze_mask = cache_33_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_33_cast_fp16")]; + tensor cache_35_begin_0 = const()[name = string("cache_35_begin_0"), val = tensor([8, 0, 0, 0])]; + tensor cache_35_end_0 = const()[name = string("cache_35_end_0"), val = tensor([9, 1, 1024, 8])]; + tensor cache_35_end_mask_0 = const()[name = string("cache_35_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_35_squeeze_mask_0 = const()[name = string("cache_35_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_35_cast_fp16 = slice_by_index(begin = cache_35_begin_0, end = cache_35_end_0, end_mask = cache_35_end_mask_0, squeeze_mask = cache_35_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_35_cast_fp16")]; + tensor input_443_axes_0 = const()[name = string("input_443_axes_0"), val = tensor([-1])]; + tensor encoder_layers_8_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_8_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170583040)))]; + tensor encoder_layers_8_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_8_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170585152)))]; + tensor input_443_cast_fp16 = layer_norm(axes = input_443_axes_0, beta = encoder_layers_8_norm_feed_forward1_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_8_norm_feed_forward1_weight_to_fp16, x = input_441_cast_fp16)[name = string("input_443_cast_fp16")]; + tensor encoder_layers_8_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170587264))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(173733056))))[name = string("encoder_layers_8_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_8_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_8_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(173733248)))]; + tensor linear_73_cast_fp16 = linear(bias = encoder_layers_8_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_8_feed_forward1_linear1_weight_to_fp16_palettized, x = input_443_cast_fp16)[name = string("linear_73_cast_fp16")]; + tensor input_447_cast_fp16 = silu(x = linear_73_cast_fp16)[name = string("input_447_cast_fp16")]; + tensor encoder_layers_8_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(173741504))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(176887296))))[name = string("encoder_layers_8_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_8_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_8_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(176887488)))]; + tensor linear_74_cast_fp16 = linear(bias = encoder_layers_8_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_8_feed_forward1_linear2_weight_to_fp16_palettized, x = input_447_cast_fp16)[name = string("linear_74_cast_fp16")]; + fp16 var_2160_to_fp16 = const()[name = string("op_2160_to_fp16"), val = fp16(0x1p-1)]; + tensor var_2161_cast_fp16 = mul(x = linear_74_cast_fp16, y = var_2160_to_fp16)[name = string("op_2161_cast_fp16")]; + tensor input_453_cast_fp16 = add(x = input_441_cast_fp16, y = var_2161_cast_fp16)[name = string("input_453_cast_fp16")]; + tensor key_17_axes_0 = const()[name = string("key_17_axes_0"), val = tensor([-1])]; + tensor encoder_layers_8_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_8_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(176889600)))]; + tensor encoder_layers_8_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_8_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(176891712)))]; + tensor key_17_cast_fp16 = layer_norm(axes = key_17_axes_0, beta = encoder_layers_8_norm_self_att_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_8_norm_self_att_weight_to_fp16, x = input_453_cast_fp16)[name = string("key_17_cast_fp16")]; + bool input_455_interleave_0 = const()[name = string("input_455_interleave_0"), val = bool(false)]; + tensor input_455_cast_fp16 = concat(axis = var_69, interleave = input_455_interleave_0, values = (cache_33_cast_fp16, key_17_cast_fp16))[name = string("input_455_cast_fp16")]; + tensor var_2183_begin_0 = const()[name = string("op_2183_begin_0"), val = tensor([0, 14, 0])]; + tensor var_2183_end_0 = const()[name = string("op_2183_end_0"), val = tensor([1, 42, 1024])]; + tensor var_2183_end_mask_0 = const()[name = string("op_2183_end_mask_0"), val = tensor([true, true, true])]; + tensor var_2183_cast_fp16 = slice_by_index(begin = var_2183_begin_0, end = var_2183_end_0, end_mask = var_2183_end_mask_0, x = cache_33_cast_fp16)[name = string("op_2183_cast_fp16")]; + bool var_2189_interleave_0 = const()[name = string("op_2189_interleave_0"), val = bool(false)]; + tensor var_2189_cast_fp16 = concat(axis = var_69, interleave = var_2189_interleave_0, values = (var_2183_cast_fp16, key_17_cast_fp16))[name = string("op_2189_cast_fp16")]; + tensor encoder_layers_8_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(176893824))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(177680320))))[name = string("encoder_layers_8_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_8_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_8_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(177680512)))]; + tensor linear_75_cast_fp16 = linear(bias = encoder_layers_8_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_8_self_attn_linear_q_weight_to_fp16_palettized, x = key_17_cast_fp16)[name = string("linear_75_cast_fp16")]; + tensor var_2194 = const()[name = string("op_2194"), val = tensor([1, -1, 8, 128])]; + tensor q_49_cast_fp16 = reshape(shape = var_2194, x = linear_75_cast_fp16)[name = string("q_49_cast_fp16")]; + tensor encoder_layers_8_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(177682624))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(178469120))))[name = string("encoder_layers_8_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_8_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_8_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(178469312)))]; + tensor linear_76_cast_fp16 = linear(bias = encoder_layers_8_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_8_self_attn_linear_k_weight_to_fp16_palettized, x = input_455_cast_fp16)[name = string("linear_76_cast_fp16")]; + tensor var_2199 = const()[name = string("op_2199"), val = tensor([1, -1, 8, 128])]; + tensor k_33_cast_fp16 = reshape(shape = var_2199, x = linear_76_cast_fp16)[name = string("k_33_cast_fp16")]; + tensor encoder_layers_8_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(178471424))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179257920))))[name = string("encoder_layers_8_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_8_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_8_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179258112)))]; + tensor linear_77_cast_fp16 = linear(bias = encoder_layers_8_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_8_self_attn_linear_v_weight_to_fp16_palettized, x = input_455_cast_fp16)[name = string("linear_77_cast_fp16")]; + tensor var_2204 = const()[name = string("op_2204"), val = tensor([1, -1, 8, 128])]; + tensor v_17_cast_fp16 = reshape(shape = var_2204, x = linear_77_cast_fp16)[name = string("v_17_cast_fp16")]; + tensor value_25_perm_0 = const()[name = string("value_25_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_8_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_8_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179260224)))]; + tensor var_2217_cast_fp16 = add(x = q_49_cast_fp16, y = encoder_layers_8_self_attn_pos_bias_u_to_fp16)[name = string("op_2217_cast_fp16")]; + tensor encoder_layers_8_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_8_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179262336)))]; + tensor var_2219_cast_fp16 = add(x = q_49_cast_fp16, y = encoder_layers_8_self_attn_pos_bias_v_to_fp16)[name = string("op_2219_cast_fp16")]; + tensor q_with_bias_v_17_perm_0 = const()[name = string("q_with_bias_v_17_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_215_transpose_x_0 = const()[name = string("x_215_transpose_x_0"), val = bool(false)]; + bool x_215_transpose_y_0 = const()[name = string("x_215_transpose_y_0"), val = bool(false)]; + tensor op_2221_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179264448))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179378176))))[name = string("op_2221_to_fp16_quantized")]; + tensor q_with_bias_v_17_cast_fp16 = transpose(perm = q_with_bias_v_17_perm_0, x = var_2219_cast_fp16)[name = string("transpose_290")]; + tensor x_215_cast_fp16 = matmul(transpose_x = x_215_transpose_x_0, transpose_y = x_215_transpose_y_0, x = q_with_bias_v_17_cast_fp16, y = op_2221_to_fp16_quantized)[name = string("x_215_cast_fp16")]; + tensor x_217_pad_0 = const()[name = string("x_217_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_217_mode_0 = const()[name = string("x_217_mode_0"), val = string("constant")]; + fp16 const_183_to_fp16 = const()[name = string("const_183_to_fp16"), val = fp16(0x0p+0)]; + tensor x_217_cast_fp16 = pad(constant_val = const_183_to_fp16, mode = x_217_mode_0, pad = x_217_pad_0, x = x_215_cast_fp16)[name = string("x_217_cast_fp16")]; + tensor var_2229 = const()[name = string("op_2229"), val = tensor([1, 8, -1, 14])]; + tensor x_219_cast_fp16 = reshape(shape = var_2229, x = x_217_cast_fp16)[name = string("x_219_cast_fp16")]; + tensor var_2233_begin_0 = const()[name = string("op_2233_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_2233_end_0 = const()[name = string("op_2233_end_0"), val = tensor([1, 8, 112, 14])]; + tensor var_2233_end_mask_0 = const()[name = string("op_2233_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_2233_cast_fp16 = slice_by_index(begin = var_2233_begin_0, end = var_2233_end_0, end_mask = var_2233_end_mask_0, x = x_219_cast_fp16)[name = string("op_2233_cast_fp16")]; + tensor var_2234 = const()[name = string("op_2234"), val = tensor([1, 8, 14, 111])]; + tensor matrix_bd_33_cast_fp16 = reshape(shape = var_2234, x = var_2233_cast_fp16)[name = string("matrix_bd_33_cast_fp16")]; + bool matrix_ac_17_transpose_x_0 = const()[name = string("matrix_ac_17_transpose_x_0"), val = bool(false)]; + bool matrix_ac_17_transpose_y_0 = const()[name = string("matrix_ac_17_transpose_y_0"), val = bool(false)]; + tensor transpose_112_perm_0 = const()[name = string("transpose_112_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_113_perm_0 = const()[name = string("transpose_113_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_113 = transpose(perm = transpose_113_perm_0, x = k_33_cast_fp16)[name = string("transpose_288")]; + tensor transpose_112 = transpose(perm = transpose_112_perm_0, x = var_2217_cast_fp16)[name = string("transpose_289")]; + tensor matrix_ac_17_cast_fp16 = matmul(transpose_x = matrix_ac_17_transpose_x_0, transpose_y = matrix_ac_17_transpose_y_0, x = transpose_112, y = transpose_113)[name = string("matrix_ac_17_cast_fp16")]; + tensor matrix_bd_35_begin_0 = const()[name = string("matrix_bd_35_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_35_end_0 = const()[name = string("matrix_bd_35_end_0"), val = tensor([1, 8, 14, 56])]; + tensor matrix_bd_35_end_mask_0 = const()[name = string("matrix_bd_35_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_35_cast_fp16 = slice_by_index(begin = matrix_bd_35_begin_0, end = matrix_bd_35_end_0, end_mask = matrix_bd_35_end_mask_0, x = matrix_bd_33_cast_fp16)[name = string("matrix_bd_35_cast_fp16")]; + tensor var_2243_cast_fp16 = add(x = matrix_ac_17_cast_fp16, y = matrix_bd_35_cast_fp16)[name = string("op_2243_cast_fp16")]; + fp16 _inversed_scores_33_y_0_to_fp16 = const()[name = string("_inversed_scores_33_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_33_cast_fp16 = mul(x = var_2243_cast_fp16, y = _inversed_scores_33_y_0_to_fp16)[name = string("_inversed_scores_33_cast_fp16")]; + tensor scores_35_cast_fp16 = select(a = var_46_to_fp16, b = _inversed_scores_33_cast_fp16, cond = mask_11)[name = string("scores_35_cast_fp16")]; + tensor var_2249_cast_fp16 = softmax(axis = var_60, x = scores_35_cast_fp16)[name = string("op_2249_cast_fp16")]; + tensor input_457_cast_fp16 = select(a = var_45_to_fp16, b = var_2249_cast_fp16, cond = mask_11)[name = string("input_457_cast_fp16")]; + bool x_221_transpose_x_0 = const()[name = string("x_221_transpose_x_0"), val = bool(false)]; + bool x_221_transpose_y_0 = const()[name = string("x_221_transpose_y_0"), val = bool(false)]; + tensor value_25_cast_fp16 = transpose(perm = value_25_perm_0, x = v_17_cast_fp16)[name = string("transpose_287")]; + tensor x_221_cast_fp16 = matmul(transpose_x = x_221_transpose_x_0, transpose_y = x_221_transpose_y_0, x = input_457_cast_fp16, y = value_25_cast_fp16)[name = string("x_221_cast_fp16")]; + tensor var_2253_perm_0 = const()[name = string("op_2253_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_2254 = const()[name = string("op_2254"), val = tensor([1, -1, 1024])]; + tensor var_2253_cast_fp16 = transpose(perm = var_2253_perm_0, x = x_221_cast_fp16)[name = string("transpose_286")]; + tensor input_459_cast_fp16 = reshape(shape = var_2254, x = var_2253_cast_fp16)[name = string("input_459_cast_fp16")]; + tensor encoder_layers_8_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179378496))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(180164992))))[name = string("encoder_layers_8_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_8_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_8_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(180165184)))]; + tensor linear_79_cast_fp16 = linear(bias = encoder_layers_8_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_8_self_attn_linear_out_weight_to_fp16_palettized, x = input_459_cast_fp16)[name = string("linear_79_cast_fp16")]; + tensor input_463_cast_fp16 = add(x = input_453_cast_fp16, y = linear_79_cast_fp16)[name = string("input_463_cast_fp16")]; + tensor x_225_axes_0 = const()[name = string("x_225_axes_0"), val = tensor([-1])]; + tensor encoder_layers_8_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_8_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(180167296)))]; + tensor encoder_layers_8_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_8_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(180169408)))]; + tensor x_225_cast_fp16 = layer_norm(axes = x_225_axes_0, beta = encoder_layers_8_norm_conv_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_8_norm_conv_weight_to_fp16, x = input_463_cast_fp16)[name = string("x_225_cast_fp16")]; + tensor input_465_perm_0 = const()[name = string("input_465_perm_0"), val = tensor([0, 2, 1])]; + string input_467_pad_type_0 = const()[name = string("input_467_pad_type_0"), val = string("valid")]; + tensor input_467_strides_0 = const()[name = string("input_467_strides_0"), val = tensor([1])]; + tensor input_467_pad_0 = const()[name = string("input_467_pad_0"), val = tensor([0, 0])]; + tensor input_467_dilations_0 = const()[name = string("input_467_dilations_0"), val = tensor([1])]; + int32 input_467_groups_0 = const()[name = string("input_467_groups_0"), val = int32(1)]; + tensor encoder_layers_8_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(180171520))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182268736))))[name = string("encoder_layers_8_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_465_cast_fp16 = transpose(perm = input_465_perm_0, x = x_225_cast_fp16)[name = string("transpose_285")]; + tensor input_467_cast_fp16 = conv(dilations = input_467_dilations_0, groups = input_467_groups_0, pad = input_467_pad_0, pad_type = input_467_pad_type_0, strides = input_467_strides_0, weight = encoder_layers_8_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_465_cast_fp16)[name = string("input_467_cast_fp16")]; + int32 x_227_split_num_splits_0 = const()[name = string("x_227_split_num_splits_0"), val = int32(2)]; + int32 x_227_split_axis_0 = const()[name = string("x_227_split_axis_0"), val = int32(1)]; + tensor x_227_split_cast_fp16_0, tensor x_227_split_cast_fp16_1 = split(axis = x_227_split_axis_0, num_splits = x_227_split_num_splits_0, x = input_467_cast_fp16)[name = string("x_227_split_cast_fp16")]; + tensor x_227_split_1_sigmoid_cast_fp16 = sigmoid(x = x_227_split_cast_fp16_1)[name = string("x_227_split_1_sigmoid_cast_fp16")]; + tensor x_227_cast_fp16 = mul(x = x_227_split_cast_fp16_0, y = x_227_split_1_sigmoid_cast_fp16)[name = string("x_227_cast_fp16")]; + tensor input_469_cast_fp16 = select(a = var_45_to_fp16, b = x_227_cast_fp16, cond = var_576)[name = string("input_469_cast_fp16")]; + bool new_x_35_interleave_0 = const()[name = string("new_x_35_interleave_0"), val = bool(false)]; + tensor new_x_35_cast_fp16 = concat(axis = var_60, interleave = new_x_35_interleave_0, values = (cache_35_cast_fp16, input_469_cast_fp16))[name = string("new_x_35_cast_fp16")]; + tensor var_2293_begin_0 = const()[name = string("op_2293_begin_0"), val = tensor([0, 0, 14])]; + tensor var_2293_end_0 = const()[name = string("op_2293_end_0"), val = tensor([1, 1024, 22])]; + tensor var_2293_end_mask_0 = const()[name = string("op_2293_end_mask_0"), val = tensor([true, true, true])]; + tensor var_2293_cast_fp16 = slice_by_index(begin = var_2293_begin_0, end = var_2293_end_0, end_mask = var_2293_end_mask_0, x = new_x_35_cast_fp16)[name = string("op_2293_cast_fp16")]; + string x_229_pad_type_0 = const()[name = string("x_229_pad_type_0"), val = string("valid")]; + int32 x_229_groups_0 = const()[name = string("x_229_groups_0"), val = int32(1024)]; + tensor x_229_strides_0 = const()[name = string("x_229_strides_0"), val = tensor([1])]; + tensor x_229_pad_0 = const()[name = string("x_229_pad_0"), val = tensor([0, 0])]; + tensor x_229_dilations_0 = const()[name = string("x_229_dilations_0"), val = tensor([1])]; + tensor encoder_layers_8_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182272896))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182282176))))[name = string("encoder_layers_8_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_229_cast_fp16 = conv(dilations = x_229_dilations_0, groups = x_229_groups_0, pad = x_229_pad_0, pad_type = x_229_pad_type_0, strides = x_229_strides_0, weight = encoder_layers_8_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_35_cast_fp16)[name = string("x_229_cast_fp16")]; + tensor input_471_perm_0 = const()[name = string("input_471_perm_0"), val = tensor([0, 2, 1])]; + tensor x_231_axes_0 = const()[name = string("x_231_axes_0"), val = tensor([-1])]; + tensor encoder_layers_8_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_8_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182284288)))]; + tensor encoder_layers_8_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_8_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182286400)))]; + tensor input_471_cast_fp16 = transpose(perm = input_471_perm_0, x = x_229_cast_fp16)[name = string("transpose_284")]; + tensor x_231_cast_fp16 = layer_norm(axes = x_231_axes_0, beta = encoder_layers_8_conv_batch_norm_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_8_conv_batch_norm_weight_to_fp16, x = input_471_cast_fp16)[name = string("x_231_cast_fp16")]; + tensor input_473_perm_0 = const()[name = string("input_473_perm_0"), val = tensor([0, 2, 1])]; + tensor input_473_cast_fp16 = transpose(perm = input_473_perm_0, x = x_231_cast_fp16)[name = string("transpose_283")]; + tensor input_475_cast_fp16 = silu(x = input_473_cast_fp16)[name = string("input_475_cast_fp16")]; + string x_233_pad_type_0 = const()[name = string("x_233_pad_type_0"), val = string("valid")]; + tensor x_233_strides_0 = const()[name = string("x_233_strides_0"), val = tensor([1])]; + tensor x_233_pad_0 = const()[name = string("x_233_pad_0"), val = tensor([0, 0])]; + tensor x_233_dilations_0 = const()[name = string("x_233_dilations_0"), val = tensor([1])]; + int32 x_233_groups_0 = const()[name = string("x_233_groups_0"), val = int32(1)]; + tensor encoder_layers_8_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182288512))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183337152))))[name = string("encoder_layers_8_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_233_cast_fp16 = conv(dilations = x_233_dilations_0, groups = x_233_groups_0, pad = x_233_pad_0, pad_type = x_233_pad_type_0, strides = x_233_strides_0, weight = encoder_layers_8_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_475_cast_fp16)[name = string("x_233_cast_fp16")]; + tensor input_477_perm_0 = const()[name = string("input_477_perm_0"), val = tensor([0, 2, 1])]; + tensor input_477_cast_fp16 = transpose(perm = input_477_perm_0, x = x_233_cast_fp16)[name = string("transpose_282")]; + tensor input_479_cast_fp16 = add(x = input_463_cast_fp16, y = input_477_cast_fp16)[name = string("input_479_cast_fp16")]; + tensor input_481_axes_0 = const()[name = string("input_481_axes_0"), val = tensor([-1])]; + tensor encoder_layers_8_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_8_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183339264)))]; + tensor encoder_layers_8_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_8_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183341376)))]; + tensor input_481_cast_fp16 = layer_norm(axes = input_481_axes_0, beta = encoder_layers_8_norm_feed_forward2_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_8_norm_feed_forward2_weight_to_fp16, x = input_479_cast_fp16)[name = string("input_481_cast_fp16")]; + tensor encoder_layers_8_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183343488))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186489280))))[name = string("encoder_layers_8_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_8_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_8_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186489472)))]; + tensor linear_80_cast_fp16 = linear(bias = encoder_layers_8_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_8_feed_forward2_linear1_weight_to_fp16_palettized, x = input_481_cast_fp16)[name = string("linear_80_cast_fp16")]; + tensor input_485_cast_fp16 = silu(x = linear_80_cast_fp16)[name = string("input_485_cast_fp16")]; + tensor encoder_layers_8_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186497728))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(189643520))))[name = string("encoder_layers_8_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_8_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_8_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(189643712)))]; + tensor linear_81_cast_fp16 = linear(bias = encoder_layers_8_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_8_feed_forward2_linear2_weight_to_fp16_palettized, x = input_485_cast_fp16)[name = string("linear_81_cast_fp16")]; + fp16 var_2336_to_fp16 = const()[name = string("op_2336_to_fp16"), val = fp16(0x1p-1)]; + tensor var_2337_cast_fp16 = mul(x = linear_81_cast_fp16, y = var_2336_to_fp16)[name = string("op_2337_cast_fp16")]; + tensor input_491_cast_fp16 = add(x = input_479_cast_fp16, y = var_2337_cast_fp16)[name = string("input_491_cast_fp16")]; + tensor input_493_axes_0 = const()[name = string("input_493_axes_0"), val = tensor([-1])]; + tensor encoder_layers_8_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_8_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(189645824)))]; + tensor encoder_layers_8_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_8_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(189647936)))]; + tensor input_493_cast_fp16 = layer_norm(axes = input_493_axes_0, beta = encoder_layers_8_norm_out_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_8_norm_out_weight_to_fp16, x = input_491_cast_fp16)[name = string("input_493_cast_fp16")]; + tensor cache_37_begin_0 = const()[name = string("cache_37_begin_0"), val = tensor([9, 0, 0, 0])]; + tensor cache_37_end_0 = const()[name = string("cache_37_end_0"), val = tensor([10, 1, 42, 1024])]; + tensor cache_37_end_mask_0 = const()[name = string("cache_37_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_37_squeeze_mask_0 = const()[name = string("cache_37_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_37_cast_fp16 = slice_by_index(begin = cache_37_begin_0, end = cache_37_end_0, end_mask = cache_37_end_mask_0, squeeze_mask = cache_37_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_37_cast_fp16")]; + tensor cache_39_begin_0 = const()[name = string("cache_39_begin_0"), val = tensor([9, 0, 0, 0])]; + tensor cache_39_end_0 = const()[name = string("cache_39_end_0"), val = tensor([10, 1, 1024, 8])]; + tensor cache_39_end_mask_0 = const()[name = string("cache_39_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_39_squeeze_mask_0 = const()[name = string("cache_39_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_39_cast_fp16 = slice_by_index(begin = cache_39_begin_0, end = cache_39_end_0, end_mask = cache_39_end_mask_0, squeeze_mask = cache_39_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_39_cast_fp16")]; + tensor input_495_axes_0 = const()[name = string("input_495_axes_0"), val = tensor([-1])]; + tensor encoder_layers_9_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_9_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(189650048)))]; + tensor encoder_layers_9_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_9_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(189652160)))]; + tensor input_495_cast_fp16 = layer_norm(axes = input_495_axes_0, beta = encoder_layers_9_norm_feed_forward1_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_9_norm_feed_forward1_weight_to_fp16, x = input_493_cast_fp16)[name = string("input_495_cast_fp16")]; + tensor encoder_layers_9_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(189654272))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(192800064))))[name = string("encoder_layers_9_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_9_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_9_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(192800256)))]; + tensor linear_82_cast_fp16 = linear(bias = encoder_layers_9_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_9_feed_forward1_linear1_weight_to_fp16_palettized, x = input_495_cast_fp16)[name = string("linear_82_cast_fp16")]; + tensor input_499_cast_fp16 = silu(x = linear_82_cast_fp16)[name = string("input_499_cast_fp16")]; + tensor encoder_layers_9_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(192808512))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(195954304))))[name = string("encoder_layers_9_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_9_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_9_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(195954496)))]; + tensor linear_83_cast_fp16 = linear(bias = encoder_layers_9_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_9_feed_forward1_linear2_weight_to_fp16_palettized, x = input_499_cast_fp16)[name = string("linear_83_cast_fp16")]; + fp16 var_2373_to_fp16 = const()[name = string("op_2373_to_fp16"), val = fp16(0x1p-1)]; + tensor var_2374_cast_fp16 = mul(x = linear_83_cast_fp16, y = var_2373_to_fp16)[name = string("op_2374_cast_fp16")]; + tensor input_505_cast_fp16 = add(x = input_493_cast_fp16, y = var_2374_cast_fp16)[name = string("input_505_cast_fp16")]; + tensor key_19_axes_0 = const()[name = string("key_19_axes_0"), val = tensor([-1])]; + tensor encoder_layers_9_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_9_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(195956608)))]; + tensor encoder_layers_9_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_9_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(195958720)))]; + tensor key_19_cast_fp16 = layer_norm(axes = key_19_axes_0, beta = encoder_layers_9_norm_self_att_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_9_norm_self_att_weight_to_fp16, x = input_505_cast_fp16)[name = string("key_19_cast_fp16")]; + bool input_507_interleave_0 = const()[name = string("input_507_interleave_0"), val = bool(false)]; + tensor input_507_cast_fp16 = concat(axis = var_69, interleave = input_507_interleave_0, values = (cache_37_cast_fp16, key_19_cast_fp16))[name = string("input_507_cast_fp16")]; + tensor var_2396_begin_0 = const()[name = string("op_2396_begin_0"), val = tensor([0, 14, 0])]; + tensor var_2396_end_0 = const()[name = string("op_2396_end_0"), val = tensor([1, 42, 1024])]; + tensor var_2396_end_mask_0 = const()[name = string("op_2396_end_mask_0"), val = tensor([true, true, true])]; + tensor var_2396_cast_fp16 = slice_by_index(begin = var_2396_begin_0, end = var_2396_end_0, end_mask = var_2396_end_mask_0, x = cache_37_cast_fp16)[name = string("op_2396_cast_fp16")]; + bool var_2402_interleave_0 = const()[name = string("op_2402_interleave_0"), val = bool(false)]; + tensor var_2402_cast_fp16 = concat(axis = var_69, interleave = var_2402_interleave_0, values = (var_2396_cast_fp16, key_19_cast_fp16))[name = string("op_2402_cast_fp16")]; + tensor encoder_layers_9_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(195960832))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196747328))))[name = string("encoder_layers_9_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_9_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_9_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196747520)))]; + tensor linear_84_cast_fp16 = linear(bias = encoder_layers_9_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_9_self_attn_linear_q_weight_to_fp16_palettized, x = key_19_cast_fp16)[name = string("linear_84_cast_fp16")]; + tensor var_2407 = const()[name = string("op_2407"), val = tensor([1, -1, 8, 128])]; + tensor q_55_cast_fp16 = reshape(shape = var_2407, x = linear_84_cast_fp16)[name = string("q_55_cast_fp16")]; + tensor encoder_layers_9_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196749632))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197536128))))[name = string("encoder_layers_9_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_9_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_9_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197536320)))]; + tensor linear_85_cast_fp16 = linear(bias = encoder_layers_9_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_9_self_attn_linear_k_weight_to_fp16_palettized, x = input_507_cast_fp16)[name = string("linear_85_cast_fp16")]; + tensor var_2412 = const()[name = string("op_2412"), val = tensor([1, -1, 8, 128])]; + tensor k_37_cast_fp16 = reshape(shape = var_2412, x = linear_85_cast_fp16)[name = string("k_37_cast_fp16")]; + tensor encoder_layers_9_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197538432))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198324928))))[name = string("encoder_layers_9_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_9_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_9_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198325120)))]; + tensor linear_86_cast_fp16 = linear(bias = encoder_layers_9_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_9_self_attn_linear_v_weight_to_fp16_palettized, x = input_507_cast_fp16)[name = string("linear_86_cast_fp16")]; + tensor var_2417 = const()[name = string("op_2417"), val = tensor([1, -1, 8, 128])]; + tensor v_19_cast_fp16 = reshape(shape = var_2417, x = linear_86_cast_fp16)[name = string("v_19_cast_fp16")]; + tensor value_27_perm_0 = const()[name = string("value_27_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_9_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_9_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198327232)))]; + tensor var_2430_cast_fp16 = add(x = q_55_cast_fp16, y = encoder_layers_9_self_attn_pos_bias_u_to_fp16)[name = string("op_2430_cast_fp16")]; + tensor encoder_layers_9_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_9_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198329344)))]; + tensor var_2432_cast_fp16 = add(x = q_55_cast_fp16, y = encoder_layers_9_self_attn_pos_bias_v_to_fp16)[name = string("op_2432_cast_fp16")]; + tensor q_with_bias_v_19_perm_0 = const()[name = string("q_with_bias_v_19_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_241_transpose_x_0 = const()[name = string("x_241_transpose_x_0"), val = bool(false)]; + bool x_241_transpose_y_0 = const()[name = string("x_241_transpose_y_0"), val = bool(false)]; + tensor op_2434_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198331456))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198445184))))[name = string("op_2434_to_fp16_quantized")]; + tensor q_with_bias_v_19_cast_fp16 = transpose(perm = q_with_bias_v_19_perm_0, x = var_2432_cast_fp16)[name = string("transpose_281")]; + tensor x_241_cast_fp16 = matmul(transpose_x = x_241_transpose_x_0, transpose_y = x_241_transpose_y_0, x = q_with_bias_v_19_cast_fp16, y = op_2434_to_fp16_quantized)[name = string("x_241_cast_fp16")]; + tensor x_243_pad_0 = const()[name = string("x_243_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_243_mode_0 = const()[name = string("x_243_mode_0"), val = string("constant")]; + fp16 const_196_to_fp16 = const()[name = string("const_196_to_fp16"), val = fp16(0x0p+0)]; + tensor x_243_cast_fp16 = pad(constant_val = const_196_to_fp16, mode = x_243_mode_0, pad = x_243_pad_0, x = x_241_cast_fp16)[name = string("x_243_cast_fp16")]; + tensor var_2442 = const()[name = string("op_2442"), val = tensor([1, 8, -1, 14])]; + tensor x_245_cast_fp16 = reshape(shape = var_2442, x = x_243_cast_fp16)[name = string("x_245_cast_fp16")]; + tensor var_2446_begin_0 = const()[name = string("op_2446_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_2446_end_0 = const()[name = string("op_2446_end_0"), val = tensor([1, 8, 112, 14])]; + tensor var_2446_end_mask_0 = const()[name = string("op_2446_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_2446_cast_fp16 = slice_by_index(begin = var_2446_begin_0, end = var_2446_end_0, end_mask = var_2446_end_mask_0, x = x_245_cast_fp16)[name = string("op_2446_cast_fp16")]; + tensor var_2447 = const()[name = string("op_2447"), val = tensor([1, 8, 14, 111])]; + tensor matrix_bd_37_cast_fp16 = reshape(shape = var_2447, x = var_2446_cast_fp16)[name = string("matrix_bd_37_cast_fp16")]; + bool matrix_ac_19_transpose_x_0 = const()[name = string("matrix_ac_19_transpose_x_0"), val = bool(false)]; + bool matrix_ac_19_transpose_y_0 = const()[name = string("matrix_ac_19_transpose_y_0"), val = bool(false)]; + tensor transpose_114_perm_0 = const()[name = string("transpose_114_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_115_perm_0 = const()[name = string("transpose_115_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_115 = transpose(perm = transpose_115_perm_0, x = k_37_cast_fp16)[name = string("transpose_279")]; + tensor transpose_114 = transpose(perm = transpose_114_perm_0, x = var_2430_cast_fp16)[name = string("transpose_280")]; + tensor matrix_ac_19_cast_fp16 = matmul(transpose_x = matrix_ac_19_transpose_x_0, transpose_y = matrix_ac_19_transpose_y_0, x = transpose_114, y = transpose_115)[name = string("matrix_ac_19_cast_fp16")]; + tensor matrix_bd_39_begin_0 = const()[name = string("matrix_bd_39_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_39_end_0 = const()[name = string("matrix_bd_39_end_0"), val = tensor([1, 8, 14, 56])]; + tensor matrix_bd_39_end_mask_0 = const()[name = string("matrix_bd_39_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_39_cast_fp16 = slice_by_index(begin = matrix_bd_39_begin_0, end = matrix_bd_39_end_0, end_mask = matrix_bd_39_end_mask_0, x = matrix_bd_37_cast_fp16)[name = string("matrix_bd_39_cast_fp16")]; + tensor var_2456_cast_fp16 = add(x = matrix_ac_19_cast_fp16, y = matrix_bd_39_cast_fp16)[name = string("op_2456_cast_fp16")]; + fp16 _inversed_scores_37_y_0_to_fp16 = const()[name = string("_inversed_scores_37_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_37_cast_fp16 = mul(x = var_2456_cast_fp16, y = _inversed_scores_37_y_0_to_fp16)[name = string("_inversed_scores_37_cast_fp16")]; + tensor scores_39_cast_fp16 = select(a = var_46_to_fp16, b = _inversed_scores_37_cast_fp16, cond = mask_11)[name = string("scores_39_cast_fp16")]; + tensor var_2462_cast_fp16 = softmax(axis = var_60, x = scores_39_cast_fp16)[name = string("op_2462_cast_fp16")]; + tensor input_509_cast_fp16 = select(a = var_45_to_fp16, b = var_2462_cast_fp16, cond = mask_11)[name = string("input_509_cast_fp16")]; + bool x_247_transpose_x_0 = const()[name = string("x_247_transpose_x_0"), val = bool(false)]; + bool x_247_transpose_y_0 = const()[name = string("x_247_transpose_y_0"), val = bool(false)]; + tensor value_27_cast_fp16 = transpose(perm = value_27_perm_0, x = v_19_cast_fp16)[name = string("transpose_278")]; + tensor x_247_cast_fp16 = matmul(transpose_x = x_247_transpose_x_0, transpose_y = x_247_transpose_y_0, x = input_509_cast_fp16, y = value_27_cast_fp16)[name = string("x_247_cast_fp16")]; + tensor var_2466_perm_0 = const()[name = string("op_2466_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_2467 = const()[name = string("op_2467"), val = tensor([1, -1, 1024])]; + tensor var_2466_cast_fp16 = transpose(perm = var_2466_perm_0, x = x_247_cast_fp16)[name = string("transpose_277")]; + tensor input_511_cast_fp16 = reshape(shape = var_2467, x = var_2466_cast_fp16)[name = string("input_511_cast_fp16")]; + tensor encoder_layers_9_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198445504))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199232000))))[name = string("encoder_layers_9_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_9_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_9_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199232192)))]; + tensor linear_88_cast_fp16 = linear(bias = encoder_layers_9_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_9_self_attn_linear_out_weight_to_fp16_palettized, x = input_511_cast_fp16)[name = string("linear_88_cast_fp16")]; + tensor input_515_cast_fp16 = add(x = input_505_cast_fp16, y = linear_88_cast_fp16)[name = string("input_515_cast_fp16")]; + tensor x_251_axes_0 = const()[name = string("x_251_axes_0"), val = tensor([-1])]; + tensor encoder_layers_9_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_9_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199234304)))]; + tensor encoder_layers_9_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_9_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199236416)))]; + tensor x_251_cast_fp16 = layer_norm(axes = x_251_axes_0, beta = encoder_layers_9_norm_conv_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_9_norm_conv_weight_to_fp16, x = input_515_cast_fp16)[name = string("x_251_cast_fp16")]; + tensor input_517_perm_0 = const()[name = string("input_517_perm_0"), val = tensor([0, 2, 1])]; + string input_519_pad_type_0 = const()[name = string("input_519_pad_type_0"), val = string("valid")]; + tensor input_519_strides_0 = const()[name = string("input_519_strides_0"), val = tensor([1])]; + tensor input_519_pad_0 = const()[name = string("input_519_pad_0"), val = tensor([0, 0])]; + tensor input_519_dilations_0 = const()[name = string("input_519_dilations_0"), val = tensor([1])]; + int32 input_519_groups_0 = const()[name = string("input_519_groups_0"), val = int32(1)]; + tensor encoder_layers_9_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199238528))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(201335744))))[name = string("encoder_layers_9_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_517_cast_fp16 = transpose(perm = input_517_perm_0, x = x_251_cast_fp16)[name = string("transpose_276")]; + tensor input_519_cast_fp16 = conv(dilations = input_519_dilations_0, groups = input_519_groups_0, pad = input_519_pad_0, pad_type = input_519_pad_type_0, strides = input_519_strides_0, weight = encoder_layers_9_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_517_cast_fp16)[name = string("input_519_cast_fp16")]; + int32 x_253_split_num_splits_0 = const()[name = string("x_253_split_num_splits_0"), val = int32(2)]; + int32 x_253_split_axis_0 = const()[name = string("x_253_split_axis_0"), val = int32(1)]; + tensor x_253_split_cast_fp16_0, tensor x_253_split_cast_fp16_1 = split(axis = x_253_split_axis_0, num_splits = x_253_split_num_splits_0, x = input_519_cast_fp16)[name = string("x_253_split_cast_fp16")]; + tensor x_253_split_1_sigmoid_cast_fp16 = sigmoid(x = x_253_split_cast_fp16_1)[name = string("x_253_split_1_sigmoid_cast_fp16")]; + tensor x_253_cast_fp16 = mul(x = x_253_split_cast_fp16_0, y = x_253_split_1_sigmoid_cast_fp16)[name = string("x_253_cast_fp16")]; + tensor input_521_cast_fp16 = select(a = var_45_to_fp16, b = x_253_cast_fp16, cond = var_576)[name = string("input_521_cast_fp16")]; + bool new_x_39_interleave_0 = const()[name = string("new_x_39_interleave_0"), val = bool(false)]; + tensor new_x_39_cast_fp16 = concat(axis = var_60, interleave = new_x_39_interleave_0, values = (cache_39_cast_fp16, input_521_cast_fp16))[name = string("new_x_39_cast_fp16")]; + tensor var_2506_begin_0 = const()[name = string("op_2506_begin_0"), val = tensor([0, 0, 14])]; + tensor var_2506_end_0 = const()[name = string("op_2506_end_0"), val = tensor([1, 1024, 22])]; + tensor var_2506_end_mask_0 = const()[name = string("op_2506_end_mask_0"), val = tensor([true, true, true])]; + tensor var_2506_cast_fp16 = slice_by_index(begin = var_2506_begin_0, end = var_2506_end_0, end_mask = var_2506_end_mask_0, x = new_x_39_cast_fp16)[name = string("op_2506_cast_fp16")]; + string x_255_pad_type_0 = const()[name = string("x_255_pad_type_0"), val = string("valid")]; + int32 x_255_groups_0 = const()[name = string("x_255_groups_0"), val = int32(1024)]; + tensor x_255_strides_0 = const()[name = string("x_255_strides_0"), val = tensor([1])]; + tensor x_255_pad_0 = const()[name = string("x_255_pad_0"), val = tensor([0, 0])]; + tensor x_255_dilations_0 = const()[name = string("x_255_dilations_0"), val = tensor([1])]; + tensor encoder_layers_9_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(201339904))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(201349184))))[name = string("encoder_layers_9_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_255_cast_fp16 = conv(dilations = x_255_dilations_0, groups = x_255_groups_0, pad = x_255_pad_0, pad_type = x_255_pad_type_0, strides = x_255_strides_0, weight = encoder_layers_9_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_39_cast_fp16)[name = string("x_255_cast_fp16")]; + tensor input_523_perm_0 = const()[name = string("input_523_perm_0"), val = tensor([0, 2, 1])]; + tensor x_257_axes_0 = const()[name = string("x_257_axes_0"), val = tensor([-1])]; + tensor encoder_layers_9_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_9_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(201351296)))]; + tensor encoder_layers_9_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_9_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(201353408)))]; + tensor input_523_cast_fp16 = transpose(perm = input_523_perm_0, x = x_255_cast_fp16)[name = string("transpose_275")]; + tensor x_257_cast_fp16 = layer_norm(axes = x_257_axes_0, beta = encoder_layers_9_conv_batch_norm_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_9_conv_batch_norm_weight_to_fp16, x = input_523_cast_fp16)[name = string("x_257_cast_fp16")]; + tensor input_525_perm_0 = const()[name = string("input_525_perm_0"), val = tensor([0, 2, 1])]; + tensor input_525_cast_fp16 = transpose(perm = input_525_perm_0, x = x_257_cast_fp16)[name = string("transpose_274")]; + tensor input_527_cast_fp16 = silu(x = input_525_cast_fp16)[name = string("input_527_cast_fp16")]; + string x_259_pad_type_0 = const()[name = string("x_259_pad_type_0"), val = string("valid")]; + tensor x_259_strides_0 = const()[name = string("x_259_strides_0"), val = tensor([1])]; + tensor x_259_pad_0 = const()[name = string("x_259_pad_0"), val = tensor([0, 0])]; + tensor x_259_dilations_0 = const()[name = string("x_259_dilations_0"), val = tensor([1])]; + int32 x_259_groups_0 = const()[name = string("x_259_groups_0"), val = int32(1)]; + tensor encoder_layers_9_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(201355520))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202404160))))[name = string("encoder_layers_9_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_259_cast_fp16 = conv(dilations = x_259_dilations_0, groups = x_259_groups_0, pad = x_259_pad_0, pad_type = x_259_pad_type_0, strides = x_259_strides_0, weight = encoder_layers_9_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_527_cast_fp16)[name = string("x_259_cast_fp16")]; + tensor input_529_perm_0 = const()[name = string("input_529_perm_0"), val = tensor([0, 2, 1])]; + tensor input_529_cast_fp16 = transpose(perm = input_529_perm_0, x = x_259_cast_fp16)[name = string("transpose_273")]; + tensor input_531_cast_fp16 = add(x = input_515_cast_fp16, y = input_529_cast_fp16)[name = string("input_531_cast_fp16")]; + tensor input_533_axes_0 = const()[name = string("input_533_axes_0"), val = tensor([-1])]; + tensor encoder_layers_9_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_9_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202406272)))]; + tensor encoder_layers_9_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_9_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202408384)))]; + tensor input_533_cast_fp16 = layer_norm(axes = input_533_axes_0, beta = encoder_layers_9_norm_feed_forward2_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_9_norm_feed_forward2_weight_to_fp16, x = input_531_cast_fp16)[name = string("input_533_cast_fp16")]; + tensor encoder_layers_9_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202410496))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(205556288))))[name = string("encoder_layers_9_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_9_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_9_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(205556480)))]; + tensor linear_89_cast_fp16 = linear(bias = encoder_layers_9_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_9_feed_forward2_linear1_weight_to_fp16_palettized, x = input_533_cast_fp16)[name = string("linear_89_cast_fp16")]; + tensor input_537_cast_fp16 = silu(x = linear_89_cast_fp16)[name = string("input_537_cast_fp16")]; + tensor encoder_layers_9_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(205564736))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(208710528))))[name = string("encoder_layers_9_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_9_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_9_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(208710720)))]; + tensor linear_90_cast_fp16 = linear(bias = encoder_layers_9_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_9_feed_forward2_linear2_weight_to_fp16_palettized, x = input_537_cast_fp16)[name = string("linear_90_cast_fp16")]; + fp16 var_2549_to_fp16 = const()[name = string("op_2549_to_fp16"), val = fp16(0x1p-1)]; + tensor var_2550_cast_fp16 = mul(x = linear_90_cast_fp16, y = var_2549_to_fp16)[name = string("op_2550_cast_fp16")]; + tensor input_543_cast_fp16 = add(x = input_531_cast_fp16, y = var_2550_cast_fp16)[name = string("input_543_cast_fp16")]; + tensor input_545_axes_0 = const()[name = string("input_545_axes_0"), val = tensor([-1])]; + tensor encoder_layers_9_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_9_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(208712832)))]; + tensor encoder_layers_9_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_9_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(208714944)))]; + tensor input_545_cast_fp16 = layer_norm(axes = input_545_axes_0, beta = encoder_layers_9_norm_out_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_9_norm_out_weight_to_fp16, x = input_543_cast_fp16)[name = string("input_545_cast_fp16")]; + tensor cache_41_begin_0 = const()[name = string("cache_41_begin_0"), val = tensor([10, 0, 0, 0])]; + tensor cache_41_end_0 = const()[name = string("cache_41_end_0"), val = tensor([11, 1, 42, 1024])]; + tensor cache_41_end_mask_0 = const()[name = string("cache_41_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_41_squeeze_mask_0 = const()[name = string("cache_41_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_41_cast_fp16 = slice_by_index(begin = cache_41_begin_0, end = cache_41_end_0, end_mask = cache_41_end_mask_0, squeeze_mask = cache_41_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_41_cast_fp16")]; + tensor cache_43_begin_0 = const()[name = string("cache_43_begin_0"), val = tensor([10, 0, 0, 0])]; + tensor cache_43_end_0 = const()[name = string("cache_43_end_0"), val = tensor([11, 1, 1024, 8])]; + tensor cache_43_end_mask_0 = const()[name = string("cache_43_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_43_squeeze_mask_0 = const()[name = string("cache_43_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_43_cast_fp16 = slice_by_index(begin = cache_43_begin_0, end = cache_43_end_0, end_mask = cache_43_end_mask_0, squeeze_mask = cache_43_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_43_cast_fp16")]; + tensor input_547_axes_0 = const()[name = string("input_547_axes_0"), val = tensor([-1])]; + tensor encoder_layers_10_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_10_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(208717056)))]; + tensor encoder_layers_10_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_10_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(208719168)))]; + tensor input_547_cast_fp16 = layer_norm(axes = input_547_axes_0, beta = encoder_layers_10_norm_feed_forward1_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_10_norm_feed_forward1_weight_to_fp16, x = input_545_cast_fp16)[name = string("input_547_cast_fp16")]; + tensor encoder_layers_10_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(208721280))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211867072))))[name = string("encoder_layers_10_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_10_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_10_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211867264)))]; + tensor linear_91_cast_fp16 = linear(bias = encoder_layers_10_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_10_feed_forward1_linear1_weight_to_fp16_palettized, x = input_547_cast_fp16)[name = string("linear_91_cast_fp16")]; + tensor input_551_cast_fp16 = silu(x = linear_91_cast_fp16)[name = string("input_551_cast_fp16")]; + tensor encoder_layers_10_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211875520))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215021312))))[name = string("encoder_layers_10_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_10_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_10_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215021504)))]; + tensor linear_92_cast_fp16 = linear(bias = encoder_layers_10_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_10_feed_forward1_linear2_weight_to_fp16_palettized, x = input_551_cast_fp16)[name = string("linear_92_cast_fp16")]; + fp16 var_2586_to_fp16 = const()[name = string("op_2586_to_fp16"), val = fp16(0x1p-1)]; + tensor var_2587_cast_fp16 = mul(x = linear_92_cast_fp16, y = var_2586_to_fp16)[name = string("op_2587_cast_fp16")]; + tensor input_557_cast_fp16 = add(x = input_545_cast_fp16, y = var_2587_cast_fp16)[name = string("input_557_cast_fp16")]; + tensor key_21_axes_0 = const()[name = string("key_21_axes_0"), val = tensor([-1])]; + tensor encoder_layers_10_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_10_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215023616)))]; + tensor encoder_layers_10_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_10_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215025728)))]; + tensor key_21_cast_fp16 = layer_norm(axes = key_21_axes_0, beta = encoder_layers_10_norm_self_att_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_10_norm_self_att_weight_to_fp16, x = input_557_cast_fp16)[name = string("key_21_cast_fp16")]; + bool input_559_interleave_0 = const()[name = string("input_559_interleave_0"), val = bool(false)]; + tensor input_559_cast_fp16 = concat(axis = var_69, interleave = input_559_interleave_0, values = (cache_41_cast_fp16, key_21_cast_fp16))[name = string("input_559_cast_fp16")]; + tensor var_2609_begin_0 = const()[name = string("op_2609_begin_0"), val = tensor([0, 14, 0])]; + tensor var_2609_end_0 = const()[name = string("op_2609_end_0"), val = tensor([1, 42, 1024])]; + tensor var_2609_end_mask_0 = const()[name = string("op_2609_end_mask_0"), val = tensor([true, true, true])]; + tensor var_2609_cast_fp16 = slice_by_index(begin = var_2609_begin_0, end = var_2609_end_0, end_mask = var_2609_end_mask_0, x = cache_41_cast_fp16)[name = string("op_2609_cast_fp16")]; + bool var_2615_interleave_0 = const()[name = string("op_2615_interleave_0"), val = bool(false)]; + tensor var_2615_cast_fp16 = concat(axis = var_69, interleave = var_2615_interleave_0, values = (var_2609_cast_fp16, key_21_cast_fp16))[name = string("op_2615_cast_fp16")]; + tensor encoder_layers_10_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215027840))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215814336))))[name = string("encoder_layers_10_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_10_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_10_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215814528)))]; + tensor linear_93_cast_fp16 = linear(bias = encoder_layers_10_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_10_self_attn_linear_q_weight_to_fp16_palettized, x = key_21_cast_fp16)[name = string("linear_93_cast_fp16")]; + tensor var_2620 = const()[name = string("op_2620"), val = tensor([1, -1, 8, 128])]; + tensor q_61_cast_fp16 = reshape(shape = var_2620, x = linear_93_cast_fp16)[name = string("q_61_cast_fp16")]; + tensor encoder_layers_10_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215816640))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216603136))))[name = string("encoder_layers_10_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_10_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_10_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216603328)))]; + tensor linear_94_cast_fp16 = linear(bias = encoder_layers_10_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_10_self_attn_linear_k_weight_to_fp16_palettized, x = input_559_cast_fp16)[name = string("linear_94_cast_fp16")]; + tensor var_2625 = const()[name = string("op_2625"), val = tensor([1, -1, 8, 128])]; + tensor k_41_cast_fp16 = reshape(shape = var_2625, x = linear_94_cast_fp16)[name = string("k_41_cast_fp16")]; + tensor encoder_layers_10_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216605440))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217391936))))[name = string("encoder_layers_10_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_10_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_10_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217392128)))]; + tensor linear_95_cast_fp16 = linear(bias = encoder_layers_10_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_10_self_attn_linear_v_weight_to_fp16_palettized, x = input_559_cast_fp16)[name = string("linear_95_cast_fp16")]; + tensor var_2630 = const()[name = string("op_2630"), val = tensor([1, -1, 8, 128])]; + tensor v_21_cast_fp16 = reshape(shape = var_2630, x = linear_95_cast_fp16)[name = string("v_21_cast_fp16")]; + tensor value_29_perm_0 = const()[name = string("value_29_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_10_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_10_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217394240)))]; + tensor var_2643_cast_fp16 = add(x = q_61_cast_fp16, y = encoder_layers_10_self_attn_pos_bias_u_to_fp16)[name = string("op_2643_cast_fp16")]; + tensor encoder_layers_10_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_10_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217396352)))]; + tensor var_2645_cast_fp16 = add(x = q_61_cast_fp16, y = encoder_layers_10_self_attn_pos_bias_v_to_fp16)[name = string("op_2645_cast_fp16")]; + tensor q_with_bias_v_21_perm_0 = const()[name = string("q_with_bias_v_21_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_267_transpose_x_0 = const()[name = string("x_267_transpose_x_0"), val = bool(false)]; + bool x_267_transpose_y_0 = const()[name = string("x_267_transpose_y_0"), val = bool(false)]; + tensor op_2647_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217398464))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217512192))))[name = string("op_2647_to_fp16_quantized")]; + tensor q_with_bias_v_21_cast_fp16 = transpose(perm = q_with_bias_v_21_perm_0, x = var_2645_cast_fp16)[name = string("transpose_272")]; + tensor x_267_cast_fp16 = matmul(transpose_x = x_267_transpose_x_0, transpose_y = x_267_transpose_y_0, x = q_with_bias_v_21_cast_fp16, y = op_2647_to_fp16_quantized)[name = string("x_267_cast_fp16")]; + tensor x_269_pad_0 = const()[name = string("x_269_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_269_mode_0 = const()[name = string("x_269_mode_0"), val = string("constant")]; + fp16 const_209_to_fp16 = const()[name = string("const_209_to_fp16"), val = fp16(0x0p+0)]; + tensor x_269_cast_fp16 = pad(constant_val = const_209_to_fp16, mode = x_269_mode_0, pad = x_269_pad_0, x = x_267_cast_fp16)[name = string("x_269_cast_fp16")]; + tensor var_2655 = const()[name = string("op_2655"), val = tensor([1, 8, -1, 14])]; + tensor x_271_cast_fp16 = reshape(shape = var_2655, x = x_269_cast_fp16)[name = string("x_271_cast_fp16")]; + tensor var_2659_begin_0 = const()[name = string("op_2659_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_2659_end_0 = const()[name = string("op_2659_end_0"), val = tensor([1, 8, 112, 14])]; + tensor var_2659_end_mask_0 = const()[name = string("op_2659_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_2659_cast_fp16 = slice_by_index(begin = var_2659_begin_0, end = var_2659_end_0, end_mask = var_2659_end_mask_0, x = x_271_cast_fp16)[name = string("op_2659_cast_fp16")]; + tensor var_2660 = const()[name = string("op_2660"), val = tensor([1, 8, 14, 111])]; + tensor matrix_bd_41_cast_fp16 = reshape(shape = var_2660, x = var_2659_cast_fp16)[name = string("matrix_bd_41_cast_fp16")]; + bool matrix_ac_21_transpose_x_0 = const()[name = string("matrix_ac_21_transpose_x_0"), val = bool(false)]; + bool matrix_ac_21_transpose_y_0 = const()[name = string("matrix_ac_21_transpose_y_0"), val = bool(false)]; + tensor transpose_116_perm_0 = const()[name = string("transpose_116_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_117_perm_0 = const()[name = string("transpose_117_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_117 = transpose(perm = transpose_117_perm_0, x = k_41_cast_fp16)[name = string("transpose_270")]; + tensor transpose_116 = transpose(perm = transpose_116_perm_0, x = var_2643_cast_fp16)[name = string("transpose_271")]; + tensor matrix_ac_21_cast_fp16 = matmul(transpose_x = matrix_ac_21_transpose_x_0, transpose_y = matrix_ac_21_transpose_y_0, x = transpose_116, y = transpose_117)[name = string("matrix_ac_21_cast_fp16")]; + tensor matrix_bd_43_begin_0 = const()[name = string("matrix_bd_43_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_43_end_0 = const()[name = string("matrix_bd_43_end_0"), val = tensor([1, 8, 14, 56])]; + tensor matrix_bd_43_end_mask_0 = const()[name = string("matrix_bd_43_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_43_cast_fp16 = slice_by_index(begin = matrix_bd_43_begin_0, end = matrix_bd_43_end_0, end_mask = matrix_bd_43_end_mask_0, x = matrix_bd_41_cast_fp16)[name = string("matrix_bd_43_cast_fp16")]; + tensor var_2669_cast_fp16 = add(x = matrix_ac_21_cast_fp16, y = matrix_bd_43_cast_fp16)[name = string("op_2669_cast_fp16")]; + fp16 _inversed_scores_41_y_0_to_fp16 = const()[name = string("_inversed_scores_41_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_41_cast_fp16 = mul(x = var_2669_cast_fp16, y = _inversed_scores_41_y_0_to_fp16)[name = string("_inversed_scores_41_cast_fp16")]; + tensor scores_43_cast_fp16 = select(a = var_46_to_fp16, b = _inversed_scores_41_cast_fp16, cond = mask_11)[name = string("scores_43_cast_fp16")]; + tensor var_2675_cast_fp16 = softmax(axis = var_60, x = scores_43_cast_fp16)[name = string("op_2675_cast_fp16")]; + tensor input_561_cast_fp16 = select(a = var_45_to_fp16, b = var_2675_cast_fp16, cond = mask_11)[name = string("input_561_cast_fp16")]; + bool x_273_transpose_x_0 = const()[name = string("x_273_transpose_x_0"), val = bool(false)]; + bool x_273_transpose_y_0 = const()[name = string("x_273_transpose_y_0"), val = bool(false)]; + tensor value_29_cast_fp16 = transpose(perm = value_29_perm_0, x = v_21_cast_fp16)[name = string("transpose_269")]; + tensor x_273_cast_fp16 = matmul(transpose_x = x_273_transpose_x_0, transpose_y = x_273_transpose_y_0, x = input_561_cast_fp16, y = value_29_cast_fp16)[name = string("x_273_cast_fp16")]; + tensor var_2679_perm_0 = const()[name = string("op_2679_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_2680 = const()[name = string("op_2680"), val = tensor([1, -1, 1024])]; + tensor var_2679_cast_fp16 = transpose(perm = var_2679_perm_0, x = x_273_cast_fp16)[name = string("transpose_268")]; + tensor input_563_cast_fp16 = reshape(shape = var_2680, x = var_2679_cast_fp16)[name = string("input_563_cast_fp16")]; + tensor encoder_layers_10_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217512512))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218299008))))[name = string("encoder_layers_10_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_10_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_10_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218299200)))]; + tensor linear_97_cast_fp16 = linear(bias = encoder_layers_10_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_10_self_attn_linear_out_weight_to_fp16_palettized, x = input_563_cast_fp16)[name = string("linear_97_cast_fp16")]; + tensor input_567_cast_fp16 = add(x = input_557_cast_fp16, y = linear_97_cast_fp16)[name = string("input_567_cast_fp16")]; + tensor x_277_axes_0 = const()[name = string("x_277_axes_0"), val = tensor([-1])]; + tensor encoder_layers_10_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_10_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218301312)))]; + tensor encoder_layers_10_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_10_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218303424)))]; + tensor x_277_cast_fp16 = layer_norm(axes = x_277_axes_0, beta = encoder_layers_10_norm_conv_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_10_norm_conv_weight_to_fp16, x = input_567_cast_fp16)[name = string("x_277_cast_fp16")]; + tensor input_569_perm_0 = const()[name = string("input_569_perm_0"), val = tensor([0, 2, 1])]; + string input_571_pad_type_0 = const()[name = string("input_571_pad_type_0"), val = string("valid")]; + tensor input_571_strides_0 = const()[name = string("input_571_strides_0"), val = tensor([1])]; + tensor input_571_pad_0 = const()[name = string("input_571_pad_0"), val = tensor([0, 0])]; + tensor input_571_dilations_0 = const()[name = string("input_571_dilations_0"), val = tensor([1])]; + int32 input_571_groups_0 = const()[name = string("input_571_groups_0"), val = int32(1)]; + tensor encoder_layers_10_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218305536))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220402752))))[name = string("encoder_layers_10_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_569_cast_fp16 = transpose(perm = input_569_perm_0, x = x_277_cast_fp16)[name = string("transpose_267")]; + tensor input_571_cast_fp16 = conv(dilations = input_571_dilations_0, groups = input_571_groups_0, pad = input_571_pad_0, pad_type = input_571_pad_type_0, strides = input_571_strides_0, weight = encoder_layers_10_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_569_cast_fp16)[name = string("input_571_cast_fp16")]; + int32 x_279_split_num_splits_0 = const()[name = string("x_279_split_num_splits_0"), val = int32(2)]; + int32 x_279_split_axis_0 = const()[name = string("x_279_split_axis_0"), val = int32(1)]; + tensor x_279_split_cast_fp16_0, tensor x_279_split_cast_fp16_1 = split(axis = x_279_split_axis_0, num_splits = x_279_split_num_splits_0, x = input_571_cast_fp16)[name = string("x_279_split_cast_fp16")]; + tensor x_279_split_1_sigmoid_cast_fp16 = sigmoid(x = x_279_split_cast_fp16_1)[name = string("x_279_split_1_sigmoid_cast_fp16")]; + tensor x_279_cast_fp16 = mul(x = x_279_split_cast_fp16_0, y = x_279_split_1_sigmoid_cast_fp16)[name = string("x_279_cast_fp16")]; + tensor input_573_cast_fp16 = select(a = var_45_to_fp16, b = x_279_cast_fp16, cond = var_576)[name = string("input_573_cast_fp16")]; + bool new_x_43_interleave_0 = const()[name = string("new_x_43_interleave_0"), val = bool(false)]; + tensor new_x_43_cast_fp16 = concat(axis = var_60, interleave = new_x_43_interleave_0, values = (cache_43_cast_fp16, input_573_cast_fp16))[name = string("new_x_43_cast_fp16")]; + tensor var_2719_begin_0 = const()[name = string("op_2719_begin_0"), val = tensor([0, 0, 14])]; + tensor var_2719_end_0 = const()[name = string("op_2719_end_0"), val = tensor([1, 1024, 22])]; + tensor var_2719_end_mask_0 = const()[name = string("op_2719_end_mask_0"), val = tensor([true, true, true])]; + tensor var_2719_cast_fp16 = slice_by_index(begin = var_2719_begin_0, end = var_2719_end_0, end_mask = var_2719_end_mask_0, x = new_x_43_cast_fp16)[name = string("op_2719_cast_fp16")]; + string x_281_pad_type_0 = const()[name = string("x_281_pad_type_0"), val = string("valid")]; + int32 x_281_groups_0 = const()[name = string("x_281_groups_0"), val = int32(1024)]; + tensor x_281_strides_0 = const()[name = string("x_281_strides_0"), val = tensor([1])]; + tensor x_281_pad_0 = const()[name = string("x_281_pad_0"), val = tensor([0, 0])]; + tensor x_281_dilations_0 = const()[name = string("x_281_dilations_0"), val = tensor([1])]; + tensor encoder_layers_10_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220406912))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220416192))))[name = string("encoder_layers_10_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_281_cast_fp16 = conv(dilations = x_281_dilations_0, groups = x_281_groups_0, pad = x_281_pad_0, pad_type = x_281_pad_type_0, strides = x_281_strides_0, weight = encoder_layers_10_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_43_cast_fp16)[name = string("x_281_cast_fp16")]; + tensor input_575_perm_0 = const()[name = string("input_575_perm_0"), val = tensor([0, 2, 1])]; + tensor x_283_axes_0 = const()[name = string("x_283_axes_0"), val = tensor([-1])]; + tensor encoder_layers_10_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_10_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220418304)))]; + tensor encoder_layers_10_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_10_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220420416)))]; + tensor input_575_cast_fp16 = transpose(perm = input_575_perm_0, x = x_281_cast_fp16)[name = string("transpose_266")]; + tensor x_283_cast_fp16 = layer_norm(axes = x_283_axes_0, beta = encoder_layers_10_conv_batch_norm_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_10_conv_batch_norm_weight_to_fp16, x = input_575_cast_fp16)[name = string("x_283_cast_fp16")]; + tensor input_577_perm_0 = const()[name = string("input_577_perm_0"), val = tensor([0, 2, 1])]; + tensor input_577_cast_fp16 = transpose(perm = input_577_perm_0, x = x_283_cast_fp16)[name = string("transpose_265")]; + tensor input_579_cast_fp16 = silu(x = input_577_cast_fp16)[name = string("input_579_cast_fp16")]; + string x_285_pad_type_0 = const()[name = string("x_285_pad_type_0"), val = string("valid")]; + tensor x_285_strides_0 = const()[name = string("x_285_strides_0"), val = tensor([1])]; + tensor x_285_pad_0 = const()[name = string("x_285_pad_0"), val = tensor([0, 0])]; + tensor x_285_dilations_0 = const()[name = string("x_285_dilations_0"), val = tensor([1])]; + int32 x_285_groups_0 = const()[name = string("x_285_groups_0"), val = int32(1)]; + tensor encoder_layers_10_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220422528))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221471168))))[name = string("encoder_layers_10_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_285_cast_fp16 = conv(dilations = x_285_dilations_0, groups = x_285_groups_0, pad = x_285_pad_0, pad_type = x_285_pad_type_0, strides = x_285_strides_0, weight = encoder_layers_10_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_579_cast_fp16)[name = string("x_285_cast_fp16")]; + tensor input_581_perm_0 = const()[name = string("input_581_perm_0"), val = tensor([0, 2, 1])]; + tensor input_581_cast_fp16 = transpose(perm = input_581_perm_0, x = x_285_cast_fp16)[name = string("transpose_264")]; + tensor input_583_cast_fp16 = add(x = input_567_cast_fp16, y = input_581_cast_fp16)[name = string("input_583_cast_fp16")]; + tensor input_585_axes_0 = const()[name = string("input_585_axes_0"), val = tensor([-1])]; + tensor encoder_layers_10_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_10_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221473280)))]; + tensor encoder_layers_10_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_10_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221475392)))]; + tensor input_585_cast_fp16 = layer_norm(axes = input_585_axes_0, beta = encoder_layers_10_norm_feed_forward2_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_10_norm_feed_forward2_weight_to_fp16, x = input_583_cast_fp16)[name = string("input_585_cast_fp16")]; + tensor encoder_layers_10_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221477504))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(224623296))))[name = string("encoder_layers_10_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_10_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_10_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(224623488)))]; + tensor linear_98_cast_fp16 = linear(bias = encoder_layers_10_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_10_feed_forward2_linear1_weight_to_fp16_palettized, x = input_585_cast_fp16)[name = string("linear_98_cast_fp16")]; + tensor input_589_cast_fp16 = silu(x = linear_98_cast_fp16)[name = string("input_589_cast_fp16")]; + tensor encoder_layers_10_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(224631744))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227777536))))[name = string("encoder_layers_10_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_10_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_10_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227777728)))]; + tensor linear_99_cast_fp16 = linear(bias = encoder_layers_10_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_10_feed_forward2_linear2_weight_to_fp16_palettized, x = input_589_cast_fp16)[name = string("linear_99_cast_fp16")]; + fp16 var_2762_to_fp16 = const()[name = string("op_2762_to_fp16"), val = fp16(0x1p-1)]; + tensor var_2763_cast_fp16 = mul(x = linear_99_cast_fp16, y = var_2762_to_fp16)[name = string("op_2763_cast_fp16")]; + tensor input_595_cast_fp16 = add(x = input_583_cast_fp16, y = var_2763_cast_fp16)[name = string("input_595_cast_fp16")]; + tensor input_597_axes_0 = const()[name = string("input_597_axes_0"), val = tensor([-1])]; + tensor encoder_layers_10_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_10_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227779840)))]; + tensor encoder_layers_10_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_10_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227781952)))]; + tensor input_597_cast_fp16 = layer_norm(axes = input_597_axes_0, beta = encoder_layers_10_norm_out_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_10_norm_out_weight_to_fp16, x = input_595_cast_fp16)[name = string("input_597_cast_fp16")]; + tensor cache_45_begin_0 = const()[name = string("cache_45_begin_0"), val = tensor([11, 0, 0, 0])]; + tensor cache_45_end_0 = const()[name = string("cache_45_end_0"), val = tensor([12, 1, 42, 1024])]; + tensor cache_45_end_mask_0 = const()[name = string("cache_45_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_45_squeeze_mask_0 = const()[name = string("cache_45_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_45_cast_fp16 = slice_by_index(begin = cache_45_begin_0, end = cache_45_end_0, end_mask = cache_45_end_mask_0, squeeze_mask = cache_45_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_45_cast_fp16")]; + tensor cache_47_begin_0 = const()[name = string("cache_47_begin_0"), val = tensor([11, 0, 0, 0])]; + tensor cache_47_end_0 = const()[name = string("cache_47_end_0"), val = tensor([12, 1, 1024, 8])]; + tensor cache_47_end_mask_0 = const()[name = string("cache_47_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_47_squeeze_mask_0 = const()[name = string("cache_47_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_47_cast_fp16 = slice_by_index(begin = cache_47_begin_0, end = cache_47_end_0, end_mask = cache_47_end_mask_0, squeeze_mask = cache_47_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_47_cast_fp16")]; + tensor input_599_axes_0 = const()[name = string("input_599_axes_0"), val = tensor([-1])]; + tensor encoder_layers_11_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_11_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227784064)))]; + tensor encoder_layers_11_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_11_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227786176)))]; + tensor input_599_cast_fp16 = layer_norm(axes = input_599_axes_0, beta = encoder_layers_11_norm_feed_forward1_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_11_norm_feed_forward1_weight_to_fp16, x = input_597_cast_fp16)[name = string("input_599_cast_fp16")]; + tensor encoder_layers_11_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227788288))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(230934080))))[name = string("encoder_layers_11_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_11_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_11_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(230934272)))]; + tensor linear_100_cast_fp16 = linear(bias = encoder_layers_11_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_11_feed_forward1_linear1_weight_to_fp16_palettized, x = input_599_cast_fp16)[name = string("linear_100_cast_fp16")]; + tensor input_603_cast_fp16 = silu(x = linear_100_cast_fp16)[name = string("input_603_cast_fp16")]; + tensor encoder_layers_11_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(230942528))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234088320))))[name = string("encoder_layers_11_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_11_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_11_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234088512)))]; + tensor linear_101_cast_fp16 = linear(bias = encoder_layers_11_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_11_feed_forward1_linear2_weight_to_fp16_palettized, x = input_603_cast_fp16)[name = string("linear_101_cast_fp16")]; + fp16 var_2799_to_fp16 = const()[name = string("op_2799_to_fp16"), val = fp16(0x1p-1)]; + tensor var_2800_cast_fp16 = mul(x = linear_101_cast_fp16, y = var_2799_to_fp16)[name = string("op_2800_cast_fp16")]; + tensor input_609_cast_fp16 = add(x = input_597_cast_fp16, y = var_2800_cast_fp16)[name = string("input_609_cast_fp16")]; + tensor key_23_axes_0 = const()[name = string("key_23_axes_0"), val = tensor([-1])]; + tensor encoder_layers_11_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_11_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234090624)))]; + tensor encoder_layers_11_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_11_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234092736)))]; + tensor key_23_cast_fp16 = layer_norm(axes = key_23_axes_0, beta = encoder_layers_11_norm_self_att_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_11_norm_self_att_weight_to_fp16, x = input_609_cast_fp16)[name = string("key_23_cast_fp16")]; + bool input_611_interleave_0 = const()[name = string("input_611_interleave_0"), val = bool(false)]; + tensor input_611_cast_fp16 = concat(axis = var_69, interleave = input_611_interleave_0, values = (cache_45_cast_fp16, key_23_cast_fp16))[name = string("input_611_cast_fp16")]; + tensor var_2822_begin_0 = const()[name = string("op_2822_begin_0"), val = tensor([0, 14, 0])]; + tensor var_2822_end_0 = const()[name = string("op_2822_end_0"), val = tensor([1, 42, 1024])]; + tensor var_2822_end_mask_0 = const()[name = string("op_2822_end_mask_0"), val = tensor([true, true, true])]; + tensor var_2822_cast_fp16 = slice_by_index(begin = var_2822_begin_0, end = var_2822_end_0, end_mask = var_2822_end_mask_0, x = cache_45_cast_fp16)[name = string("op_2822_cast_fp16")]; + bool var_2828_interleave_0 = const()[name = string("op_2828_interleave_0"), val = bool(false)]; + tensor var_2828_cast_fp16 = concat(axis = var_69, interleave = var_2828_interleave_0, values = (var_2822_cast_fp16, key_23_cast_fp16))[name = string("op_2828_cast_fp16")]; + tensor encoder_layers_11_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234094848))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234881344))))[name = string("encoder_layers_11_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_11_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_11_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234881536)))]; + tensor linear_102_cast_fp16 = linear(bias = encoder_layers_11_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_11_self_attn_linear_q_weight_to_fp16_palettized, x = key_23_cast_fp16)[name = string("linear_102_cast_fp16")]; + tensor var_2833 = const()[name = string("op_2833"), val = tensor([1, -1, 8, 128])]; + tensor q_67_cast_fp16 = reshape(shape = var_2833, x = linear_102_cast_fp16)[name = string("q_67_cast_fp16")]; + tensor encoder_layers_11_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234883648))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(235670144))))[name = string("encoder_layers_11_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_11_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_11_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(235670336)))]; + tensor linear_103_cast_fp16 = linear(bias = encoder_layers_11_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_11_self_attn_linear_k_weight_to_fp16_palettized, x = input_611_cast_fp16)[name = string("linear_103_cast_fp16")]; + tensor var_2838 = const()[name = string("op_2838"), val = tensor([1, -1, 8, 128])]; + tensor k_45_cast_fp16 = reshape(shape = var_2838, x = linear_103_cast_fp16)[name = string("k_45_cast_fp16")]; + tensor encoder_layers_11_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(235672448))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236458944))))[name = string("encoder_layers_11_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_11_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_11_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236459136)))]; + tensor linear_104_cast_fp16 = linear(bias = encoder_layers_11_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_11_self_attn_linear_v_weight_to_fp16_palettized, x = input_611_cast_fp16)[name = string("linear_104_cast_fp16")]; + tensor var_2843 = const()[name = string("op_2843"), val = tensor([1, -1, 8, 128])]; + tensor v_23_cast_fp16 = reshape(shape = var_2843, x = linear_104_cast_fp16)[name = string("v_23_cast_fp16")]; + tensor value_31_perm_0 = const()[name = string("value_31_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_11_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_11_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236461248)))]; + tensor var_2856_cast_fp16 = add(x = q_67_cast_fp16, y = encoder_layers_11_self_attn_pos_bias_u_to_fp16)[name = string("op_2856_cast_fp16")]; + tensor encoder_layers_11_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_11_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236463360)))]; + tensor var_2858_cast_fp16 = add(x = q_67_cast_fp16, y = encoder_layers_11_self_attn_pos_bias_v_to_fp16)[name = string("op_2858_cast_fp16")]; + tensor q_with_bias_v_23_perm_0 = const()[name = string("q_with_bias_v_23_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_293_transpose_x_0 = const()[name = string("x_293_transpose_x_0"), val = bool(false)]; + bool x_293_transpose_y_0 = const()[name = string("x_293_transpose_y_0"), val = bool(false)]; + tensor op_2860_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236465472))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236579200))))[name = string("op_2860_to_fp16_quantized")]; + tensor q_with_bias_v_23_cast_fp16 = transpose(perm = q_with_bias_v_23_perm_0, x = var_2858_cast_fp16)[name = string("transpose_263")]; + tensor x_293_cast_fp16 = matmul(transpose_x = x_293_transpose_x_0, transpose_y = x_293_transpose_y_0, x = q_with_bias_v_23_cast_fp16, y = op_2860_to_fp16_quantized)[name = string("x_293_cast_fp16")]; + tensor x_295_pad_0 = const()[name = string("x_295_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_295_mode_0 = const()[name = string("x_295_mode_0"), val = string("constant")]; + fp16 const_222_to_fp16 = const()[name = string("const_222_to_fp16"), val = fp16(0x0p+0)]; + tensor x_295_cast_fp16 = pad(constant_val = const_222_to_fp16, mode = x_295_mode_0, pad = x_295_pad_0, x = x_293_cast_fp16)[name = string("x_295_cast_fp16")]; + tensor var_2868 = const()[name = string("op_2868"), val = tensor([1, 8, -1, 14])]; + tensor x_297_cast_fp16 = reshape(shape = var_2868, x = x_295_cast_fp16)[name = string("x_297_cast_fp16")]; + tensor var_2872_begin_0 = const()[name = string("op_2872_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_2872_end_0 = const()[name = string("op_2872_end_0"), val = tensor([1, 8, 112, 14])]; + tensor var_2872_end_mask_0 = const()[name = string("op_2872_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_2872_cast_fp16 = slice_by_index(begin = var_2872_begin_0, end = var_2872_end_0, end_mask = var_2872_end_mask_0, x = x_297_cast_fp16)[name = string("op_2872_cast_fp16")]; + tensor var_2873 = const()[name = string("op_2873"), val = tensor([1, 8, 14, 111])]; + tensor matrix_bd_45_cast_fp16 = reshape(shape = var_2873, x = var_2872_cast_fp16)[name = string("matrix_bd_45_cast_fp16")]; + bool matrix_ac_23_transpose_x_0 = const()[name = string("matrix_ac_23_transpose_x_0"), val = bool(false)]; + bool matrix_ac_23_transpose_y_0 = const()[name = string("matrix_ac_23_transpose_y_0"), val = bool(false)]; + tensor transpose_118_perm_0 = const()[name = string("transpose_118_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_119_perm_0 = const()[name = string("transpose_119_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_119 = transpose(perm = transpose_119_perm_0, x = k_45_cast_fp16)[name = string("transpose_261")]; + tensor transpose_118 = transpose(perm = transpose_118_perm_0, x = var_2856_cast_fp16)[name = string("transpose_262")]; + tensor matrix_ac_23_cast_fp16 = matmul(transpose_x = matrix_ac_23_transpose_x_0, transpose_y = matrix_ac_23_transpose_y_0, x = transpose_118, y = transpose_119)[name = string("matrix_ac_23_cast_fp16")]; + tensor matrix_bd_47_begin_0 = const()[name = string("matrix_bd_47_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_47_end_0 = const()[name = string("matrix_bd_47_end_0"), val = tensor([1, 8, 14, 56])]; + tensor matrix_bd_47_end_mask_0 = const()[name = string("matrix_bd_47_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_47_cast_fp16 = slice_by_index(begin = matrix_bd_47_begin_0, end = matrix_bd_47_end_0, end_mask = matrix_bd_47_end_mask_0, x = matrix_bd_45_cast_fp16)[name = string("matrix_bd_47_cast_fp16")]; + tensor var_2882_cast_fp16 = add(x = matrix_ac_23_cast_fp16, y = matrix_bd_47_cast_fp16)[name = string("op_2882_cast_fp16")]; + fp16 _inversed_scores_45_y_0_to_fp16 = const()[name = string("_inversed_scores_45_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_45_cast_fp16 = mul(x = var_2882_cast_fp16, y = _inversed_scores_45_y_0_to_fp16)[name = string("_inversed_scores_45_cast_fp16")]; + tensor scores_47_cast_fp16 = select(a = var_46_to_fp16, b = _inversed_scores_45_cast_fp16, cond = mask_11)[name = string("scores_47_cast_fp16")]; + tensor var_2888_cast_fp16 = softmax(axis = var_60, x = scores_47_cast_fp16)[name = string("op_2888_cast_fp16")]; + tensor input_613_cast_fp16 = select(a = var_45_to_fp16, b = var_2888_cast_fp16, cond = mask_11)[name = string("input_613_cast_fp16")]; + bool x_299_transpose_x_0 = const()[name = string("x_299_transpose_x_0"), val = bool(false)]; + bool x_299_transpose_y_0 = const()[name = string("x_299_transpose_y_0"), val = bool(false)]; + tensor value_31_cast_fp16 = transpose(perm = value_31_perm_0, x = v_23_cast_fp16)[name = string("transpose_260")]; + tensor x_299_cast_fp16 = matmul(transpose_x = x_299_transpose_x_0, transpose_y = x_299_transpose_y_0, x = input_613_cast_fp16, y = value_31_cast_fp16)[name = string("x_299_cast_fp16")]; + tensor var_2892_perm_0 = const()[name = string("op_2892_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_2893 = const()[name = string("op_2893"), val = tensor([1, -1, 1024])]; + tensor var_2892_cast_fp16 = transpose(perm = var_2892_perm_0, x = x_299_cast_fp16)[name = string("transpose_259")]; + tensor input_615_cast_fp16 = reshape(shape = var_2893, x = var_2892_cast_fp16)[name = string("input_615_cast_fp16")]; + tensor encoder_layers_11_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236579520))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(237366016))))[name = string("encoder_layers_11_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_11_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_11_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(237366208)))]; + tensor linear_106_cast_fp16 = linear(bias = encoder_layers_11_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_11_self_attn_linear_out_weight_to_fp16_palettized, x = input_615_cast_fp16)[name = string("linear_106_cast_fp16")]; + tensor input_619_cast_fp16 = add(x = input_609_cast_fp16, y = linear_106_cast_fp16)[name = string("input_619_cast_fp16")]; + tensor x_303_axes_0 = const()[name = string("x_303_axes_0"), val = tensor([-1])]; + tensor encoder_layers_11_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_11_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(237368320)))]; + tensor encoder_layers_11_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_11_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(237370432)))]; + tensor x_303_cast_fp16 = layer_norm(axes = x_303_axes_0, beta = encoder_layers_11_norm_conv_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_11_norm_conv_weight_to_fp16, x = input_619_cast_fp16)[name = string("x_303_cast_fp16")]; + tensor input_621_perm_0 = const()[name = string("input_621_perm_0"), val = tensor([0, 2, 1])]; + string input_623_pad_type_0 = const()[name = string("input_623_pad_type_0"), val = string("valid")]; + tensor input_623_strides_0 = const()[name = string("input_623_strides_0"), val = tensor([1])]; + tensor input_623_pad_0 = const()[name = string("input_623_pad_0"), val = tensor([0, 0])]; + tensor input_623_dilations_0 = const()[name = string("input_623_dilations_0"), val = tensor([1])]; + int32 input_623_groups_0 = const()[name = string("input_623_groups_0"), val = int32(1)]; + tensor encoder_layers_11_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(237372544))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(239469760))))[name = string("encoder_layers_11_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_621_cast_fp16 = transpose(perm = input_621_perm_0, x = x_303_cast_fp16)[name = string("transpose_258")]; + tensor input_623_cast_fp16 = conv(dilations = input_623_dilations_0, groups = input_623_groups_0, pad = input_623_pad_0, pad_type = input_623_pad_type_0, strides = input_623_strides_0, weight = encoder_layers_11_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_621_cast_fp16)[name = string("input_623_cast_fp16")]; + int32 x_305_split_num_splits_0 = const()[name = string("x_305_split_num_splits_0"), val = int32(2)]; + int32 x_305_split_axis_0 = const()[name = string("x_305_split_axis_0"), val = int32(1)]; + tensor x_305_split_cast_fp16_0, tensor x_305_split_cast_fp16_1 = split(axis = x_305_split_axis_0, num_splits = x_305_split_num_splits_0, x = input_623_cast_fp16)[name = string("x_305_split_cast_fp16")]; + tensor x_305_split_1_sigmoid_cast_fp16 = sigmoid(x = x_305_split_cast_fp16_1)[name = string("x_305_split_1_sigmoid_cast_fp16")]; + tensor x_305_cast_fp16 = mul(x = x_305_split_cast_fp16_0, y = x_305_split_1_sigmoid_cast_fp16)[name = string("x_305_cast_fp16")]; + tensor input_625_cast_fp16 = select(a = var_45_to_fp16, b = x_305_cast_fp16, cond = var_576)[name = string("input_625_cast_fp16")]; + bool new_x_47_interleave_0 = const()[name = string("new_x_47_interleave_0"), val = bool(false)]; + tensor new_x_47_cast_fp16 = concat(axis = var_60, interleave = new_x_47_interleave_0, values = (cache_47_cast_fp16, input_625_cast_fp16))[name = string("new_x_47_cast_fp16")]; + tensor var_2932_begin_0 = const()[name = string("op_2932_begin_0"), val = tensor([0, 0, 14])]; + tensor var_2932_end_0 = const()[name = string("op_2932_end_0"), val = tensor([1, 1024, 22])]; + tensor var_2932_end_mask_0 = const()[name = string("op_2932_end_mask_0"), val = tensor([true, true, true])]; + tensor var_2932_cast_fp16 = slice_by_index(begin = var_2932_begin_0, end = var_2932_end_0, end_mask = var_2932_end_mask_0, x = new_x_47_cast_fp16)[name = string("op_2932_cast_fp16")]; + string x_307_pad_type_0 = const()[name = string("x_307_pad_type_0"), val = string("valid")]; + int32 x_307_groups_0 = const()[name = string("x_307_groups_0"), val = int32(1024)]; + tensor x_307_strides_0 = const()[name = string("x_307_strides_0"), val = tensor([1])]; + tensor x_307_pad_0 = const()[name = string("x_307_pad_0"), val = tensor([0, 0])]; + tensor x_307_dilations_0 = const()[name = string("x_307_dilations_0"), val = tensor([1])]; + tensor encoder_layers_11_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(239473920))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(239483200))))[name = string("encoder_layers_11_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_307_cast_fp16 = conv(dilations = x_307_dilations_0, groups = x_307_groups_0, pad = x_307_pad_0, pad_type = x_307_pad_type_0, strides = x_307_strides_0, weight = encoder_layers_11_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_47_cast_fp16)[name = string("x_307_cast_fp16")]; + tensor input_627_perm_0 = const()[name = string("input_627_perm_0"), val = tensor([0, 2, 1])]; + tensor x_309_axes_0 = const()[name = string("x_309_axes_0"), val = tensor([-1])]; + tensor encoder_layers_11_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_11_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(239485312)))]; + tensor encoder_layers_11_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_11_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(239487424)))]; + tensor input_627_cast_fp16 = transpose(perm = input_627_perm_0, x = x_307_cast_fp16)[name = string("transpose_257")]; + tensor x_309_cast_fp16 = layer_norm(axes = x_309_axes_0, beta = encoder_layers_11_conv_batch_norm_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_11_conv_batch_norm_weight_to_fp16, x = input_627_cast_fp16)[name = string("x_309_cast_fp16")]; + tensor input_629_perm_0 = const()[name = string("input_629_perm_0"), val = tensor([0, 2, 1])]; + tensor input_629_cast_fp16 = transpose(perm = input_629_perm_0, x = x_309_cast_fp16)[name = string("transpose_256")]; + tensor input_631_cast_fp16 = silu(x = input_629_cast_fp16)[name = string("input_631_cast_fp16")]; + string x_311_pad_type_0 = const()[name = string("x_311_pad_type_0"), val = string("valid")]; + tensor x_311_strides_0 = const()[name = string("x_311_strides_0"), val = tensor([1])]; + tensor x_311_pad_0 = const()[name = string("x_311_pad_0"), val = tensor([0, 0])]; + tensor x_311_dilations_0 = const()[name = string("x_311_dilations_0"), val = tensor([1])]; + int32 x_311_groups_0 = const()[name = string("x_311_groups_0"), val = int32(1)]; + tensor encoder_layers_11_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(239489536))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240538176))))[name = string("encoder_layers_11_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_311_cast_fp16 = conv(dilations = x_311_dilations_0, groups = x_311_groups_0, pad = x_311_pad_0, pad_type = x_311_pad_type_0, strides = x_311_strides_0, weight = encoder_layers_11_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_631_cast_fp16)[name = string("x_311_cast_fp16")]; + tensor input_633_perm_0 = const()[name = string("input_633_perm_0"), val = tensor([0, 2, 1])]; + tensor input_633_cast_fp16 = transpose(perm = input_633_perm_0, x = x_311_cast_fp16)[name = string("transpose_255")]; + tensor input_635_cast_fp16 = add(x = input_619_cast_fp16, y = input_633_cast_fp16)[name = string("input_635_cast_fp16")]; + tensor input_637_axes_0 = const()[name = string("input_637_axes_0"), val = tensor([-1])]; + tensor encoder_layers_11_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_11_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240540288)))]; + tensor encoder_layers_11_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_11_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240542400)))]; + tensor input_637_cast_fp16 = layer_norm(axes = input_637_axes_0, beta = encoder_layers_11_norm_feed_forward2_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_11_norm_feed_forward2_weight_to_fp16, x = input_635_cast_fp16)[name = string("input_637_cast_fp16")]; + tensor encoder_layers_11_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240544512))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(243690304))))[name = string("encoder_layers_11_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_11_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_11_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(243690496)))]; + tensor linear_107_cast_fp16 = linear(bias = encoder_layers_11_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_11_feed_forward2_linear1_weight_to_fp16_palettized, x = input_637_cast_fp16)[name = string("linear_107_cast_fp16")]; + tensor input_641_cast_fp16 = silu(x = linear_107_cast_fp16)[name = string("input_641_cast_fp16")]; + tensor encoder_layers_11_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(243698752))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(246844544))))[name = string("encoder_layers_11_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_11_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_11_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(246844736)))]; + tensor linear_108_cast_fp16 = linear(bias = encoder_layers_11_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_11_feed_forward2_linear2_weight_to_fp16_palettized, x = input_641_cast_fp16)[name = string("linear_108_cast_fp16")]; + fp16 var_2975_to_fp16 = const()[name = string("op_2975_to_fp16"), val = fp16(0x1p-1)]; + tensor var_2976_cast_fp16 = mul(x = linear_108_cast_fp16, y = var_2975_to_fp16)[name = string("op_2976_cast_fp16")]; + tensor input_647_cast_fp16 = add(x = input_635_cast_fp16, y = var_2976_cast_fp16)[name = string("input_647_cast_fp16")]; + tensor input_649_axes_0 = const()[name = string("input_649_axes_0"), val = tensor([-1])]; + tensor encoder_layers_11_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_11_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(246846848)))]; + tensor encoder_layers_11_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_11_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(246848960)))]; + tensor input_649_cast_fp16 = layer_norm(axes = input_649_axes_0, beta = encoder_layers_11_norm_out_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_11_norm_out_weight_to_fp16, x = input_647_cast_fp16)[name = string("input_649_cast_fp16")]; + tensor cache_49_begin_0 = const()[name = string("cache_49_begin_0"), val = tensor([12, 0, 0, 0])]; + tensor cache_49_end_0 = const()[name = string("cache_49_end_0"), val = tensor([13, 1, 42, 1024])]; + tensor cache_49_end_mask_0 = const()[name = string("cache_49_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_49_squeeze_mask_0 = const()[name = string("cache_49_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_49_cast_fp16 = slice_by_index(begin = cache_49_begin_0, end = cache_49_end_0, end_mask = cache_49_end_mask_0, squeeze_mask = cache_49_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_49_cast_fp16")]; + tensor cache_51_begin_0 = const()[name = string("cache_51_begin_0"), val = tensor([12, 0, 0, 0])]; + tensor cache_51_end_0 = const()[name = string("cache_51_end_0"), val = tensor([13, 1, 1024, 8])]; + tensor cache_51_end_mask_0 = const()[name = string("cache_51_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_51_squeeze_mask_0 = const()[name = string("cache_51_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_51_cast_fp16 = slice_by_index(begin = cache_51_begin_0, end = cache_51_end_0, end_mask = cache_51_end_mask_0, squeeze_mask = cache_51_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_51_cast_fp16")]; + tensor input_651_axes_0 = const()[name = string("input_651_axes_0"), val = tensor([-1])]; + tensor encoder_layers_12_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_12_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(246851072)))]; + tensor encoder_layers_12_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_12_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(246853184)))]; + tensor input_651_cast_fp16 = layer_norm(axes = input_651_axes_0, beta = encoder_layers_12_norm_feed_forward1_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_12_norm_feed_forward1_weight_to_fp16, x = input_649_cast_fp16)[name = string("input_651_cast_fp16")]; + tensor encoder_layers_12_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(246855296))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(250001088))))[name = string("encoder_layers_12_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_12_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_12_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(250001280)))]; + tensor linear_109_cast_fp16 = linear(bias = encoder_layers_12_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_12_feed_forward1_linear1_weight_to_fp16_palettized, x = input_651_cast_fp16)[name = string("linear_109_cast_fp16")]; + tensor input_655_cast_fp16 = silu(x = linear_109_cast_fp16)[name = string("input_655_cast_fp16")]; + tensor encoder_layers_12_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(250009536))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253155328))))[name = string("encoder_layers_12_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_12_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_12_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253155520)))]; + tensor linear_110_cast_fp16 = linear(bias = encoder_layers_12_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_12_feed_forward1_linear2_weight_to_fp16_palettized, x = input_655_cast_fp16)[name = string("linear_110_cast_fp16")]; + fp16 var_3012_to_fp16 = const()[name = string("op_3012_to_fp16"), val = fp16(0x1p-1)]; + tensor var_3013_cast_fp16 = mul(x = linear_110_cast_fp16, y = var_3012_to_fp16)[name = string("op_3013_cast_fp16")]; + tensor input_661_cast_fp16 = add(x = input_649_cast_fp16, y = var_3013_cast_fp16)[name = string("input_661_cast_fp16")]; + tensor key_25_axes_0 = const()[name = string("key_25_axes_0"), val = tensor([-1])]; + tensor encoder_layers_12_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_12_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253157632)))]; + tensor encoder_layers_12_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_12_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253159744)))]; + tensor key_25_cast_fp16 = layer_norm(axes = key_25_axes_0, beta = encoder_layers_12_norm_self_att_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_12_norm_self_att_weight_to_fp16, x = input_661_cast_fp16)[name = string("key_25_cast_fp16")]; + bool input_663_interleave_0 = const()[name = string("input_663_interleave_0"), val = bool(false)]; + tensor input_663_cast_fp16 = concat(axis = var_69, interleave = input_663_interleave_0, values = (cache_49_cast_fp16, key_25_cast_fp16))[name = string("input_663_cast_fp16")]; + tensor var_3035_begin_0 = const()[name = string("op_3035_begin_0"), val = tensor([0, 14, 0])]; + tensor var_3035_end_0 = const()[name = string("op_3035_end_0"), val = tensor([1, 42, 1024])]; + tensor var_3035_end_mask_0 = const()[name = string("op_3035_end_mask_0"), val = tensor([true, true, true])]; + tensor var_3035_cast_fp16 = slice_by_index(begin = var_3035_begin_0, end = var_3035_end_0, end_mask = var_3035_end_mask_0, x = cache_49_cast_fp16)[name = string("op_3035_cast_fp16")]; + bool var_3041_interleave_0 = const()[name = string("op_3041_interleave_0"), val = bool(false)]; + tensor var_3041_cast_fp16 = concat(axis = var_69, interleave = var_3041_interleave_0, values = (var_3035_cast_fp16, key_25_cast_fp16))[name = string("op_3041_cast_fp16")]; + tensor encoder_layers_12_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253161856))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253948352))))[name = string("encoder_layers_12_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_12_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_12_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253948544)))]; + tensor linear_111_cast_fp16 = linear(bias = encoder_layers_12_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_12_self_attn_linear_q_weight_to_fp16_palettized, x = key_25_cast_fp16)[name = string("linear_111_cast_fp16")]; + tensor var_3046 = const()[name = string("op_3046"), val = tensor([1, -1, 8, 128])]; + tensor q_73_cast_fp16 = reshape(shape = var_3046, x = linear_111_cast_fp16)[name = string("q_73_cast_fp16")]; + tensor encoder_layers_12_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253950656))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(254737152))))[name = string("encoder_layers_12_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_12_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_12_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(254737344)))]; + tensor linear_112_cast_fp16 = linear(bias = encoder_layers_12_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_12_self_attn_linear_k_weight_to_fp16_palettized, x = input_663_cast_fp16)[name = string("linear_112_cast_fp16")]; + tensor var_3051 = const()[name = string("op_3051"), val = tensor([1, -1, 8, 128])]; + tensor k_49_cast_fp16 = reshape(shape = var_3051, x = linear_112_cast_fp16)[name = string("k_49_cast_fp16")]; + tensor encoder_layers_12_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(254739456))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255525952))))[name = string("encoder_layers_12_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_12_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_12_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255526144)))]; + tensor linear_113_cast_fp16 = linear(bias = encoder_layers_12_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_12_self_attn_linear_v_weight_to_fp16_palettized, x = input_663_cast_fp16)[name = string("linear_113_cast_fp16")]; + tensor var_3056 = const()[name = string("op_3056"), val = tensor([1, -1, 8, 128])]; + tensor v_25_cast_fp16 = reshape(shape = var_3056, x = linear_113_cast_fp16)[name = string("v_25_cast_fp16")]; + tensor value_33_perm_0 = const()[name = string("value_33_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_12_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_12_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255528256)))]; + tensor var_3069_cast_fp16 = add(x = q_73_cast_fp16, y = encoder_layers_12_self_attn_pos_bias_u_to_fp16)[name = string("op_3069_cast_fp16")]; + tensor encoder_layers_12_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_12_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255530368)))]; + tensor var_3071_cast_fp16 = add(x = q_73_cast_fp16, y = encoder_layers_12_self_attn_pos_bias_v_to_fp16)[name = string("op_3071_cast_fp16")]; + tensor q_with_bias_v_25_perm_0 = const()[name = string("q_with_bias_v_25_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_319_transpose_x_0 = const()[name = string("x_319_transpose_x_0"), val = bool(false)]; + bool x_319_transpose_y_0 = const()[name = string("x_319_transpose_y_0"), val = bool(false)]; + tensor op_3073_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255532480))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255646208))))[name = string("op_3073_to_fp16_quantized")]; + tensor q_with_bias_v_25_cast_fp16 = transpose(perm = q_with_bias_v_25_perm_0, x = var_3071_cast_fp16)[name = string("transpose_254")]; + tensor x_319_cast_fp16 = matmul(transpose_x = x_319_transpose_x_0, transpose_y = x_319_transpose_y_0, x = q_with_bias_v_25_cast_fp16, y = op_3073_to_fp16_quantized)[name = string("x_319_cast_fp16")]; + tensor x_321_pad_0 = const()[name = string("x_321_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_321_mode_0 = const()[name = string("x_321_mode_0"), val = string("constant")]; + fp16 const_235_to_fp16 = const()[name = string("const_235_to_fp16"), val = fp16(0x0p+0)]; + tensor x_321_cast_fp16 = pad(constant_val = const_235_to_fp16, mode = x_321_mode_0, pad = x_321_pad_0, x = x_319_cast_fp16)[name = string("x_321_cast_fp16")]; + tensor var_3081 = const()[name = string("op_3081"), val = tensor([1, 8, -1, 14])]; + tensor x_323_cast_fp16 = reshape(shape = var_3081, x = x_321_cast_fp16)[name = string("x_323_cast_fp16")]; + tensor var_3085_begin_0 = const()[name = string("op_3085_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_3085_end_0 = const()[name = string("op_3085_end_0"), val = tensor([1, 8, 112, 14])]; + tensor var_3085_end_mask_0 = const()[name = string("op_3085_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_3085_cast_fp16 = slice_by_index(begin = var_3085_begin_0, end = var_3085_end_0, end_mask = var_3085_end_mask_0, x = x_323_cast_fp16)[name = string("op_3085_cast_fp16")]; + tensor var_3086 = const()[name = string("op_3086"), val = tensor([1, 8, 14, 111])]; + tensor matrix_bd_49_cast_fp16 = reshape(shape = var_3086, x = var_3085_cast_fp16)[name = string("matrix_bd_49_cast_fp16")]; + bool matrix_ac_25_transpose_x_0 = const()[name = string("matrix_ac_25_transpose_x_0"), val = bool(false)]; + bool matrix_ac_25_transpose_y_0 = const()[name = string("matrix_ac_25_transpose_y_0"), val = bool(false)]; + tensor transpose_120_perm_0 = const()[name = string("transpose_120_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_121_perm_0 = const()[name = string("transpose_121_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_121 = transpose(perm = transpose_121_perm_0, x = k_49_cast_fp16)[name = string("transpose_252")]; + tensor transpose_120 = transpose(perm = transpose_120_perm_0, x = var_3069_cast_fp16)[name = string("transpose_253")]; + tensor matrix_ac_25_cast_fp16 = matmul(transpose_x = matrix_ac_25_transpose_x_0, transpose_y = matrix_ac_25_transpose_y_0, x = transpose_120, y = transpose_121)[name = string("matrix_ac_25_cast_fp16")]; + tensor matrix_bd_51_begin_0 = const()[name = string("matrix_bd_51_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_51_end_0 = const()[name = string("matrix_bd_51_end_0"), val = tensor([1, 8, 14, 56])]; + tensor matrix_bd_51_end_mask_0 = const()[name = string("matrix_bd_51_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_51_cast_fp16 = slice_by_index(begin = matrix_bd_51_begin_0, end = matrix_bd_51_end_0, end_mask = matrix_bd_51_end_mask_0, x = matrix_bd_49_cast_fp16)[name = string("matrix_bd_51_cast_fp16")]; + tensor var_3095_cast_fp16 = add(x = matrix_ac_25_cast_fp16, y = matrix_bd_51_cast_fp16)[name = string("op_3095_cast_fp16")]; + fp16 _inversed_scores_49_y_0_to_fp16 = const()[name = string("_inversed_scores_49_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_49_cast_fp16 = mul(x = var_3095_cast_fp16, y = _inversed_scores_49_y_0_to_fp16)[name = string("_inversed_scores_49_cast_fp16")]; + tensor scores_51_cast_fp16 = select(a = var_46_to_fp16, b = _inversed_scores_49_cast_fp16, cond = mask_11)[name = string("scores_51_cast_fp16")]; + tensor var_3101_cast_fp16 = softmax(axis = var_60, x = scores_51_cast_fp16)[name = string("op_3101_cast_fp16")]; + tensor input_665_cast_fp16 = select(a = var_45_to_fp16, b = var_3101_cast_fp16, cond = mask_11)[name = string("input_665_cast_fp16")]; + bool x_325_transpose_x_0 = const()[name = string("x_325_transpose_x_0"), val = bool(false)]; + bool x_325_transpose_y_0 = const()[name = string("x_325_transpose_y_0"), val = bool(false)]; + tensor value_33_cast_fp16 = transpose(perm = value_33_perm_0, x = v_25_cast_fp16)[name = string("transpose_251")]; + tensor x_325_cast_fp16 = matmul(transpose_x = x_325_transpose_x_0, transpose_y = x_325_transpose_y_0, x = input_665_cast_fp16, y = value_33_cast_fp16)[name = string("x_325_cast_fp16")]; + tensor var_3105_perm_0 = const()[name = string("op_3105_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_3106 = const()[name = string("op_3106"), val = tensor([1, -1, 1024])]; + tensor var_3105_cast_fp16 = transpose(perm = var_3105_perm_0, x = x_325_cast_fp16)[name = string("transpose_250")]; + tensor input_667_cast_fp16 = reshape(shape = var_3106, x = var_3105_cast_fp16)[name = string("input_667_cast_fp16")]; + tensor encoder_layers_12_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255646528))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256433024))))[name = string("encoder_layers_12_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_12_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_12_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256433216)))]; + tensor linear_115_cast_fp16 = linear(bias = encoder_layers_12_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_12_self_attn_linear_out_weight_to_fp16_palettized, x = input_667_cast_fp16)[name = string("linear_115_cast_fp16")]; + tensor input_671_cast_fp16 = add(x = input_661_cast_fp16, y = linear_115_cast_fp16)[name = string("input_671_cast_fp16")]; + tensor x_329_axes_0 = const()[name = string("x_329_axes_0"), val = tensor([-1])]; + tensor encoder_layers_12_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_12_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256435328)))]; + tensor encoder_layers_12_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_12_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256437440)))]; + tensor x_329_cast_fp16 = layer_norm(axes = x_329_axes_0, beta = encoder_layers_12_norm_conv_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_12_norm_conv_weight_to_fp16, x = input_671_cast_fp16)[name = string("x_329_cast_fp16")]; + tensor input_673_perm_0 = const()[name = string("input_673_perm_0"), val = tensor([0, 2, 1])]; + string input_675_pad_type_0 = const()[name = string("input_675_pad_type_0"), val = string("valid")]; + tensor input_675_strides_0 = const()[name = string("input_675_strides_0"), val = tensor([1])]; + tensor input_675_pad_0 = const()[name = string("input_675_pad_0"), val = tensor([0, 0])]; + tensor input_675_dilations_0 = const()[name = string("input_675_dilations_0"), val = tensor([1])]; + int32 input_675_groups_0 = const()[name = string("input_675_groups_0"), val = int32(1)]; + tensor encoder_layers_12_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256439552))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(258536768))))[name = string("encoder_layers_12_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_673_cast_fp16 = transpose(perm = input_673_perm_0, x = x_329_cast_fp16)[name = string("transpose_249")]; + tensor input_675_cast_fp16 = conv(dilations = input_675_dilations_0, groups = input_675_groups_0, pad = input_675_pad_0, pad_type = input_675_pad_type_0, strides = input_675_strides_0, weight = encoder_layers_12_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_673_cast_fp16)[name = string("input_675_cast_fp16")]; + int32 x_331_split_num_splits_0 = const()[name = string("x_331_split_num_splits_0"), val = int32(2)]; + int32 x_331_split_axis_0 = const()[name = string("x_331_split_axis_0"), val = int32(1)]; + tensor x_331_split_cast_fp16_0, tensor x_331_split_cast_fp16_1 = split(axis = x_331_split_axis_0, num_splits = x_331_split_num_splits_0, x = input_675_cast_fp16)[name = string("x_331_split_cast_fp16")]; + tensor x_331_split_1_sigmoid_cast_fp16 = sigmoid(x = x_331_split_cast_fp16_1)[name = string("x_331_split_1_sigmoid_cast_fp16")]; + tensor x_331_cast_fp16 = mul(x = x_331_split_cast_fp16_0, y = x_331_split_1_sigmoid_cast_fp16)[name = string("x_331_cast_fp16")]; + tensor input_677_cast_fp16 = select(a = var_45_to_fp16, b = x_331_cast_fp16, cond = var_576)[name = string("input_677_cast_fp16")]; + bool new_x_51_interleave_0 = const()[name = string("new_x_51_interleave_0"), val = bool(false)]; + tensor new_x_51_cast_fp16 = concat(axis = var_60, interleave = new_x_51_interleave_0, values = (cache_51_cast_fp16, input_677_cast_fp16))[name = string("new_x_51_cast_fp16")]; + tensor var_3145_begin_0 = const()[name = string("op_3145_begin_0"), val = tensor([0, 0, 14])]; + tensor var_3145_end_0 = const()[name = string("op_3145_end_0"), val = tensor([1, 1024, 22])]; + tensor var_3145_end_mask_0 = const()[name = string("op_3145_end_mask_0"), val = tensor([true, true, true])]; + tensor var_3145_cast_fp16 = slice_by_index(begin = var_3145_begin_0, end = var_3145_end_0, end_mask = var_3145_end_mask_0, x = new_x_51_cast_fp16)[name = string("op_3145_cast_fp16")]; + string x_333_pad_type_0 = const()[name = string("x_333_pad_type_0"), val = string("valid")]; + int32 x_333_groups_0 = const()[name = string("x_333_groups_0"), val = int32(1024)]; + tensor x_333_strides_0 = const()[name = string("x_333_strides_0"), val = tensor([1])]; + tensor x_333_pad_0 = const()[name = string("x_333_pad_0"), val = tensor([0, 0])]; + tensor x_333_dilations_0 = const()[name = string("x_333_dilations_0"), val = tensor([1])]; + tensor encoder_layers_12_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(258540928))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(258550208))))[name = string("encoder_layers_12_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_333_cast_fp16 = conv(dilations = x_333_dilations_0, groups = x_333_groups_0, pad = x_333_pad_0, pad_type = x_333_pad_type_0, strides = x_333_strides_0, weight = encoder_layers_12_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_51_cast_fp16)[name = string("x_333_cast_fp16")]; + tensor input_679_perm_0 = const()[name = string("input_679_perm_0"), val = tensor([0, 2, 1])]; + tensor x_335_axes_0 = const()[name = string("x_335_axes_0"), val = tensor([-1])]; + tensor encoder_layers_12_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_12_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(258552320)))]; + tensor encoder_layers_12_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_12_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(258554432)))]; + tensor input_679_cast_fp16 = transpose(perm = input_679_perm_0, x = x_333_cast_fp16)[name = string("transpose_248")]; + tensor x_335_cast_fp16 = layer_norm(axes = x_335_axes_0, beta = encoder_layers_12_conv_batch_norm_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_12_conv_batch_norm_weight_to_fp16, x = input_679_cast_fp16)[name = string("x_335_cast_fp16")]; + tensor input_681_perm_0 = const()[name = string("input_681_perm_0"), val = tensor([0, 2, 1])]; + tensor input_681_cast_fp16 = transpose(perm = input_681_perm_0, x = x_335_cast_fp16)[name = string("transpose_247")]; + tensor input_683_cast_fp16 = silu(x = input_681_cast_fp16)[name = string("input_683_cast_fp16")]; + string x_337_pad_type_0 = const()[name = string("x_337_pad_type_0"), val = string("valid")]; + tensor x_337_strides_0 = const()[name = string("x_337_strides_0"), val = tensor([1])]; + tensor x_337_pad_0 = const()[name = string("x_337_pad_0"), val = tensor([0, 0])]; + tensor x_337_dilations_0 = const()[name = string("x_337_dilations_0"), val = tensor([1])]; + int32 x_337_groups_0 = const()[name = string("x_337_groups_0"), val = int32(1)]; + tensor encoder_layers_12_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(258556544))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259605184))))[name = string("encoder_layers_12_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_337_cast_fp16 = conv(dilations = x_337_dilations_0, groups = x_337_groups_0, pad = x_337_pad_0, pad_type = x_337_pad_type_0, strides = x_337_strides_0, weight = encoder_layers_12_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_683_cast_fp16)[name = string("x_337_cast_fp16")]; + tensor input_685_perm_0 = const()[name = string("input_685_perm_0"), val = tensor([0, 2, 1])]; + tensor input_685_cast_fp16 = transpose(perm = input_685_perm_0, x = x_337_cast_fp16)[name = string("transpose_246")]; + tensor input_687_cast_fp16 = add(x = input_671_cast_fp16, y = input_685_cast_fp16)[name = string("input_687_cast_fp16")]; + tensor input_689_axes_0 = const()[name = string("input_689_axes_0"), val = tensor([-1])]; + tensor encoder_layers_12_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_12_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259607296)))]; + tensor encoder_layers_12_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_12_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259609408)))]; + tensor input_689_cast_fp16 = layer_norm(axes = input_689_axes_0, beta = encoder_layers_12_norm_feed_forward2_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_12_norm_feed_forward2_weight_to_fp16, x = input_687_cast_fp16)[name = string("input_689_cast_fp16")]; + tensor encoder_layers_12_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259611520))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(262757312))))[name = string("encoder_layers_12_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_12_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_12_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(262757504)))]; + tensor linear_116_cast_fp16 = linear(bias = encoder_layers_12_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_12_feed_forward2_linear1_weight_to_fp16_palettized, x = input_689_cast_fp16)[name = string("linear_116_cast_fp16")]; + tensor input_693_cast_fp16 = silu(x = linear_116_cast_fp16)[name = string("input_693_cast_fp16")]; + tensor encoder_layers_12_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(262765760))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(265911552))))[name = string("encoder_layers_12_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_12_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_12_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(265911744)))]; + tensor linear_117_cast_fp16 = linear(bias = encoder_layers_12_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_12_feed_forward2_linear2_weight_to_fp16_palettized, x = input_693_cast_fp16)[name = string("linear_117_cast_fp16")]; + fp16 var_3188_to_fp16 = const()[name = string("op_3188_to_fp16"), val = fp16(0x1p-1)]; + tensor var_3189_cast_fp16 = mul(x = linear_117_cast_fp16, y = var_3188_to_fp16)[name = string("op_3189_cast_fp16")]; + tensor input_699_cast_fp16 = add(x = input_687_cast_fp16, y = var_3189_cast_fp16)[name = string("input_699_cast_fp16")]; + tensor input_701_axes_0 = const()[name = string("input_701_axes_0"), val = tensor([-1])]; + tensor encoder_layers_12_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_12_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(265913856)))]; + tensor encoder_layers_12_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_12_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(265915968)))]; + tensor input_701_cast_fp16 = layer_norm(axes = input_701_axes_0, beta = encoder_layers_12_norm_out_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_12_norm_out_weight_to_fp16, x = input_699_cast_fp16)[name = string("input_701_cast_fp16")]; + tensor cache_53_begin_0 = const()[name = string("cache_53_begin_0"), val = tensor([13, 0, 0, 0])]; + tensor cache_53_end_0 = const()[name = string("cache_53_end_0"), val = tensor([14, 1, 42, 1024])]; + tensor cache_53_end_mask_0 = const()[name = string("cache_53_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_53_squeeze_mask_0 = const()[name = string("cache_53_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_53_cast_fp16 = slice_by_index(begin = cache_53_begin_0, end = cache_53_end_0, end_mask = cache_53_end_mask_0, squeeze_mask = cache_53_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_53_cast_fp16")]; + tensor cache_55_begin_0 = const()[name = string("cache_55_begin_0"), val = tensor([13, 0, 0, 0])]; + tensor cache_55_end_0 = const()[name = string("cache_55_end_0"), val = tensor([14, 1, 1024, 8])]; + tensor cache_55_end_mask_0 = const()[name = string("cache_55_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_55_squeeze_mask_0 = const()[name = string("cache_55_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_55_cast_fp16 = slice_by_index(begin = cache_55_begin_0, end = cache_55_end_0, end_mask = cache_55_end_mask_0, squeeze_mask = cache_55_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_55_cast_fp16")]; + tensor input_703_axes_0 = const()[name = string("input_703_axes_0"), val = tensor([-1])]; + tensor encoder_layers_13_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_13_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(265918080)))]; + tensor encoder_layers_13_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_13_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(265920192)))]; + tensor input_703_cast_fp16 = layer_norm(axes = input_703_axes_0, beta = encoder_layers_13_norm_feed_forward1_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_13_norm_feed_forward1_weight_to_fp16, x = input_701_cast_fp16)[name = string("input_703_cast_fp16")]; + tensor encoder_layers_13_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(265922304))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(269068096))))[name = string("encoder_layers_13_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_13_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_13_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(269068288)))]; + tensor linear_118_cast_fp16 = linear(bias = encoder_layers_13_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_13_feed_forward1_linear1_weight_to_fp16_palettized, x = input_703_cast_fp16)[name = string("linear_118_cast_fp16")]; + tensor input_707_cast_fp16 = silu(x = linear_118_cast_fp16)[name = string("input_707_cast_fp16")]; + tensor encoder_layers_13_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(269076544))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272222336))))[name = string("encoder_layers_13_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_13_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_13_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272222528)))]; + tensor linear_119_cast_fp16 = linear(bias = encoder_layers_13_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_13_feed_forward1_linear2_weight_to_fp16_palettized, x = input_707_cast_fp16)[name = string("linear_119_cast_fp16")]; + fp16 var_3225_to_fp16 = const()[name = string("op_3225_to_fp16"), val = fp16(0x1p-1)]; + tensor var_3226_cast_fp16 = mul(x = linear_119_cast_fp16, y = var_3225_to_fp16)[name = string("op_3226_cast_fp16")]; + tensor input_713_cast_fp16 = add(x = input_701_cast_fp16, y = var_3226_cast_fp16)[name = string("input_713_cast_fp16")]; + tensor key_27_axes_0 = const()[name = string("key_27_axes_0"), val = tensor([-1])]; + tensor encoder_layers_13_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_13_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272224640)))]; + tensor encoder_layers_13_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_13_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272226752)))]; + tensor key_27_cast_fp16 = layer_norm(axes = key_27_axes_0, beta = encoder_layers_13_norm_self_att_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_13_norm_self_att_weight_to_fp16, x = input_713_cast_fp16)[name = string("key_27_cast_fp16")]; + bool input_715_interleave_0 = const()[name = string("input_715_interleave_0"), val = bool(false)]; + tensor input_715_cast_fp16 = concat(axis = var_69, interleave = input_715_interleave_0, values = (cache_53_cast_fp16, key_27_cast_fp16))[name = string("input_715_cast_fp16")]; + tensor var_3248_begin_0 = const()[name = string("op_3248_begin_0"), val = tensor([0, 14, 0])]; + tensor var_3248_end_0 = const()[name = string("op_3248_end_0"), val = tensor([1, 42, 1024])]; + tensor var_3248_end_mask_0 = const()[name = string("op_3248_end_mask_0"), val = tensor([true, true, true])]; + tensor var_3248_cast_fp16 = slice_by_index(begin = var_3248_begin_0, end = var_3248_end_0, end_mask = var_3248_end_mask_0, x = cache_53_cast_fp16)[name = string("op_3248_cast_fp16")]; + bool var_3254_interleave_0 = const()[name = string("op_3254_interleave_0"), val = bool(false)]; + tensor var_3254_cast_fp16 = concat(axis = var_69, interleave = var_3254_interleave_0, values = (var_3248_cast_fp16, key_27_cast_fp16))[name = string("op_3254_cast_fp16")]; + tensor encoder_layers_13_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272228864))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(273015360))))[name = string("encoder_layers_13_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_13_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_13_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(273015552)))]; + tensor linear_120_cast_fp16 = linear(bias = encoder_layers_13_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_13_self_attn_linear_q_weight_to_fp16_palettized, x = key_27_cast_fp16)[name = string("linear_120_cast_fp16")]; + tensor var_3259 = const()[name = string("op_3259"), val = tensor([1, -1, 8, 128])]; + tensor q_79_cast_fp16 = reshape(shape = var_3259, x = linear_120_cast_fp16)[name = string("q_79_cast_fp16")]; + tensor encoder_layers_13_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(273017664))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(273804160))))[name = string("encoder_layers_13_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_13_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_13_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(273804352)))]; + tensor linear_121_cast_fp16 = linear(bias = encoder_layers_13_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_13_self_attn_linear_k_weight_to_fp16_palettized, x = input_715_cast_fp16)[name = string("linear_121_cast_fp16")]; + tensor var_3264 = const()[name = string("op_3264"), val = tensor([1, -1, 8, 128])]; + tensor k_53_cast_fp16 = reshape(shape = var_3264, x = linear_121_cast_fp16)[name = string("k_53_cast_fp16")]; + tensor encoder_layers_13_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(273806464))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274592960))))[name = string("encoder_layers_13_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_13_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_13_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274593152)))]; + tensor linear_122_cast_fp16 = linear(bias = encoder_layers_13_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_13_self_attn_linear_v_weight_to_fp16_palettized, x = input_715_cast_fp16)[name = string("linear_122_cast_fp16")]; + tensor var_3269 = const()[name = string("op_3269"), val = tensor([1, -1, 8, 128])]; + tensor v_27_cast_fp16 = reshape(shape = var_3269, x = linear_122_cast_fp16)[name = string("v_27_cast_fp16")]; + tensor value_35_perm_0 = const()[name = string("value_35_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_13_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_13_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274595264)))]; + tensor var_3282_cast_fp16 = add(x = q_79_cast_fp16, y = encoder_layers_13_self_attn_pos_bias_u_to_fp16)[name = string("op_3282_cast_fp16")]; + tensor encoder_layers_13_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_13_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274597376)))]; + tensor var_3284_cast_fp16 = add(x = q_79_cast_fp16, y = encoder_layers_13_self_attn_pos_bias_v_to_fp16)[name = string("op_3284_cast_fp16")]; + tensor q_with_bias_v_27_perm_0 = const()[name = string("q_with_bias_v_27_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_345_transpose_x_0 = const()[name = string("x_345_transpose_x_0"), val = bool(false)]; + bool x_345_transpose_y_0 = const()[name = string("x_345_transpose_y_0"), val = bool(false)]; + tensor op_3286_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274599488))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274713216))))[name = string("op_3286_to_fp16_quantized")]; + tensor q_with_bias_v_27_cast_fp16 = transpose(perm = q_with_bias_v_27_perm_0, x = var_3284_cast_fp16)[name = string("transpose_245")]; + tensor x_345_cast_fp16 = matmul(transpose_x = x_345_transpose_x_0, transpose_y = x_345_transpose_y_0, x = q_with_bias_v_27_cast_fp16, y = op_3286_to_fp16_quantized)[name = string("x_345_cast_fp16")]; + tensor x_347_pad_0 = const()[name = string("x_347_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_347_mode_0 = const()[name = string("x_347_mode_0"), val = string("constant")]; + fp16 const_248_to_fp16 = const()[name = string("const_248_to_fp16"), val = fp16(0x0p+0)]; + tensor x_347_cast_fp16 = pad(constant_val = const_248_to_fp16, mode = x_347_mode_0, pad = x_347_pad_0, x = x_345_cast_fp16)[name = string("x_347_cast_fp16")]; + tensor var_3294 = const()[name = string("op_3294"), val = tensor([1, 8, -1, 14])]; + tensor x_349_cast_fp16 = reshape(shape = var_3294, x = x_347_cast_fp16)[name = string("x_349_cast_fp16")]; + tensor var_3298_begin_0 = const()[name = string("op_3298_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_3298_end_0 = const()[name = string("op_3298_end_0"), val = tensor([1, 8, 112, 14])]; + tensor var_3298_end_mask_0 = const()[name = string("op_3298_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_3298_cast_fp16 = slice_by_index(begin = var_3298_begin_0, end = var_3298_end_0, end_mask = var_3298_end_mask_0, x = x_349_cast_fp16)[name = string("op_3298_cast_fp16")]; + tensor var_3299 = const()[name = string("op_3299"), val = tensor([1, 8, 14, 111])]; + tensor matrix_bd_53_cast_fp16 = reshape(shape = var_3299, x = var_3298_cast_fp16)[name = string("matrix_bd_53_cast_fp16")]; + bool matrix_ac_27_transpose_x_0 = const()[name = string("matrix_ac_27_transpose_x_0"), val = bool(false)]; + bool matrix_ac_27_transpose_y_0 = const()[name = string("matrix_ac_27_transpose_y_0"), val = bool(false)]; + tensor transpose_122_perm_0 = const()[name = string("transpose_122_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_123_perm_0 = const()[name = string("transpose_123_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_123 = transpose(perm = transpose_123_perm_0, x = k_53_cast_fp16)[name = string("transpose_243")]; + tensor transpose_122 = transpose(perm = transpose_122_perm_0, x = var_3282_cast_fp16)[name = string("transpose_244")]; + tensor matrix_ac_27_cast_fp16 = matmul(transpose_x = matrix_ac_27_transpose_x_0, transpose_y = matrix_ac_27_transpose_y_0, x = transpose_122, y = transpose_123)[name = string("matrix_ac_27_cast_fp16")]; + tensor matrix_bd_55_begin_0 = const()[name = string("matrix_bd_55_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_55_end_0 = const()[name = string("matrix_bd_55_end_0"), val = tensor([1, 8, 14, 56])]; + tensor matrix_bd_55_end_mask_0 = const()[name = string("matrix_bd_55_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_55_cast_fp16 = slice_by_index(begin = matrix_bd_55_begin_0, end = matrix_bd_55_end_0, end_mask = matrix_bd_55_end_mask_0, x = matrix_bd_53_cast_fp16)[name = string("matrix_bd_55_cast_fp16")]; + tensor var_3308_cast_fp16 = add(x = matrix_ac_27_cast_fp16, y = matrix_bd_55_cast_fp16)[name = string("op_3308_cast_fp16")]; + fp16 _inversed_scores_53_y_0_to_fp16 = const()[name = string("_inversed_scores_53_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_53_cast_fp16 = mul(x = var_3308_cast_fp16, y = _inversed_scores_53_y_0_to_fp16)[name = string("_inversed_scores_53_cast_fp16")]; + tensor scores_55_cast_fp16 = select(a = var_46_to_fp16, b = _inversed_scores_53_cast_fp16, cond = mask_11)[name = string("scores_55_cast_fp16")]; + tensor var_3314_cast_fp16 = softmax(axis = var_60, x = scores_55_cast_fp16)[name = string("op_3314_cast_fp16")]; + tensor input_717_cast_fp16 = select(a = var_45_to_fp16, b = var_3314_cast_fp16, cond = mask_11)[name = string("input_717_cast_fp16")]; + bool x_351_transpose_x_0 = const()[name = string("x_351_transpose_x_0"), val = bool(false)]; + bool x_351_transpose_y_0 = const()[name = string("x_351_transpose_y_0"), val = bool(false)]; + tensor value_35_cast_fp16 = transpose(perm = value_35_perm_0, x = v_27_cast_fp16)[name = string("transpose_242")]; + tensor x_351_cast_fp16 = matmul(transpose_x = x_351_transpose_x_0, transpose_y = x_351_transpose_y_0, x = input_717_cast_fp16, y = value_35_cast_fp16)[name = string("x_351_cast_fp16")]; + tensor var_3318_perm_0 = const()[name = string("op_3318_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_3319 = const()[name = string("op_3319"), val = tensor([1, -1, 1024])]; + tensor var_3318_cast_fp16 = transpose(perm = var_3318_perm_0, x = x_351_cast_fp16)[name = string("transpose_241")]; + tensor input_719_cast_fp16 = reshape(shape = var_3319, x = var_3318_cast_fp16)[name = string("input_719_cast_fp16")]; + tensor encoder_layers_13_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274713536))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275500032))))[name = string("encoder_layers_13_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_13_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_13_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275500224)))]; + tensor linear_124_cast_fp16 = linear(bias = encoder_layers_13_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_13_self_attn_linear_out_weight_to_fp16_palettized, x = input_719_cast_fp16)[name = string("linear_124_cast_fp16")]; + tensor input_723_cast_fp16 = add(x = input_713_cast_fp16, y = linear_124_cast_fp16)[name = string("input_723_cast_fp16")]; + tensor x_355_axes_0 = const()[name = string("x_355_axes_0"), val = tensor([-1])]; + tensor encoder_layers_13_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_13_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275502336)))]; + tensor encoder_layers_13_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_13_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275504448)))]; + tensor x_355_cast_fp16 = layer_norm(axes = x_355_axes_0, beta = encoder_layers_13_norm_conv_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_13_norm_conv_weight_to_fp16, x = input_723_cast_fp16)[name = string("x_355_cast_fp16")]; + tensor input_725_perm_0 = const()[name = string("input_725_perm_0"), val = tensor([0, 2, 1])]; + string input_727_pad_type_0 = const()[name = string("input_727_pad_type_0"), val = string("valid")]; + tensor input_727_strides_0 = const()[name = string("input_727_strides_0"), val = tensor([1])]; + tensor input_727_pad_0 = const()[name = string("input_727_pad_0"), val = tensor([0, 0])]; + tensor input_727_dilations_0 = const()[name = string("input_727_dilations_0"), val = tensor([1])]; + int32 input_727_groups_0 = const()[name = string("input_727_groups_0"), val = int32(1)]; + tensor encoder_layers_13_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275506560))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(277603776))))[name = string("encoder_layers_13_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_725_cast_fp16 = transpose(perm = input_725_perm_0, x = x_355_cast_fp16)[name = string("transpose_240")]; + tensor input_727_cast_fp16 = conv(dilations = input_727_dilations_0, groups = input_727_groups_0, pad = input_727_pad_0, pad_type = input_727_pad_type_0, strides = input_727_strides_0, weight = encoder_layers_13_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_725_cast_fp16)[name = string("input_727_cast_fp16")]; + int32 x_357_split_num_splits_0 = const()[name = string("x_357_split_num_splits_0"), val = int32(2)]; + int32 x_357_split_axis_0 = const()[name = string("x_357_split_axis_0"), val = int32(1)]; + tensor x_357_split_cast_fp16_0, tensor x_357_split_cast_fp16_1 = split(axis = x_357_split_axis_0, num_splits = x_357_split_num_splits_0, x = input_727_cast_fp16)[name = string("x_357_split_cast_fp16")]; + tensor x_357_split_1_sigmoid_cast_fp16 = sigmoid(x = x_357_split_cast_fp16_1)[name = string("x_357_split_1_sigmoid_cast_fp16")]; + tensor x_357_cast_fp16 = mul(x = x_357_split_cast_fp16_0, y = x_357_split_1_sigmoid_cast_fp16)[name = string("x_357_cast_fp16")]; + tensor input_729_cast_fp16 = select(a = var_45_to_fp16, b = x_357_cast_fp16, cond = var_576)[name = string("input_729_cast_fp16")]; + bool new_x_55_interleave_0 = const()[name = string("new_x_55_interleave_0"), val = bool(false)]; + tensor new_x_55_cast_fp16 = concat(axis = var_60, interleave = new_x_55_interleave_0, values = (cache_55_cast_fp16, input_729_cast_fp16))[name = string("new_x_55_cast_fp16")]; + tensor var_3358_begin_0 = const()[name = string("op_3358_begin_0"), val = tensor([0, 0, 14])]; + tensor var_3358_end_0 = const()[name = string("op_3358_end_0"), val = tensor([1, 1024, 22])]; + tensor var_3358_end_mask_0 = const()[name = string("op_3358_end_mask_0"), val = tensor([true, true, true])]; + tensor var_3358_cast_fp16 = slice_by_index(begin = var_3358_begin_0, end = var_3358_end_0, end_mask = var_3358_end_mask_0, x = new_x_55_cast_fp16)[name = string("op_3358_cast_fp16")]; + string x_359_pad_type_0 = const()[name = string("x_359_pad_type_0"), val = string("valid")]; + int32 x_359_groups_0 = const()[name = string("x_359_groups_0"), val = int32(1024)]; + tensor x_359_strides_0 = const()[name = string("x_359_strides_0"), val = tensor([1])]; + tensor x_359_pad_0 = const()[name = string("x_359_pad_0"), val = tensor([0, 0])]; + tensor x_359_dilations_0 = const()[name = string("x_359_dilations_0"), val = tensor([1])]; + tensor encoder_layers_13_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(277607936))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(277617216))))[name = string("encoder_layers_13_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_359_cast_fp16 = conv(dilations = x_359_dilations_0, groups = x_359_groups_0, pad = x_359_pad_0, pad_type = x_359_pad_type_0, strides = x_359_strides_0, weight = encoder_layers_13_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_55_cast_fp16)[name = string("x_359_cast_fp16")]; + tensor input_731_perm_0 = const()[name = string("input_731_perm_0"), val = tensor([0, 2, 1])]; + tensor x_361_axes_0 = const()[name = string("x_361_axes_0"), val = tensor([-1])]; + tensor encoder_layers_13_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_13_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(277619328)))]; + tensor encoder_layers_13_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_13_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(277621440)))]; + tensor input_731_cast_fp16 = transpose(perm = input_731_perm_0, x = x_359_cast_fp16)[name = string("transpose_239")]; + tensor x_361_cast_fp16 = layer_norm(axes = x_361_axes_0, beta = encoder_layers_13_conv_batch_norm_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_13_conv_batch_norm_weight_to_fp16, x = input_731_cast_fp16)[name = string("x_361_cast_fp16")]; + tensor input_733_perm_0 = const()[name = string("input_733_perm_0"), val = tensor([0, 2, 1])]; + tensor input_733_cast_fp16 = transpose(perm = input_733_perm_0, x = x_361_cast_fp16)[name = string("transpose_238")]; + tensor input_735_cast_fp16 = silu(x = input_733_cast_fp16)[name = string("input_735_cast_fp16")]; + string x_363_pad_type_0 = const()[name = string("x_363_pad_type_0"), val = string("valid")]; + tensor x_363_strides_0 = const()[name = string("x_363_strides_0"), val = tensor([1])]; + tensor x_363_pad_0 = const()[name = string("x_363_pad_0"), val = tensor([0, 0])]; + tensor x_363_dilations_0 = const()[name = string("x_363_dilations_0"), val = tensor([1])]; + int32 x_363_groups_0 = const()[name = string("x_363_groups_0"), val = int32(1)]; + tensor encoder_layers_13_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(277623552))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278672192))))[name = string("encoder_layers_13_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_363_cast_fp16 = conv(dilations = x_363_dilations_0, groups = x_363_groups_0, pad = x_363_pad_0, pad_type = x_363_pad_type_0, strides = x_363_strides_0, weight = encoder_layers_13_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_735_cast_fp16)[name = string("x_363_cast_fp16")]; + tensor input_737_perm_0 = const()[name = string("input_737_perm_0"), val = tensor([0, 2, 1])]; + tensor input_737_cast_fp16 = transpose(perm = input_737_perm_0, x = x_363_cast_fp16)[name = string("transpose_237")]; + tensor input_739_cast_fp16 = add(x = input_723_cast_fp16, y = input_737_cast_fp16)[name = string("input_739_cast_fp16")]; + tensor input_741_axes_0 = const()[name = string("input_741_axes_0"), val = tensor([-1])]; + tensor encoder_layers_13_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_13_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278674304)))]; + tensor encoder_layers_13_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_13_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278676416)))]; + tensor input_741_cast_fp16 = layer_norm(axes = input_741_axes_0, beta = encoder_layers_13_norm_feed_forward2_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_13_norm_feed_forward2_weight_to_fp16, x = input_739_cast_fp16)[name = string("input_741_cast_fp16")]; + tensor encoder_layers_13_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278678528))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(281824320))))[name = string("encoder_layers_13_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_13_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_13_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(281824512)))]; + tensor linear_125_cast_fp16 = linear(bias = encoder_layers_13_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_13_feed_forward2_linear1_weight_to_fp16_palettized, x = input_741_cast_fp16)[name = string("linear_125_cast_fp16")]; + tensor input_745_cast_fp16 = silu(x = linear_125_cast_fp16)[name = string("input_745_cast_fp16")]; + tensor encoder_layers_13_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(281832768))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(284978560))))[name = string("encoder_layers_13_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_13_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_13_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(284978752)))]; + tensor linear_126_cast_fp16 = linear(bias = encoder_layers_13_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_13_feed_forward2_linear2_weight_to_fp16_palettized, x = input_745_cast_fp16)[name = string("linear_126_cast_fp16")]; + fp16 var_3401_to_fp16 = const()[name = string("op_3401_to_fp16"), val = fp16(0x1p-1)]; + tensor var_3402_cast_fp16 = mul(x = linear_126_cast_fp16, y = var_3401_to_fp16)[name = string("op_3402_cast_fp16")]; + tensor input_751_cast_fp16 = add(x = input_739_cast_fp16, y = var_3402_cast_fp16)[name = string("input_751_cast_fp16")]; + tensor input_753_axes_0 = const()[name = string("input_753_axes_0"), val = tensor([-1])]; + tensor encoder_layers_13_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_13_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(284980864)))]; + tensor encoder_layers_13_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_13_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(284982976)))]; + tensor input_753_cast_fp16 = layer_norm(axes = input_753_axes_0, beta = encoder_layers_13_norm_out_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_13_norm_out_weight_to_fp16, x = input_751_cast_fp16)[name = string("input_753_cast_fp16")]; + tensor cache_57_begin_0 = const()[name = string("cache_57_begin_0"), val = tensor([14, 0, 0, 0])]; + tensor cache_57_end_0 = const()[name = string("cache_57_end_0"), val = tensor([15, 1, 42, 1024])]; + tensor cache_57_end_mask_0 = const()[name = string("cache_57_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_57_squeeze_mask_0 = const()[name = string("cache_57_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_57_cast_fp16 = slice_by_index(begin = cache_57_begin_0, end = cache_57_end_0, end_mask = cache_57_end_mask_0, squeeze_mask = cache_57_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_57_cast_fp16")]; + tensor cache_59_begin_0 = const()[name = string("cache_59_begin_0"), val = tensor([14, 0, 0, 0])]; + tensor cache_59_end_0 = const()[name = string("cache_59_end_0"), val = tensor([15, 1, 1024, 8])]; + tensor cache_59_end_mask_0 = const()[name = string("cache_59_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_59_squeeze_mask_0 = const()[name = string("cache_59_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_59_cast_fp16 = slice_by_index(begin = cache_59_begin_0, end = cache_59_end_0, end_mask = cache_59_end_mask_0, squeeze_mask = cache_59_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_59_cast_fp16")]; + tensor input_755_axes_0 = const()[name = string("input_755_axes_0"), val = tensor([-1])]; + tensor encoder_layers_14_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_14_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(284985088)))]; + tensor encoder_layers_14_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_14_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(284987200)))]; + tensor input_755_cast_fp16 = layer_norm(axes = input_755_axes_0, beta = encoder_layers_14_norm_feed_forward1_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_14_norm_feed_forward1_weight_to_fp16, x = input_753_cast_fp16)[name = string("input_755_cast_fp16")]; + tensor encoder_layers_14_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(284989312))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(288135104))))[name = string("encoder_layers_14_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_14_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_14_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(288135296)))]; + tensor linear_127_cast_fp16 = linear(bias = encoder_layers_14_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_14_feed_forward1_linear1_weight_to_fp16_palettized, x = input_755_cast_fp16)[name = string("linear_127_cast_fp16")]; + tensor input_759_cast_fp16 = silu(x = linear_127_cast_fp16)[name = string("input_759_cast_fp16")]; + tensor encoder_layers_14_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(288143552))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291289344))))[name = string("encoder_layers_14_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_14_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_14_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291289536)))]; + tensor linear_128_cast_fp16 = linear(bias = encoder_layers_14_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_14_feed_forward1_linear2_weight_to_fp16_palettized, x = input_759_cast_fp16)[name = string("linear_128_cast_fp16")]; + fp16 var_3438_to_fp16 = const()[name = string("op_3438_to_fp16"), val = fp16(0x1p-1)]; + tensor var_3439_cast_fp16 = mul(x = linear_128_cast_fp16, y = var_3438_to_fp16)[name = string("op_3439_cast_fp16")]; + tensor input_765_cast_fp16 = add(x = input_753_cast_fp16, y = var_3439_cast_fp16)[name = string("input_765_cast_fp16")]; + tensor key_29_axes_0 = const()[name = string("key_29_axes_0"), val = tensor([-1])]; + tensor encoder_layers_14_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_14_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291291648)))]; + tensor encoder_layers_14_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_14_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291293760)))]; + tensor key_29_cast_fp16 = layer_norm(axes = key_29_axes_0, beta = encoder_layers_14_norm_self_att_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_14_norm_self_att_weight_to_fp16, x = input_765_cast_fp16)[name = string("key_29_cast_fp16")]; + bool input_767_interleave_0 = const()[name = string("input_767_interleave_0"), val = bool(false)]; + tensor input_767_cast_fp16 = concat(axis = var_69, interleave = input_767_interleave_0, values = (cache_57_cast_fp16, key_29_cast_fp16))[name = string("input_767_cast_fp16")]; + tensor var_3461_begin_0 = const()[name = string("op_3461_begin_0"), val = tensor([0, 14, 0])]; + tensor var_3461_end_0 = const()[name = string("op_3461_end_0"), val = tensor([1, 42, 1024])]; + tensor var_3461_end_mask_0 = const()[name = string("op_3461_end_mask_0"), val = tensor([true, true, true])]; + tensor var_3461_cast_fp16 = slice_by_index(begin = var_3461_begin_0, end = var_3461_end_0, end_mask = var_3461_end_mask_0, x = cache_57_cast_fp16)[name = string("op_3461_cast_fp16")]; + bool var_3467_interleave_0 = const()[name = string("op_3467_interleave_0"), val = bool(false)]; + tensor var_3467_cast_fp16 = concat(axis = var_69, interleave = var_3467_interleave_0, values = (var_3461_cast_fp16, key_29_cast_fp16))[name = string("op_3467_cast_fp16")]; + tensor encoder_layers_14_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291295872))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(292082368))))[name = string("encoder_layers_14_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_14_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_14_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(292082560)))]; + tensor linear_129_cast_fp16 = linear(bias = encoder_layers_14_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_14_self_attn_linear_q_weight_to_fp16_palettized, x = key_29_cast_fp16)[name = string("linear_129_cast_fp16")]; + tensor var_3472 = const()[name = string("op_3472"), val = tensor([1, -1, 8, 128])]; + tensor q_85_cast_fp16 = reshape(shape = var_3472, x = linear_129_cast_fp16)[name = string("q_85_cast_fp16")]; + tensor encoder_layers_14_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(292084672))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(292871168))))[name = string("encoder_layers_14_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_14_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_14_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(292871360)))]; + tensor linear_130_cast_fp16 = linear(bias = encoder_layers_14_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_14_self_attn_linear_k_weight_to_fp16_palettized, x = input_767_cast_fp16)[name = string("linear_130_cast_fp16")]; + tensor var_3477 = const()[name = string("op_3477"), val = tensor([1, -1, 8, 128])]; + tensor k_57_cast_fp16 = reshape(shape = var_3477, x = linear_130_cast_fp16)[name = string("k_57_cast_fp16")]; + tensor encoder_layers_14_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(292873472))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293659968))))[name = string("encoder_layers_14_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_14_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_14_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293660160)))]; + tensor linear_131_cast_fp16 = linear(bias = encoder_layers_14_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_14_self_attn_linear_v_weight_to_fp16_palettized, x = input_767_cast_fp16)[name = string("linear_131_cast_fp16")]; + tensor var_3482 = const()[name = string("op_3482"), val = tensor([1, -1, 8, 128])]; + tensor v_29_cast_fp16 = reshape(shape = var_3482, x = linear_131_cast_fp16)[name = string("v_29_cast_fp16")]; + tensor value_37_perm_0 = const()[name = string("value_37_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_14_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_14_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293662272)))]; + tensor var_3495_cast_fp16 = add(x = q_85_cast_fp16, y = encoder_layers_14_self_attn_pos_bias_u_to_fp16)[name = string("op_3495_cast_fp16")]; + tensor encoder_layers_14_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_14_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293664384)))]; + tensor var_3497_cast_fp16 = add(x = q_85_cast_fp16, y = encoder_layers_14_self_attn_pos_bias_v_to_fp16)[name = string("op_3497_cast_fp16")]; + tensor q_with_bias_v_29_perm_0 = const()[name = string("q_with_bias_v_29_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_371_transpose_x_0 = const()[name = string("x_371_transpose_x_0"), val = bool(false)]; + bool x_371_transpose_y_0 = const()[name = string("x_371_transpose_y_0"), val = bool(false)]; + tensor op_3499_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293666496))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293780224))))[name = string("op_3499_to_fp16_quantized")]; + tensor q_with_bias_v_29_cast_fp16 = transpose(perm = q_with_bias_v_29_perm_0, x = var_3497_cast_fp16)[name = string("transpose_236")]; + tensor x_371_cast_fp16 = matmul(transpose_x = x_371_transpose_x_0, transpose_y = x_371_transpose_y_0, x = q_with_bias_v_29_cast_fp16, y = op_3499_to_fp16_quantized)[name = string("x_371_cast_fp16")]; + tensor x_373_pad_0 = const()[name = string("x_373_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_373_mode_0 = const()[name = string("x_373_mode_0"), val = string("constant")]; + fp16 const_261_to_fp16 = const()[name = string("const_261_to_fp16"), val = fp16(0x0p+0)]; + tensor x_373_cast_fp16 = pad(constant_val = const_261_to_fp16, mode = x_373_mode_0, pad = x_373_pad_0, x = x_371_cast_fp16)[name = string("x_373_cast_fp16")]; + tensor var_3507 = const()[name = string("op_3507"), val = tensor([1, 8, -1, 14])]; + tensor x_375_cast_fp16 = reshape(shape = var_3507, x = x_373_cast_fp16)[name = string("x_375_cast_fp16")]; + tensor var_3511_begin_0 = const()[name = string("op_3511_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_3511_end_0 = const()[name = string("op_3511_end_0"), val = tensor([1, 8, 112, 14])]; + tensor var_3511_end_mask_0 = const()[name = string("op_3511_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_3511_cast_fp16 = slice_by_index(begin = var_3511_begin_0, end = var_3511_end_0, end_mask = var_3511_end_mask_0, x = x_375_cast_fp16)[name = string("op_3511_cast_fp16")]; + tensor var_3512 = const()[name = string("op_3512"), val = tensor([1, 8, 14, 111])]; + tensor matrix_bd_57_cast_fp16 = reshape(shape = var_3512, x = var_3511_cast_fp16)[name = string("matrix_bd_57_cast_fp16")]; + bool matrix_ac_29_transpose_x_0 = const()[name = string("matrix_ac_29_transpose_x_0"), val = bool(false)]; + bool matrix_ac_29_transpose_y_0 = const()[name = string("matrix_ac_29_transpose_y_0"), val = bool(false)]; + tensor transpose_124_perm_0 = const()[name = string("transpose_124_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_125_perm_0 = const()[name = string("transpose_125_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_125 = transpose(perm = transpose_125_perm_0, x = k_57_cast_fp16)[name = string("transpose_234")]; + tensor transpose_124 = transpose(perm = transpose_124_perm_0, x = var_3495_cast_fp16)[name = string("transpose_235")]; + tensor matrix_ac_29_cast_fp16 = matmul(transpose_x = matrix_ac_29_transpose_x_0, transpose_y = matrix_ac_29_transpose_y_0, x = transpose_124, y = transpose_125)[name = string("matrix_ac_29_cast_fp16")]; + tensor matrix_bd_59_begin_0 = const()[name = string("matrix_bd_59_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_59_end_0 = const()[name = string("matrix_bd_59_end_0"), val = tensor([1, 8, 14, 56])]; + tensor matrix_bd_59_end_mask_0 = const()[name = string("matrix_bd_59_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_59_cast_fp16 = slice_by_index(begin = matrix_bd_59_begin_0, end = matrix_bd_59_end_0, end_mask = matrix_bd_59_end_mask_0, x = matrix_bd_57_cast_fp16)[name = string("matrix_bd_59_cast_fp16")]; + tensor var_3521_cast_fp16 = add(x = matrix_ac_29_cast_fp16, y = matrix_bd_59_cast_fp16)[name = string("op_3521_cast_fp16")]; + fp16 _inversed_scores_57_y_0_to_fp16 = const()[name = string("_inversed_scores_57_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_57_cast_fp16 = mul(x = var_3521_cast_fp16, y = _inversed_scores_57_y_0_to_fp16)[name = string("_inversed_scores_57_cast_fp16")]; + tensor scores_59_cast_fp16 = select(a = var_46_to_fp16, b = _inversed_scores_57_cast_fp16, cond = mask_11)[name = string("scores_59_cast_fp16")]; + tensor var_3527_cast_fp16 = softmax(axis = var_60, x = scores_59_cast_fp16)[name = string("op_3527_cast_fp16")]; + tensor input_769_cast_fp16 = select(a = var_45_to_fp16, b = var_3527_cast_fp16, cond = mask_11)[name = string("input_769_cast_fp16")]; + bool x_377_transpose_x_0 = const()[name = string("x_377_transpose_x_0"), val = bool(false)]; + bool x_377_transpose_y_0 = const()[name = string("x_377_transpose_y_0"), val = bool(false)]; + tensor value_37_cast_fp16 = transpose(perm = value_37_perm_0, x = v_29_cast_fp16)[name = string("transpose_233")]; + tensor x_377_cast_fp16 = matmul(transpose_x = x_377_transpose_x_0, transpose_y = x_377_transpose_y_0, x = input_769_cast_fp16, y = value_37_cast_fp16)[name = string("x_377_cast_fp16")]; + tensor var_3531_perm_0 = const()[name = string("op_3531_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_3532 = const()[name = string("op_3532"), val = tensor([1, -1, 1024])]; + tensor var_3531_cast_fp16 = transpose(perm = var_3531_perm_0, x = x_377_cast_fp16)[name = string("transpose_232")]; + tensor input_771_cast_fp16 = reshape(shape = var_3532, x = var_3531_cast_fp16)[name = string("input_771_cast_fp16")]; + tensor encoder_layers_14_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293780544))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294567040))))[name = string("encoder_layers_14_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_14_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_14_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294567232)))]; + tensor linear_133_cast_fp16 = linear(bias = encoder_layers_14_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_14_self_attn_linear_out_weight_to_fp16_palettized, x = input_771_cast_fp16)[name = string("linear_133_cast_fp16")]; + tensor input_775_cast_fp16 = add(x = input_765_cast_fp16, y = linear_133_cast_fp16)[name = string("input_775_cast_fp16")]; + tensor x_381_axes_0 = const()[name = string("x_381_axes_0"), val = tensor([-1])]; + tensor encoder_layers_14_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_14_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294569344)))]; + tensor encoder_layers_14_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_14_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294571456)))]; + tensor x_381_cast_fp16 = layer_norm(axes = x_381_axes_0, beta = encoder_layers_14_norm_conv_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_14_norm_conv_weight_to_fp16, x = input_775_cast_fp16)[name = string("x_381_cast_fp16")]; + tensor input_777_perm_0 = const()[name = string("input_777_perm_0"), val = tensor([0, 2, 1])]; + string input_779_pad_type_0 = const()[name = string("input_779_pad_type_0"), val = string("valid")]; + tensor input_779_strides_0 = const()[name = string("input_779_strides_0"), val = tensor([1])]; + tensor input_779_pad_0 = const()[name = string("input_779_pad_0"), val = tensor([0, 0])]; + tensor input_779_dilations_0 = const()[name = string("input_779_dilations_0"), val = tensor([1])]; + int32 input_779_groups_0 = const()[name = string("input_779_groups_0"), val = int32(1)]; + tensor encoder_layers_14_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294573568))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296670784))))[name = string("encoder_layers_14_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_777_cast_fp16 = transpose(perm = input_777_perm_0, x = x_381_cast_fp16)[name = string("transpose_231")]; + tensor input_779_cast_fp16 = conv(dilations = input_779_dilations_0, groups = input_779_groups_0, pad = input_779_pad_0, pad_type = input_779_pad_type_0, strides = input_779_strides_0, weight = encoder_layers_14_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_777_cast_fp16)[name = string("input_779_cast_fp16")]; + int32 x_383_split_num_splits_0 = const()[name = string("x_383_split_num_splits_0"), val = int32(2)]; + int32 x_383_split_axis_0 = const()[name = string("x_383_split_axis_0"), val = int32(1)]; + tensor x_383_split_cast_fp16_0, tensor x_383_split_cast_fp16_1 = split(axis = x_383_split_axis_0, num_splits = x_383_split_num_splits_0, x = input_779_cast_fp16)[name = string("x_383_split_cast_fp16")]; + tensor x_383_split_1_sigmoid_cast_fp16 = sigmoid(x = x_383_split_cast_fp16_1)[name = string("x_383_split_1_sigmoid_cast_fp16")]; + tensor x_383_cast_fp16 = mul(x = x_383_split_cast_fp16_0, y = x_383_split_1_sigmoid_cast_fp16)[name = string("x_383_cast_fp16")]; + tensor input_781_cast_fp16 = select(a = var_45_to_fp16, b = x_383_cast_fp16, cond = var_576)[name = string("input_781_cast_fp16")]; + bool new_x_59_interleave_0 = const()[name = string("new_x_59_interleave_0"), val = bool(false)]; + tensor new_x_59_cast_fp16 = concat(axis = var_60, interleave = new_x_59_interleave_0, values = (cache_59_cast_fp16, input_781_cast_fp16))[name = string("new_x_59_cast_fp16")]; + tensor var_3571_begin_0 = const()[name = string("op_3571_begin_0"), val = tensor([0, 0, 14])]; + tensor var_3571_end_0 = const()[name = string("op_3571_end_0"), val = tensor([1, 1024, 22])]; + tensor var_3571_end_mask_0 = const()[name = string("op_3571_end_mask_0"), val = tensor([true, true, true])]; + tensor var_3571_cast_fp16 = slice_by_index(begin = var_3571_begin_0, end = var_3571_end_0, end_mask = var_3571_end_mask_0, x = new_x_59_cast_fp16)[name = string("op_3571_cast_fp16")]; + string x_385_pad_type_0 = const()[name = string("x_385_pad_type_0"), val = string("valid")]; + int32 x_385_groups_0 = const()[name = string("x_385_groups_0"), val = int32(1024)]; + tensor x_385_strides_0 = const()[name = string("x_385_strides_0"), val = tensor([1])]; + tensor x_385_pad_0 = const()[name = string("x_385_pad_0"), val = tensor([0, 0])]; + tensor x_385_dilations_0 = const()[name = string("x_385_dilations_0"), val = tensor([1])]; + tensor encoder_layers_14_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296674944))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296684224))))[name = string("encoder_layers_14_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_385_cast_fp16 = conv(dilations = x_385_dilations_0, groups = x_385_groups_0, pad = x_385_pad_0, pad_type = x_385_pad_type_0, strides = x_385_strides_0, weight = encoder_layers_14_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_59_cast_fp16)[name = string("x_385_cast_fp16")]; + tensor input_783_perm_0 = const()[name = string("input_783_perm_0"), val = tensor([0, 2, 1])]; + tensor x_387_axes_0 = const()[name = string("x_387_axes_0"), val = tensor([-1])]; + tensor encoder_layers_14_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_14_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296686336)))]; + tensor encoder_layers_14_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_14_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296688448)))]; + tensor input_783_cast_fp16 = transpose(perm = input_783_perm_0, x = x_385_cast_fp16)[name = string("transpose_230")]; + tensor x_387_cast_fp16 = layer_norm(axes = x_387_axes_0, beta = encoder_layers_14_conv_batch_norm_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_14_conv_batch_norm_weight_to_fp16, x = input_783_cast_fp16)[name = string("x_387_cast_fp16")]; + tensor input_785_perm_0 = const()[name = string("input_785_perm_0"), val = tensor([0, 2, 1])]; + tensor input_785_cast_fp16 = transpose(perm = input_785_perm_0, x = x_387_cast_fp16)[name = string("transpose_229")]; + tensor input_787_cast_fp16 = silu(x = input_785_cast_fp16)[name = string("input_787_cast_fp16")]; + string x_389_pad_type_0 = const()[name = string("x_389_pad_type_0"), val = string("valid")]; + tensor x_389_strides_0 = const()[name = string("x_389_strides_0"), val = tensor([1])]; + tensor x_389_pad_0 = const()[name = string("x_389_pad_0"), val = tensor([0, 0])]; + tensor x_389_dilations_0 = const()[name = string("x_389_dilations_0"), val = tensor([1])]; + int32 x_389_groups_0 = const()[name = string("x_389_groups_0"), val = int32(1)]; + tensor encoder_layers_14_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296690560))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297739200))))[name = string("encoder_layers_14_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_389_cast_fp16 = conv(dilations = x_389_dilations_0, groups = x_389_groups_0, pad = x_389_pad_0, pad_type = x_389_pad_type_0, strides = x_389_strides_0, weight = encoder_layers_14_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_787_cast_fp16)[name = string("x_389_cast_fp16")]; + tensor input_789_perm_0 = const()[name = string("input_789_perm_0"), val = tensor([0, 2, 1])]; + tensor input_789_cast_fp16 = transpose(perm = input_789_perm_0, x = x_389_cast_fp16)[name = string("transpose_228")]; + tensor input_791_cast_fp16 = add(x = input_775_cast_fp16, y = input_789_cast_fp16)[name = string("input_791_cast_fp16")]; + tensor input_793_axes_0 = const()[name = string("input_793_axes_0"), val = tensor([-1])]; + tensor encoder_layers_14_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_14_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297741312)))]; + tensor encoder_layers_14_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_14_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297743424)))]; + tensor input_793_cast_fp16 = layer_norm(axes = input_793_axes_0, beta = encoder_layers_14_norm_feed_forward2_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_14_norm_feed_forward2_weight_to_fp16, x = input_791_cast_fp16)[name = string("input_793_cast_fp16")]; + tensor encoder_layers_14_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297745536))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(300891328))))[name = string("encoder_layers_14_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_14_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_14_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(300891520)))]; + tensor linear_134_cast_fp16 = linear(bias = encoder_layers_14_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_14_feed_forward2_linear1_weight_to_fp16_palettized, x = input_793_cast_fp16)[name = string("linear_134_cast_fp16")]; + tensor input_797_cast_fp16 = silu(x = linear_134_cast_fp16)[name = string("input_797_cast_fp16")]; + tensor encoder_layers_14_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(300899776))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(304045568))))[name = string("encoder_layers_14_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_14_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_14_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(304045760)))]; + tensor linear_135_cast_fp16 = linear(bias = encoder_layers_14_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_14_feed_forward2_linear2_weight_to_fp16_palettized, x = input_797_cast_fp16)[name = string("linear_135_cast_fp16")]; + fp16 var_3614_to_fp16 = const()[name = string("op_3614_to_fp16"), val = fp16(0x1p-1)]; + tensor var_3615_cast_fp16 = mul(x = linear_135_cast_fp16, y = var_3614_to_fp16)[name = string("op_3615_cast_fp16")]; + tensor input_803_cast_fp16 = add(x = input_791_cast_fp16, y = var_3615_cast_fp16)[name = string("input_803_cast_fp16")]; + tensor input_805_axes_0 = const()[name = string("input_805_axes_0"), val = tensor([-1])]; + tensor encoder_layers_14_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_14_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(304047872)))]; + tensor encoder_layers_14_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_14_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(304049984)))]; + tensor input_805_cast_fp16 = layer_norm(axes = input_805_axes_0, beta = encoder_layers_14_norm_out_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_14_norm_out_weight_to_fp16, x = input_803_cast_fp16)[name = string("input_805_cast_fp16")]; + tensor cache_61_begin_0 = const()[name = string("cache_61_begin_0"), val = tensor([15, 0, 0, 0])]; + tensor cache_61_end_0 = const()[name = string("cache_61_end_0"), val = tensor([16, 1, 42, 1024])]; + tensor cache_61_end_mask_0 = const()[name = string("cache_61_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_61_squeeze_mask_0 = const()[name = string("cache_61_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_61_cast_fp16 = slice_by_index(begin = cache_61_begin_0, end = cache_61_end_0, end_mask = cache_61_end_mask_0, squeeze_mask = cache_61_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_61_cast_fp16")]; + tensor cache_63_begin_0 = const()[name = string("cache_63_begin_0"), val = tensor([15, 0, 0, 0])]; + tensor cache_63_end_0 = const()[name = string("cache_63_end_0"), val = tensor([16, 1, 1024, 8])]; + tensor cache_63_end_mask_0 = const()[name = string("cache_63_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_63_squeeze_mask_0 = const()[name = string("cache_63_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_63_cast_fp16 = slice_by_index(begin = cache_63_begin_0, end = cache_63_end_0, end_mask = cache_63_end_mask_0, squeeze_mask = cache_63_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_63_cast_fp16")]; + tensor input_807_axes_0 = const()[name = string("input_807_axes_0"), val = tensor([-1])]; + tensor encoder_layers_15_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_15_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(304052096)))]; + tensor encoder_layers_15_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_15_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(304054208)))]; + tensor input_807_cast_fp16 = layer_norm(axes = input_807_axes_0, beta = encoder_layers_15_norm_feed_forward1_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_15_norm_feed_forward1_weight_to_fp16, x = input_805_cast_fp16)[name = string("input_807_cast_fp16")]; + tensor encoder_layers_15_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(304056320))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307202112))))[name = string("encoder_layers_15_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_15_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_15_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307202304)))]; + tensor linear_136_cast_fp16 = linear(bias = encoder_layers_15_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_15_feed_forward1_linear1_weight_to_fp16_palettized, x = input_807_cast_fp16)[name = string("linear_136_cast_fp16")]; + tensor input_811_cast_fp16 = silu(x = linear_136_cast_fp16)[name = string("input_811_cast_fp16")]; + tensor encoder_layers_15_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307210560))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310356352))))[name = string("encoder_layers_15_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_15_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_15_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310356544)))]; + tensor linear_137_cast_fp16 = linear(bias = encoder_layers_15_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_15_feed_forward1_linear2_weight_to_fp16_palettized, x = input_811_cast_fp16)[name = string("linear_137_cast_fp16")]; + fp16 var_3651_to_fp16 = const()[name = string("op_3651_to_fp16"), val = fp16(0x1p-1)]; + tensor var_3652_cast_fp16 = mul(x = linear_137_cast_fp16, y = var_3651_to_fp16)[name = string("op_3652_cast_fp16")]; + tensor input_817_cast_fp16 = add(x = input_805_cast_fp16, y = var_3652_cast_fp16)[name = string("input_817_cast_fp16")]; + tensor key_31_axes_0 = const()[name = string("key_31_axes_0"), val = tensor([-1])]; + tensor encoder_layers_15_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_15_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310358656)))]; + tensor encoder_layers_15_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_15_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310360768)))]; + tensor key_31_cast_fp16 = layer_norm(axes = key_31_axes_0, beta = encoder_layers_15_norm_self_att_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_15_norm_self_att_weight_to_fp16, x = input_817_cast_fp16)[name = string("key_31_cast_fp16")]; + bool input_819_interleave_0 = const()[name = string("input_819_interleave_0"), val = bool(false)]; + tensor input_819_cast_fp16 = concat(axis = var_69, interleave = input_819_interleave_0, values = (cache_61_cast_fp16, key_31_cast_fp16))[name = string("input_819_cast_fp16")]; + tensor var_3674_begin_0 = const()[name = string("op_3674_begin_0"), val = tensor([0, 14, 0])]; + tensor var_3674_end_0 = const()[name = string("op_3674_end_0"), val = tensor([1, 42, 1024])]; + tensor var_3674_end_mask_0 = const()[name = string("op_3674_end_mask_0"), val = tensor([true, true, true])]; + tensor var_3674_cast_fp16 = slice_by_index(begin = var_3674_begin_0, end = var_3674_end_0, end_mask = var_3674_end_mask_0, x = cache_61_cast_fp16)[name = string("op_3674_cast_fp16")]; + bool var_3680_interleave_0 = const()[name = string("op_3680_interleave_0"), val = bool(false)]; + tensor var_3680_cast_fp16 = concat(axis = var_69, interleave = var_3680_interleave_0, values = (var_3674_cast_fp16, key_31_cast_fp16))[name = string("op_3680_cast_fp16")]; + tensor encoder_layers_15_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310362880))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311149376))))[name = string("encoder_layers_15_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_15_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_15_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311149568)))]; + tensor linear_138_cast_fp16 = linear(bias = encoder_layers_15_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_15_self_attn_linear_q_weight_to_fp16_palettized, x = key_31_cast_fp16)[name = string("linear_138_cast_fp16")]; + tensor var_3685 = const()[name = string("op_3685"), val = tensor([1, -1, 8, 128])]; + tensor q_91_cast_fp16 = reshape(shape = var_3685, x = linear_138_cast_fp16)[name = string("q_91_cast_fp16")]; + tensor encoder_layers_15_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311151680))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311938176))))[name = string("encoder_layers_15_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_15_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_15_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311938368)))]; + tensor linear_139_cast_fp16 = linear(bias = encoder_layers_15_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_15_self_attn_linear_k_weight_to_fp16_palettized, x = input_819_cast_fp16)[name = string("linear_139_cast_fp16")]; + tensor var_3690 = const()[name = string("op_3690"), val = tensor([1, -1, 8, 128])]; + tensor k_61_cast_fp16 = reshape(shape = var_3690, x = linear_139_cast_fp16)[name = string("k_61_cast_fp16")]; + tensor encoder_layers_15_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311940480))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(312726976))))[name = string("encoder_layers_15_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_15_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_15_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(312727168)))]; + tensor linear_140_cast_fp16 = linear(bias = encoder_layers_15_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_15_self_attn_linear_v_weight_to_fp16_palettized, x = input_819_cast_fp16)[name = string("linear_140_cast_fp16")]; + tensor var_3695 = const()[name = string("op_3695"), val = tensor([1, -1, 8, 128])]; + tensor v_31_cast_fp16 = reshape(shape = var_3695, x = linear_140_cast_fp16)[name = string("v_31_cast_fp16")]; + tensor value_39_perm_0 = const()[name = string("value_39_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_15_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_15_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(312729280)))]; + tensor var_3708_cast_fp16 = add(x = q_91_cast_fp16, y = encoder_layers_15_self_attn_pos_bias_u_to_fp16)[name = string("op_3708_cast_fp16")]; + tensor encoder_layers_15_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_15_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(312731392)))]; + tensor var_3710_cast_fp16 = add(x = q_91_cast_fp16, y = encoder_layers_15_self_attn_pos_bias_v_to_fp16)[name = string("op_3710_cast_fp16")]; + tensor q_with_bias_v_31_perm_0 = const()[name = string("q_with_bias_v_31_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_397_transpose_x_0 = const()[name = string("x_397_transpose_x_0"), val = bool(false)]; + bool x_397_transpose_y_0 = const()[name = string("x_397_transpose_y_0"), val = bool(false)]; + tensor op_3712_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(312733504))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(312847232))))[name = string("op_3712_to_fp16_quantized")]; + tensor q_with_bias_v_31_cast_fp16 = transpose(perm = q_with_bias_v_31_perm_0, x = var_3710_cast_fp16)[name = string("transpose_227")]; + tensor x_397_cast_fp16 = matmul(transpose_x = x_397_transpose_x_0, transpose_y = x_397_transpose_y_0, x = q_with_bias_v_31_cast_fp16, y = op_3712_to_fp16_quantized)[name = string("x_397_cast_fp16")]; + tensor x_399_pad_0 = const()[name = string("x_399_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_399_mode_0 = const()[name = string("x_399_mode_0"), val = string("constant")]; + fp16 const_274_to_fp16 = const()[name = string("const_274_to_fp16"), val = fp16(0x0p+0)]; + tensor x_399_cast_fp16 = pad(constant_val = const_274_to_fp16, mode = x_399_mode_0, pad = x_399_pad_0, x = x_397_cast_fp16)[name = string("x_399_cast_fp16")]; + tensor var_3720 = const()[name = string("op_3720"), val = tensor([1, 8, -1, 14])]; + tensor x_401_cast_fp16 = reshape(shape = var_3720, x = x_399_cast_fp16)[name = string("x_401_cast_fp16")]; + tensor var_3724_begin_0 = const()[name = string("op_3724_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_3724_end_0 = const()[name = string("op_3724_end_0"), val = tensor([1, 8, 112, 14])]; + tensor var_3724_end_mask_0 = const()[name = string("op_3724_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_3724_cast_fp16 = slice_by_index(begin = var_3724_begin_0, end = var_3724_end_0, end_mask = var_3724_end_mask_0, x = x_401_cast_fp16)[name = string("op_3724_cast_fp16")]; + tensor var_3725 = const()[name = string("op_3725"), val = tensor([1, 8, 14, 111])]; + tensor matrix_bd_61_cast_fp16 = reshape(shape = var_3725, x = var_3724_cast_fp16)[name = string("matrix_bd_61_cast_fp16")]; + bool matrix_ac_31_transpose_x_0 = const()[name = string("matrix_ac_31_transpose_x_0"), val = bool(false)]; + bool matrix_ac_31_transpose_y_0 = const()[name = string("matrix_ac_31_transpose_y_0"), val = bool(false)]; + tensor transpose_126_perm_0 = const()[name = string("transpose_126_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_127_perm_0 = const()[name = string("transpose_127_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_127 = transpose(perm = transpose_127_perm_0, x = k_61_cast_fp16)[name = string("transpose_225")]; + tensor transpose_126 = transpose(perm = transpose_126_perm_0, x = var_3708_cast_fp16)[name = string("transpose_226")]; + tensor matrix_ac_31_cast_fp16 = matmul(transpose_x = matrix_ac_31_transpose_x_0, transpose_y = matrix_ac_31_transpose_y_0, x = transpose_126, y = transpose_127)[name = string("matrix_ac_31_cast_fp16")]; + tensor matrix_bd_63_begin_0 = const()[name = string("matrix_bd_63_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_63_end_0 = const()[name = string("matrix_bd_63_end_0"), val = tensor([1, 8, 14, 56])]; + tensor matrix_bd_63_end_mask_0 = const()[name = string("matrix_bd_63_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_63_cast_fp16 = slice_by_index(begin = matrix_bd_63_begin_0, end = matrix_bd_63_end_0, end_mask = matrix_bd_63_end_mask_0, x = matrix_bd_61_cast_fp16)[name = string("matrix_bd_63_cast_fp16")]; + tensor var_3734_cast_fp16 = add(x = matrix_ac_31_cast_fp16, y = matrix_bd_63_cast_fp16)[name = string("op_3734_cast_fp16")]; + fp16 _inversed_scores_61_y_0_to_fp16 = const()[name = string("_inversed_scores_61_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_61_cast_fp16 = mul(x = var_3734_cast_fp16, y = _inversed_scores_61_y_0_to_fp16)[name = string("_inversed_scores_61_cast_fp16")]; + tensor scores_63_cast_fp16 = select(a = var_46_to_fp16, b = _inversed_scores_61_cast_fp16, cond = mask_11)[name = string("scores_63_cast_fp16")]; + tensor var_3740_cast_fp16 = softmax(axis = var_60, x = scores_63_cast_fp16)[name = string("op_3740_cast_fp16")]; + tensor input_821_cast_fp16 = select(a = var_45_to_fp16, b = var_3740_cast_fp16, cond = mask_11)[name = string("input_821_cast_fp16")]; + bool x_403_transpose_x_0 = const()[name = string("x_403_transpose_x_0"), val = bool(false)]; + bool x_403_transpose_y_0 = const()[name = string("x_403_transpose_y_0"), val = bool(false)]; + tensor value_39_cast_fp16 = transpose(perm = value_39_perm_0, x = v_31_cast_fp16)[name = string("transpose_224")]; + tensor x_403_cast_fp16 = matmul(transpose_x = x_403_transpose_x_0, transpose_y = x_403_transpose_y_0, x = input_821_cast_fp16, y = value_39_cast_fp16)[name = string("x_403_cast_fp16")]; + tensor var_3744_perm_0 = const()[name = string("op_3744_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_3745 = const()[name = string("op_3745"), val = tensor([1, -1, 1024])]; + tensor var_3744_cast_fp16 = transpose(perm = var_3744_perm_0, x = x_403_cast_fp16)[name = string("transpose_223")]; + tensor input_823_cast_fp16 = reshape(shape = var_3745, x = var_3744_cast_fp16)[name = string("input_823_cast_fp16")]; + tensor encoder_layers_15_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(312847552))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(313634048))))[name = string("encoder_layers_15_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_15_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_15_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(313634240)))]; + tensor linear_142_cast_fp16 = linear(bias = encoder_layers_15_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_15_self_attn_linear_out_weight_to_fp16_palettized, x = input_823_cast_fp16)[name = string("linear_142_cast_fp16")]; + tensor input_827_cast_fp16 = add(x = input_817_cast_fp16, y = linear_142_cast_fp16)[name = string("input_827_cast_fp16")]; + tensor x_407_axes_0 = const()[name = string("x_407_axes_0"), val = tensor([-1])]; + tensor encoder_layers_15_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_15_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(313636352)))]; + tensor encoder_layers_15_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_15_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(313638464)))]; + tensor x_407_cast_fp16 = layer_norm(axes = x_407_axes_0, beta = encoder_layers_15_norm_conv_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_15_norm_conv_weight_to_fp16, x = input_827_cast_fp16)[name = string("x_407_cast_fp16")]; + tensor input_829_perm_0 = const()[name = string("input_829_perm_0"), val = tensor([0, 2, 1])]; + string input_831_pad_type_0 = const()[name = string("input_831_pad_type_0"), val = string("valid")]; + tensor input_831_strides_0 = const()[name = string("input_831_strides_0"), val = tensor([1])]; + tensor input_831_pad_0 = const()[name = string("input_831_pad_0"), val = tensor([0, 0])]; + tensor input_831_dilations_0 = const()[name = string("input_831_dilations_0"), val = tensor([1])]; + int32 input_831_groups_0 = const()[name = string("input_831_groups_0"), val = int32(1)]; + tensor encoder_layers_15_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(313640576))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315737792))))[name = string("encoder_layers_15_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_829_cast_fp16 = transpose(perm = input_829_perm_0, x = x_407_cast_fp16)[name = string("transpose_222")]; + tensor input_831_cast_fp16 = conv(dilations = input_831_dilations_0, groups = input_831_groups_0, pad = input_831_pad_0, pad_type = input_831_pad_type_0, strides = input_831_strides_0, weight = encoder_layers_15_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_829_cast_fp16)[name = string("input_831_cast_fp16")]; + int32 x_409_split_num_splits_0 = const()[name = string("x_409_split_num_splits_0"), val = int32(2)]; + int32 x_409_split_axis_0 = const()[name = string("x_409_split_axis_0"), val = int32(1)]; + tensor x_409_split_cast_fp16_0, tensor x_409_split_cast_fp16_1 = split(axis = x_409_split_axis_0, num_splits = x_409_split_num_splits_0, x = input_831_cast_fp16)[name = string("x_409_split_cast_fp16")]; + tensor x_409_split_1_sigmoid_cast_fp16 = sigmoid(x = x_409_split_cast_fp16_1)[name = string("x_409_split_1_sigmoid_cast_fp16")]; + tensor x_409_cast_fp16 = mul(x = x_409_split_cast_fp16_0, y = x_409_split_1_sigmoid_cast_fp16)[name = string("x_409_cast_fp16")]; + tensor input_833_cast_fp16 = select(a = var_45_to_fp16, b = x_409_cast_fp16, cond = var_576)[name = string("input_833_cast_fp16")]; + bool new_x_63_interleave_0 = const()[name = string("new_x_63_interleave_0"), val = bool(false)]; + tensor new_x_63_cast_fp16 = concat(axis = var_60, interleave = new_x_63_interleave_0, values = (cache_63_cast_fp16, input_833_cast_fp16))[name = string("new_x_63_cast_fp16")]; + tensor var_3784_begin_0 = const()[name = string("op_3784_begin_0"), val = tensor([0, 0, 14])]; + tensor var_3784_end_0 = const()[name = string("op_3784_end_0"), val = tensor([1, 1024, 22])]; + tensor var_3784_end_mask_0 = const()[name = string("op_3784_end_mask_0"), val = tensor([true, true, true])]; + tensor var_3784_cast_fp16 = slice_by_index(begin = var_3784_begin_0, end = var_3784_end_0, end_mask = var_3784_end_mask_0, x = new_x_63_cast_fp16)[name = string("op_3784_cast_fp16")]; + string x_411_pad_type_0 = const()[name = string("x_411_pad_type_0"), val = string("valid")]; + int32 x_411_groups_0 = const()[name = string("x_411_groups_0"), val = int32(1024)]; + tensor x_411_strides_0 = const()[name = string("x_411_strides_0"), val = tensor([1])]; + tensor x_411_pad_0 = const()[name = string("x_411_pad_0"), val = tensor([0, 0])]; + tensor x_411_dilations_0 = const()[name = string("x_411_dilations_0"), val = tensor([1])]; + tensor encoder_layers_15_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315741952))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315751232))))[name = string("encoder_layers_15_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_411_cast_fp16 = conv(dilations = x_411_dilations_0, groups = x_411_groups_0, pad = x_411_pad_0, pad_type = x_411_pad_type_0, strides = x_411_strides_0, weight = encoder_layers_15_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_63_cast_fp16)[name = string("x_411_cast_fp16")]; + tensor input_835_perm_0 = const()[name = string("input_835_perm_0"), val = tensor([0, 2, 1])]; + tensor x_413_axes_0 = const()[name = string("x_413_axes_0"), val = tensor([-1])]; + tensor encoder_layers_15_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_15_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315753344)))]; + tensor encoder_layers_15_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_15_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315755456)))]; + tensor input_835_cast_fp16 = transpose(perm = input_835_perm_0, x = x_411_cast_fp16)[name = string("transpose_221")]; + tensor x_413_cast_fp16 = layer_norm(axes = x_413_axes_0, beta = encoder_layers_15_conv_batch_norm_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_15_conv_batch_norm_weight_to_fp16, x = input_835_cast_fp16)[name = string("x_413_cast_fp16")]; + tensor input_837_perm_0 = const()[name = string("input_837_perm_0"), val = tensor([0, 2, 1])]; + tensor input_837_cast_fp16 = transpose(perm = input_837_perm_0, x = x_413_cast_fp16)[name = string("transpose_220")]; + tensor input_839_cast_fp16 = silu(x = input_837_cast_fp16)[name = string("input_839_cast_fp16")]; + string x_415_pad_type_0 = const()[name = string("x_415_pad_type_0"), val = string("valid")]; + tensor x_415_strides_0 = const()[name = string("x_415_strides_0"), val = tensor([1])]; + tensor x_415_pad_0 = const()[name = string("x_415_pad_0"), val = tensor([0, 0])]; + tensor x_415_dilations_0 = const()[name = string("x_415_dilations_0"), val = tensor([1])]; + int32 x_415_groups_0 = const()[name = string("x_415_groups_0"), val = int32(1)]; + tensor encoder_layers_15_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315757568))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316806208))))[name = string("encoder_layers_15_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_415_cast_fp16 = conv(dilations = x_415_dilations_0, groups = x_415_groups_0, pad = x_415_pad_0, pad_type = x_415_pad_type_0, strides = x_415_strides_0, weight = encoder_layers_15_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_839_cast_fp16)[name = string("x_415_cast_fp16")]; + tensor input_841_perm_0 = const()[name = string("input_841_perm_0"), val = tensor([0, 2, 1])]; + tensor input_841_cast_fp16 = transpose(perm = input_841_perm_0, x = x_415_cast_fp16)[name = string("transpose_219")]; + tensor input_843_cast_fp16 = add(x = input_827_cast_fp16, y = input_841_cast_fp16)[name = string("input_843_cast_fp16")]; + tensor input_845_axes_0 = const()[name = string("input_845_axes_0"), val = tensor([-1])]; + tensor encoder_layers_15_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_15_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316808320)))]; + tensor encoder_layers_15_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_15_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316810432)))]; + tensor input_845_cast_fp16 = layer_norm(axes = input_845_axes_0, beta = encoder_layers_15_norm_feed_forward2_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_15_norm_feed_forward2_weight_to_fp16, x = input_843_cast_fp16)[name = string("input_845_cast_fp16")]; + tensor encoder_layers_15_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316812544))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(319958336))))[name = string("encoder_layers_15_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_15_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_15_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(319958528)))]; + tensor linear_143_cast_fp16 = linear(bias = encoder_layers_15_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_15_feed_forward2_linear1_weight_to_fp16_palettized, x = input_845_cast_fp16)[name = string("linear_143_cast_fp16")]; + tensor input_849_cast_fp16 = silu(x = linear_143_cast_fp16)[name = string("input_849_cast_fp16")]; + tensor encoder_layers_15_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(319966784))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(323112576))))[name = string("encoder_layers_15_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_15_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_15_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(323112768)))]; + tensor linear_144_cast_fp16 = linear(bias = encoder_layers_15_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_15_feed_forward2_linear2_weight_to_fp16_palettized, x = input_849_cast_fp16)[name = string("linear_144_cast_fp16")]; + fp16 var_3827_to_fp16 = const()[name = string("op_3827_to_fp16"), val = fp16(0x1p-1)]; + tensor var_3828_cast_fp16 = mul(x = linear_144_cast_fp16, y = var_3827_to_fp16)[name = string("op_3828_cast_fp16")]; + tensor input_855_cast_fp16 = add(x = input_843_cast_fp16, y = var_3828_cast_fp16)[name = string("input_855_cast_fp16")]; + tensor input_857_axes_0 = const()[name = string("input_857_axes_0"), val = tensor([-1])]; + tensor encoder_layers_15_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_15_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(323114880)))]; + tensor encoder_layers_15_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_15_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(323116992)))]; + tensor input_857_cast_fp16 = layer_norm(axes = input_857_axes_0, beta = encoder_layers_15_norm_out_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_15_norm_out_weight_to_fp16, x = input_855_cast_fp16)[name = string("input_857_cast_fp16")]; + tensor cache_65_begin_0 = const()[name = string("cache_65_begin_0"), val = tensor([16, 0, 0, 0])]; + tensor cache_65_end_0 = const()[name = string("cache_65_end_0"), val = tensor([17, 1, 42, 1024])]; + tensor cache_65_end_mask_0 = const()[name = string("cache_65_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_65_squeeze_mask_0 = const()[name = string("cache_65_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_65_cast_fp16 = slice_by_index(begin = cache_65_begin_0, end = cache_65_end_0, end_mask = cache_65_end_mask_0, squeeze_mask = cache_65_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_65_cast_fp16")]; + tensor cache_67_begin_0 = const()[name = string("cache_67_begin_0"), val = tensor([16, 0, 0, 0])]; + tensor cache_67_end_0 = const()[name = string("cache_67_end_0"), val = tensor([17, 1, 1024, 8])]; + tensor cache_67_end_mask_0 = const()[name = string("cache_67_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_67_squeeze_mask_0 = const()[name = string("cache_67_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_67_cast_fp16 = slice_by_index(begin = cache_67_begin_0, end = cache_67_end_0, end_mask = cache_67_end_mask_0, squeeze_mask = cache_67_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_67_cast_fp16")]; + tensor input_859_axes_0 = const()[name = string("input_859_axes_0"), val = tensor([-1])]; + tensor encoder_layers_16_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_16_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(323119104)))]; + tensor encoder_layers_16_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_16_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(323121216)))]; + tensor input_859_cast_fp16 = layer_norm(axes = input_859_axes_0, beta = encoder_layers_16_norm_feed_forward1_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_16_norm_feed_forward1_weight_to_fp16, x = input_857_cast_fp16)[name = string("input_859_cast_fp16")]; + tensor encoder_layers_16_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(323123328))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(326269120))))[name = string("encoder_layers_16_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_16_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_16_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(326269312)))]; + tensor linear_145_cast_fp16 = linear(bias = encoder_layers_16_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_16_feed_forward1_linear1_weight_to_fp16_palettized, x = input_859_cast_fp16)[name = string("linear_145_cast_fp16")]; + tensor input_863_cast_fp16 = silu(x = linear_145_cast_fp16)[name = string("input_863_cast_fp16")]; + tensor encoder_layers_16_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(326277568))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(329423360))))[name = string("encoder_layers_16_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_16_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_16_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(329423552)))]; + tensor linear_146_cast_fp16 = linear(bias = encoder_layers_16_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_16_feed_forward1_linear2_weight_to_fp16_palettized, x = input_863_cast_fp16)[name = string("linear_146_cast_fp16")]; + fp16 var_3864_to_fp16 = const()[name = string("op_3864_to_fp16"), val = fp16(0x1p-1)]; + tensor var_3865_cast_fp16 = mul(x = linear_146_cast_fp16, y = var_3864_to_fp16)[name = string("op_3865_cast_fp16")]; + tensor input_869_cast_fp16 = add(x = input_857_cast_fp16, y = var_3865_cast_fp16)[name = string("input_869_cast_fp16")]; + tensor key_33_axes_0 = const()[name = string("key_33_axes_0"), val = tensor([-1])]; + tensor encoder_layers_16_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_16_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(329425664)))]; + tensor encoder_layers_16_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_16_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(329427776)))]; + tensor key_33_cast_fp16 = layer_norm(axes = key_33_axes_0, beta = encoder_layers_16_norm_self_att_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_16_norm_self_att_weight_to_fp16, x = input_869_cast_fp16)[name = string("key_33_cast_fp16")]; + bool input_871_interleave_0 = const()[name = string("input_871_interleave_0"), val = bool(false)]; + tensor input_871_cast_fp16 = concat(axis = var_69, interleave = input_871_interleave_0, values = (cache_65_cast_fp16, key_33_cast_fp16))[name = string("input_871_cast_fp16")]; + tensor var_3887_begin_0 = const()[name = string("op_3887_begin_0"), val = tensor([0, 14, 0])]; + tensor var_3887_end_0 = const()[name = string("op_3887_end_0"), val = tensor([1, 42, 1024])]; + tensor var_3887_end_mask_0 = const()[name = string("op_3887_end_mask_0"), val = tensor([true, true, true])]; + tensor var_3887_cast_fp16 = slice_by_index(begin = var_3887_begin_0, end = var_3887_end_0, end_mask = var_3887_end_mask_0, x = cache_65_cast_fp16)[name = string("op_3887_cast_fp16")]; + bool var_3893_interleave_0 = const()[name = string("op_3893_interleave_0"), val = bool(false)]; + tensor var_3893_cast_fp16 = concat(axis = var_69, interleave = var_3893_interleave_0, values = (var_3887_cast_fp16, key_33_cast_fp16))[name = string("op_3893_cast_fp16")]; + tensor encoder_layers_16_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(329429888))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(330216384))))[name = string("encoder_layers_16_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_16_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_16_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(330216576)))]; + tensor linear_147_cast_fp16 = linear(bias = encoder_layers_16_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_16_self_attn_linear_q_weight_to_fp16_palettized, x = key_33_cast_fp16)[name = string("linear_147_cast_fp16")]; + tensor var_3898 = const()[name = string("op_3898"), val = tensor([1, -1, 8, 128])]; + tensor q_97_cast_fp16 = reshape(shape = var_3898, x = linear_147_cast_fp16)[name = string("q_97_cast_fp16")]; + tensor encoder_layers_16_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(330218688))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331005184))))[name = string("encoder_layers_16_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_16_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_16_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331005376)))]; + tensor linear_148_cast_fp16 = linear(bias = encoder_layers_16_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_16_self_attn_linear_k_weight_to_fp16_palettized, x = input_871_cast_fp16)[name = string("linear_148_cast_fp16")]; + tensor var_3903 = const()[name = string("op_3903"), val = tensor([1, -1, 8, 128])]; + tensor k_65_cast_fp16 = reshape(shape = var_3903, x = linear_148_cast_fp16)[name = string("k_65_cast_fp16")]; + tensor encoder_layers_16_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331007488))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331793984))))[name = string("encoder_layers_16_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_16_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_16_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331794176)))]; + tensor linear_149_cast_fp16 = linear(bias = encoder_layers_16_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_16_self_attn_linear_v_weight_to_fp16_palettized, x = input_871_cast_fp16)[name = string("linear_149_cast_fp16")]; + tensor var_3908 = const()[name = string("op_3908"), val = tensor([1, -1, 8, 128])]; + tensor v_33_cast_fp16 = reshape(shape = var_3908, x = linear_149_cast_fp16)[name = string("v_33_cast_fp16")]; + tensor value_41_perm_0 = const()[name = string("value_41_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_16_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_16_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331796288)))]; + tensor var_3921_cast_fp16 = add(x = q_97_cast_fp16, y = encoder_layers_16_self_attn_pos_bias_u_to_fp16)[name = string("op_3921_cast_fp16")]; + tensor encoder_layers_16_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_16_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331798400)))]; + tensor var_3923_cast_fp16 = add(x = q_97_cast_fp16, y = encoder_layers_16_self_attn_pos_bias_v_to_fp16)[name = string("op_3923_cast_fp16")]; + tensor q_with_bias_v_33_perm_0 = const()[name = string("q_with_bias_v_33_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_423_transpose_x_0 = const()[name = string("x_423_transpose_x_0"), val = bool(false)]; + bool x_423_transpose_y_0 = const()[name = string("x_423_transpose_y_0"), val = bool(false)]; + tensor op_3925_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331800512))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331914240))))[name = string("op_3925_to_fp16_quantized")]; + tensor q_with_bias_v_33_cast_fp16 = transpose(perm = q_with_bias_v_33_perm_0, x = var_3923_cast_fp16)[name = string("transpose_218")]; + tensor x_423_cast_fp16 = matmul(transpose_x = x_423_transpose_x_0, transpose_y = x_423_transpose_y_0, x = q_with_bias_v_33_cast_fp16, y = op_3925_to_fp16_quantized)[name = string("x_423_cast_fp16")]; + tensor x_425_pad_0 = const()[name = string("x_425_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_425_mode_0 = const()[name = string("x_425_mode_0"), val = string("constant")]; + fp16 const_287_to_fp16 = const()[name = string("const_287_to_fp16"), val = fp16(0x0p+0)]; + tensor x_425_cast_fp16 = pad(constant_val = const_287_to_fp16, mode = x_425_mode_0, pad = x_425_pad_0, x = x_423_cast_fp16)[name = string("x_425_cast_fp16")]; + tensor var_3933 = const()[name = string("op_3933"), val = tensor([1, 8, -1, 14])]; + tensor x_427_cast_fp16 = reshape(shape = var_3933, x = x_425_cast_fp16)[name = string("x_427_cast_fp16")]; + tensor var_3937_begin_0 = const()[name = string("op_3937_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_3937_end_0 = const()[name = string("op_3937_end_0"), val = tensor([1, 8, 112, 14])]; + tensor var_3937_end_mask_0 = const()[name = string("op_3937_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_3937_cast_fp16 = slice_by_index(begin = var_3937_begin_0, end = var_3937_end_0, end_mask = var_3937_end_mask_0, x = x_427_cast_fp16)[name = string("op_3937_cast_fp16")]; + tensor var_3938 = const()[name = string("op_3938"), val = tensor([1, 8, 14, 111])]; + tensor matrix_bd_65_cast_fp16 = reshape(shape = var_3938, x = var_3937_cast_fp16)[name = string("matrix_bd_65_cast_fp16")]; + bool matrix_ac_33_transpose_x_0 = const()[name = string("matrix_ac_33_transpose_x_0"), val = bool(false)]; + bool matrix_ac_33_transpose_y_0 = const()[name = string("matrix_ac_33_transpose_y_0"), val = bool(false)]; + tensor transpose_128_perm_0 = const()[name = string("transpose_128_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_129_perm_0 = const()[name = string("transpose_129_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_129 = transpose(perm = transpose_129_perm_0, x = k_65_cast_fp16)[name = string("transpose_216")]; + tensor transpose_128 = transpose(perm = transpose_128_perm_0, x = var_3921_cast_fp16)[name = string("transpose_217")]; + tensor matrix_ac_33_cast_fp16 = matmul(transpose_x = matrix_ac_33_transpose_x_0, transpose_y = matrix_ac_33_transpose_y_0, x = transpose_128, y = transpose_129)[name = string("matrix_ac_33_cast_fp16")]; + tensor matrix_bd_67_begin_0 = const()[name = string("matrix_bd_67_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_67_end_0 = const()[name = string("matrix_bd_67_end_0"), val = tensor([1, 8, 14, 56])]; + tensor matrix_bd_67_end_mask_0 = const()[name = string("matrix_bd_67_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_67_cast_fp16 = slice_by_index(begin = matrix_bd_67_begin_0, end = matrix_bd_67_end_0, end_mask = matrix_bd_67_end_mask_0, x = matrix_bd_65_cast_fp16)[name = string("matrix_bd_67_cast_fp16")]; + tensor var_3947_cast_fp16 = add(x = matrix_ac_33_cast_fp16, y = matrix_bd_67_cast_fp16)[name = string("op_3947_cast_fp16")]; + fp16 _inversed_scores_65_y_0_to_fp16 = const()[name = string("_inversed_scores_65_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_65_cast_fp16 = mul(x = var_3947_cast_fp16, y = _inversed_scores_65_y_0_to_fp16)[name = string("_inversed_scores_65_cast_fp16")]; + tensor scores_67_cast_fp16 = select(a = var_46_to_fp16, b = _inversed_scores_65_cast_fp16, cond = mask_11)[name = string("scores_67_cast_fp16")]; + tensor var_3953_cast_fp16 = softmax(axis = var_60, x = scores_67_cast_fp16)[name = string("op_3953_cast_fp16")]; + tensor input_873_cast_fp16 = select(a = var_45_to_fp16, b = var_3953_cast_fp16, cond = mask_11)[name = string("input_873_cast_fp16")]; + bool x_429_transpose_x_0 = const()[name = string("x_429_transpose_x_0"), val = bool(false)]; + bool x_429_transpose_y_0 = const()[name = string("x_429_transpose_y_0"), val = bool(false)]; + tensor value_41_cast_fp16 = transpose(perm = value_41_perm_0, x = v_33_cast_fp16)[name = string("transpose_215")]; + tensor x_429_cast_fp16 = matmul(transpose_x = x_429_transpose_x_0, transpose_y = x_429_transpose_y_0, x = input_873_cast_fp16, y = value_41_cast_fp16)[name = string("x_429_cast_fp16")]; + tensor var_3957_perm_0 = const()[name = string("op_3957_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_3958 = const()[name = string("op_3958"), val = tensor([1, -1, 1024])]; + tensor var_3957_cast_fp16 = transpose(perm = var_3957_perm_0, x = x_429_cast_fp16)[name = string("transpose_214")]; + tensor input_875_cast_fp16 = reshape(shape = var_3958, x = var_3957_cast_fp16)[name = string("input_875_cast_fp16")]; + tensor encoder_layers_16_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331914560))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332701056))))[name = string("encoder_layers_16_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_16_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_16_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332701248)))]; + tensor linear_151_cast_fp16 = linear(bias = encoder_layers_16_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_16_self_attn_linear_out_weight_to_fp16_palettized, x = input_875_cast_fp16)[name = string("linear_151_cast_fp16")]; + tensor input_879_cast_fp16 = add(x = input_869_cast_fp16, y = linear_151_cast_fp16)[name = string("input_879_cast_fp16")]; + tensor x_433_axes_0 = const()[name = string("x_433_axes_0"), val = tensor([-1])]; + tensor encoder_layers_16_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_16_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332703360)))]; + tensor encoder_layers_16_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_16_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332705472)))]; + tensor x_433_cast_fp16 = layer_norm(axes = x_433_axes_0, beta = encoder_layers_16_norm_conv_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_16_norm_conv_weight_to_fp16, x = input_879_cast_fp16)[name = string("x_433_cast_fp16")]; + tensor input_881_perm_0 = const()[name = string("input_881_perm_0"), val = tensor([0, 2, 1])]; + string input_883_pad_type_0 = const()[name = string("input_883_pad_type_0"), val = string("valid")]; + tensor input_883_strides_0 = const()[name = string("input_883_strides_0"), val = tensor([1])]; + tensor input_883_pad_0 = const()[name = string("input_883_pad_0"), val = tensor([0, 0])]; + tensor input_883_dilations_0 = const()[name = string("input_883_dilations_0"), val = tensor([1])]; + int32 input_883_groups_0 = const()[name = string("input_883_groups_0"), val = int32(1)]; + tensor encoder_layers_16_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332707584))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(334804800))))[name = string("encoder_layers_16_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_881_cast_fp16 = transpose(perm = input_881_perm_0, x = x_433_cast_fp16)[name = string("transpose_213")]; + tensor input_883_cast_fp16 = conv(dilations = input_883_dilations_0, groups = input_883_groups_0, pad = input_883_pad_0, pad_type = input_883_pad_type_0, strides = input_883_strides_0, weight = encoder_layers_16_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_881_cast_fp16)[name = string("input_883_cast_fp16")]; + int32 x_435_split_num_splits_0 = const()[name = string("x_435_split_num_splits_0"), val = int32(2)]; + int32 x_435_split_axis_0 = const()[name = string("x_435_split_axis_0"), val = int32(1)]; + tensor x_435_split_cast_fp16_0, tensor x_435_split_cast_fp16_1 = split(axis = x_435_split_axis_0, num_splits = x_435_split_num_splits_0, x = input_883_cast_fp16)[name = string("x_435_split_cast_fp16")]; + tensor x_435_split_1_sigmoid_cast_fp16 = sigmoid(x = x_435_split_cast_fp16_1)[name = string("x_435_split_1_sigmoid_cast_fp16")]; + tensor x_435_cast_fp16 = mul(x = x_435_split_cast_fp16_0, y = x_435_split_1_sigmoid_cast_fp16)[name = string("x_435_cast_fp16")]; + tensor input_885_cast_fp16 = select(a = var_45_to_fp16, b = x_435_cast_fp16, cond = var_576)[name = string("input_885_cast_fp16")]; + bool new_x_67_interleave_0 = const()[name = string("new_x_67_interleave_0"), val = bool(false)]; + tensor new_x_67_cast_fp16 = concat(axis = var_60, interleave = new_x_67_interleave_0, values = (cache_67_cast_fp16, input_885_cast_fp16))[name = string("new_x_67_cast_fp16")]; + tensor var_3997_begin_0 = const()[name = string("op_3997_begin_0"), val = tensor([0, 0, 14])]; + tensor var_3997_end_0 = const()[name = string("op_3997_end_0"), val = tensor([1, 1024, 22])]; + tensor var_3997_end_mask_0 = const()[name = string("op_3997_end_mask_0"), val = tensor([true, true, true])]; + tensor var_3997_cast_fp16 = slice_by_index(begin = var_3997_begin_0, end = var_3997_end_0, end_mask = var_3997_end_mask_0, x = new_x_67_cast_fp16)[name = string("op_3997_cast_fp16")]; + string x_437_pad_type_0 = const()[name = string("x_437_pad_type_0"), val = string("valid")]; + int32 x_437_groups_0 = const()[name = string("x_437_groups_0"), val = int32(1024)]; + tensor x_437_strides_0 = const()[name = string("x_437_strides_0"), val = tensor([1])]; + tensor x_437_pad_0 = const()[name = string("x_437_pad_0"), val = tensor([0, 0])]; + tensor x_437_dilations_0 = const()[name = string("x_437_dilations_0"), val = tensor([1])]; + tensor encoder_layers_16_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(334808960))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(334818240))))[name = string("encoder_layers_16_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_437_cast_fp16 = conv(dilations = x_437_dilations_0, groups = x_437_groups_0, pad = x_437_pad_0, pad_type = x_437_pad_type_0, strides = x_437_strides_0, weight = encoder_layers_16_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_67_cast_fp16)[name = string("x_437_cast_fp16")]; + tensor input_887_perm_0 = const()[name = string("input_887_perm_0"), val = tensor([0, 2, 1])]; + tensor x_439_axes_0 = const()[name = string("x_439_axes_0"), val = tensor([-1])]; + tensor encoder_layers_16_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_16_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(334820352)))]; + tensor encoder_layers_16_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_16_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(334822464)))]; + tensor input_887_cast_fp16 = transpose(perm = input_887_perm_0, x = x_437_cast_fp16)[name = string("transpose_212")]; + tensor x_439_cast_fp16 = layer_norm(axes = x_439_axes_0, beta = encoder_layers_16_conv_batch_norm_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_16_conv_batch_norm_weight_to_fp16, x = input_887_cast_fp16)[name = string("x_439_cast_fp16")]; + tensor input_889_perm_0 = const()[name = string("input_889_perm_0"), val = tensor([0, 2, 1])]; + tensor input_889_cast_fp16 = transpose(perm = input_889_perm_0, x = x_439_cast_fp16)[name = string("transpose_211")]; + tensor input_891_cast_fp16 = silu(x = input_889_cast_fp16)[name = string("input_891_cast_fp16")]; + string x_441_pad_type_0 = const()[name = string("x_441_pad_type_0"), val = string("valid")]; + tensor x_441_strides_0 = const()[name = string("x_441_strides_0"), val = tensor([1])]; + tensor x_441_pad_0 = const()[name = string("x_441_pad_0"), val = tensor([0, 0])]; + tensor x_441_dilations_0 = const()[name = string("x_441_dilations_0"), val = tensor([1])]; + int32 x_441_groups_0 = const()[name = string("x_441_groups_0"), val = int32(1)]; + tensor encoder_layers_16_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(334824576))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(335873216))))[name = string("encoder_layers_16_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_441_cast_fp16 = conv(dilations = x_441_dilations_0, groups = x_441_groups_0, pad = x_441_pad_0, pad_type = x_441_pad_type_0, strides = x_441_strides_0, weight = encoder_layers_16_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_891_cast_fp16)[name = string("x_441_cast_fp16")]; + tensor input_893_perm_0 = const()[name = string("input_893_perm_0"), val = tensor([0, 2, 1])]; + tensor input_893_cast_fp16 = transpose(perm = input_893_perm_0, x = x_441_cast_fp16)[name = string("transpose_210")]; + tensor input_895_cast_fp16 = add(x = input_879_cast_fp16, y = input_893_cast_fp16)[name = string("input_895_cast_fp16")]; + tensor input_897_axes_0 = const()[name = string("input_897_axes_0"), val = tensor([-1])]; + tensor encoder_layers_16_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_16_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(335875328)))]; + tensor encoder_layers_16_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_16_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(335877440)))]; + tensor input_897_cast_fp16 = layer_norm(axes = input_897_axes_0, beta = encoder_layers_16_norm_feed_forward2_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_16_norm_feed_forward2_weight_to_fp16, x = input_895_cast_fp16)[name = string("input_897_cast_fp16")]; + tensor encoder_layers_16_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(335879552))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(339025344))))[name = string("encoder_layers_16_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_16_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_16_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(339025536)))]; + tensor linear_152_cast_fp16 = linear(bias = encoder_layers_16_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_16_feed_forward2_linear1_weight_to_fp16_palettized, x = input_897_cast_fp16)[name = string("linear_152_cast_fp16")]; + tensor input_901_cast_fp16 = silu(x = linear_152_cast_fp16)[name = string("input_901_cast_fp16")]; + tensor encoder_layers_16_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(339033792))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(342179584))))[name = string("encoder_layers_16_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_16_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_16_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(342179776)))]; + tensor linear_153_cast_fp16 = linear(bias = encoder_layers_16_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_16_feed_forward2_linear2_weight_to_fp16_palettized, x = input_901_cast_fp16)[name = string("linear_153_cast_fp16")]; + fp16 var_4040_to_fp16 = const()[name = string("op_4040_to_fp16"), val = fp16(0x1p-1)]; + tensor var_4041_cast_fp16 = mul(x = linear_153_cast_fp16, y = var_4040_to_fp16)[name = string("op_4041_cast_fp16")]; + tensor input_907_cast_fp16 = add(x = input_895_cast_fp16, y = var_4041_cast_fp16)[name = string("input_907_cast_fp16")]; + tensor input_909_axes_0 = const()[name = string("input_909_axes_0"), val = tensor([-1])]; + tensor encoder_layers_16_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_16_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(342181888)))]; + tensor encoder_layers_16_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_16_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(342184000)))]; + tensor input_909_cast_fp16 = layer_norm(axes = input_909_axes_0, beta = encoder_layers_16_norm_out_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_16_norm_out_weight_to_fp16, x = input_907_cast_fp16)[name = string("input_909_cast_fp16")]; + tensor cache_69_begin_0 = const()[name = string("cache_69_begin_0"), val = tensor([17, 0, 0, 0])]; + tensor cache_69_end_0 = const()[name = string("cache_69_end_0"), val = tensor([18, 1, 42, 1024])]; + tensor cache_69_end_mask_0 = const()[name = string("cache_69_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_69_squeeze_mask_0 = const()[name = string("cache_69_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_69_cast_fp16 = slice_by_index(begin = cache_69_begin_0, end = cache_69_end_0, end_mask = cache_69_end_mask_0, squeeze_mask = cache_69_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_69_cast_fp16")]; + tensor cache_71_begin_0 = const()[name = string("cache_71_begin_0"), val = tensor([17, 0, 0, 0])]; + tensor cache_71_end_0 = const()[name = string("cache_71_end_0"), val = tensor([18, 1, 1024, 8])]; + tensor cache_71_end_mask_0 = const()[name = string("cache_71_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_71_squeeze_mask_0 = const()[name = string("cache_71_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_71_cast_fp16 = slice_by_index(begin = cache_71_begin_0, end = cache_71_end_0, end_mask = cache_71_end_mask_0, squeeze_mask = cache_71_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_71_cast_fp16")]; + tensor input_911_axes_0 = const()[name = string("input_911_axes_0"), val = tensor([-1])]; + tensor encoder_layers_17_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_17_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(342186112)))]; + tensor encoder_layers_17_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_17_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(342188224)))]; + tensor input_911_cast_fp16 = layer_norm(axes = input_911_axes_0, beta = encoder_layers_17_norm_feed_forward1_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_17_norm_feed_forward1_weight_to_fp16, x = input_909_cast_fp16)[name = string("input_911_cast_fp16")]; + tensor encoder_layers_17_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(342190336))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(345336128))))[name = string("encoder_layers_17_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_17_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_17_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(345336320)))]; + tensor linear_154_cast_fp16 = linear(bias = encoder_layers_17_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_17_feed_forward1_linear1_weight_to_fp16_palettized, x = input_911_cast_fp16)[name = string("linear_154_cast_fp16")]; + tensor input_915_cast_fp16 = silu(x = linear_154_cast_fp16)[name = string("input_915_cast_fp16")]; + tensor encoder_layers_17_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(345344576))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(348490368))))[name = string("encoder_layers_17_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_17_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_17_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(348490560)))]; + tensor linear_155_cast_fp16 = linear(bias = encoder_layers_17_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_17_feed_forward1_linear2_weight_to_fp16_palettized, x = input_915_cast_fp16)[name = string("linear_155_cast_fp16")]; + fp16 var_4077_to_fp16 = const()[name = string("op_4077_to_fp16"), val = fp16(0x1p-1)]; + tensor var_4078_cast_fp16 = mul(x = linear_155_cast_fp16, y = var_4077_to_fp16)[name = string("op_4078_cast_fp16")]; + tensor input_921_cast_fp16 = add(x = input_909_cast_fp16, y = var_4078_cast_fp16)[name = string("input_921_cast_fp16")]; + tensor key_35_axes_0 = const()[name = string("key_35_axes_0"), val = tensor([-1])]; + tensor encoder_layers_17_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_17_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(348492672)))]; + tensor encoder_layers_17_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_17_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(348494784)))]; + tensor key_35_cast_fp16 = layer_norm(axes = key_35_axes_0, beta = encoder_layers_17_norm_self_att_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_17_norm_self_att_weight_to_fp16, x = input_921_cast_fp16)[name = string("key_35_cast_fp16")]; + bool input_923_interleave_0 = const()[name = string("input_923_interleave_0"), val = bool(false)]; + tensor input_923_cast_fp16 = concat(axis = var_69, interleave = input_923_interleave_0, values = (cache_69_cast_fp16, key_35_cast_fp16))[name = string("input_923_cast_fp16")]; + tensor var_4100_begin_0 = const()[name = string("op_4100_begin_0"), val = tensor([0, 14, 0])]; + tensor var_4100_end_0 = const()[name = string("op_4100_end_0"), val = tensor([1, 42, 1024])]; + tensor var_4100_end_mask_0 = const()[name = string("op_4100_end_mask_0"), val = tensor([true, true, true])]; + tensor var_4100_cast_fp16 = slice_by_index(begin = var_4100_begin_0, end = var_4100_end_0, end_mask = var_4100_end_mask_0, x = cache_69_cast_fp16)[name = string("op_4100_cast_fp16")]; + bool var_4106_interleave_0 = const()[name = string("op_4106_interleave_0"), val = bool(false)]; + tensor var_4106_cast_fp16 = concat(axis = var_69, interleave = var_4106_interleave_0, values = (var_4100_cast_fp16, key_35_cast_fp16))[name = string("op_4106_cast_fp16")]; + tensor encoder_layers_17_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(348496896))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(349283392))))[name = string("encoder_layers_17_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_17_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_17_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(349283584)))]; + tensor linear_156_cast_fp16 = linear(bias = encoder_layers_17_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_17_self_attn_linear_q_weight_to_fp16_palettized, x = key_35_cast_fp16)[name = string("linear_156_cast_fp16")]; + tensor var_4111 = const()[name = string("op_4111"), val = tensor([1, -1, 8, 128])]; + tensor q_103_cast_fp16 = reshape(shape = var_4111, x = linear_156_cast_fp16)[name = string("q_103_cast_fp16")]; + tensor encoder_layers_17_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(349285696))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350072192))))[name = string("encoder_layers_17_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_17_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_17_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350072384)))]; + tensor linear_157_cast_fp16 = linear(bias = encoder_layers_17_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_17_self_attn_linear_k_weight_to_fp16_palettized, x = input_923_cast_fp16)[name = string("linear_157_cast_fp16")]; + tensor var_4116 = const()[name = string("op_4116"), val = tensor([1, -1, 8, 128])]; + tensor k_69_cast_fp16 = reshape(shape = var_4116, x = linear_157_cast_fp16)[name = string("k_69_cast_fp16")]; + tensor encoder_layers_17_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350074496))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350860992))))[name = string("encoder_layers_17_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_17_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_17_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350861184)))]; + tensor linear_158_cast_fp16 = linear(bias = encoder_layers_17_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_17_self_attn_linear_v_weight_to_fp16_palettized, x = input_923_cast_fp16)[name = string("linear_158_cast_fp16")]; + tensor var_4121 = const()[name = string("op_4121"), val = tensor([1, -1, 8, 128])]; + tensor v_35_cast_fp16 = reshape(shape = var_4121, x = linear_158_cast_fp16)[name = string("v_35_cast_fp16")]; + tensor value_43_perm_0 = const()[name = string("value_43_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_17_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_17_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350863296)))]; + tensor var_4134_cast_fp16 = add(x = q_103_cast_fp16, y = encoder_layers_17_self_attn_pos_bias_u_to_fp16)[name = string("op_4134_cast_fp16")]; + tensor encoder_layers_17_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_17_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350865408)))]; + tensor var_4136_cast_fp16 = add(x = q_103_cast_fp16, y = encoder_layers_17_self_attn_pos_bias_v_to_fp16)[name = string("op_4136_cast_fp16")]; + tensor q_with_bias_v_35_perm_0 = const()[name = string("q_with_bias_v_35_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_449_transpose_x_0 = const()[name = string("x_449_transpose_x_0"), val = bool(false)]; + bool x_449_transpose_y_0 = const()[name = string("x_449_transpose_y_0"), val = bool(false)]; + tensor op_4138_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350867520))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350981248))))[name = string("op_4138_to_fp16_quantized")]; + tensor q_with_bias_v_35_cast_fp16 = transpose(perm = q_with_bias_v_35_perm_0, x = var_4136_cast_fp16)[name = string("transpose_209")]; + tensor x_449_cast_fp16 = matmul(transpose_x = x_449_transpose_x_0, transpose_y = x_449_transpose_y_0, x = q_with_bias_v_35_cast_fp16, y = op_4138_to_fp16_quantized)[name = string("x_449_cast_fp16")]; + tensor x_451_pad_0 = const()[name = string("x_451_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_451_mode_0 = const()[name = string("x_451_mode_0"), val = string("constant")]; + fp16 const_300_to_fp16 = const()[name = string("const_300_to_fp16"), val = fp16(0x0p+0)]; + tensor x_451_cast_fp16 = pad(constant_val = const_300_to_fp16, mode = x_451_mode_0, pad = x_451_pad_0, x = x_449_cast_fp16)[name = string("x_451_cast_fp16")]; + tensor var_4146 = const()[name = string("op_4146"), val = tensor([1, 8, -1, 14])]; + tensor x_453_cast_fp16 = reshape(shape = var_4146, x = x_451_cast_fp16)[name = string("x_453_cast_fp16")]; + tensor var_4150_begin_0 = const()[name = string("op_4150_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_4150_end_0 = const()[name = string("op_4150_end_0"), val = tensor([1, 8, 112, 14])]; + tensor var_4150_end_mask_0 = const()[name = string("op_4150_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_4150_cast_fp16 = slice_by_index(begin = var_4150_begin_0, end = var_4150_end_0, end_mask = var_4150_end_mask_0, x = x_453_cast_fp16)[name = string("op_4150_cast_fp16")]; + tensor var_4151 = const()[name = string("op_4151"), val = tensor([1, 8, 14, 111])]; + tensor matrix_bd_69_cast_fp16 = reshape(shape = var_4151, x = var_4150_cast_fp16)[name = string("matrix_bd_69_cast_fp16")]; + bool matrix_ac_35_transpose_x_0 = const()[name = string("matrix_ac_35_transpose_x_0"), val = bool(false)]; + bool matrix_ac_35_transpose_y_0 = const()[name = string("matrix_ac_35_transpose_y_0"), val = bool(false)]; + tensor transpose_130_perm_0 = const()[name = string("transpose_130_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_131_perm_0 = const()[name = string("transpose_131_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_131 = transpose(perm = transpose_131_perm_0, x = k_69_cast_fp16)[name = string("transpose_207")]; + tensor transpose_130 = transpose(perm = transpose_130_perm_0, x = var_4134_cast_fp16)[name = string("transpose_208")]; + tensor matrix_ac_35_cast_fp16 = matmul(transpose_x = matrix_ac_35_transpose_x_0, transpose_y = matrix_ac_35_transpose_y_0, x = transpose_130, y = transpose_131)[name = string("matrix_ac_35_cast_fp16")]; + tensor matrix_bd_71_begin_0 = const()[name = string("matrix_bd_71_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_71_end_0 = const()[name = string("matrix_bd_71_end_0"), val = tensor([1, 8, 14, 56])]; + tensor matrix_bd_71_end_mask_0 = const()[name = string("matrix_bd_71_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_71_cast_fp16 = slice_by_index(begin = matrix_bd_71_begin_0, end = matrix_bd_71_end_0, end_mask = matrix_bd_71_end_mask_0, x = matrix_bd_69_cast_fp16)[name = string("matrix_bd_71_cast_fp16")]; + tensor var_4160_cast_fp16 = add(x = matrix_ac_35_cast_fp16, y = matrix_bd_71_cast_fp16)[name = string("op_4160_cast_fp16")]; + fp16 _inversed_scores_69_y_0_to_fp16 = const()[name = string("_inversed_scores_69_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_69_cast_fp16 = mul(x = var_4160_cast_fp16, y = _inversed_scores_69_y_0_to_fp16)[name = string("_inversed_scores_69_cast_fp16")]; + tensor scores_71_cast_fp16 = select(a = var_46_to_fp16, b = _inversed_scores_69_cast_fp16, cond = mask_11)[name = string("scores_71_cast_fp16")]; + tensor var_4166_cast_fp16 = softmax(axis = var_60, x = scores_71_cast_fp16)[name = string("op_4166_cast_fp16")]; + tensor input_925_cast_fp16 = select(a = var_45_to_fp16, b = var_4166_cast_fp16, cond = mask_11)[name = string("input_925_cast_fp16")]; + bool x_455_transpose_x_0 = const()[name = string("x_455_transpose_x_0"), val = bool(false)]; + bool x_455_transpose_y_0 = const()[name = string("x_455_transpose_y_0"), val = bool(false)]; + tensor value_43_cast_fp16 = transpose(perm = value_43_perm_0, x = v_35_cast_fp16)[name = string("transpose_206")]; + tensor x_455_cast_fp16 = matmul(transpose_x = x_455_transpose_x_0, transpose_y = x_455_transpose_y_0, x = input_925_cast_fp16, y = value_43_cast_fp16)[name = string("x_455_cast_fp16")]; + tensor var_4170_perm_0 = const()[name = string("op_4170_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_4171 = const()[name = string("op_4171"), val = tensor([1, -1, 1024])]; + tensor var_4170_cast_fp16 = transpose(perm = var_4170_perm_0, x = x_455_cast_fp16)[name = string("transpose_205")]; + tensor input_927_cast_fp16 = reshape(shape = var_4171, x = var_4170_cast_fp16)[name = string("input_927_cast_fp16")]; + tensor encoder_layers_17_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350981568))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351768064))))[name = string("encoder_layers_17_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_17_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_17_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351768256)))]; + tensor linear_160_cast_fp16 = linear(bias = encoder_layers_17_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_17_self_attn_linear_out_weight_to_fp16_palettized, x = input_927_cast_fp16)[name = string("linear_160_cast_fp16")]; + tensor input_931_cast_fp16 = add(x = input_921_cast_fp16, y = linear_160_cast_fp16)[name = string("input_931_cast_fp16")]; + tensor x_459_axes_0 = const()[name = string("x_459_axes_0"), val = tensor([-1])]; + tensor encoder_layers_17_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_17_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351770368)))]; + tensor encoder_layers_17_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_17_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351772480)))]; + tensor x_459_cast_fp16 = layer_norm(axes = x_459_axes_0, beta = encoder_layers_17_norm_conv_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_17_norm_conv_weight_to_fp16, x = input_931_cast_fp16)[name = string("x_459_cast_fp16")]; + tensor input_933_perm_0 = const()[name = string("input_933_perm_0"), val = tensor([0, 2, 1])]; + string input_935_pad_type_0 = const()[name = string("input_935_pad_type_0"), val = string("valid")]; + tensor input_935_strides_0 = const()[name = string("input_935_strides_0"), val = tensor([1])]; + tensor input_935_pad_0 = const()[name = string("input_935_pad_0"), val = tensor([0, 0])]; + tensor input_935_dilations_0 = const()[name = string("input_935_dilations_0"), val = tensor([1])]; + int32 input_935_groups_0 = const()[name = string("input_935_groups_0"), val = int32(1)]; + tensor encoder_layers_17_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351774592))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(353871808))))[name = string("encoder_layers_17_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_933_cast_fp16 = transpose(perm = input_933_perm_0, x = x_459_cast_fp16)[name = string("transpose_204")]; + tensor input_935_cast_fp16 = conv(dilations = input_935_dilations_0, groups = input_935_groups_0, pad = input_935_pad_0, pad_type = input_935_pad_type_0, strides = input_935_strides_0, weight = encoder_layers_17_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_933_cast_fp16)[name = string("input_935_cast_fp16")]; + int32 x_461_split_num_splits_0 = const()[name = string("x_461_split_num_splits_0"), val = int32(2)]; + int32 x_461_split_axis_0 = const()[name = string("x_461_split_axis_0"), val = int32(1)]; + tensor x_461_split_cast_fp16_0, tensor x_461_split_cast_fp16_1 = split(axis = x_461_split_axis_0, num_splits = x_461_split_num_splits_0, x = input_935_cast_fp16)[name = string("x_461_split_cast_fp16")]; + tensor x_461_split_1_sigmoid_cast_fp16 = sigmoid(x = x_461_split_cast_fp16_1)[name = string("x_461_split_1_sigmoid_cast_fp16")]; + tensor x_461_cast_fp16 = mul(x = x_461_split_cast_fp16_0, y = x_461_split_1_sigmoid_cast_fp16)[name = string("x_461_cast_fp16")]; + tensor input_937_cast_fp16 = select(a = var_45_to_fp16, b = x_461_cast_fp16, cond = var_576)[name = string("input_937_cast_fp16")]; + bool new_x_71_interleave_0 = const()[name = string("new_x_71_interleave_0"), val = bool(false)]; + tensor new_x_71_cast_fp16 = concat(axis = var_60, interleave = new_x_71_interleave_0, values = (cache_71_cast_fp16, input_937_cast_fp16))[name = string("new_x_71_cast_fp16")]; + tensor var_4210_begin_0 = const()[name = string("op_4210_begin_0"), val = tensor([0, 0, 14])]; + tensor var_4210_end_0 = const()[name = string("op_4210_end_0"), val = tensor([1, 1024, 22])]; + tensor var_4210_end_mask_0 = const()[name = string("op_4210_end_mask_0"), val = tensor([true, true, true])]; + tensor var_4210_cast_fp16 = slice_by_index(begin = var_4210_begin_0, end = var_4210_end_0, end_mask = var_4210_end_mask_0, x = new_x_71_cast_fp16)[name = string("op_4210_cast_fp16")]; + string x_463_pad_type_0 = const()[name = string("x_463_pad_type_0"), val = string("valid")]; + int32 x_463_groups_0 = const()[name = string("x_463_groups_0"), val = int32(1024)]; + tensor x_463_strides_0 = const()[name = string("x_463_strides_0"), val = tensor([1])]; + tensor x_463_pad_0 = const()[name = string("x_463_pad_0"), val = tensor([0, 0])]; + tensor x_463_dilations_0 = const()[name = string("x_463_dilations_0"), val = tensor([1])]; + tensor encoder_layers_17_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(353875968))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(353885248))))[name = string("encoder_layers_17_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_463_cast_fp16 = conv(dilations = x_463_dilations_0, groups = x_463_groups_0, pad = x_463_pad_0, pad_type = x_463_pad_type_0, strides = x_463_strides_0, weight = encoder_layers_17_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_71_cast_fp16)[name = string("x_463_cast_fp16")]; + tensor input_939_perm_0 = const()[name = string("input_939_perm_0"), val = tensor([0, 2, 1])]; + tensor x_465_axes_0 = const()[name = string("x_465_axes_0"), val = tensor([-1])]; + tensor encoder_layers_17_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_17_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(353887360)))]; + tensor encoder_layers_17_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_17_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(353889472)))]; + tensor input_939_cast_fp16 = transpose(perm = input_939_perm_0, x = x_463_cast_fp16)[name = string("transpose_203")]; + tensor x_465_cast_fp16 = layer_norm(axes = x_465_axes_0, beta = encoder_layers_17_conv_batch_norm_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_17_conv_batch_norm_weight_to_fp16, x = input_939_cast_fp16)[name = string("x_465_cast_fp16")]; + tensor input_941_perm_0 = const()[name = string("input_941_perm_0"), val = tensor([0, 2, 1])]; + tensor input_941_cast_fp16 = transpose(perm = input_941_perm_0, x = x_465_cast_fp16)[name = string("transpose_202")]; + tensor input_943_cast_fp16 = silu(x = input_941_cast_fp16)[name = string("input_943_cast_fp16")]; + string x_467_pad_type_0 = const()[name = string("x_467_pad_type_0"), val = string("valid")]; + tensor x_467_strides_0 = const()[name = string("x_467_strides_0"), val = tensor([1])]; + tensor x_467_pad_0 = const()[name = string("x_467_pad_0"), val = tensor([0, 0])]; + tensor x_467_dilations_0 = const()[name = string("x_467_dilations_0"), val = tensor([1])]; + int32 x_467_groups_0 = const()[name = string("x_467_groups_0"), val = int32(1)]; + tensor encoder_layers_17_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(353891584))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(354940224))))[name = string("encoder_layers_17_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_467_cast_fp16 = conv(dilations = x_467_dilations_0, groups = x_467_groups_0, pad = x_467_pad_0, pad_type = x_467_pad_type_0, strides = x_467_strides_0, weight = encoder_layers_17_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_943_cast_fp16)[name = string("x_467_cast_fp16")]; + tensor input_945_perm_0 = const()[name = string("input_945_perm_0"), val = tensor([0, 2, 1])]; + tensor input_945_cast_fp16 = transpose(perm = input_945_perm_0, x = x_467_cast_fp16)[name = string("transpose_201")]; + tensor input_947_cast_fp16 = add(x = input_931_cast_fp16, y = input_945_cast_fp16)[name = string("input_947_cast_fp16")]; + tensor input_949_axes_0 = const()[name = string("input_949_axes_0"), val = tensor([-1])]; + tensor encoder_layers_17_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_17_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(354942336)))]; + tensor encoder_layers_17_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_17_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(354944448)))]; + tensor input_949_cast_fp16 = layer_norm(axes = input_949_axes_0, beta = encoder_layers_17_norm_feed_forward2_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_17_norm_feed_forward2_weight_to_fp16, x = input_947_cast_fp16)[name = string("input_949_cast_fp16")]; + tensor encoder_layers_17_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(354946560))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(358092352))))[name = string("encoder_layers_17_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_17_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_17_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(358092544)))]; + tensor linear_161_cast_fp16 = linear(bias = encoder_layers_17_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_17_feed_forward2_linear1_weight_to_fp16_palettized, x = input_949_cast_fp16)[name = string("linear_161_cast_fp16")]; + tensor input_953_cast_fp16 = silu(x = linear_161_cast_fp16)[name = string("input_953_cast_fp16")]; + tensor encoder_layers_17_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(358100800))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361246592))))[name = string("encoder_layers_17_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_17_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_17_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361246784)))]; + tensor linear_162_cast_fp16 = linear(bias = encoder_layers_17_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_17_feed_forward2_linear2_weight_to_fp16_palettized, x = input_953_cast_fp16)[name = string("linear_162_cast_fp16")]; + fp16 var_4253_to_fp16 = const()[name = string("op_4253_to_fp16"), val = fp16(0x1p-1)]; + tensor var_4254_cast_fp16 = mul(x = linear_162_cast_fp16, y = var_4253_to_fp16)[name = string("op_4254_cast_fp16")]; + tensor input_959_cast_fp16 = add(x = input_947_cast_fp16, y = var_4254_cast_fp16)[name = string("input_959_cast_fp16")]; + tensor input_961_axes_0 = const()[name = string("input_961_axes_0"), val = tensor([-1])]; + tensor encoder_layers_17_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_17_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361248896)))]; + tensor encoder_layers_17_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_17_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361251008)))]; + tensor input_961_cast_fp16 = layer_norm(axes = input_961_axes_0, beta = encoder_layers_17_norm_out_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_17_norm_out_weight_to_fp16, x = input_959_cast_fp16)[name = string("input_961_cast_fp16")]; + tensor cache_73_begin_0 = const()[name = string("cache_73_begin_0"), val = tensor([18, 0, 0, 0])]; + tensor cache_73_end_0 = const()[name = string("cache_73_end_0"), val = tensor([19, 1, 42, 1024])]; + tensor cache_73_end_mask_0 = const()[name = string("cache_73_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_73_squeeze_mask_0 = const()[name = string("cache_73_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_73_cast_fp16 = slice_by_index(begin = cache_73_begin_0, end = cache_73_end_0, end_mask = cache_73_end_mask_0, squeeze_mask = cache_73_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_73_cast_fp16")]; + tensor cache_75_begin_0 = const()[name = string("cache_75_begin_0"), val = tensor([18, 0, 0, 0])]; + tensor cache_75_end_0 = const()[name = string("cache_75_end_0"), val = tensor([19, 1, 1024, 8])]; + tensor cache_75_end_mask_0 = const()[name = string("cache_75_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_75_squeeze_mask_0 = const()[name = string("cache_75_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_75_cast_fp16 = slice_by_index(begin = cache_75_begin_0, end = cache_75_end_0, end_mask = cache_75_end_mask_0, squeeze_mask = cache_75_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_75_cast_fp16")]; + tensor input_963_axes_0 = const()[name = string("input_963_axes_0"), val = tensor([-1])]; + tensor encoder_layers_18_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_18_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361253120)))]; + tensor encoder_layers_18_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_18_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361255232)))]; + tensor input_963_cast_fp16 = layer_norm(axes = input_963_axes_0, beta = encoder_layers_18_norm_feed_forward1_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_18_norm_feed_forward1_weight_to_fp16, x = input_961_cast_fp16)[name = string("input_963_cast_fp16")]; + tensor encoder_layers_18_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361257344))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(364403136))))[name = string("encoder_layers_18_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_18_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_18_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(364403328)))]; + tensor linear_163_cast_fp16 = linear(bias = encoder_layers_18_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_18_feed_forward1_linear1_weight_to_fp16_palettized, x = input_963_cast_fp16)[name = string("linear_163_cast_fp16")]; + tensor input_967_cast_fp16 = silu(x = linear_163_cast_fp16)[name = string("input_967_cast_fp16")]; + tensor encoder_layers_18_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(364411584))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(367557376))))[name = string("encoder_layers_18_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_18_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_18_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(367557568)))]; + tensor linear_164_cast_fp16 = linear(bias = encoder_layers_18_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_18_feed_forward1_linear2_weight_to_fp16_palettized, x = input_967_cast_fp16)[name = string("linear_164_cast_fp16")]; + fp16 var_4290_to_fp16 = const()[name = string("op_4290_to_fp16"), val = fp16(0x1p-1)]; + tensor var_4291_cast_fp16 = mul(x = linear_164_cast_fp16, y = var_4290_to_fp16)[name = string("op_4291_cast_fp16")]; + tensor input_973_cast_fp16 = add(x = input_961_cast_fp16, y = var_4291_cast_fp16)[name = string("input_973_cast_fp16")]; + tensor key_37_axes_0 = const()[name = string("key_37_axes_0"), val = tensor([-1])]; + tensor encoder_layers_18_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_18_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(367559680)))]; + tensor encoder_layers_18_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_18_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(367561792)))]; + tensor key_37_cast_fp16 = layer_norm(axes = key_37_axes_0, beta = encoder_layers_18_norm_self_att_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_18_norm_self_att_weight_to_fp16, x = input_973_cast_fp16)[name = string("key_37_cast_fp16")]; + bool input_975_interleave_0 = const()[name = string("input_975_interleave_0"), val = bool(false)]; + tensor input_975_cast_fp16 = concat(axis = var_69, interleave = input_975_interleave_0, values = (cache_73_cast_fp16, key_37_cast_fp16))[name = string("input_975_cast_fp16")]; + tensor var_4313_begin_0 = const()[name = string("op_4313_begin_0"), val = tensor([0, 14, 0])]; + tensor var_4313_end_0 = const()[name = string("op_4313_end_0"), val = tensor([1, 42, 1024])]; + tensor var_4313_end_mask_0 = const()[name = string("op_4313_end_mask_0"), val = tensor([true, true, true])]; + tensor var_4313_cast_fp16 = slice_by_index(begin = var_4313_begin_0, end = var_4313_end_0, end_mask = var_4313_end_mask_0, x = cache_73_cast_fp16)[name = string("op_4313_cast_fp16")]; + bool var_4319_interleave_0 = const()[name = string("op_4319_interleave_0"), val = bool(false)]; + tensor var_4319_cast_fp16 = concat(axis = var_69, interleave = var_4319_interleave_0, values = (var_4313_cast_fp16, key_37_cast_fp16))[name = string("op_4319_cast_fp16")]; + tensor encoder_layers_18_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(367563904))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368350400))))[name = string("encoder_layers_18_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_18_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_18_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368350592)))]; + tensor linear_165_cast_fp16 = linear(bias = encoder_layers_18_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_18_self_attn_linear_q_weight_to_fp16_palettized, x = key_37_cast_fp16)[name = string("linear_165_cast_fp16")]; + tensor var_4324 = const()[name = string("op_4324"), val = tensor([1, -1, 8, 128])]; + tensor q_109_cast_fp16 = reshape(shape = var_4324, x = linear_165_cast_fp16)[name = string("q_109_cast_fp16")]; + tensor encoder_layers_18_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368352704))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369139200))))[name = string("encoder_layers_18_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_18_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_18_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369139392)))]; + tensor linear_166_cast_fp16 = linear(bias = encoder_layers_18_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_18_self_attn_linear_k_weight_to_fp16_palettized, x = input_975_cast_fp16)[name = string("linear_166_cast_fp16")]; + tensor var_4329 = const()[name = string("op_4329"), val = tensor([1, -1, 8, 128])]; + tensor k_73_cast_fp16 = reshape(shape = var_4329, x = linear_166_cast_fp16)[name = string("k_73_cast_fp16")]; + tensor encoder_layers_18_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369141504))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369928000))))[name = string("encoder_layers_18_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_18_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_18_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369928192)))]; + tensor linear_167_cast_fp16 = linear(bias = encoder_layers_18_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_18_self_attn_linear_v_weight_to_fp16_palettized, x = input_975_cast_fp16)[name = string("linear_167_cast_fp16")]; + tensor var_4334 = const()[name = string("op_4334"), val = tensor([1, -1, 8, 128])]; + tensor v_37_cast_fp16 = reshape(shape = var_4334, x = linear_167_cast_fp16)[name = string("v_37_cast_fp16")]; + tensor value_45_perm_0 = const()[name = string("value_45_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_18_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_18_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369930304)))]; + tensor var_4347_cast_fp16 = add(x = q_109_cast_fp16, y = encoder_layers_18_self_attn_pos_bias_u_to_fp16)[name = string("op_4347_cast_fp16")]; + tensor encoder_layers_18_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_18_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369932416)))]; + tensor var_4349_cast_fp16 = add(x = q_109_cast_fp16, y = encoder_layers_18_self_attn_pos_bias_v_to_fp16)[name = string("op_4349_cast_fp16")]; + tensor q_with_bias_v_37_perm_0 = const()[name = string("q_with_bias_v_37_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_475_transpose_x_0 = const()[name = string("x_475_transpose_x_0"), val = bool(false)]; + bool x_475_transpose_y_0 = const()[name = string("x_475_transpose_y_0"), val = bool(false)]; + tensor op_4351_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369934528))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(370048256))))[name = string("op_4351_to_fp16_quantized")]; + tensor q_with_bias_v_37_cast_fp16 = transpose(perm = q_with_bias_v_37_perm_0, x = var_4349_cast_fp16)[name = string("transpose_200")]; + tensor x_475_cast_fp16 = matmul(transpose_x = x_475_transpose_x_0, transpose_y = x_475_transpose_y_0, x = q_with_bias_v_37_cast_fp16, y = op_4351_to_fp16_quantized)[name = string("x_475_cast_fp16")]; + tensor x_477_pad_0 = const()[name = string("x_477_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_477_mode_0 = const()[name = string("x_477_mode_0"), val = string("constant")]; + fp16 const_313_to_fp16 = const()[name = string("const_313_to_fp16"), val = fp16(0x0p+0)]; + tensor x_477_cast_fp16 = pad(constant_val = const_313_to_fp16, mode = x_477_mode_0, pad = x_477_pad_0, x = x_475_cast_fp16)[name = string("x_477_cast_fp16")]; + tensor var_4359 = const()[name = string("op_4359"), val = tensor([1, 8, -1, 14])]; + tensor x_479_cast_fp16 = reshape(shape = var_4359, x = x_477_cast_fp16)[name = string("x_479_cast_fp16")]; + tensor var_4363_begin_0 = const()[name = string("op_4363_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_4363_end_0 = const()[name = string("op_4363_end_0"), val = tensor([1, 8, 112, 14])]; + tensor var_4363_end_mask_0 = const()[name = string("op_4363_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_4363_cast_fp16 = slice_by_index(begin = var_4363_begin_0, end = var_4363_end_0, end_mask = var_4363_end_mask_0, x = x_479_cast_fp16)[name = string("op_4363_cast_fp16")]; + tensor var_4364 = const()[name = string("op_4364"), val = tensor([1, 8, 14, 111])]; + tensor matrix_bd_73_cast_fp16 = reshape(shape = var_4364, x = var_4363_cast_fp16)[name = string("matrix_bd_73_cast_fp16")]; + bool matrix_ac_37_transpose_x_0 = const()[name = string("matrix_ac_37_transpose_x_0"), val = bool(false)]; + bool matrix_ac_37_transpose_y_0 = const()[name = string("matrix_ac_37_transpose_y_0"), val = bool(false)]; + tensor transpose_132_perm_0 = const()[name = string("transpose_132_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_133_perm_0 = const()[name = string("transpose_133_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_133 = transpose(perm = transpose_133_perm_0, x = k_73_cast_fp16)[name = string("transpose_198")]; + tensor transpose_132 = transpose(perm = transpose_132_perm_0, x = var_4347_cast_fp16)[name = string("transpose_199")]; + tensor matrix_ac_37_cast_fp16 = matmul(transpose_x = matrix_ac_37_transpose_x_0, transpose_y = matrix_ac_37_transpose_y_0, x = transpose_132, y = transpose_133)[name = string("matrix_ac_37_cast_fp16")]; + tensor matrix_bd_75_begin_0 = const()[name = string("matrix_bd_75_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_75_end_0 = const()[name = string("matrix_bd_75_end_0"), val = tensor([1, 8, 14, 56])]; + tensor matrix_bd_75_end_mask_0 = const()[name = string("matrix_bd_75_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_75_cast_fp16 = slice_by_index(begin = matrix_bd_75_begin_0, end = matrix_bd_75_end_0, end_mask = matrix_bd_75_end_mask_0, x = matrix_bd_73_cast_fp16)[name = string("matrix_bd_75_cast_fp16")]; + tensor var_4373_cast_fp16 = add(x = matrix_ac_37_cast_fp16, y = matrix_bd_75_cast_fp16)[name = string("op_4373_cast_fp16")]; + fp16 _inversed_scores_73_y_0_to_fp16 = const()[name = string("_inversed_scores_73_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_73_cast_fp16 = mul(x = var_4373_cast_fp16, y = _inversed_scores_73_y_0_to_fp16)[name = string("_inversed_scores_73_cast_fp16")]; + tensor scores_75_cast_fp16 = select(a = var_46_to_fp16, b = _inversed_scores_73_cast_fp16, cond = mask_11)[name = string("scores_75_cast_fp16")]; + tensor var_4379_cast_fp16 = softmax(axis = var_60, x = scores_75_cast_fp16)[name = string("op_4379_cast_fp16")]; + tensor input_977_cast_fp16 = select(a = var_45_to_fp16, b = var_4379_cast_fp16, cond = mask_11)[name = string("input_977_cast_fp16")]; + bool x_481_transpose_x_0 = const()[name = string("x_481_transpose_x_0"), val = bool(false)]; + bool x_481_transpose_y_0 = const()[name = string("x_481_transpose_y_0"), val = bool(false)]; + tensor value_45_cast_fp16 = transpose(perm = value_45_perm_0, x = v_37_cast_fp16)[name = string("transpose_197")]; + tensor x_481_cast_fp16 = matmul(transpose_x = x_481_transpose_x_0, transpose_y = x_481_transpose_y_0, x = input_977_cast_fp16, y = value_45_cast_fp16)[name = string("x_481_cast_fp16")]; + tensor var_4383_perm_0 = const()[name = string("op_4383_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_4384 = const()[name = string("op_4384"), val = tensor([1, -1, 1024])]; + tensor var_4383_cast_fp16 = transpose(perm = var_4383_perm_0, x = x_481_cast_fp16)[name = string("transpose_196")]; + tensor input_979_cast_fp16 = reshape(shape = var_4384, x = var_4383_cast_fp16)[name = string("input_979_cast_fp16")]; + tensor encoder_layers_18_self_attn_linear_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(370048576))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371097216))))[name = string("encoder_layers_18_self_attn_linear_out_weight_to_fp16_quantized")]; + tensor encoder_layers_18_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_18_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371099328)))]; + tensor linear_169_cast_fp16 = linear(bias = encoder_layers_18_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_18_self_attn_linear_out_weight_to_fp16_quantized, x = input_979_cast_fp16)[name = string("linear_169_cast_fp16")]; + tensor input_983_cast_fp16 = add(x = input_973_cast_fp16, y = linear_169_cast_fp16)[name = string("input_983_cast_fp16")]; + tensor x_485_axes_0 = const()[name = string("x_485_axes_0"), val = tensor([-1])]; + tensor encoder_layers_18_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_18_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371101440)))]; + tensor encoder_layers_18_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_18_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371103552)))]; + tensor x_485_cast_fp16 = layer_norm(axes = x_485_axes_0, beta = encoder_layers_18_norm_conv_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_18_norm_conv_weight_to_fp16, x = input_983_cast_fp16)[name = string("x_485_cast_fp16")]; + tensor input_985_perm_0 = const()[name = string("input_985_perm_0"), val = tensor([0, 2, 1])]; + string input_987_pad_type_0 = const()[name = string("input_987_pad_type_0"), val = string("valid")]; + tensor input_987_strides_0 = const()[name = string("input_987_strides_0"), val = tensor([1])]; + tensor input_987_pad_0 = const()[name = string("input_987_pad_0"), val = tensor([0, 0])]; + tensor input_987_dilations_0 = const()[name = string("input_987_dilations_0"), val = tensor([1])]; + int32 input_987_groups_0 = const()[name = string("input_987_groups_0"), val = int32(1)]; + tensor encoder_layers_18_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371105664))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(373202880))))[name = string("encoder_layers_18_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_985_cast_fp16 = transpose(perm = input_985_perm_0, x = x_485_cast_fp16)[name = string("transpose_195")]; + tensor input_987_cast_fp16 = conv(dilations = input_987_dilations_0, groups = input_987_groups_0, pad = input_987_pad_0, pad_type = input_987_pad_type_0, strides = input_987_strides_0, weight = encoder_layers_18_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_985_cast_fp16)[name = string("input_987_cast_fp16")]; + int32 x_487_split_num_splits_0 = const()[name = string("x_487_split_num_splits_0"), val = int32(2)]; + int32 x_487_split_axis_0 = const()[name = string("x_487_split_axis_0"), val = int32(1)]; + tensor x_487_split_cast_fp16_0, tensor x_487_split_cast_fp16_1 = split(axis = x_487_split_axis_0, num_splits = x_487_split_num_splits_0, x = input_987_cast_fp16)[name = string("x_487_split_cast_fp16")]; + tensor x_487_split_1_sigmoid_cast_fp16 = sigmoid(x = x_487_split_cast_fp16_1)[name = string("x_487_split_1_sigmoid_cast_fp16")]; + tensor x_487_cast_fp16 = mul(x = x_487_split_cast_fp16_0, y = x_487_split_1_sigmoid_cast_fp16)[name = string("x_487_cast_fp16")]; + tensor input_989_cast_fp16 = select(a = var_45_to_fp16, b = x_487_cast_fp16, cond = var_576)[name = string("input_989_cast_fp16")]; + bool new_x_75_interleave_0 = const()[name = string("new_x_75_interleave_0"), val = bool(false)]; + tensor new_x_75_cast_fp16 = concat(axis = var_60, interleave = new_x_75_interleave_0, values = (cache_75_cast_fp16, input_989_cast_fp16))[name = string("new_x_75_cast_fp16")]; + tensor var_4423_begin_0 = const()[name = string("op_4423_begin_0"), val = tensor([0, 0, 14])]; + tensor var_4423_end_0 = const()[name = string("op_4423_end_0"), val = tensor([1, 1024, 22])]; + tensor var_4423_end_mask_0 = const()[name = string("op_4423_end_mask_0"), val = tensor([true, true, true])]; + tensor var_4423_cast_fp16 = slice_by_index(begin = var_4423_begin_0, end = var_4423_end_0, end_mask = var_4423_end_mask_0, x = new_x_75_cast_fp16)[name = string("op_4423_cast_fp16")]; + string x_489_pad_type_0 = const()[name = string("x_489_pad_type_0"), val = string("valid")]; + int32 x_489_groups_0 = const()[name = string("x_489_groups_0"), val = int32(1024)]; + tensor x_489_strides_0 = const()[name = string("x_489_strides_0"), val = tensor([1])]; + tensor x_489_pad_0 = const()[name = string("x_489_pad_0"), val = tensor([0, 0])]; + tensor x_489_dilations_0 = const()[name = string("x_489_dilations_0"), val = tensor([1])]; + tensor encoder_layers_18_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(373207040))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(373216320))))[name = string("encoder_layers_18_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_489_cast_fp16 = conv(dilations = x_489_dilations_0, groups = x_489_groups_0, pad = x_489_pad_0, pad_type = x_489_pad_type_0, strides = x_489_strides_0, weight = encoder_layers_18_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_75_cast_fp16)[name = string("x_489_cast_fp16")]; + tensor input_991_perm_0 = const()[name = string("input_991_perm_0"), val = tensor([0, 2, 1])]; + tensor x_491_axes_0 = const()[name = string("x_491_axes_0"), val = tensor([-1])]; + tensor encoder_layers_18_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_18_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(373218432)))]; + tensor encoder_layers_18_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_18_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(373220544)))]; + tensor input_991_cast_fp16 = transpose(perm = input_991_perm_0, x = x_489_cast_fp16)[name = string("transpose_194")]; + tensor x_491_cast_fp16 = layer_norm(axes = x_491_axes_0, beta = encoder_layers_18_conv_batch_norm_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_18_conv_batch_norm_weight_to_fp16, x = input_991_cast_fp16)[name = string("x_491_cast_fp16")]; + tensor input_993_perm_0 = const()[name = string("input_993_perm_0"), val = tensor([0, 2, 1])]; + tensor input_993_cast_fp16 = transpose(perm = input_993_perm_0, x = x_491_cast_fp16)[name = string("transpose_193")]; + tensor input_995_cast_fp16 = silu(x = input_993_cast_fp16)[name = string("input_995_cast_fp16")]; + string x_493_pad_type_0 = const()[name = string("x_493_pad_type_0"), val = string("valid")]; + tensor x_493_strides_0 = const()[name = string("x_493_strides_0"), val = tensor([1])]; + tensor x_493_pad_0 = const()[name = string("x_493_pad_0"), val = tensor([0, 0])]; + tensor x_493_dilations_0 = const()[name = string("x_493_dilations_0"), val = tensor([1])]; + int32 x_493_groups_0 = const()[name = string("x_493_groups_0"), val = int32(1)]; + tensor encoder_layers_18_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(373222656))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(374271296))))[name = string("encoder_layers_18_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_493_cast_fp16 = conv(dilations = x_493_dilations_0, groups = x_493_groups_0, pad = x_493_pad_0, pad_type = x_493_pad_type_0, strides = x_493_strides_0, weight = encoder_layers_18_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_995_cast_fp16)[name = string("x_493_cast_fp16")]; + tensor input_997_perm_0 = const()[name = string("input_997_perm_0"), val = tensor([0, 2, 1])]; + tensor input_997_cast_fp16 = transpose(perm = input_997_perm_0, x = x_493_cast_fp16)[name = string("transpose_192")]; + tensor input_999_cast_fp16 = add(x = input_983_cast_fp16, y = input_997_cast_fp16)[name = string("input_999_cast_fp16")]; + tensor input_1001_axes_0 = const()[name = string("input_1001_axes_0"), val = tensor([-1])]; + tensor encoder_layers_18_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_18_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(374273408)))]; + tensor encoder_layers_18_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_18_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(374275520)))]; + tensor input_1001_cast_fp16 = layer_norm(axes = input_1001_axes_0, beta = encoder_layers_18_norm_feed_forward2_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_18_norm_feed_forward2_weight_to_fp16, x = input_999_cast_fp16)[name = string("input_1001_cast_fp16")]; + tensor encoder_layers_18_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(374277632))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(378472000))))[name = string("encoder_layers_18_feed_forward2_linear1_weight_to_fp16_quantized")]; + tensor encoder_layers_18_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_18_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(378480256)))]; + tensor linear_170_cast_fp16 = linear(bias = encoder_layers_18_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_18_feed_forward2_linear1_weight_to_fp16_quantized, x = input_1001_cast_fp16)[name = string("linear_170_cast_fp16")]; + tensor input_1005_cast_fp16 = silu(x = linear_170_cast_fp16)[name = string("input_1005_cast_fp16")]; + tensor encoder_layers_18_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(378488512))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(382682880))))[name = string("encoder_layers_18_feed_forward2_linear2_weight_to_fp16_quantized")]; + tensor encoder_layers_18_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_18_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(382684992)))]; + tensor linear_171_cast_fp16 = linear(bias = encoder_layers_18_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_18_feed_forward2_linear2_weight_to_fp16_quantized, x = input_1005_cast_fp16)[name = string("linear_171_cast_fp16")]; + fp16 var_4466_to_fp16 = const()[name = string("op_4466_to_fp16"), val = fp16(0x1p-1)]; + tensor var_4467_cast_fp16 = mul(x = linear_171_cast_fp16, y = var_4466_to_fp16)[name = string("op_4467_cast_fp16")]; + tensor input_1011_cast_fp16 = add(x = input_999_cast_fp16, y = var_4467_cast_fp16)[name = string("input_1011_cast_fp16")]; + tensor input_1013_axes_0 = const()[name = string("input_1013_axes_0"), val = tensor([-1])]; + tensor encoder_layers_18_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_18_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(382687104)))]; + tensor encoder_layers_18_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_18_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(382689216)))]; + tensor input_1013_cast_fp16 = layer_norm(axes = input_1013_axes_0, beta = encoder_layers_18_norm_out_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_18_norm_out_weight_to_fp16, x = input_1011_cast_fp16)[name = string("input_1013_cast_fp16")]; + tensor cache_77_begin_0 = const()[name = string("cache_77_begin_0"), val = tensor([19, 0, 0, 0])]; + tensor cache_77_end_0 = const()[name = string("cache_77_end_0"), val = tensor([20, 1, 42, 1024])]; + tensor cache_77_end_mask_0 = const()[name = string("cache_77_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_77_squeeze_mask_0 = const()[name = string("cache_77_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_77_cast_fp16 = slice_by_index(begin = cache_77_begin_0, end = cache_77_end_0, end_mask = cache_77_end_mask_0, squeeze_mask = cache_77_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_77_cast_fp16")]; + tensor cache_79_begin_0 = const()[name = string("cache_79_begin_0"), val = tensor([19, 0, 0, 0])]; + tensor cache_79_end_0 = const()[name = string("cache_79_end_0"), val = tensor([20, 1, 1024, 8])]; + tensor cache_79_end_mask_0 = const()[name = string("cache_79_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_79_squeeze_mask_0 = const()[name = string("cache_79_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_79_cast_fp16 = slice_by_index(begin = cache_79_begin_0, end = cache_79_end_0, end_mask = cache_79_end_mask_0, squeeze_mask = cache_79_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_79_cast_fp16")]; + tensor input_1015_axes_0 = const()[name = string("input_1015_axes_0"), val = tensor([-1])]; + tensor encoder_layers_19_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_19_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(382691328)))]; + tensor encoder_layers_19_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_19_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(382693440)))]; + tensor input_1015_cast_fp16 = layer_norm(axes = input_1015_axes_0, beta = encoder_layers_19_norm_feed_forward1_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_19_norm_feed_forward1_weight_to_fp16, x = input_1013_cast_fp16)[name = string("input_1015_cast_fp16")]; + tensor encoder_layers_19_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(382695552))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(386889920))))[name = string("encoder_layers_19_feed_forward1_linear1_weight_to_fp16_quantized")]; + tensor encoder_layers_19_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_19_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(386898176)))]; + tensor linear_172_cast_fp16 = linear(bias = encoder_layers_19_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_19_feed_forward1_linear1_weight_to_fp16_quantized, x = input_1015_cast_fp16)[name = string("linear_172_cast_fp16")]; + tensor input_1019_cast_fp16 = silu(x = linear_172_cast_fp16)[name = string("input_1019_cast_fp16")]; + tensor encoder_layers_19_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(386906432))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391100800))))[name = string("encoder_layers_19_feed_forward1_linear2_weight_to_fp16_quantized")]; + tensor encoder_layers_19_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_19_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391102912)))]; + tensor linear_173_cast_fp16 = linear(bias = encoder_layers_19_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_19_feed_forward1_linear2_weight_to_fp16_quantized, x = input_1019_cast_fp16)[name = string("linear_173_cast_fp16")]; + fp16 var_4503_to_fp16 = const()[name = string("op_4503_to_fp16"), val = fp16(0x1p-1)]; + tensor var_4504_cast_fp16 = mul(x = linear_173_cast_fp16, y = var_4503_to_fp16)[name = string("op_4504_cast_fp16")]; + tensor input_1025_cast_fp16 = add(x = input_1013_cast_fp16, y = var_4504_cast_fp16)[name = string("input_1025_cast_fp16")]; + tensor key_39_axes_0 = const()[name = string("key_39_axes_0"), val = tensor([-1])]; + tensor encoder_layers_19_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_19_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391105024)))]; + tensor encoder_layers_19_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_19_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391107136)))]; + tensor key_39_cast_fp16 = layer_norm(axes = key_39_axes_0, beta = encoder_layers_19_norm_self_att_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_19_norm_self_att_weight_to_fp16, x = input_1025_cast_fp16)[name = string("key_39_cast_fp16")]; + bool input_1027_interleave_0 = const()[name = string("input_1027_interleave_0"), val = bool(false)]; + tensor input_1027_cast_fp16 = concat(axis = var_69, interleave = input_1027_interleave_0, values = (cache_77_cast_fp16, key_39_cast_fp16))[name = string("input_1027_cast_fp16")]; + tensor var_4526_begin_0 = const()[name = string("op_4526_begin_0"), val = tensor([0, 14, 0])]; + tensor var_4526_end_0 = const()[name = string("op_4526_end_0"), val = tensor([1, 42, 1024])]; + tensor var_4526_end_mask_0 = const()[name = string("op_4526_end_mask_0"), val = tensor([true, true, true])]; + tensor var_4526_cast_fp16 = slice_by_index(begin = var_4526_begin_0, end = var_4526_end_0, end_mask = var_4526_end_mask_0, x = cache_77_cast_fp16)[name = string("op_4526_cast_fp16")]; + bool var_4532_interleave_0 = const()[name = string("op_4532_interleave_0"), val = bool(false)]; + tensor var_4532_cast_fp16 = concat(axis = var_69, interleave = var_4532_interleave_0, values = (var_4526_cast_fp16, key_39_cast_fp16))[name = string("op_4532_cast_fp16")]; + tensor encoder_layers_19_self_attn_linear_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391109248))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(392157888))))[name = string("encoder_layers_19_self_attn_linear_q_weight_to_fp16_quantized")]; + tensor encoder_layers_19_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_19_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(392160000)))]; + tensor linear_174_cast_fp16 = linear(bias = encoder_layers_19_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_19_self_attn_linear_q_weight_to_fp16_quantized, x = key_39_cast_fp16)[name = string("linear_174_cast_fp16")]; + tensor var_4537 = const()[name = string("op_4537"), val = tensor([1, -1, 8, 128])]; + tensor q_115_cast_fp16 = reshape(shape = var_4537, x = linear_174_cast_fp16)[name = string("q_115_cast_fp16")]; + tensor encoder_layers_19_self_attn_linear_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(392162112))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(393210752))))[name = string("encoder_layers_19_self_attn_linear_k_weight_to_fp16_quantized")]; + tensor encoder_layers_19_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_19_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(393212864)))]; + tensor linear_175_cast_fp16 = linear(bias = encoder_layers_19_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_19_self_attn_linear_k_weight_to_fp16_quantized, x = input_1027_cast_fp16)[name = string("linear_175_cast_fp16")]; + tensor var_4542 = const()[name = string("op_4542"), val = tensor([1, -1, 8, 128])]; + tensor k_77_cast_fp16 = reshape(shape = var_4542, x = linear_175_cast_fp16)[name = string("k_77_cast_fp16")]; + tensor encoder_layers_19_self_attn_linear_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(393214976))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(394263616))))[name = string("encoder_layers_19_self_attn_linear_v_weight_to_fp16_quantized")]; + tensor encoder_layers_19_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_19_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(394265728)))]; + tensor linear_176_cast_fp16 = linear(bias = encoder_layers_19_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_19_self_attn_linear_v_weight_to_fp16_quantized, x = input_1027_cast_fp16)[name = string("linear_176_cast_fp16")]; + tensor var_4547 = const()[name = string("op_4547"), val = tensor([1, -1, 8, 128])]; + tensor v_39_cast_fp16 = reshape(shape = var_4547, x = linear_176_cast_fp16)[name = string("v_39_cast_fp16")]; + tensor value_47_perm_0 = const()[name = string("value_47_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_19_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_19_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(394267840)))]; + tensor var_4560_cast_fp16 = add(x = q_115_cast_fp16, y = encoder_layers_19_self_attn_pos_bias_u_to_fp16)[name = string("op_4560_cast_fp16")]; + tensor encoder_layers_19_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_19_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(394269952)))]; + tensor var_4562_cast_fp16 = add(x = q_115_cast_fp16, y = encoder_layers_19_self_attn_pos_bias_v_to_fp16)[name = string("op_4562_cast_fp16")]; + tensor q_with_bias_v_39_perm_0 = const()[name = string("q_with_bias_v_39_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_501_transpose_x_0 = const()[name = string("x_501_transpose_x_0"), val = bool(false)]; + bool x_501_transpose_y_0 = const()[name = string("x_501_transpose_y_0"), val = bool(false)]; + tensor op_4564_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(394272064))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(394385792))))[name = string("op_4564_to_fp16_quantized")]; + tensor q_with_bias_v_39_cast_fp16 = transpose(perm = q_with_bias_v_39_perm_0, x = var_4562_cast_fp16)[name = string("transpose_191")]; + tensor x_501_cast_fp16 = matmul(transpose_x = x_501_transpose_x_0, transpose_y = x_501_transpose_y_0, x = q_with_bias_v_39_cast_fp16, y = op_4564_to_fp16_quantized)[name = string("x_501_cast_fp16")]; + tensor x_503_pad_0 = const()[name = string("x_503_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_503_mode_0 = const()[name = string("x_503_mode_0"), val = string("constant")]; + fp16 const_326_to_fp16 = const()[name = string("const_326_to_fp16"), val = fp16(0x0p+0)]; + tensor x_503_cast_fp16 = pad(constant_val = const_326_to_fp16, mode = x_503_mode_0, pad = x_503_pad_0, x = x_501_cast_fp16)[name = string("x_503_cast_fp16")]; + tensor var_4572 = const()[name = string("op_4572"), val = tensor([1, 8, -1, 14])]; + tensor x_505_cast_fp16 = reshape(shape = var_4572, x = x_503_cast_fp16)[name = string("x_505_cast_fp16")]; + tensor var_4576_begin_0 = const()[name = string("op_4576_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_4576_end_0 = const()[name = string("op_4576_end_0"), val = tensor([1, 8, 112, 14])]; + tensor var_4576_end_mask_0 = const()[name = string("op_4576_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_4576_cast_fp16 = slice_by_index(begin = var_4576_begin_0, end = var_4576_end_0, end_mask = var_4576_end_mask_0, x = x_505_cast_fp16)[name = string("op_4576_cast_fp16")]; + tensor var_4577 = const()[name = string("op_4577"), val = tensor([1, 8, 14, 111])]; + tensor matrix_bd_77_cast_fp16 = reshape(shape = var_4577, x = var_4576_cast_fp16)[name = string("matrix_bd_77_cast_fp16")]; + bool matrix_ac_39_transpose_x_0 = const()[name = string("matrix_ac_39_transpose_x_0"), val = bool(false)]; + bool matrix_ac_39_transpose_y_0 = const()[name = string("matrix_ac_39_transpose_y_0"), val = bool(false)]; + tensor transpose_134_perm_0 = const()[name = string("transpose_134_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_135_perm_0 = const()[name = string("transpose_135_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_135 = transpose(perm = transpose_135_perm_0, x = k_77_cast_fp16)[name = string("transpose_189")]; + tensor transpose_134 = transpose(perm = transpose_134_perm_0, x = var_4560_cast_fp16)[name = string("transpose_190")]; + tensor matrix_ac_39_cast_fp16 = matmul(transpose_x = matrix_ac_39_transpose_x_0, transpose_y = matrix_ac_39_transpose_y_0, x = transpose_134, y = transpose_135)[name = string("matrix_ac_39_cast_fp16")]; + tensor matrix_bd_79_begin_0 = const()[name = string("matrix_bd_79_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_79_end_0 = const()[name = string("matrix_bd_79_end_0"), val = tensor([1, 8, 14, 56])]; + tensor matrix_bd_79_end_mask_0 = const()[name = string("matrix_bd_79_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_79_cast_fp16 = slice_by_index(begin = matrix_bd_79_begin_0, end = matrix_bd_79_end_0, end_mask = matrix_bd_79_end_mask_0, x = matrix_bd_77_cast_fp16)[name = string("matrix_bd_79_cast_fp16")]; + tensor var_4586_cast_fp16 = add(x = matrix_ac_39_cast_fp16, y = matrix_bd_79_cast_fp16)[name = string("op_4586_cast_fp16")]; + fp16 _inversed_scores_77_y_0_to_fp16 = const()[name = string("_inversed_scores_77_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_77_cast_fp16 = mul(x = var_4586_cast_fp16, y = _inversed_scores_77_y_0_to_fp16)[name = string("_inversed_scores_77_cast_fp16")]; + tensor scores_79_cast_fp16 = select(a = var_46_to_fp16, b = _inversed_scores_77_cast_fp16, cond = mask_11)[name = string("scores_79_cast_fp16")]; + tensor var_4592_cast_fp16 = softmax(axis = var_60, x = scores_79_cast_fp16)[name = string("op_4592_cast_fp16")]; + tensor input_1029_cast_fp16 = select(a = var_45_to_fp16, b = var_4592_cast_fp16, cond = mask_11)[name = string("input_1029_cast_fp16")]; + bool x_507_transpose_x_0 = const()[name = string("x_507_transpose_x_0"), val = bool(false)]; + bool x_507_transpose_y_0 = const()[name = string("x_507_transpose_y_0"), val = bool(false)]; + tensor value_47_cast_fp16 = transpose(perm = value_47_perm_0, x = v_39_cast_fp16)[name = string("transpose_188")]; + tensor x_507_cast_fp16 = matmul(transpose_x = x_507_transpose_x_0, transpose_y = x_507_transpose_y_0, x = input_1029_cast_fp16, y = value_47_cast_fp16)[name = string("x_507_cast_fp16")]; + tensor var_4596_perm_0 = const()[name = string("op_4596_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_4597 = const()[name = string("op_4597"), val = tensor([1, -1, 1024])]; + tensor var_4596_cast_fp16 = transpose(perm = var_4596_perm_0, x = x_507_cast_fp16)[name = string("transpose_187")]; + tensor input_1031_cast_fp16 = reshape(shape = var_4597, x = var_4596_cast_fp16)[name = string("input_1031_cast_fp16")]; + tensor encoder_layers_19_self_attn_linear_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(394386112))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(395434752))))[name = string("encoder_layers_19_self_attn_linear_out_weight_to_fp16_quantized")]; + tensor encoder_layers_19_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_19_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(395436864)))]; + tensor linear_178_cast_fp16 = linear(bias = encoder_layers_19_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_19_self_attn_linear_out_weight_to_fp16_quantized, x = input_1031_cast_fp16)[name = string("linear_178_cast_fp16")]; + tensor input_1035_cast_fp16 = add(x = input_1025_cast_fp16, y = linear_178_cast_fp16)[name = string("input_1035_cast_fp16")]; + tensor x_511_axes_0 = const()[name = string("x_511_axes_0"), val = tensor([-1])]; + tensor encoder_layers_19_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_19_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(395438976)))]; + tensor encoder_layers_19_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_19_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(395441088)))]; + tensor x_511_cast_fp16 = layer_norm(axes = x_511_axes_0, beta = encoder_layers_19_norm_conv_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_19_norm_conv_weight_to_fp16, x = input_1035_cast_fp16)[name = string("x_511_cast_fp16")]; + tensor input_1037_perm_0 = const()[name = string("input_1037_perm_0"), val = tensor([0, 2, 1])]; + string input_1039_pad_type_0 = const()[name = string("input_1039_pad_type_0"), val = string("valid")]; + tensor input_1039_strides_0 = const()[name = string("input_1039_strides_0"), val = tensor([1])]; + tensor input_1039_pad_0 = const()[name = string("input_1039_pad_0"), val = tensor([0, 0])]; + tensor input_1039_dilations_0 = const()[name = string("input_1039_dilations_0"), val = tensor([1])]; + int32 input_1039_groups_0 = const()[name = string("input_1039_groups_0"), val = int32(1)]; + tensor encoder_layers_19_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(395443200))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397540416))))[name = string("encoder_layers_19_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_1037_cast_fp16 = transpose(perm = input_1037_perm_0, x = x_511_cast_fp16)[name = string("transpose_186")]; + tensor input_1039_cast_fp16 = conv(dilations = input_1039_dilations_0, groups = input_1039_groups_0, pad = input_1039_pad_0, pad_type = input_1039_pad_type_0, strides = input_1039_strides_0, weight = encoder_layers_19_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_1037_cast_fp16)[name = string("input_1039_cast_fp16")]; + int32 x_513_split_num_splits_0 = const()[name = string("x_513_split_num_splits_0"), val = int32(2)]; + int32 x_513_split_axis_0 = const()[name = string("x_513_split_axis_0"), val = int32(1)]; + tensor x_513_split_cast_fp16_0, tensor x_513_split_cast_fp16_1 = split(axis = x_513_split_axis_0, num_splits = x_513_split_num_splits_0, x = input_1039_cast_fp16)[name = string("x_513_split_cast_fp16")]; + tensor x_513_split_1_sigmoid_cast_fp16 = sigmoid(x = x_513_split_cast_fp16_1)[name = string("x_513_split_1_sigmoid_cast_fp16")]; + tensor x_513_cast_fp16 = mul(x = x_513_split_cast_fp16_0, y = x_513_split_1_sigmoid_cast_fp16)[name = string("x_513_cast_fp16")]; + tensor input_1041_cast_fp16 = select(a = var_45_to_fp16, b = x_513_cast_fp16, cond = var_576)[name = string("input_1041_cast_fp16")]; + bool new_x_79_interleave_0 = const()[name = string("new_x_79_interleave_0"), val = bool(false)]; + tensor new_x_79_cast_fp16 = concat(axis = var_60, interleave = new_x_79_interleave_0, values = (cache_79_cast_fp16, input_1041_cast_fp16))[name = string("new_x_79_cast_fp16")]; + tensor var_4636_begin_0 = const()[name = string("op_4636_begin_0"), val = tensor([0, 0, 14])]; + tensor var_4636_end_0 = const()[name = string("op_4636_end_0"), val = tensor([1, 1024, 22])]; + tensor var_4636_end_mask_0 = const()[name = string("op_4636_end_mask_0"), val = tensor([true, true, true])]; + tensor var_4636_cast_fp16 = slice_by_index(begin = var_4636_begin_0, end = var_4636_end_0, end_mask = var_4636_end_mask_0, x = new_x_79_cast_fp16)[name = string("op_4636_cast_fp16")]; + string x_515_pad_type_0 = const()[name = string("x_515_pad_type_0"), val = string("valid")]; + int32 x_515_groups_0 = const()[name = string("x_515_groups_0"), val = int32(1024)]; + tensor x_515_strides_0 = const()[name = string("x_515_strides_0"), val = tensor([1])]; + tensor x_515_pad_0 = const()[name = string("x_515_pad_0"), val = tensor([0, 0])]; + tensor x_515_dilations_0 = const()[name = string("x_515_dilations_0"), val = tensor([1])]; + tensor encoder_layers_19_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397544576))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397553856))))[name = string("encoder_layers_19_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_515_cast_fp16 = conv(dilations = x_515_dilations_0, groups = x_515_groups_0, pad = x_515_pad_0, pad_type = x_515_pad_type_0, strides = x_515_strides_0, weight = encoder_layers_19_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_79_cast_fp16)[name = string("x_515_cast_fp16")]; + tensor input_1043_perm_0 = const()[name = string("input_1043_perm_0"), val = tensor([0, 2, 1])]; + tensor x_517_axes_0 = const()[name = string("x_517_axes_0"), val = tensor([-1])]; + tensor encoder_layers_19_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_19_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397555968)))]; + tensor encoder_layers_19_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_19_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397558080)))]; + tensor input_1043_cast_fp16 = transpose(perm = input_1043_perm_0, x = x_515_cast_fp16)[name = string("transpose_185")]; + tensor x_517_cast_fp16 = layer_norm(axes = x_517_axes_0, beta = encoder_layers_19_conv_batch_norm_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_19_conv_batch_norm_weight_to_fp16, x = input_1043_cast_fp16)[name = string("x_517_cast_fp16")]; + tensor input_1045_perm_0 = const()[name = string("input_1045_perm_0"), val = tensor([0, 2, 1])]; + tensor input_1045_cast_fp16 = transpose(perm = input_1045_perm_0, x = x_517_cast_fp16)[name = string("transpose_184")]; + tensor input_1047_cast_fp16 = silu(x = input_1045_cast_fp16)[name = string("input_1047_cast_fp16")]; + string x_519_pad_type_0 = const()[name = string("x_519_pad_type_0"), val = string("valid")]; + tensor x_519_strides_0 = const()[name = string("x_519_strides_0"), val = tensor([1])]; + tensor x_519_pad_0 = const()[name = string("x_519_pad_0"), val = tensor([0, 0])]; + tensor x_519_dilations_0 = const()[name = string("x_519_dilations_0"), val = tensor([1])]; + int32 x_519_groups_0 = const()[name = string("x_519_groups_0"), val = int32(1)]; + tensor encoder_layers_19_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397560192))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(398608832))))[name = string("encoder_layers_19_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_519_cast_fp16 = conv(dilations = x_519_dilations_0, groups = x_519_groups_0, pad = x_519_pad_0, pad_type = x_519_pad_type_0, strides = x_519_strides_0, weight = encoder_layers_19_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_1047_cast_fp16)[name = string("x_519_cast_fp16")]; + tensor input_1049_perm_0 = const()[name = string("input_1049_perm_0"), val = tensor([0, 2, 1])]; + tensor input_1049_cast_fp16 = transpose(perm = input_1049_perm_0, x = x_519_cast_fp16)[name = string("transpose_183")]; + tensor input_1051_cast_fp16 = add(x = input_1035_cast_fp16, y = input_1049_cast_fp16)[name = string("input_1051_cast_fp16")]; + tensor input_1053_axes_0 = const()[name = string("input_1053_axes_0"), val = tensor([-1])]; + tensor encoder_layers_19_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_19_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(398610944)))]; + tensor encoder_layers_19_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_19_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(398613056)))]; + tensor input_1053_cast_fp16 = layer_norm(axes = input_1053_axes_0, beta = encoder_layers_19_norm_feed_forward2_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_19_norm_feed_forward2_weight_to_fp16, x = input_1051_cast_fp16)[name = string("input_1053_cast_fp16")]; + tensor encoder_layers_19_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(398615168))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(402809536))))[name = string("encoder_layers_19_feed_forward2_linear1_weight_to_fp16_quantized")]; + tensor encoder_layers_19_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_19_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(402817792)))]; + tensor linear_179_cast_fp16 = linear(bias = encoder_layers_19_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_19_feed_forward2_linear1_weight_to_fp16_quantized, x = input_1053_cast_fp16)[name = string("linear_179_cast_fp16")]; + tensor input_1057_cast_fp16 = silu(x = linear_179_cast_fp16)[name = string("input_1057_cast_fp16")]; + tensor encoder_layers_19_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(402826048))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407020416))))[name = string("encoder_layers_19_feed_forward2_linear2_weight_to_fp16_quantized")]; + tensor encoder_layers_19_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_19_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407022528)))]; + tensor linear_180_cast_fp16 = linear(bias = encoder_layers_19_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_19_feed_forward2_linear2_weight_to_fp16_quantized, x = input_1057_cast_fp16)[name = string("linear_180_cast_fp16")]; + fp16 var_4679_to_fp16 = const()[name = string("op_4679_to_fp16"), val = fp16(0x1p-1)]; + tensor var_4680_cast_fp16 = mul(x = linear_180_cast_fp16, y = var_4679_to_fp16)[name = string("op_4680_cast_fp16")]; + tensor input_1063_cast_fp16 = add(x = input_1051_cast_fp16, y = var_4680_cast_fp16)[name = string("input_1063_cast_fp16")]; + tensor input_1065_axes_0 = const()[name = string("input_1065_axes_0"), val = tensor([-1])]; + tensor encoder_layers_19_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_19_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407024640)))]; + tensor encoder_layers_19_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_19_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407026752)))]; + tensor input_1065_cast_fp16 = layer_norm(axes = input_1065_axes_0, beta = encoder_layers_19_norm_out_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_19_norm_out_weight_to_fp16, x = input_1063_cast_fp16)[name = string("input_1065_cast_fp16")]; + tensor cache_81_begin_0 = const()[name = string("cache_81_begin_0"), val = tensor([20, 0, 0, 0])]; + tensor cache_81_end_0 = const()[name = string("cache_81_end_0"), val = tensor([21, 1, 42, 1024])]; + tensor cache_81_end_mask_0 = const()[name = string("cache_81_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_81_squeeze_mask_0 = const()[name = string("cache_81_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_81_cast_fp16 = slice_by_index(begin = cache_81_begin_0, end = cache_81_end_0, end_mask = cache_81_end_mask_0, squeeze_mask = cache_81_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_81_cast_fp16")]; + tensor cache_83_begin_0 = const()[name = string("cache_83_begin_0"), val = tensor([20, 0, 0, 0])]; + tensor cache_83_end_0 = const()[name = string("cache_83_end_0"), val = tensor([21, 1, 1024, 8])]; + tensor cache_83_end_mask_0 = const()[name = string("cache_83_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_83_squeeze_mask_0 = const()[name = string("cache_83_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_83_cast_fp16 = slice_by_index(begin = cache_83_begin_0, end = cache_83_end_0, end_mask = cache_83_end_mask_0, squeeze_mask = cache_83_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_83_cast_fp16")]; + tensor input_1067_axes_0 = const()[name = string("input_1067_axes_0"), val = tensor([-1])]; + tensor encoder_layers_20_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_20_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407028864)))]; + tensor encoder_layers_20_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_20_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407030976)))]; + tensor input_1067_cast_fp16 = layer_norm(axes = input_1067_axes_0, beta = encoder_layers_20_norm_feed_forward1_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_20_norm_feed_forward1_weight_to_fp16, x = input_1065_cast_fp16)[name = string("input_1067_cast_fp16")]; + tensor encoder_layers_20_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407033088))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(411227456))))[name = string("encoder_layers_20_feed_forward1_linear1_weight_to_fp16_quantized")]; + tensor encoder_layers_20_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_20_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(411235712)))]; + tensor linear_181_cast_fp16 = linear(bias = encoder_layers_20_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_20_feed_forward1_linear1_weight_to_fp16_quantized, x = input_1067_cast_fp16)[name = string("linear_181_cast_fp16")]; + tensor input_1071_cast_fp16 = silu(x = linear_181_cast_fp16)[name = string("input_1071_cast_fp16")]; + tensor encoder_layers_20_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(411243968))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(415438336))))[name = string("encoder_layers_20_feed_forward1_linear2_weight_to_fp16_quantized")]; + tensor encoder_layers_20_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_20_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(415440448)))]; + tensor linear_182_cast_fp16 = linear(bias = encoder_layers_20_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_20_feed_forward1_linear2_weight_to_fp16_quantized, x = input_1071_cast_fp16)[name = string("linear_182_cast_fp16")]; + fp16 var_4716_to_fp16 = const()[name = string("op_4716_to_fp16"), val = fp16(0x1p-1)]; + tensor var_4717_cast_fp16 = mul(x = linear_182_cast_fp16, y = var_4716_to_fp16)[name = string("op_4717_cast_fp16")]; + tensor input_1077_cast_fp16 = add(x = input_1065_cast_fp16, y = var_4717_cast_fp16)[name = string("input_1077_cast_fp16")]; + tensor key_41_axes_0 = const()[name = string("key_41_axes_0"), val = tensor([-1])]; + tensor encoder_layers_20_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_20_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(415442560)))]; + tensor encoder_layers_20_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_20_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(415444672)))]; + tensor key_41_cast_fp16 = layer_norm(axes = key_41_axes_0, beta = encoder_layers_20_norm_self_att_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_20_norm_self_att_weight_to_fp16, x = input_1077_cast_fp16)[name = string("key_41_cast_fp16")]; + bool input_1079_interleave_0 = const()[name = string("input_1079_interleave_0"), val = bool(false)]; + tensor input_1079_cast_fp16 = concat(axis = var_69, interleave = input_1079_interleave_0, values = (cache_81_cast_fp16, key_41_cast_fp16))[name = string("input_1079_cast_fp16")]; + tensor var_4739_begin_0 = const()[name = string("op_4739_begin_0"), val = tensor([0, 14, 0])]; + tensor var_4739_end_0 = const()[name = string("op_4739_end_0"), val = tensor([1, 42, 1024])]; + tensor var_4739_end_mask_0 = const()[name = string("op_4739_end_mask_0"), val = tensor([true, true, true])]; + tensor var_4739_cast_fp16 = slice_by_index(begin = var_4739_begin_0, end = var_4739_end_0, end_mask = var_4739_end_mask_0, x = cache_81_cast_fp16)[name = string("op_4739_cast_fp16")]; + bool var_4745_interleave_0 = const()[name = string("op_4745_interleave_0"), val = bool(false)]; + tensor var_4745_cast_fp16 = concat(axis = var_69, interleave = var_4745_interleave_0, values = (var_4739_cast_fp16, key_41_cast_fp16))[name = string("op_4745_cast_fp16")]; + tensor encoder_layers_20_self_attn_linear_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(415446784))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(416495424))))[name = string("encoder_layers_20_self_attn_linear_q_weight_to_fp16_quantized")]; + tensor encoder_layers_20_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_20_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(416497536)))]; + tensor linear_183_cast_fp16 = linear(bias = encoder_layers_20_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_20_self_attn_linear_q_weight_to_fp16_quantized, x = key_41_cast_fp16)[name = string("linear_183_cast_fp16")]; + tensor var_4750 = const()[name = string("op_4750"), val = tensor([1, -1, 8, 128])]; + tensor q_121_cast_fp16 = reshape(shape = var_4750, x = linear_183_cast_fp16)[name = string("q_121_cast_fp16")]; + tensor encoder_layers_20_self_attn_linear_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(416499648))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(417548288))))[name = string("encoder_layers_20_self_attn_linear_k_weight_to_fp16_quantized")]; + tensor encoder_layers_20_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_20_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(417550400)))]; + tensor linear_184_cast_fp16 = linear(bias = encoder_layers_20_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_20_self_attn_linear_k_weight_to_fp16_quantized, x = input_1079_cast_fp16)[name = string("linear_184_cast_fp16")]; + tensor var_4755 = const()[name = string("op_4755"), val = tensor([1, -1, 8, 128])]; + tensor k_81_cast_fp16 = reshape(shape = var_4755, x = linear_184_cast_fp16)[name = string("k_81_cast_fp16")]; + tensor encoder_layers_20_self_attn_linear_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(417552512))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(418601152))))[name = string("encoder_layers_20_self_attn_linear_v_weight_to_fp16_quantized")]; + tensor encoder_layers_20_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_20_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(418603264)))]; + tensor linear_185_cast_fp16 = linear(bias = encoder_layers_20_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_20_self_attn_linear_v_weight_to_fp16_quantized, x = input_1079_cast_fp16)[name = string("linear_185_cast_fp16")]; + tensor var_4760 = const()[name = string("op_4760"), val = tensor([1, -1, 8, 128])]; + tensor v_41_cast_fp16 = reshape(shape = var_4760, x = linear_185_cast_fp16)[name = string("v_41_cast_fp16")]; + tensor value_49_perm_0 = const()[name = string("value_49_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_20_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_20_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(418605376)))]; + tensor var_4773_cast_fp16 = add(x = q_121_cast_fp16, y = encoder_layers_20_self_attn_pos_bias_u_to_fp16)[name = string("op_4773_cast_fp16")]; + tensor encoder_layers_20_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_20_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(418607488)))]; + tensor var_4775_cast_fp16 = add(x = q_121_cast_fp16, y = encoder_layers_20_self_attn_pos_bias_v_to_fp16)[name = string("op_4775_cast_fp16")]; + tensor q_with_bias_v_41_perm_0 = const()[name = string("q_with_bias_v_41_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_527_transpose_x_0 = const()[name = string("x_527_transpose_x_0"), val = bool(false)]; + bool x_527_transpose_y_0 = const()[name = string("x_527_transpose_y_0"), val = bool(false)]; + tensor op_4777_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(418609600))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(418723328))))[name = string("op_4777_to_fp16_quantized")]; + tensor q_with_bias_v_41_cast_fp16 = transpose(perm = q_with_bias_v_41_perm_0, x = var_4775_cast_fp16)[name = string("transpose_182")]; + tensor x_527_cast_fp16 = matmul(transpose_x = x_527_transpose_x_0, transpose_y = x_527_transpose_y_0, x = q_with_bias_v_41_cast_fp16, y = op_4777_to_fp16_quantized)[name = string("x_527_cast_fp16")]; + tensor x_529_pad_0 = const()[name = string("x_529_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_529_mode_0 = const()[name = string("x_529_mode_0"), val = string("constant")]; + fp16 const_339_to_fp16 = const()[name = string("const_339_to_fp16"), val = fp16(0x0p+0)]; + tensor x_529_cast_fp16 = pad(constant_val = const_339_to_fp16, mode = x_529_mode_0, pad = x_529_pad_0, x = x_527_cast_fp16)[name = string("x_529_cast_fp16")]; + tensor var_4785 = const()[name = string("op_4785"), val = tensor([1, 8, -1, 14])]; + tensor x_531_cast_fp16 = reshape(shape = var_4785, x = x_529_cast_fp16)[name = string("x_531_cast_fp16")]; + tensor var_4789_begin_0 = const()[name = string("op_4789_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_4789_end_0 = const()[name = string("op_4789_end_0"), val = tensor([1, 8, 112, 14])]; + tensor var_4789_end_mask_0 = const()[name = string("op_4789_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_4789_cast_fp16 = slice_by_index(begin = var_4789_begin_0, end = var_4789_end_0, end_mask = var_4789_end_mask_0, x = x_531_cast_fp16)[name = string("op_4789_cast_fp16")]; + tensor var_4790 = const()[name = string("op_4790"), val = tensor([1, 8, 14, 111])]; + tensor matrix_bd_81_cast_fp16 = reshape(shape = var_4790, x = var_4789_cast_fp16)[name = string("matrix_bd_81_cast_fp16")]; + bool matrix_ac_41_transpose_x_0 = const()[name = string("matrix_ac_41_transpose_x_0"), val = bool(false)]; + bool matrix_ac_41_transpose_y_0 = const()[name = string("matrix_ac_41_transpose_y_0"), val = bool(false)]; + tensor transpose_136_perm_0 = const()[name = string("transpose_136_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_137_perm_0 = const()[name = string("transpose_137_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_137 = transpose(perm = transpose_137_perm_0, x = k_81_cast_fp16)[name = string("transpose_180")]; + tensor transpose_136 = transpose(perm = transpose_136_perm_0, x = var_4773_cast_fp16)[name = string("transpose_181")]; + tensor matrix_ac_41_cast_fp16 = matmul(transpose_x = matrix_ac_41_transpose_x_0, transpose_y = matrix_ac_41_transpose_y_0, x = transpose_136, y = transpose_137)[name = string("matrix_ac_41_cast_fp16")]; + tensor matrix_bd_83_begin_0 = const()[name = string("matrix_bd_83_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_83_end_0 = const()[name = string("matrix_bd_83_end_0"), val = tensor([1, 8, 14, 56])]; + tensor matrix_bd_83_end_mask_0 = const()[name = string("matrix_bd_83_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_83_cast_fp16 = slice_by_index(begin = matrix_bd_83_begin_0, end = matrix_bd_83_end_0, end_mask = matrix_bd_83_end_mask_0, x = matrix_bd_81_cast_fp16)[name = string("matrix_bd_83_cast_fp16")]; + tensor var_4799_cast_fp16 = add(x = matrix_ac_41_cast_fp16, y = matrix_bd_83_cast_fp16)[name = string("op_4799_cast_fp16")]; + fp16 _inversed_scores_81_y_0_to_fp16 = const()[name = string("_inversed_scores_81_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_81_cast_fp16 = mul(x = var_4799_cast_fp16, y = _inversed_scores_81_y_0_to_fp16)[name = string("_inversed_scores_81_cast_fp16")]; + tensor scores_83_cast_fp16 = select(a = var_46_to_fp16, b = _inversed_scores_81_cast_fp16, cond = mask_11)[name = string("scores_83_cast_fp16")]; + tensor var_4805_cast_fp16 = softmax(axis = var_60, x = scores_83_cast_fp16)[name = string("op_4805_cast_fp16")]; + tensor input_1081_cast_fp16 = select(a = var_45_to_fp16, b = var_4805_cast_fp16, cond = mask_11)[name = string("input_1081_cast_fp16")]; + bool x_533_transpose_x_0 = const()[name = string("x_533_transpose_x_0"), val = bool(false)]; + bool x_533_transpose_y_0 = const()[name = string("x_533_transpose_y_0"), val = bool(false)]; + tensor value_49_cast_fp16 = transpose(perm = value_49_perm_0, x = v_41_cast_fp16)[name = string("transpose_179")]; + tensor x_533_cast_fp16 = matmul(transpose_x = x_533_transpose_x_0, transpose_y = x_533_transpose_y_0, x = input_1081_cast_fp16, y = value_49_cast_fp16)[name = string("x_533_cast_fp16")]; + tensor var_4809_perm_0 = const()[name = string("op_4809_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_4810 = const()[name = string("op_4810"), val = tensor([1, -1, 1024])]; + tensor var_4809_cast_fp16 = transpose(perm = var_4809_perm_0, x = x_533_cast_fp16)[name = string("transpose_178")]; + tensor input_1083_cast_fp16 = reshape(shape = var_4810, x = var_4809_cast_fp16)[name = string("input_1083_cast_fp16")]; + tensor encoder_layers_20_self_attn_linear_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(418723648))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(419772288))))[name = string("encoder_layers_20_self_attn_linear_out_weight_to_fp16_quantized")]; + tensor encoder_layers_20_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_20_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(419774400)))]; + tensor linear_187_cast_fp16 = linear(bias = encoder_layers_20_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_20_self_attn_linear_out_weight_to_fp16_quantized, x = input_1083_cast_fp16)[name = string("linear_187_cast_fp16")]; + tensor input_1087_cast_fp16 = add(x = input_1077_cast_fp16, y = linear_187_cast_fp16)[name = string("input_1087_cast_fp16")]; + tensor x_537_axes_0 = const()[name = string("x_537_axes_0"), val = tensor([-1])]; + tensor encoder_layers_20_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_20_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(419776512)))]; + tensor encoder_layers_20_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_20_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(419778624)))]; + tensor x_537_cast_fp16 = layer_norm(axes = x_537_axes_0, beta = encoder_layers_20_norm_conv_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_20_norm_conv_weight_to_fp16, x = input_1087_cast_fp16)[name = string("x_537_cast_fp16")]; + tensor input_1089_perm_0 = const()[name = string("input_1089_perm_0"), val = tensor([0, 2, 1])]; + string input_1091_pad_type_0 = const()[name = string("input_1091_pad_type_0"), val = string("valid")]; + tensor input_1091_strides_0 = const()[name = string("input_1091_strides_0"), val = tensor([1])]; + tensor input_1091_pad_0 = const()[name = string("input_1091_pad_0"), val = tensor([0, 0])]; + tensor input_1091_dilations_0 = const()[name = string("input_1091_dilations_0"), val = tensor([1])]; + int32 input_1091_groups_0 = const()[name = string("input_1091_groups_0"), val = int32(1)]; + tensor encoder_layers_20_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(419780736))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421877952))))[name = string("encoder_layers_20_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_1089_cast_fp16 = transpose(perm = input_1089_perm_0, x = x_537_cast_fp16)[name = string("transpose_177")]; + tensor input_1091_cast_fp16 = conv(dilations = input_1091_dilations_0, groups = input_1091_groups_0, pad = input_1091_pad_0, pad_type = input_1091_pad_type_0, strides = input_1091_strides_0, weight = encoder_layers_20_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_1089_cast_fp16)[name = string("input_1091_cast_fp16")]; + int32 x_539_split_num_splits_0 = const()[name = string("x_539_split_num_splits_0"), val = int32(2)]; + int32 x_539_split_axis_0 = const()[name = string("x_539_split_axis_0"), val = int32(1)]; + tensor x_539_split_cast_fp16_0, tensor x_539_split_cast_fp16_1 = split(axis = x_539_split_axis_0, num_splits = x_539_split_num_splits_0, x = input_1091_cast_fp16)[name = string("x_539_split_cast_fp16")]; + tensor x_539_split_1_sigmoid_cast_fp16 = sigmoid(x = x_539_split_cast_fp16_1)[name = string("x_539_split_1_sigmoid_cast_fp16")]; + tensor x_539_cast_fp16 = mul(x = x_539_split_cast_fp16_0, y = x_539_split_1_sigmoid_cast_fp16)[name = string("x_539_cast_fp16")]; + tensor input_1093_cast_fp16 = select(a = var_45_to_fp16, b = x_539_cast_fp16, cond = var_576)[name = string("input_1093_cast_fp16")]; + bool new_x_83_interleave_0 = const()[name = string("new_x_83_interleave_0"), val = bool(false)]; + tensor new_x_83_cast_fp16 = concat(axis = var_60, interleave = new_x_83_interleave_0, values = (cache_83_cast_fp16, input_1093_cast_fp16))[name = string("new_x_83_cast_fp16")]; + tensor var_4849_begin_0 = const()[name = string("op_4849_begin_0"), val = tensor([0, 0, 14])]; + tensor var_4849_end_0 = const()[name = string("op_4849_end_0"), val = tensor([1, 1024, 22])]; + tensor var_4849_end_mask_0 = const()[name = string("op_4849_end_mask_0"), val = tensor([true, true, true])]; + tensor var_4849_cast_fp16 = slice_by_index(begin = var_4849_begin_0, end = var_4849_end_0, end_mask = var_4849_end_mask_0, x = new_x_83_cast_fp16)[name = string("op_4849_cast_fp16")]; + string x_541_pad_type_0 = const()[name = string("x_541_pad_type_0"), val = string("valid")]; + int32 x_541_groups_0 = const()[name = string("x_541_groups_0"), val = int32(1024)]; + tensor x_541_strides_0 = const()[name = string("x_541_strides_0"), val = tensor([1])]; + tensor x_541_pad_0 = const()[name = string("x_541_pad_0"), val = tensor([0, 0])]; + tensor x_541_dilations_0 = const()[name = string("x_541_dilations_0"), val = tensor([1])]; + tensor encoder_layers_20_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421882112))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421891392))))[name = string("encoder_layers_20_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_541_cast_fp16 = conv(dilations = x_541_dilations_0, groups = x_541_groups_0, pad = x_541_pad_0, pad_type = x_541_pad_type_0, strides = x_541_strides_0, weight = encoder_layers_20_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_83_cast_fp16)[name = string("x_541_cast_fp16")]; + tensor input_1095_perm_0 = const()[name = string("input_1095_perm_0"), val = tensor([0, 2, 1])]; + tensor x_543_axes_0 = const()[name = string("x_543_axes_0"), val = tensor([-1])]; + tensor encoder_layers_20_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_20_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421893504)))]; + tensor encoder_layers_20_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_20_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421895616)))]; + tensor input_1095_cast_fp16 = transpose(perm = input_1095_perm_0, x = x_541_cast_fp16)[name = string("transpose_176")]; + tensor x_543_cast_fp16 = layer_norm(axes = x_543_axes_0, beta = encoder_layers_20_conv_batch_norm_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_20_conv_batch_norm_weight_to_fp16, x = input_1095_cast_fp16)[name = string("x_543_cast_fp16")]; + tensor input_1097_perm_0 = const()[name = string("input_1097_perm_0"), val = tensor([0, 2, 1])]; + tensor input_1097_cast_fp16 = transpose(perm = input_1097_perm_0, x = x_543_cast_fp16)[name = string("transpose_175")]; + tensor input_1099_cast_fp16 = silu(x = input_1097_cast_fp16)[name = string("input_1099_cast_fp16")]; + string x_545_pad_type_0 = const()[name = string("x_545_pad_type_0"), val = string("valid")]; + tensor x_545_strides_0 = const()[name = string("x_545_strides_0"), val = tensor([1])]; + tensor x_545_pad_0 = const()[name = string("x_545_pad_0"), val = tensor([0, 0])]; + tensor x_545_dilations_0 = const()[name = string("x_545_dilations_0"), val = tensor([1])]; + int32 x_545_groups_0 = const()[name = string("x_545_groups_0"), val = int32(1)]; + tensor encoder_layers_20_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421897728))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(422946368))))[name = string("encoder_layers_20_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_545_cast_fp16 = conv(dilations = x_545_dilations_0, groups = x_545_groups_0, pad = x_545_pad_0, pad_type = x_545_pad_type_0, strides = x_545_strides_0, weight = encoder_layers_20_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_1099_cast_fp16)[name = string("x_545_cast_fp16")]; + tensor input_1101_perm_0 = const()[name = string("input_1101_perm_0"), val = tensor([0, 2, 1])]; + tensor input_1101_cast_fp16 = transpose(perm = input_1101_perm_0, x = x_545_cast_fp16)[name = string("transpose_174")]; + tensor input_1103_cast_fp16 = add(x = input_1087_cast_fp16, y = input_1101_cast_fp16)[name = string("input_1103_cast_fp16")]; + tensor input_1105_axes_0 = const()[name = string("input_1105_axes_0"), val = tensor([-1])]; + tensor encoder_layers_20_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_20_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(422948480)))]; + tensor encoder_layers_20_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_20_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(422950592)))]; + tensor input_1105_cast_fp16 = layer_norm(axes = input_1105_axes_0, beta = encoder_layers_20_norm_feed_forward2_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_20_norm_feed_forward2_weight_to_fp16, x = input_1103_cast_fp16)[name = string("input_1105_cast_fp16")]; + tensor encoder_layers_20_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(422952704))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427147072))))[name = string("encoder_layers_20_feed_forward2_linear1_weight_to_fp16_quantized")]; + tensor encoder_layers_20_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_20_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427155328)))]; + tensor linear_188_cast_fp16 = linear(bias = encoder_layers_20_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_20_feed_forward2_linear1_weight_to_fp16_quantized, x = input_1105_cast_fp16)[name = string("linear_188_cast_fp16")]; + tensor input_1109_cast_fp16 = silu(x = linear_188_cast_fp16)[name = string("input_1109_cast_fp16")]; + tensor encoder_layers_20_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427163584))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(431357952))))[name = string("encoder_layers_20_feed_forward2_linear2_weight_to_fp16_quantized")]; + tensor encoder_layers_20_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_20_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(431360064)))]; + tensor linear_189_cast_fp16 = linear(bias = encoder_layers_20_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_20_feed_forward2_linear2_weight_to_fp16_quantized, x = input_1109_cast_fp16)[name = string("linear_189_cast_fp16")]; + fp16 var_4892_to_fp16 = const()[name = string("op_4892_to_fp16"), val = fp16(0x1p-1)]; + tensor var_4893_cast_fp16 = mul(x = linear_189_cast_fp16, y = var_4892_to_fp16)[name = string("op_4893_cast_fp16")]; + tensor input_1115_cast_fp16 = add(x = input_1103_cast_fp16, y = var_4893_cast_fp16)[name = string("input_1115_cast_fp16")]; + tensor input_1117_axes_0 = const()[name = string("input_1117_axes_0"), val = tensor([-1])]; + tensor encoder_layers_20_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_20_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(431362176)))]; + tensor encoder_layers_20_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_20_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(431364288)))]; + tensor input_1117_cast_fp16 = layer_norm(axes = input_1117_axes_0, beta = encoder_layers_20_norm_out_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_20_norm_out_weight_to_fp16, x = input_1115_cast_fp16)[name = string("input_1117_cast_fp16")]; + tensor cache_85_begin_0 = const()[name = string("cache_85_begin_0"), val = tensor([21, 0, 0, 0])]; + tensor cache_85_end_0 = const()[name = string("cache_85_end_0"), val = tensor([22, 1, 42, 1024])]; + tensor cache_85_end_mask_0 = const()[name = string("cache_85_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_85_squeeze_mask_0 = const()[name = string("cache_85_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_85_cast_fp16 = slice_by_index(begin = cache_85_begin_0, end = cache_85_end_0, end_mask = cache_85_end_mask_0, squeeze_mask = cache_85_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_85_cast_fp16")]; + tensor cache_87_begin_0 = const()[name = string("cache_87_begin_0"), val = tensor([21, 0, 0, 0])]; + tensor cache_87_end_0 = const()[name = string("cache_87_end_0"), val = tensor([22, 1, 1024, 8])]; + tensor cache_87_end_mask_0 = const()[name = string("cache_87_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_87_squeeze_mask_0 = const()[name = string("cache_87_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_87_cast_fp16 = slice_by_index(begin = cache_87_begin_0, end = cache_87_end_0, end_mask = cache_87_end_mask_0, squeeze_mask = cache_87_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_87_cast_fp16")]; + tensor input_1119_axes_0 = const()[name = string("input_1119_axes_0"), val = tensor([-1])]; + tensor encoder_layers_21_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_21_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(431366400)))]; + tensor encoder_layers_21_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_21_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(431368512)))]; + tensor input_1119_cast_fp16 = layer_norm(axes = input_1119_axes_0, beta = encoder_layers_21_norm_feed_forward1_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_21_norm_feed_forward1_weight_to_fp16, x = input_1117_cast_fp16)[name = string("input_1119_cast_fp16")]; + tensor encoder_layers_21_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(431370624))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(435564992))))[name = string("encoder_layers_21_feed_forward1_linear1_weight_to_fp16_quantized")]; + tensor encoder_layers_21_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_21_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(435573248)))]; + tensor linear_190_cast_fp16 = linear(bias = encoder_layers_21_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_21_feed_forward1_linear1_weight_to_fp16_quantized, x = input_1119_cast_fp16)[name = string("linear_190_cast_fp16")]; + tensor input_1123_cast_fp16 = silu(x = linear_190_cast_fp16)[name = string("input_1123_cast_fp16")]; + tensor encoder_layers_21_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(435581504))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(439775872))))[name = string("encoder_layers_21_feed_forward1_linear2_weight_to_fp16_quantized")]; + tensor encoder_layers_21_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_21_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(439777984)))]; + tensor linear_191_cast_fp16 = linear(bias = encoder_layers_21_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_21_feed_forward1_linear2_weight_to_fp16_quantized, x = input_1123_cast_fp16)[name = string("linear_191_cast_fp16")]; + fp16 var_4929_to_fp16 = const()[name = string("op_4929_to_fp16"), val = fp16(0x1p-1)]; + tensor var_4930_cast_fp16 = mul(x = linear_191_cast_fp16, y = var_4929_to_fp16)[name = string("op_4930_cast_fp16")]; + tensor input_1129_cast_fp16 = add(x = input_1117_cast_fp16, y = var_4930_cast_fp16)[name = string("input_1129_cast_fp16")]; + tensor key_43_axes_0 = const()[name = string("key_43_axes_0"), val = tensor([-1])]; + tensor encoder_layers_21_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_21_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(439780096)))]; + tensor encoder_layers_21_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_21_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(439782208)))]; + tensor key_43_cast_fp16 = layer_norm(axes = key_43_axes_0, beta = encoder_layers_21_norm_self_att_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_21_norm_self_att_weight_to_fp16, x = input_1129_cast_fp16)[name = string("key_43_cast_fp16")]; + bool input_1131_interleave_0 = const()[name = string("input_1131_interleave_0"), val = bool(false)]; + tensor input_1131_cast_fp16 = concat(axis = var_69, interleave = input_1131_interleave_0, values = (cache_85_cast_fp16, key_43_cast_fp16))[name = string("input_1131_cast_fp16")]; + tensor var_4952_begin_0 = const()[name = string("op_4952_begin_0"), val = tensor([0, 14, 0])]; + tensor var_4952_end_0 = const()[name = string("op_4952_end_0"), val = tensor([1, 42, 1024])]; + tensor var_4952_end_mask_0 = const()[name = string("op_4952_end_mask_0"), val = tensor([true, true, true])]; + tensor var_4952_cast_fp16 = slice_by_index(begin = var_4952_begin_0, end = var_4952_end_0, end_mask = var_4952_end_mask_0, x = cache_85_cast_fp16)[name = string("op_4952_cast_fp16")]; + bool var_4958_interleave_0 = const()[name = string("op_4958_interleave_0"), val = bool(false)]; + tensor var_4958_cast_fp16 = concat(axis = var_69, interleave = var_4958_interleave_0, values = (var_4952_cast_fp16, key_43_cast_fp16))[name = string("op_4958_cast_fp16")]; + tensor encoder_layers_21_self_attn_linear_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(439784320))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(440832960))))[name = string("encoder_layers_21_self_attn_linear_q_weight_to_fp16_quantized")]; + tensor encoder_layers_21_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_21_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(440835072)))]; + tensor linear_192_cast_fp16 = linear(bias = encoder_layers_21_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_21_self_attn_linear_q_weight_to_fp16_quantized, x = key_43_cast_fp16)[name = string("linear_192_cast_fp16")]; + tensor var_4963 = const()[name = string("op_4963"), val = tensor([1, -1, 8, 128])]; + tensor q_127_cast_fp16 = reshape(shape = var_4963, x = linear_192_cast_fp16)[name = string("q_127_cast_fp16")]; + tensor encoder_layers_21_self_attn_linear_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(440837184))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(441885824))))[name = string("encoder_layers_21_self_attn_linear_k_weight_to_fp16_quantized")]; + tensor encoder_layers_21_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_21_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(441887936)))]; + tensor linear_193_cast_fp16 = linear(bias = encoder_layers_21_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_21_self_attn_linear_k_weight_to_fp16_quantized, x = input_1131_cast_fp16)[name = string("linear_193_cast_fp16")]; + tensor var_4968 = const()[name = string("op_4968"), val = tensor([1, -1, 8, 128])]; + tensor k_85_cast_fp16 = reshape(shape = var_4968, x = linear_193_cast_fp16)[name = string("k_85_cast_fp16")]; + tensor encoder_layers_21_self_attn_linear_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(441890048))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(442938688))))[name = string("encoder_layers_21_self_attn_linear_v_weight_to_fp16_quantized")]; + tensor encoder_layers_21_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_21_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(442940800)))]; + tensor linear_194_cast_fp16 = linear(bias = encoder_layers_21_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_21_self_attn_linear_v_weight_to_fp16_quantized, x = input_1131_cast_fp16)[name = string("linear_194_cast_fp16")]; + tensor var_4973 = const()[name = string("op_4973"), val = tensor([1, -1, 8, 128])]; + tensor v_43_cast_fp16 = reshape(shape = var_4973, x = linear_194_cast_fp16)[name = string("v_43_cast_fp16")]; + tensor value_51_perm_0 = const()[name = string("value_51_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_21_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_21_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(442942912)))]; + tensor var_4986_cast_fp16 = add(x = q_127_cast_fp16, y = encoder_layers_21_self_attn_pos_bias_u_to_fp16)[name = string("op_4986_cast_fp16")]; + tensor encoder_layers_21_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_21_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(442945024)))]; + tensor var_4988_cast_fp16 = add(x = q_127_cast_fp16, y = encoder_layers_21_self_attn_pos_bias_v_to_fp16)[name = string("op_4988_cast_fp16")]; + tensor q_with_bias_v_43_perm_0 = const()[name = string("q_with_bias_v_43_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_553_transpose_x_0 = const()[name = string("x_553_transpose_x_0"), val = bool(false)]; + bool x_553_transpose_y_0 = const()[name = string("x_553_transpose_y_0"), val = bool(false)]; + tensor op_4990_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(442947136))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(443060864))))[name = string("op_4990_to_fp16_quantized")]; + tensor q_with_bias_v_43_cast_fp16 = transpose(perm = q_with_bias_v_43_perm_0, x = var_4988_cast_fp16)[name = string("transpose_173")]; + tensor x_553_cast_fp16 = matmul(transpose_x = x_553_transpose_x_0, transpose_y = x_553_transpose_y_0, x = q_with_bias_v_43_cast_fp16, y = op_4990_to_fp16_quantized)[name = string("x_553_cast_fp16")]; + tensor x_555_pad_0 = const()[name = string("x_555_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_555_mode_0 = const()[name = string("x_555_mode_0"), val = string("constant")]; + fp16 const_352_to_fp16 = const()[name = string("const_352_to_fp16"), val = fp16(0x0p+0)]; + tensor x_555_cast_fp16 = pad(constant_val = const_352_to_fp16, mode = x_555_mode_0, pad = x_555_pad_0, x = x_553_cast_fp16)[name = string("x_555_cast_fp16")]; + tensor var_4998 = const()[name = string("op_4998"), val = tensor([1, 8, -1, 14])]; + tensor x_557_cast_fp16 = reshape(shape = var_4998, x = x_555_cast_fp16)[name = string("x_557_cast_fp16")]; + tensor var_5002_begin_0 = const()[name = string("op_5002_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_5002_end_0 = const()[name = string("op_5002_end_0"), val = tensor([1, 8, 112, 14])]; + tensor var_5002_end_mask_0 = const()[name = string("op_5002_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_5002_cast_fp16 = slice_by_index(begin = var_5002_begin_0, end = var_5002_end_0, end_mask = var_5002_end_mask_0, x = x_557_cast_fp16)[name = string("op_5002_cast_fp16")]; + tensor var_5003 = const()[name = string("op_5003"), val = tensor([1, 8, 14, 111])]; + tensor matrix_bd_85_cast_fp16 = reshape(shape = var_5003, x = var_5002_cast_fp16)[name = string("matrix_bd_85_cast_fp16")]; + bool matrix_ac_43_transpose_x_0 = const()[name = string("matrix_ac_43_transpose_x_0"), val = bool(false)]; + bool matrix_ac_43_transpose_y_0 = const()[name = string("matrix_ac_43_transpose_y_0"), val = bool(false)]; + tensor transpose_138_perm_0 = const()[name = string("transpose_138_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_139_perm_0 = const()[name = string("transpose_139_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_139 = transpose(perm = transpose_139_perm_0, x = k_85_cast_fp16)[name = string("transpose_171")]; + tensor transpose_138 = transpose(perm = transpose_138_perm_0, x = var_4986_cast_fp16)[name = string("transpose_172")]; + tensor matrix_ac_43_cast_fp16 = matmul(transpose_x = matrix_ac_43_transpose_x_0, transpose_y = matrix_ac_43_transpose_y_0, x = transpose_138, y = transpose_139)[name = string("matrix_ac_43_cast_fp16")]; + tensor matrix_bd_87_begin_0 = const()[name = string("matrix_bd_87_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_87_end_0 = const()[name = string("matrix_bd_87_end_0"), val = tensor([1, 8, 14, 56])]; + tensor matrix_bd_87_end_mask_0 = const()[name = string("matrix_bd_87_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_87_cast_fp16 = slice_by_index(begin = matrix_bd_87_begin_0, end = matrix_bd_87_end_0, end_mask = matrix_bd_87_end_mask_0, x = matrix_bd_85_cast_fp16)[name = string("matrix_bd_87_cast_fp16")]; + tensor var_5012_cast_fp16 = add(x = matrix_ac_43_cast_fp16, y = matrix_bd_87_cast_fp16)[name = string("op_5012_cast_fp16")]; + fp16 _inversed_scores_85_y_0_to_fp16 = const()[name = string("_inversed_scores_85_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_85_cast_fp16 = mul(x = var_5012_cast_fp16, y = _inversed_scores_85_y_0_to_fp16)[name = string("_inversed_scores_85_cast_fp16")]; + tensor scores_87_cast_fp16 = select(a = var_46_to_fp16, b = _inversed_scores_85_cast_fp16, cond = mask_11)[name = string("scores_87_cast_fp16")]; + tensor var_5018_cast_fp16 = softmax(axis = var_60, x = scores_87_cast_fp16)[name = string("op_5018_cast_fp16")]; + tensor input_1133_cast_fp16 = select(a = var_45_to_fp16, b = var_5018_cast_fp16, cond = mask_11)[name = string("input_1133_cast_fp16")]; + bool x_559_transpose_x_0 = const()[name = string("x_559_transpose_x_0"), val = bool(false)]; + bool x_559_transpose_y_0 = const()[name = string("x_559_transpose_y_0"), val = bool(false)]; + tensor value_51_cast_fp16 = transpose(perm = value_51_perm_0, x = v_43_cast_fp16)[name = string("transpose_170")]; + tensor x_559_cast_fp16 = matmul(transpose_x = x_559_transpose_x_0, transpose_y = x_559_transpose_y_0, x = input_1133_cast_fp16, y = value_51_cast_fp16)[name = string("x_559_cast_fp16")]; + tensor var_5022_perm_0 = const()[name = string("op_5022_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_5023 = const()[name = string("op_5023"), val = tensor([1, -1, 1024])]; + tensor var_5022_cast_fp16 = transpose(perm = var_5022_perm_0, x = x_559_cast_fp16)[name = string("transpose_169")]; + tensor input_1135_cast_fp16 = reshape(shape = var_5023, x = var_5022_cast_fp16)[name = string("input_1135_cast_fp16")]; + tensor encoder_layers_21_self_attn_linear_out_weight_to_fp16 = const()[name = string("encoder_layers_21_self_attn_linear_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(443061184)))]; + tensor encoder_layers_21_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_21_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(445158400)))]; + tensor linear_196_cast_fp16 = linear(bias = encoder_layers_21_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_21_self_attn_linear_out_weight_to_fp16, x = input_1135_cast_fp16)[name = string("linear_196_cast_fp16")]; + tensor input_1139_cast_fp16 = add(x = input_1129_cast_fp16, y = linear_196_cast_fp16)[name = string("input_1139_cast_fp16")]; + tensor x_563_axes_0 = const()[name = string("x_563_axes_0"), val = tensor([-1])]; + tensor encoder_layers_21_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_21_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(445160512)))]; + tensor encoder_layers_21_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_21_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(445162624)))]; + tensor x_563_cast_fp16 = layer_norm(axes = x_563_axes_0, beta = encoder_layers_21_norm_conv_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_21_norm_conv_weight_to_fp16, x = input_1139_cast_fp16)[name = string("x_563_cast_fp16")]; + tensor input_1141_perm_0 = const()[name = string("input_1141_perm_0"), val = tensor([0, 2, 1])]; + string input_1143_pad_type_0 = const()[name = string("input_1143_pad_type_0"), val = string("valid")]; + tensor input_1143_strides_0 = const()[name = string("input_1143_strides_0"), val = tensor([1])]; + tensor input_1143_pad_0 = const()[name = string("input_1143_pad_0"), val = tensor([0, 0])]; + tensor input_1143_dilations_0 = const()[name = string("input_1143_dilations_0"), val = tensor([1])]; + int32 input_1143_groups_0 = const()[name = string("input_1143_groups_0"), val = int32(1)]; + tensor encoder_layers_21_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(445164736))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(447261952))))[name = string("encoder_layers_21_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_1141_cast_fp16 = transpose(perm = input_1141_perm_0, x = x_563_cast_fp16)[name = string("transpose_168")]; + tensor input_1143_cast_fp16 = conv(dilations = input_1143_dilations_0, groups = input_1143_groups_0, pad = input_1143_pad_0, pad_type = input_1143_pad_type_0, strides = input_1143_strides_0, weight = encoder_layers_21_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_1141_cast_fp16)[name = string("input_1143_cast_fp16")]; + int32 x_565_split_num_splits_0 = const()[name = string("x_565_split_num_splits_0"), val = int32(2)]; + int32 x_565_split_axis_0 = const()[name = string("x_565_split_axis_0"), val = int32(1)]; + tensor x_565_split_cast_fp16_0, tensor x_565_split_cast_fp16_1 = split(axis = x_565_split_axis_0, num_splits = x_565_split_num_splits_0, x = input_1143_cast_fp16)[name = string("x_565_split_cast_fp16")]; + tensor x_565_split_1_sigmoid_cast_fp16 = sigmoid(x = x_565_split_cast_fp16_1)[name = string("x_565_split_1_sigmoid_cast_fp16")]; + tensor x_565_cast_fp16 = mul(x = x_565_split_cast_fp16_0, y = x_565_split_1_sigmoid_cast_fp16)[name = string("x_565_cast_fp16")]; + tensor input_1145_cast_fp16 = select(a = var_45_to_fp16, b = x_565_cast_fp16, cond = var_576)[name = string("input_1145_cast_fp16")]; + bool new_x_87_interleave_0 = const()[name = string("new_x_87_interleave_0"), val = bool(false)]; + tensor new_x_87_cast_fp16 = concat(axis = var_60, interleave = new_x_87_interleave_0, values = (cache_87_cast_fp16, input_1145_cast_fp16))[name = string("new_x_87_cast_fp16")]; + tensor var_5062_begin_0 = const()[name = string("op_5062_begin_0"), val = tensor([0, 0, 14])]; + tensor var_5062_end_0 = const()[name = string("op_5062_end_0"), val = tensor([1, 1024, 22])]; + tensor var_5062_end_mask_0 = const()[name = string("op_5062_end_mask_0"), val = tensor([true, true, true])]; + tensor var_5062_cast_fp16 = slice_by_index(begin = var_5062_begin_0, end = var_5062_end_0, end_mask = var_5062_end_mask_0, x = new_x_87_cast_fp16)[name = string("op_5062_cast_fp16")]; + string x_567_pad_type_0 = const()[name = string("x_567_pad_type_0"), val = string("valid")]; + int32 x_567_groups_0 = const()[name = string("x_567_groups_0"), val = int32(1024)]; + tensor x_567_strides_0 = const()[name = string("x_567_strides_0"), val = tensor([1])]; + tensor x_567_pad_0 = const()[name = string("x_567_pad_0"), val = tensor([0, 0])]; + tensor x_567_dilations_0 = const()[name = string("x_567_dilations_0"), val = tensor([1])]; + tensor encoder_layers_21_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(447266112))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(447275392))))[name = string("encoder_layers_21_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_567_cast_fp16 = conv(dilations = x_567_dilations_0, groups = x_567_groups_0, pad = x_567_pad_0, pad_type = x_567_pad_type_0, strides = x_567_strides_0, weight = encoder_layers_21_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_87_cast_fp16)[name = string("x_567_cast_fp16")]; + tensor input_1147_perm_0 = const()[name = string("input_1147_perm_0"), val = tensor([0, 2, 1])]; + tensor x_569_axes_0 = const()[name = string("x_569_axes_0"), val = tensor([-1])]; + tensor encoder_layers_21_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_21_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(447277504)))]; + tensor encoder_layers_21_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_21_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(447279616)))]; + tensor input_1147_cast_fp16 = transpose(perm = input_1147_perm_0, x = x_567_cast_fp16)[name = string("transpose_167")]; + tensor x_569_cast_fp16 = layer_norm(axes = x_569_axes_0, beta = encoder_layers_21_conv_batch_norm_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_21_conv_batch_norm_weight_to_fp16, x = input_1147_cast_fp16)[name = string("x_569_cast_fp16")]; + tensor input_1149_perm_0 = const()[name = string("input_1149_perm_0"), val = tensor([0, 2, 1])]; + tensor input_1149_cast_fp16 = transpose(perm = input_1149_perm_0, x = x_569_cast_fp16)[name = string("transpose_166")]; + tensor input_1151_cast_fp16 = silu(x = input_1149_cast_fp16)[name = string("input_1151_cast_fp16")]; + string x_571_pad_type_0 = const()[name = string("x_571_pad_type_0"), val = string("valid")]; + tensor x_571_strides_0 = const()[name = string("x_571_strides_0"), val = tensor([1])]; + tensor x_571_pad_0 = const()[name = string("x_571_pad_0"), val = tensor([0, 0])]; + tensor x_571_dilations_0 = const()[name = string("x_571_dilations_0"), val = tensor([1])]; + int32 x_571_groups_0 = const()[name = string("x_571_groups_0"), val = int32(1)]; + tensor encoder_layers_21_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(447281728))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(448330368))))[name = string("encoder_layers_21_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_571_cast_fp16 = conv(dilations = x_571_dilations_0, groups = x_571_groups_0, pad = x_571_pad_0, pad_type = x_571_pad_type_0, strides = x_571_strides_0, weight = encoder_layers_21_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_1151_cast_fp16)[name = string("x_571_cast_fp16")]; + tensor input_1153_perm_0 = const()[name = string("input_1153_perm_0"), val = tensor([0, 2, 1])]; + tensor input_1153_cast_fp16 = transpose(perm = input_1153_perm_0, x = x_571_cast_fp16)[name = string("transpose_165")]; + tensor input_1155_cast_fp16 = add(x = input_1139_cast_fp16, y = input_1153_cast_fp16)[name = string("input_1155_cast_fp16")]; + tensor input_1157_axes_0 = const()[name = string("input_1157_axes_0"), val = tensor([-1])]; + tensor encoder_layers_21_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_21_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(448332480)))]; + tensor encoder_layers_21_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_21_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(448334592)))]; + tensor input_1157_cast_fp16 = layer_norm(axes = input_1157_axes_0, beta = encoder_layers_21_norm_feed_forward2_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_21_norm_feed_forward2_weight_to_fp16, x = input_1155_cast_fp16)[name = string("input_1157_cast_fp16")]; + tensor encoder_layers_21_feed_forward2_linear1_weight_to_fp16 = const()[name = string("encoder_layers_21_feed_forward2_linear1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(448336704)))]; + tensor encoder_layers_21_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_21_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(456725376)))]; + tensor linear_197_cast_fp16 = linear(bias = encoder_layers_21_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_21_feed_forward2_linear1_weight_to_fp16, x = input_1157_cast_fp16)[name = string("linear_197_cast_fp16")]; + tensor input_1161_cast_fp16 = silu(x = linear_197_cast_fp16)[name = string("input_1161_cast_fp16")]; + tensor encoder_layers_21_feed_forward2_linear2_weight_to_fp16 = const()[name = string("encoder_layers_21_feed_forward2_linear2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(456733632)))]; + tensor encoder_layers_21_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_21_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465122304)))]; + tensor linear_198_cast_fp16 = linear(bias = encoder_layers_21_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_21_feed_forward2_linear2_weight_to_fp16, x = input_1161_cast_fp16)[name = string("linear_198_cast_fp16")]; + fp16 var_5105_to_fp16 = const()[name = string("op_5105_to_fp16"), val = fp16(0x1p-1)]; + tensor var_5106_cast_fp16 = mul(x = linear_198_cast_fp16, y = var_5105_to_fp16)[name = string("op_5106_cast_fp16")]; + tensor input_1167_cast_fp16 = add(x = input_1155_cast_fp16, y = var_5106_cast_fp16)[name = string("input_1167_cast_fp16")]; + tensor input_1169_axes_0 = const()[name = string("input_1169_axes_0"), val = tensor([-1])]; + tensor encoder_layers_21_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_21_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465124416)))]; + tensor encoder_layers_21_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_21_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465126528)))]; + tensor input_1169_cast_fp16 = layer_norm(axes = input_1169_axes_0, beta = encoder_layers_21_norm_out_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_21_norm_out_weight_to_fp16, x = input_1167_cast_fp16)[name = string("input_1169_cast_fp16")]; + tensor cache_89_begin_0 = const()[name = string("cache_89_begin_0"), val = tensor([22, 0, 0, 0])]; + tensor cache_89_end_0 = const()[name = string("cache_89_end_0"), val = tensor([23, 1, 42, 1024])]; + tensor cache_89_end_mask_0 = const()[name = string("cache_89_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_89_squeeze_mask_0 = const()[name = string("cache_89_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_89_cast_fp16 = slice_by_index(begin = cache_89_begin_0, end = cache_89_end_0, end_mask = cache_89_end_mask_0, squeeze_mask = cache_89_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_89_cast_fp16")]; + tensor cache_91_begin_0 = const()[name = string("cache_91_begin_0"), val = tensor([22, 0, 0, 0])]; + tensor cache_91_end_0 = const()[name = string("cache_91_end_0"), val = tensor([23, 1, 1024, 8])]; + tensor cache_91_end_mask_0 = const()[name = string("cache_91_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_91_squeeze_mask_0 = const()[name = string("cache_91_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_91_cast_fp16 = slice_by_index(begin = cache_91_begin_0, end = cache_91_end_0, end_mask = cache_91_end_mask_0, squeeze_mask = cache_91_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_91_cast_fp16")]; + tensor input_1171_axes_0 = const()[name = string("input_1171_axes_0"), val = tensor([-1])]; + tensor encoder_layers_22_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_22_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465128640)))]; + tensor encoder_layers_22_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_22_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465130752)))]; + tensor input_1171_cast_fp16 = layer_norm(axes = input_1171_axes_0, beta = encoder_layers_22_norm_feed_forward1_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_22_norm_feed_forward1_weight_to_fp16, x = input_1169_cast_fp16)[name = string("input_1171_cast_fp16")]; + tensor encoder_layers_22_feed_forward1_linear1_weight_to_fp16 = const()[name = string("encoder_layers_22_feed_forward1_linear1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465132864)))]; + tensor encoder_layers_22_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_22_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(473521536)))]; + tensor linear_199_cast_fp16 = linear(bias = encoder_layers_22_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_22_feed_forward1_linear1_weight_to_fp16, x = input_1171_cast_fp16)[name = string("linear_199_cast_fp16")]; + tensor input_1175_cast_fp16 = silu(x = linear_199_cast_fp16)[name = string("input_1175_cast_fp16")]; + tensor encoder_layers_22_feed_forward1_linear2_weight_to_fp16 = const()[name = string("encoder_layers_22_feed_forward1_linear2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(473529792)))]; + tensor encoder_layers_22_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_22_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481918464)))]; + tensor linear_200_cast_fp16 = linear(bias = encoder_layers_22_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_22_feed_forward1_linear2_weight_to_fp16, x = input_1175_cast_fp16)[name = string("linear_200_cast_fp16")]; + fp16 var_5142_to_fp16 = const()[name = string("op_5142_to_fp16"), val = fp16(0x1p-1)]; + tensor var_5143_cast_fp16 = mul(x = linear_200_cast_fp16, y = var_5142_to_fp16)[name = string("op_5143_cast_fp16")]; + tensor input_1181_cast_fp16 = add(x = input_1169_cast_fp16, y = var_5143_cast_fp16)[name = string("input_1181_cast_fp16")]; + tensor key_45_axes_0 = const()[name = string("key_45_axes_0"), val = tensor([-1])]; + tensor encoder_layers_22_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_22_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481920576)))]; + tensor encoder_layers_22_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_22_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481922688)))]; + tensor key_45_cast_fp16 = layer_norm(axes = key_45_axes_0, beta = encoder_layers_22_norm_self_att_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_22_norm_self_att_weight_to_fp16, x = input_1181_cast_fp16)[name = string("key_45_cast_fp16")]; + bool input_1183_interleave_0 = const()[name = string("input_1183_interleave_0"), val = bool(false)]; + tensor input_1183_cast_fp16 = concat(axis = var_69, interleave = input_1183_interleave_0, values = (cache_89_cast_fp16, key_45_cast_fp16))[name = string("input_1183_cast_fp16")]; + tensor var_5165_begin_0 = const()[name = string("op_5165_begin_0"), val = tensor([0, 14, 0])]; + tensor var_5165_end_0 = const()[name = string("op_5165_end_0"), val = tensor([1, 42, 1024])]; + tensor var_5165_end_mask_0 = const()[name = string("op_5165_end_mask_0"), val = tensor([true, true, true])]; + tensor var_5165_cast_fp16 = slice_by_index(begin = var_5165_begin_0, end = var_5165_end_0, end_mask = var_5165_end_mask_0, x = cache_89_cast_fp16)[name = string("op_5165_cast_fp16")]; + bool var_5171_interleave_0 = const()[name = string("op_5171_interleave_0"), val = bool(false)]; + tensor var_5171_cast_fp16 = concat(axis = var_69, interleave = var_5171_interleave_0, values = (var_5165_cast_fp16, key_45_cast_fp16))[name = string("op_5171_cast_fp16")]; + tensor encoder_layers_22_self_attn_linear_q_weight_to_fp16 = const()[name = string("encoder_layers_22_self_attn_linear_q_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481924800)))]; + tensor encoder_layers_22_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_22_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(484022016)))]; + tensor linear_201_cast_fp16 = linear(bias = encoder_layers_22_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_22_self_attn_linear_q_weight_to_fp16, x = key_45_cast_fp16)[name = string("linear_201_cast_fp16")]; + tensor var_5176 = const()[name = string("op_5176"), val = tensor([1, -1, 8, 128])]; + tensor q_133_cast_fp16 = reshape(shape = var_5176, x = linear_201_cast_fp16)[name = string("q_133_cast_fp16")]; + tensor encoder_layers_22_self_attn_linear_k_weight_to_fp16 = const()[name = string("encoder_layers_22_self_attn_linear_k_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(484024128)))]; + tensor encoder_layers_22_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_22_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(486121344)))]; + tensor linear_202_cast_fp16 = linear(bias = encoder_layers_22_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_22_self_attn_linear_k_weight_to_fp16, x = input_1183_cast_fp16)[name = string("linear_202_cast_fp16")]; + tensor var_5181 = const()[name = string("op_5181"), val = tensor([1, -1, 8, 128])]; + tensor k_89_cast_fp16 = reshape(shape = var_5181, x = linear_202_cast_fp16)[name = string("k_89_cast_fp16")]; + tensor encoder_layers_22_self_attn_linear_v_weight_to_fp16 = const()[name = string("encoder_layers_22_self_attn_linear_v_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(486123456)))]; + tensor encoder_layers_22_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_22_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(488220672)))]; + tensor linear_203_cast_fp16 = linear(bias = encoder_layers_22_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_22_self_attn_linear_v_weight_to_fp16, x = input_1183_cast_fp16)[name = string("linear_203_cast_fp16")]; + tensor var_5186 = const()[name = string("op_5186"), val = tensor([1, -1, 8, 128])]; + tensor v_45_cast_fp16 = reshape(shape = var_5186, x = linear_203_cast_fp16)[name = string("v_45_cast_fp16")]; + tensor value_53_perm_0 = const()[name = string("value_53_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_22_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_22_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(488222784)))]; + tensor var_5199_cast_fp16 = add(x = q_133_cast_fp16, y = encoder_layers_22_self_attn_pos_bias_u_to_fp16)[name = string("op_5199_cast_fp16")]; + tensor encoder_layers_22_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_22_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(488224896)))]; + tensor var_5201_cast_fp16 = add(x = q_133_cast_fp16, y = encoder_layers_22_self_attn_pos_bias_v_to_fp16)[name = string("op_5201_cast_fp16")]; + tensor q_with_bias_v_45_perm_0 = const()[name = string("q_with_bias_v_45_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_579_transpose_x_0 = const()[name = string("x_579_transpose_x_0"), val = bool(false)]; + bool x_579_transpose_y_0 = const()[name = string("x_579_transpose_y_0"), val = bool(false)]; + tensor op_5203_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(488227008))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(488340736))))[name = string("op_5203_to_fp16_quantized")]; + tensor q_with_bias_v_45_cast_fp16 = transpose(perm = q_with_bias_v_45_perm_0, x = var_5201_cast_fp16)[name = string("transpose_164")]; + tensor x_579_cast_fp16 = matmul(transpose_x = x_579_transpose_x_0, transpose_y = x_579_transpose_y_0, x = q_with_bias_v_45_cast_fp16, y = op_5203_to_fp16_quantized)[name = string("x_579_cast_fp16")]; + tensor x_581_pad_0 = const()[name = string("x_581_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_581_mode_0 = const()[name = string("x_581_mode_0"), val = string("constant")]; + fp16 const_365_to_fp16 = const()[name = string("const_365_to_fp16"), val = fp16(0x0p+0)]; + tensor x_581_cast_fp16 = pad(constant_val = const_365_to_fp16, mode = x_581_mode_0, pad = x_581_pad_0, x = x_579_cast_fp16)[name = string("x_581_cast_fp16")]; + tensor var_5211 = const()[name = string("op_5211"), val = tensor([1, 8, -1, 14])]; + tensor x_583_cast_fp16 = reshape(shape = var_5211, x = x_581_cast_fp16)[name = string("x_583_cast_fp16")]; + tensor var_5215_begin_0 = const()[name = string("op_5215_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_5215_end_0 = const()[name = string("op_5215_end_0"), val = tensor([1, 8, 112, 14])]; + tensor var_5215_end_mask_0 = const()[name = string("op_5215_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_5215_cast_fp16 = slice_by_index(begin = var_5215_begin_0, end = var_5215_end_0, end_mask = var_5215_end_mask_0, x = x_583_cast_fp16)[name = string("op_5215_cast_fp16")]; + tensor var_5216 = const()[name = string("op_5216"), val = tensor([1, 8, 14, 111])]; + tensor matrix_bd_89_cast_fp16 = reshape(shape = var_5216, x = var_5215_cast_fp16)[name = string("matrix_bd_89_cast_fp16")]; + bool matrix_ac_45_transpose_x_0 = const()[name = string("matrix_ac_45_transpose_x_0"), val = bool(false)]; + bool matrix_ac_45_transpose_y_0 = const()[name = string("matrix_ac_45_transpose_y_0"), val = bool(false)]; + tensor transpose_140_perm_0 = const()[name = string("transpose_140_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_141_perm_0 = const()[name = string("transpose_141_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_141 = transpose(perm = transpose_141_perm_0, x = k_89_cast_fp16)[name = string("transpose_162")]; + tensor transpose_140 = transpose(perm = transpose_140_perm_0, x = var_5199_cast_fp16)[name = string("transpose_163")]; + tensor matrix_ac_45_cast_fp16 = matmul(transpose_x = matrix_ac_45_transpose_x_0, transpose_y = matrix_ac_45_transpose_y_0, x = transpose_140, y = transpose_141)[name = string("matrix_ac_45_cast_fp16")]; + tensor matrix_bd_91_begin_0 = const()[name = string("matrix_bd_91_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_91_end_0 = const()[name = string("matrix_bd_91_end_0"), val = tensor([1, 8, 14, 56])]; + tensor matrix_bd_91_end_mask_0 = const()[name = string("matrix_bd_91_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_91_cast_fp16 = slice_by_index(begin = matrix_bd_91_begin_0, end = matrix_bd_91_end_0, end_mask = matrix_bd_91_end_mask_0, x = matrix_bd_89_cast_fp16)[name = string("matrix_bd_91_cast_fp16")]; + tensor var_5225_cast_fp16 = add(x = matrix_ac_45_cast_fp16, y = matrix_bd_91_cast_fp16)[name = string("op_5225_cast_fp16")]; + fp16 _inversed_scores_89_y_0_to_fp16 = const()[name = string("_inversed_scores_89_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_89_cast_fp16 = mul(x = var_5225_cast_fp16, y = _inversed_scores_89_y_0_to_fp16)[name = string("_inversed_scores_89_cast_fp16")]; + tensor scores_91_cast_fp16 = select(a = var_46_to_fp16, b = _inversed_scores_89_cast_fp16, cond = mask_11)[name = string("scores_91_cast_fp16")]; + tensor var_5231_cast_fp16 = softmax(axis = var_60, x = scores_91_cast_fp16)[name = string("op_5231_cast_fp16")]; + tensor input_1185_cast_fp16 = select(a = var_45_to_fp16, b = var_5231_cast_fp16, cond = mask_11)[name = string("input_1185_cast_fp16")]; + bool x_585_transpose_x_0 = const()[name = string("x_585_transpose_x_0"), val = bool(false)]; + bool x_585_transpose_y_0 = const()[name = string("x_585_transpose_y_0"), val = bool(false)]; + tensor value_53_cast_fp16 = transpose(perm = value_53_perm_0, x = v_45_cast_fp16)[name = string("transpose_161")]; + tensor x_585_cast_fp16 = matmul(transpose_x = x_585_transpose_x_0, transpose_y = x_585_transpose_y_0, x = input_1185_cast_fp16, y = value_53_cast_fp16)[name = string("x_585_cast_fp16")]; + tensor var_5235_perm_0 = const()[name = string("op_5235_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_5236 = const()[name = string("op_5236"), val = tensor([1, -1, 1024])]; + tensor var_5235_cast_fp16 = transpose(perm = var_5235_perm_0, x = x_585_cast_fp16)[name = string("transpose_160")]; + tensor input_1187_cast_fp16 = reshape(shape = var_5236, x = var_5235_cast_fp16)[name = string("input_1187_cast_fp16")]; + tensor encoder_layers_22_self_attn_linear_out_weight_to_fp16 = const()[name = string("encoder_layers_22_self_attn_linear_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(488341056)))]; + tensor encoder_layers_22_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_22_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(490438272)))]; + tensor linear_205_cast_fp16 = linear(bias = encoder_layers_22_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_22_self_attn_linear_out_weight_to_fp16, x = input_1187_cast_fp16)[name = string("linear_205_cast_fp16")]; + tensor input_1191_cast_fp16 = add(x = input_1181_cast_fp16, y = linear_205_cast_fp16)[name = string("input_1191_cast_fp16")]; + tensor x_589_axes_0 = const()[name = string("x_589_axes_0"), val = tensor([-1])]; + tensor encoder_layers_22_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_22_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(490440384)))]; + tensor encoder_layers_22_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_22_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(490442496)))]; + tensor x_589_cast_fp16 = layer_norm(axes = x_589_axes_0, beta = encoder_layers_22_norm_conv_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_22_norm_conv_weight_to_fp16, x = input_1191_cast_fp16)[name = string("x_589_cast_fp16")]; + tensor input_1193_perm_0 = const()[name = string("input_1193_perm_0"), val = tensor([0, 2, 1])]; + string input_1195_pad_type_0 = const()[name = string("input_1195_pad_type_0"), val = string("valid")]; + tensor input_1195_strides_0 = const()[name = string("input_1195_strides_0"), val = tensor([1])]; + tensor input_1195_pad_0 = const()[name = string("input_1195_pad_0"), val = tensor([0, 0])]; + tensor input_1195_dilations_0 = const()[name = string("input_1195_dilations_0"), val = tensor([1])]; + int32 input_1195_groups_0 = const()[name = string("input_1195_groups_0"), val = int32(1)]; + tensor encoder_layers_22_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(490444608))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(492541824))))[name = string("encoder_layers_22_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_1193_cast_fp16 = transpose(perm = input_1193_perm_0, x = x_589_cast_fp16)[name = string("transpose_159")]; + tensor input_1195_cast_fp16 = conv(dilations = input_1195_dilations_0, groups = input_1195_groups_0, pad = input_1195_pad_0, pad_type = input_1195_pad_type_0, strides = input_1195_strides_0, weight = encoder_layers_22_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_1193_cast_fp16)[name = string("input_1195_cast_fp16")]; + int32 x_591_split_num_splits_0 = const()[name = string("x_591_split_num_splits_0"), val = int32(2)]; + int32 x_591_split_axis_0 = const()[name = string("x_591_split_axis_0"), val = int32(1)]; + tensor x_591_split_cast_fp16_0, tensor x_591_split_cast_fp16_1 = split(axis = x_591_split_axis_0, num_splits = x_591_split_num_splits_0, x = input_1195_cast_fp16)[name = string("x_591_split_cast_fp16")]; + tensor x_591_split_1_sigmoid_cast_fp16 = sigmoid(x = x_591_split_cast_fp16_1)[name = string("x_591_split_1_sigmoid_cast_fp16")]; + tensor x_591_cast_fp16 = mul(x = x_591_split_cast_fp16_0, y = x_591_split_1_sigmoid_cast_fp16)[name = string("x_591_cast_fp16")]; + tensor input_1197_cast_fp16 = select(a = var_45_to_fp16, b = x_591_cast_fp16, cond = var_576)[name = string("input_1197_cast_fp16")]; + bool new_x_91_interleave_0 = const()[name = string("new_x_91_interleave_0"), val = bool(false)]; + tensor new_x_91_cast_fp16 = concat(axis = var_60, interleave = new_x_91_interleave_0, values = (cache_91_cast_fp16, input_1197_cast_fp16))[name = string("new_x_91_cast_fp16")]; + tensor var_5275_begin_0 = const()[name = string("op_5275_begin_0"), val = tensor([0, 0, 14])]; + tensor var_5275_end_0 = const()[name = string("op_5275_end_0"), val = tensor([1, 1024, 22])]; + tensor var_5275_end_mask_0 = const()[name = string("op_5275_end_mask_0"), val = tensor([true, true, true])]; + tensor var_5275_cast_fp16 = slice_by_index(begin = var_5275_begin_0, end = var_5275_end_0, end_mask = var_5275_end_mask_0, x = new_x_91_cast_fp16)[name = string("op_5275_cast_fp16")]; + string x_593_pad_type_0 = const()[name = string("x_593_pad_type_0"), val = string("valid")]; + int32 x_593_groups_0 = const()[name = string("x_593_groups_0"), val = int32(1024)]; + tensor x_593_strides_0 = const()[name = string("x_593_strides_0"), val = tensor([1])]; + tensor x_593_pad_0 = const()[name = string("x_593_pad_0"), val = tensor([0, 0])]; + tensor x_593_dilations_0 = const()[name = string("x_593_dilations_0"), val = tensor([1])]; + tensor encoder_layers_22_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(492545984))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(492555264))))[name = string("encoder_layers_22_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_593_cast_fp16 = conv(dilations = x_593_dilations_0, groups = x_593_groups_0, pad = x_593_pad_0, pad_type = x_593_pad_type_0, strides = x_593_strides_0, weight = encoder_layers_22_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_91_cast_fp16)[name = string("x_593_cast_fp16")]; + tensor input_1199_perm_0 = const()[name = string("input_1199_perm_0"), val = tensor([0, 2, 1])]; + tensor x_595_axes_0 = const()[name = string("x_595_axes_0"), val = tensor([-1])]; + tensor encoder_layers_22_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_22_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(492557376)))]; + tensor encoder_layers_22_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_22_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(492559488)))]; + tensor input_1199_cast_fp16 = transpose(perm = input_1199_perm_0, x = x_593_cast_fp16)[name = string("transpose_158")]; + tensor x_595_cast_fp16 = layer_norm(axes = x_595_axes_0, beta = encoder_layers_22_conv_batch_norm_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_22_conv_batch_norm_weight_to_fp16, x = input_1199_cast_fp16)[name = string("x_595_cast_fp16")]; + tensor input_1201_perm_0 = const()[name = string("input_1201_perm_0"), val = tensor([0, 2, 1])]; + tensor input_1201_cast_fp16 = transpose(perm = input_1201_perm_0, x = x_595_cast_fp16)[name = string("transpose_157")]; + tensor input_1203_cast_fp16 = silu(x = input_1201_cast_fp16)[name = string("input_1203_cast_fp16")]; + string x_597_pad_type_0 = const()[name = string("x_597_pad_type_0"), val = string("valid")]; + tensor x_597_strides_0 = const()[name = string("x_597_strides_0"), val = tensor([1])]; + tensor x_597_pad_0 = const()[name = string("x_597_pad_0"), val = tensor([0, 0])]; + tensor x_597_dilations_0 = const()[name = string("x_597_dilations_0"), val = tensor([1])]; + int32 x_597_groups_0 = const()[name = string("x_597_groups_0"), val = int32(1)]; + tensor encoder_layers_22_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(492561600))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(493610240))))[name = string("encoder_layers_22_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_597_cast_fp16 = conv(dilations = x_597_dilations_0, groups = x_597_groups_0, pad = x_597_pad_0, pad_type = x_597_pad_type_0, strides = x_597_strides_0, weight = encoder_layers_22_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_1203_cast_fp16)[name = string("x_597_cast_fp16")]; + tensor input_1205_perm_0 = const()[name = string("input_1205_perm_0"), val = tensor([0, 2, 1])]; + tensor input_1205_cast_fp16 = transpose(perm = input_1205_perm_0, x = x_597_cast_fp16)[name = string("transpose_156")]; + tensor input_1207_cast_fp16 = add(x = input_1191_cast_fp16, y = input_1205_cast_fp16)[name = string("input_1207_cast_fp16")]; + tensor input_1209_axes_0 = const()[name = string("input_1209_axes_0"), val = tensor([-1])]; + tensor encoder_layers_22_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_22_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(493612352)))]; + tensor encoder_layers_22_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_22_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(493614464)))]; + tensor input_1209_cast_fp16 = layer_norm(axes = input_1209_axes_0, beta = encoder_layers_22_norm_feed_forward2_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_22_norm_feed_forward2_weight_to_fp16, x = input_1207_cast_fp16)[name = string("input_1209_cast_fp16")]; + tensor encoder_layers_22_feed_forward2_linear1_weight_to_fp16 = const()[name = string("encoder_layers_22_feed_forward2_linear1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(493616576)))]; + tensor encoder_layers_22_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_22_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(502005248)))]; + tensor linear_206_cast_fp16 = linear(bias = encoder_layers_22_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_22_feed_forward2_linear1_weight_to_fp16, x = input_1209_cast_fp16)[name = string("linear_206_cast_fp16")]; + tensor input_1213_cast_fp16 = silu(x = linear_206_cast_fp16)[name = string("input_1213_cast_fp16")]; + tensor encoder_layers_22_feed_forward2_linear2_weight_to_fp16 = const()[name = string("encoder_layers_22_feed_forward2_linear2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(502013504)))]; + tensor encoder_layers_22_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_22_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(510402176)))]; + tensor linear_207_cast_fp16 = linear(bias = encoder_layers_22_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_22_feed_forward2_linear2_weight_to_fp16, x = input_1213_cast_fp16)[name = string("linear_207_cast_fp16")]; + fp16 var_5318_to_fp16 = const()[name = string("op_5318_to_fp16"), val = fp16(0x1p-1)]; + tensor var_5319_cast_fp16 = mul(x = linear_207_cast_fp16, y = var_5318_to_fp16)[name = string("op_5319_cast_fp16")]; + tensor input_1219_cast_fp16 = add(x = input_1207_cast_fp16, y = var_5319_cast_fp16)[name = string("input_1219_cast_fp16")]; + tensor input_1221_axes_0 = const()[name = string("input_1221_axes_0"), val = tensor([-1])]; + tensor encoder_layers_22_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_22_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(510404288)))]; + tensor encoder_layers_22_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_22_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(510406400)))]; + tensor input_1221_cast_fp16 = layer_norm(axes = input_1221_axes_0, beta = encoder_layers_22_norm_out_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_22_norm_out_weight_to_fp16, x = input_1219_cast_fp16)[name = string("input_1221_cast_fp16")]; + tensor cache_93_begin_0 = const()[name = string("cache_93_begin_0"), val = tensor([23, 0, 0, 0])]; + tensor cache_93_end_0 = const()[name = string("cache_93_end_0"), val = tensor([24, 1, 42, 1024])]; + tensor cache_93_end_mask_0 = const()[name = string("cache_93_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_93_squeeze_mask_0 = const()[name = string("cache_93_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_93_cast_fp16 = slice_by_index(begin = cache_93_begin_0, end = cache_93_end_0, end_mask = cache_93_end_mask_0, squeeze_mask = cache_93_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_93_cast_fp16")]; + tensor cache_begin_0 = const()[name = string("cache_begin_0"), val = tensor([23, 0, 0, 0])]; + tensor cache_end_0 = const()[name = string("cache_end_0"), val = tensor([24, 1, 1024, 8])]; + tensor cache_end_mask_0 = const()[name = string("cache_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_squeeze_mask_0 = const()[name = string("cache_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_cast_fp16 = slice_by_index(begin = cache_begin_0, end = cache_end_0, end_mask = cache_end_mask_0, squeeze_mask = cache_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_cast_fp16")]; + tensor input_1223_axes_0 = const()[name = string("input_1223_axes_0"), val = tensor([-1])]; + tensor encoder_layers_23_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_23_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(510408512)))]; + tensor encoder_layers_23_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_23_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(510410624)))]; + tensor input_1223_cast_fp16 = layer_norm(axes = input_1223_axes_0, beta = encoder_layers_23_norm_feed_forward1_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_23_norm_feed_forward1_weight_to_fp16, x = input_1221_cast_fp16)[name = string("input_1223_cast_fp16")]; + tensor encoder_layers_23_feed_forward1_linear1_weight_to_fp16 = const()[name = string("encoder_layers_23_feed_forward1_linear1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(510412736)))]; + tensor encoder_layers_23_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_23_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(518801408)))]; + tensor linear_208_cast_fp16 = linear(bias = encoder_layers_23_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_23_feed_forward1_linear1_weight_to_fp16, x = input_1223_cast_fp16)[name = string("linear_208_cast_fp16")]; + tensor input_1227_cast_fp16 = silu(x = linear_208_cast_fp16)[name = string("input_1227_cast_fp16")]; + tensor encoder_layers_23_feed_forward1_linear2_weight_to_fp16 = const()[name = string("encoder_layers_23_feed_forward1_linear2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(518809664)))]; + tensor encoder_layers_23_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_23_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(527198336)))]; + tensor linear_209_cast_fp16 = linear(bias = encoder_layers_23_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_23_feed_forward1_linear2_weight_to_fp16, x = input_1227_cast_fp16)[name = string("linear_209_cast_fp16")]; + fp16 var_5355_to_fp16 = const()[name = string("op_5355_to_fp16"), val = fp16(0x1p-1)]; + tensor var_5356_cast_fp16 = mul(x = linear_209_cast_fp16, y = var_5355_to_fp16)[name = string("op_5356_cast_fp16")]; + tensor input_1233_cast_fp16 = add(x = input_1221_cast_fp16, y = var_5356_cast_fp16)[name = string("input_1233_cast_fp16")]; + tensor key_axes_0 = const()[name = string("key_axes_0"), val = tensor([-1])]; + tensor encoder_layers_23_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_23_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(527200448)))]; + tensor encoder_layers_23_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_23_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(527202560)))]; + tensor key_cast_fp16 = layer_norm(axes = key_axes_0, beta = encoder_layers_23_norm_self_att_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_23_norm_self_att_weight_to_fp16, x = input_1233_cast_fp16)[name = string("key_cast_fp16")]; + bool input_1235_interleave_0 = const()[name = string("input_1235_interleave_0"), val = bool(false)]; + tensor input_1235_cast_fp16 = concat(axis = var_69, interleave = input_1235_interleave_0, values = (cache_93_cast_fp16, key_cast_fp16))[name = string("input_1235_cast_fp16")]; + tensor var_5378_begin_0 = const()[name = string("op_5378_begin_0"), val = tensor([0, 14, 0])]; + tensor var_5378_end_0 = const()[name = string("op_5378_end_0"), val = tensor([1, 42, 1024])]; + tensor var_5378_end_mask_0 = const()[name = string("op_5378_end_mask_0"), val = tensor([true, true, true])]; + tensor var_5378_cast_fp16 = slice_by_index(begin = var_5378_begin_0, end = var_5378_end_0, end_mask = var_5378_end_mask_0, x = cache_93_cast_fp16)[name = string("op_5378_cast_fp16")]; + bool cache_last_channel_cur_interleave_0 = const()[name = string("cache_last_channel_cur_interleave_0"), val = bool(false)]; + tensor cache_last_channel_cur_cast_fp16 = concat(axis = var_69, interleave = cache_last_channel_cur_interleave_0, values = (var_5378_cast_fp16, key_cast_fp16))[name = string("cache_last_channel_cur_cast_fp16")]; + tensor encoder_layers_23_self_attn_linear_q_weight_to_fp16 = const()[name = string("encoder_layers_23_self_attn_linear_q_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(527204672)))]; + tensor encoder_layers_23_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_23_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(529301888)))]; + tensor linear_210_cast_fp16 = linear(bias = encoder_layers_23_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_23_self_attn_linear_q_weight_to_fp16, x = key_cast_fp16)[name = string("linear_210_cast_fp16")]; + tensor var_5389 = const()[name = string("op_5389"), val = tensor([1, -1, 8, 128])]; + tensor q_139_cast_fp16 = reshape(shape = var_5389, x = linear_210_cast_fp16)[name = string("q_139_cast_fp16")]; + tensor encoder_layers_23_self_attn_linear_k_weight_to_fp16 = const()[name = string("encoder_layers_23_self_attn_linear_k_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(529304000)))]; + tensor encoder_layers_23_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_23_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(531401216)))]; + tensor linear_211_cast_fp16 = linear(bias = encoder_layers_23_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_23_self_attn_linear_k_weight_to_fp16, x = input_1235_cast_fp16)[name = string("linear_211_cast_fp16")]; + tensor var_5394 = const()[name = string("op_5394"), val = tensor([1, -1, 8, 128])]; + tensor k_93_cast_fp16 = reshape(shape = var_5394, x = linear_211_cast_fp16)[name = string("k_93_cast_fp16")]; + tensor encoder_layers_23_self_attn_linear_v_weight_to_fp16 = const()[name = string("encoder_layers_23_self_attn_linear_v_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(531403328)))]; + tensor encoder_layers_23_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_23_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(533500544)))]; + tensor linear_212_cast_fp16 = linear(bias = encoder_layers_23_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_23_self_attn_linear_v_weight_to_fp16, x = input_1235_cast_fp16)[name = string("linear_212_cast_fp16")]; + tensor var_5399 = const()[name = string("op_5399"), val = tensor([1, -1, 8, 128])]; + tensor v_cast_fp16 = reshape(shape = var_5399, x = linear_212_cast_fp16)[name = string("v_cast_fp16")]; + tensor value_perm_0 = const()[name = string("value_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_23_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_23_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(533502656)))]; + tensor var_5412_cast_fp16 = add(x = q_139_cast_fp16, y = encoder_layers_23_self_attn_pos_bias_u_to_fp16)[name = string("op_5412_cast_fp16")]; + tensor encoder_layers_23_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_23_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(533504768)))]; + tensor var_5414_cast_fp16 = add(x = q_139_cast_fp16, y = encoder_layers_23_self_attn_pos_bias_v_to_fp16)[name = string("op_5414_cast_fp16")]; + tensor q_with_bias_v_perm_0 = const()[name = string("q_with_bias_v_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_605_transpose_x_0 = const()[name = string("x_605_transpose_x_0"), val = bool(false)]; + bool x_605_transpose_y_0 = const()[name = string("x_605_transpose_y_0"), val = bool(false)]; + tensor op_5416_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(533506880))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(533620608))))[name = string("op_5416_to_fp16_quantized")]; + tensor q_with_bias_v_cast_fp16 = transpose(perm = q_with_bias_v_perm_0, x = var_5414_cast_fp16)[name = string("transpose_155")]; + tensor x_605_cast_fp16 = matmul(transpose_x = x_605_transpose_x_0, transpose_y = x_605_transpose_y_0, x = q_with_bias_v_cast_fp16, y = op_5416_to_fp16_quantized)[name = string("x_605_cast_fp16")]; + tensor x_607_pad_0 = const()[name = string("x_607_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_607_mode_0 = const()[name = string("x_607_mode_0"), val = string("constant")]; + fp16 const_378_to_fp16 = const()[name = string("const_378_to_fp16"), val = fp16(0x0p+0)]; + tensor x_607_cast_fp16 = pad(constant_val = const_378_to_fp16, mode = x_607_mode_0, pad = x_607_pad_0, x = x_605_cast_fp16)[name = string("x_607_cast_fp16")]; + tensor var_5424 = const()[name = string("op_5424"), val = tensor([1, 8, -1, 14])]; + tensor x_609_cast_fp16 = reshape(shape = var_5424, x = x_607_cast_fp16)[name = string("x_609_cast_fp16")]; + tensor var_5428_begin_0 = const()[name = string("op_5428_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_5428_end_0 = const()[name = string("op_5428_end_0"), val = tensor([1, 8, 112, 14])]; + tensor var_5428_end_mask_0 = const()[name = string("op_5428_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_5428_cast_fp16 = slice_by_index(begin = var_5428_begin_0, end = var_5428_end_0, end_mask = var_5428_end_mask_0, x = x_609_cast_fp16)[name = string("op_5428_cast_fp16")]; + tensor var_5429 = const()[name = string("op_5429"), val = tensor([1, 8, 14, 111])]; + tensor matrix_bd_93_cast_fp16 = reshape(shape = var_5429, x = var_5428_cast_fp16)[name = string("matrix_bd_93_cast_fp16")]; + bool matrix_ac_transpose_x_0 = const()[name = string("matrix_ac_transpose_x_0"), val = bool(false)]; + bool matrix_ac_transpose_y_0 = const()[name = string("matrix_ac_transpose_y_0"), val = bool(false)]; + tensor transpose_142_perm_0 = const()[name = string("transpose_142_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_143_perm_0 = const()[name = string("transpose_143_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_143 = transpose(perm = transpose_143_perm_0, x = k_93_cast_fp16)[name = string("transpose_153")]; + tensor transpose_142 = transpose(perm = transpose_142_perm_0, x = var_5412_cast_fp16)[name = string("transpose_154")]; + tensor matrix_ac_cast_fp16 = matmul(transpose_x = matrix_ac_transpose_x_0, transpose_y = matrix_ac_transpose_y_0, x = transpose_142, y = transpose_143)[name = string("matrix_ac_cast_fp16")]; + tensor matrix_bd_begin_0 = const()[name = string("matrix_bd_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_end_0 = const()[name = string("matrix_bd_end_0"), val = tensor([1, 8, 14, 56])]; + tensor matrix_bd_end_mask_0 = const()[name = string("matrix_bd_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_cast_fp16 = slice_by_index(begin = matrix_bd_begin_0, end = matrix_bd_end_0, end_mask = matrix_bd_end_mask_0, x = matrix_bd_93_cast_fp16)[name = string("matrix_bd_cast_fp16")]; + tensor var_5438_cast_fp16 = add(x = matrix_ac_cast_fp16, y = matrix_bd_cast_fp16)[name = string("op_5438_cast_fp16")]; + fp16 _inversed_scores_93_y_0_to_fp16 = const()[name = string("_inversed_scores_93_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_93_cast_fp16 = mul(x = var_5438_cast_fp16, y = _inversed_scores_93_y_0_to_fp16)[name = string("_inversed_scores_93_cast_fp16")]; + tensor scores_cast_fp16 = select(a = var_46_to_fp16, b = _inversed_scores_93_cast_fp16, cond = mask_11)[name = string("scores_cast_fp16")]; + tensor var_5444_cast_fp16 = softmax(axis = var_60, x = scores_cast_fp16)[name = string("op_5444_cast_fp16")]; + tensor input_1237_cast_fp16 = select(a = var_45_to_fp16, b = var_5444_cast_fp16, cond = mask_11)[name = string("input_1237_cast_fp16")]; + bool x_611_transpose_x_0 = const()[name = string("x_611_transpose_x_0"), val = bool(false)]; + bool x_611_transpose_y_0 = const()[name = string("x_611_transpose_y_0"), val = bool(false)]; + tensor value_cast_fp16 = transpose(perm = value_perm_0, x = v_cast_fp16)[name = string("transpose_152")]; + tensor x_611_cast_fp16 = matmul(transpose_x = x_611_transpose_x_0, transpose_y = x_611_transpose_y_0, x = input_1237_cast_fp16, y = value_cast_fp16)[name = string("x_611_cast_fp16")]; + tensor var_5448_perm_0 = const()[name = string("op_5448_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_5449 = const()[name = string("op_5449"), val = tensor([1, -1, 1024])]; + tensor var_5448_cast_fp16 = transpose(perm = var_5448_perm_0, x = x_611_cast_fp16)[name = string("transpose_151")]; + tensor input_1239_cast_fp16 = reshape(shape = var_5449, x = var_5448_cast_fp16)[name = string("input_1239_cast_fp16")]; + tensor encoder_layers_23_self_attn_linear_out_weight_to_fp16 = const()[name = string("encoder_layers_23_self_attn_linear_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(533620928)))]; + tensor encoder_layers_23_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_23_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(535718144)))]; + tensor linear_214_cast_fp16 = linear(bias = encoder_layers_23_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_23_self_attn_linear_out_weight_to_fp16, x = input_1239_cast_fp16)[name = string("linear_214_cast_fp16")]; + tensor input_1243_cast_fp16 = add(x = input_1233_cast_fp16, y = linear_214_cast_fp16)[name = string("input_1243_cast_fp16")]; + tensor x_615_axes_0 = const()[name = string("x_615_axes_0"), val = tensor([-1])]; + tensor encoder_layers_23_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_23_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(535720256)))]; + tensor encoder_layers_23_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_23_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(535722368)))]; + tensor x_615_cast_fp16 = layer_norm(axes = x_615_axes_0, beta = encoder_layers_23_norm_conv_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_23_norm_conv_weight_to_fp16, x = input_1243_cast_fp16)[name = string("x_615_cast_fp16")]; + tensor input_1245_perm_0 = const()[name = string("input_1245_perm_0"), val = tensor([0, 2, 1])]; + string input_1247_pad_type_0 = const()[name = string("input_1247_pad_type_0"), val = string("valid")]; + tensor input_1247_strides_0 = const()[name = string("input_1247_strides_0"), val = tensor([1])]; + tensor input_1247_pad_0 = const()[name = string("input_1247_pad_0"), val = tensor([0, 0])]; + tensor input_1247_dilations_0 = const()[name = string("input_1247_dilations_0"), val = tensor([1])]; + int32 input_1247_groups_0 = const()[name = string("input_1247_groups_0"), val = int32(1)]; + tensor encoder_layers_23_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(535724480))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(537821696))))[name = string("encoder_layers_23_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_1245_cast_fp16 = transpose(perm = input_1245_perm_0, x = x_615_cast_fp16)[name = string("transpose_150")]; + tensor input_1247_cast_fp16 = conv(dilations = input_1247_dilations_0, groups = input_1247_groups_0, pad = input_1247_pad_0, pad_type = input_1247_pad_type_0, strides = input_1247_strides_0, weight = encoder_layers_23_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_1245_cast_fp16)[name = string("input_1247_cast_fp16")]; + int32 x_617_split_num_splits_0 = const()[name = string("x_617_split_num_splits_0"), val = int32(2)]; + int32 x_617_split_axis_0 = const()[name = string("x_617_split_axis_0"), val = int32(1)]; + tensor x_617_split_cast_fp16_0, tensor x_617_split_cast_fp16_1 = split(axis = x_617_split_axis_0, num_splits = x_617_split_num_splits_0, x = input_1247_cast_fp16)[name = string("x_617_split_cast_fp16")]; + tensor x_617_split_1_sigmoid_cast_fp16 = sigmoid(x = x_617_split_cast_fp16_1)[name = string("x_617_split_1_sigmoid_cast_fp16")]; + tensor x_617_cast_fp16 = mul(x = x_617_split_cast_fp16_0, y = x_617_split_1_sigmoid_cast_fp16)[name = string("x_617_cast_fp16")]; + tensor input_1249_cast_fp16 = select(a = var_45_to_fp16, b = x_617_cast_fp16, cond = var_576)[name = string("input_1249_cast_fp16")]; + bool new_x_interleave_0 = const()[name = string("new_x_interleave_0"), val = bool(false)]; + tensor new_x_cast_fp16 = concat(axis = var_60, interleave = new_x_interleave_0, values = (cache_cast_fp16, input_1249_cast_fp16))[name = string("new_x_cast_fp16")]; + tensor cache_last_time_cur_begin_0 = const()[name = string("cache_last_time_cur_begin_0"), val = tensor([0, 0, 14])]; + tensor cache_last_time_cur_end_0 = const()[name = string("cache_last_time_cur_end_0"), val = tensor([1, 1024, 22])]; + tensor cache_last_time_cur_end_mask_0 = const()[name = string("cache_last_time_cur_end_mask_0"), val = tensor([true, true, true])]; + tensor cache_last_time_cur_cast_fp16 = slice_by_index(begin = cache_last_time_cur_begin_0, end = cache_last_time_cur_end_0, end_mask = cache_last_time_cur_end_mask_0, x = new_x_cast_fp16)[name = string("cache_last_time_cur_cast_fp16")]; + string x_619_pad_type_0 = const()[name = string("x_619_pad_type_0"), val = string("valid")]; + int32 x_619_groups_0 = const()[name = string("x_619_groups_0"), val = int32(1024)]; + tensor x_619_strides_0 = const()[name = string("x_619_strides_0"), val = tensor([1])]; + tensor x_619_pad_0 = const()[name = string("x_619_pad_0"), val = tensor([0, 0])]; + tensor x_619_dilations_0 = const()[name = string("x_619_dilations_0"), val = tensor([1])]; + tensor encoder_layers_23_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(537825856))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(537835136))))[name = string("encoder_layers_23_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_619_cast_fp16 = conv(dilations = x_619_dilations_0, groups = x_619_groups_0, pad = x_619_pad_0, pad_type = x_619_pad_type_0, strides = x_619_strides_0, weight = encoder_layers_23_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_cast_fp16)[name = string("x_619_cast_fp16")]; + tensor input_1251_perm_0 = const()[name = string("input_1251_perm_0"), val = tensor([0, 2, 1])]; + tensor x_621_axes_0 = const()[name = string("x_621_axes_0"), val = tensor([-1])]; + tensor encoder_layers_23_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_23_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(537837248)))]; + tensor encoder_layers_23_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_23_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(537839360)))]; + tensor input_1251_cast_fp16 = transpose(perm = input_1251_perm_0, x = x_619_cast_fp16)[name = string("transpose_149")]; + tensor x_621_cast_fp16 = layer_norm(axes = x_621_axes_0, beta = encoder_layers_23_conv_batch_norm_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_23_conv_batch_norm_weight_to_fp16, x = input_1251_cast_fp16)[name = string("x_621_cast_fp16")]; + tensor input_1253_perm_0 = const()[name = string("input_1253_perm_0"), val = tensor([0, 2, 1])]; + tensor input_1253_cast_fp16 = transpose(perm = input_1253_perm_0, x = x_621_cast_fp16)[name = string("transpose_148")]; + tensor input_1255_cast_fp16 = silu(x = input_1253_cast_fp16)[name = string("input_1255_cast_fp16")]; + string x_623_pad_type_0 = const()[name = string("x_623_pad_type_0"), val = string("valid")]; + tensor x_623_strides_0 = const()[name = string("x_623_strides_0"), val = tensor([1])]; + tensor x_623_pad_0 = const()[name = string("x_623_pad_0"), val = tensor([0, 0])]; + tensor x_623_dilations_0 = const()[name = string("x_623_dilations_0"), val = tensor([1])]; + int32 x_623_groups_0 = const()[name = string("x_623_groups_0"), val = int32(1)]; + tensor encoder_layers_23_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(537841472))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(538890112))))[name = string("encoder_layers_23_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_623_cast_fp16 = conv(dilations = x_623_dilations_0, groups = x_623_groups_0, pad = x_623_pad_0, pad_type = x_623_pad_type_0, strides = x_623_strides_0, weight = encoder_layers_23_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_1255_cast_fp16)[name = string("x_623_cast_fp16")]; + tensor input_1257_perm_0 = const()[name = string("input_1257_perm_0"), val = tensor([0, 2, 1])]; + tensor input_1257_cast_fp16 = transpose(perm = input_1257_perm_0, x = x_623_cast_fp16)[name = string("transpose_147")]; + tensor input_1259_cast_fp16 = add(x = input_1243_cast_fp16, y = input_1257_cast_fp16)[name = string("input_1259_cast_fp16")]; + tensor input_1261_axes_0 = const()[name = string("input_1261_axes_0"), val = tensor([-1])]; + tensor encoder_layers_23_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_23_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(538892224)))]; + tensor encoder_layers_23_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_23_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(538894336)))]; + tensor input_1261_cast_fp16 = layer_norm(axes = input_1261_axes_0, beta = encoder_layers_23_norm_feed_forward2_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_23_norm_feed_forward2_weight_to_fp16, x = input_1259_cast_fp16)[name = string("input_1261_cast_fp16")]; + tensor encoder_layers_23_feed_forward2_linear1_weight_to_fp16 = const()[name = string("encoder_layers_23_feed_forward2_linear1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(538896448)))]; + tensor encoder_layers_23_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_23_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(547285120)))]; + tensor linear_215_cast_fp16 = linear(bias = encoder_layers_23_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_23_feed_forward2_linear1_weight_to_fp16, x = input_1261_cast_fp16)[name = string("linear_215_cast_fp16")]; + tensor input_1265_cast_fp16 = silu(x = linear_215_cast_fp16)[name = string("input_1265_cast_fp16")]; + tensor encoder_layers_23_feed_forward2_linear2_weight_to_fp16 = const()[name = string("encoder_layers_23_feed_forward2_linear2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(547293376)))]; + tensor encoder_layers_23_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_23_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(555682048)))]; + tensor linear_216_cast_fp16 = linear(bias = encoder_layers_23_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_23_feed_forward2_linear2_weight_to_fp16, x = input_1265_cast_fp16)[name = string("linear_216_cast_fp16")]; + fp16 var_5531_to_fp16 = const()[name = string("op_5531_to_fp16"), val = fp16(0x1p-1)]; + tensor var_5532_cast_fp16 = mul(x = linear_216_cast_fp16, y = var_5531_to_fp16)[name = string("op_5532_cast_fp16")]; + tensor input_1271_cast_fp16 = add(x = input_1259_cast_fp16, y = var_5532_cast_fp16)[name = string("input_1271_cast_fp16")]; + tensor audio_signal_axes_0 = const()[name = string("audio_signal_axes_0"), val = tensor([-1])]; + tensor encoder_layers_23_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_23_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(555684160)))]; + tensor encoder_layers_23_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_23_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(555686272)))]; + tensor audio_signal_cast_fp16 = layer_norm(axes = audio_signal_axes_0, beta = encoder_layers_23_norm_out_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_23_norm_out_weight_to_fp16, x = input_1271_cast_fp16)[name = string("audio_signal_cast_fp16")]; + int32 obj_5_axis_0 = const()[name = string("obj_5_axis_0"), val = int32(0)]; + tensor obj_5_cast_fp16 = stack(axis = obj_5_axis_0, values = (var_485_cast_fp16, var_698_cast_fp16, var_911_cast_fp16, var_1124_cast_fp16, var_1337_cast_fp16, var_1550_cast_fp16, var_1763_cast_fp16, var_1976_cast_fp16, var_2189_cast_fp16, var_2402_cast_fp16, var_2615_cast_fp16, var_2828_cast_fp16, var_3041_cast_fp16, var_3254_cast_fp16, var_3467_cast_fp16, var_3680_cast_fp16, var_3893_cast_fp16, var_4106_cast_fp16, var_4319_cast_fp16, var_4532_cast_fp16, var_4745_cast_fp16, var_4958_cast_fp16, var_5171_cast_fp16, cache_last_channel_cur_cast_fp16))[name = string("obj_5_cast_fp16")]; + int32 obj_7_axis_0 = const()[name = string("obj_7_axis_0"), val = int32(0)]; + tensor obj_7_cast_fp16 = stack(axis = obj_7_axis_0, values = (var_589_cast_fp16, var_802_cast_fp16, var_1015_cast_fp16, var_1228_cast_fp16, var_1441_cast_fp16, var_1654_cast_fp16, var_1867_cast_fp16, var_2080_cast_fp16, var_2293_cast_fp16, var_2506_cast_fp16, var_2719_cast_fp16, var_2932_cast_fp16, var_3145_cast_fp16, var_3358_cast_fp16, var_3571_cast_fp16, var_3784_cast_fp16, var_3997_cast_fp16, var_4210_cast_fp16, var_4423_cast_fp16, var_4636_cast_fp16, var_4849_cast_fp16, var_5062_cast_fp16, var_5275_cast_fp16, cache_last_time_cur_cast_fp16))[name = string("obj_7_cast_fp16")]; + tensor var_5548 = add(x = cache_len, y = max_audio_length_1)[name = string("op_5548")]; + string var_5548_promoted_to_fp16_dtype_0 = const()[name = string("op_5548_promoted_to_fp16_dtype_0"), val = string("fp16")]; + fp16 const_384_to_fp16 = const()[name = string("const_384_to_fp16"), val = fp16(-inf)]; + fp16 var_50_promoted_to_fp16 = const()[name = string("op_50_promoted_to_fp16"), val = fp16(0x1.5p+5)]; + tensor var_5548_to_fp16 = cast(dtype = var_5548_promoted_to_fp16_dtype_0, x = var_5548)[name = string("cast_10")]; + tensor clip_1_cast_fp16 = clip(alpha = const_384_to_fp16, beta = var_50_promoted_to_fp16, x = var_5548_to_fp16)[name = string("clip_1_cast_fp16")]; + int32 one_hot_1_batch_dims_0 = const()[name = string("one_hot_1_batch_dims_0"), val = int32(0)]; + bool one_hot_1_validate_indices_0 = const()[name = string("one_hot_1_validate_indices_0"), val = bool(false)]; + tensor to_onehot_identity_table_to_fp16 = const()[name = string("to_onehot_identity_table_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(555688384)))]; + string prompt_id_to_int16_dtype_0 = const()[name = string("prompt_id_to_int16_dtype_0"), val = string("int16")]; + string cast_230_dtype_0 = const()[name = string("cast_230_dtype_0"), val = string("int32")]; + int32 greater_equal_0_y_0 = const()[name = string("greater_equal_0_y_0"), val = int32(0)]; + tensor prompt_id_to_int16 = cast(dtype = prompt_id_to_int16_dtype_0, x = prompt_id)[name = string("cast_9")]; + tensor cast_230 = cast(dtype = cast_230_dtype_0, x = prompt_id_to_int16)[name = string("cast_8")]; + tensor greater_equal_0 = greater_equal(x = cast_230, y = greater_equal_0_y_0)[name = string("greater_equal_0")]; + int32 slice_by_index_2 = const()[name = string("slice_by_index_2"), val = int32(128)]; + tensor add_0 = add(x = cast_230, y = slice_by_index_2)[name = string("add_0")]; + tensor select_0 = select(a = cast_230, b = add_0, cond = greater_equal_0)[name = string("select_0")]; + string select_0_to_int16_dtype_0 = const()[name = string("select_0_to_int16_dtype_0"), val = string("int16")]; + string cast_0_dtype_0 = const()[name = string("cast_0_dtype_0"), val = string("int32")]; + int32 greater_equal_0_y_0_1 = const()[name = string("greater_equal_0_y_0_1"), val = int32(0)]; + tensor select_0_to_int16 = cast(dtype = select_0_to_int16_dtype_0, x = select_0)[name = string("cast_7")]; + tensor cast_0 = cast(dtype = cast_0_dtype_0, x = select_0_to_int16)[name = string("cast_6")]; + tensor greater_equal_0_1 = greater_equal(x = cast_0, y = greater_equal_0_y_0_1)[name = string("greater_equal_0_1")]; + int32 slice_by_index_0 = const()[name = string("slice_by_index_0"), val = int32(128)]; + tensor add_0_1 = add(x = cast_0, y = slice_by_index_0)[name = string("add_0_1")]; + tensor select_0_1 = select(a = cast_0, b = add_0_1, cond = greater_equal_0_1)[name = string("select_0_1")]; + int32 greater_equal_0_y_0_2 = const()[name = string("greater_equal_0_y_0_2"), val = int32(0)]; + tensor greater_equal_0_2 = greater_equal(x = select_0_1, y = greater_equal_0_y_0_2)[name = string("greater_equal_0_2")]; + int32 slice_by_index_0_1 = const()[name = string("slice_by_index_0_1"), val = int32(128)]; + tensor add_0_2 = add(x = select_0_1, y = slice_by_index_0_1)[name = string("add_0_2")]; + tensor select_0_2 = select(a = select_0_1, b = add_0_2, cond = greater_equal_0_2)[name = string("select_0_2")]; + int32 one_hot_1_cast_fp16_cast_uint16_cast_uint16_axis_0 = const()[name = string("one_hot_1_cast_fp16_cast_uint16_cast_uint16_axis_0"), val = int32(0)]; + tensor one_hot_1_cast_fp16_cast_uint16_cast_uint16 = gather(axis = one_hot_1_cast_fp16_cast_uint16_cast_uint16_axis_0, batch_dims = one_hot_1_batch_dims_0, indices = select_0_2, validate_indices = one_hot_1_validate_indices_0, x = to_onehot_identity_table_to_fp16)[name = string("one_hot_1_cast_fp16_cast_uint16_cast_uint16")]; + tensor var_5594_axes_0 = const()[name = string("op_5594_axes_0"), val = tensor([1])]; + tensor var_5594_cast_fp16 = expand_dims(axes = var_5594_axes_0, x = one_hot_1_cast_fp16_cast_uint16_cast_uint16)[name = string("op_5594_cast_fp16")]; + tensor one_hot_reps_0 = const()[name = string("one_hot_reps_0"), val = tensor([1, 14, 1])]; + tensor one_hot_cast_fp16 = tile(reps = one_hot_reps_0, x = var_5594_cast_fp16)[name = string("one_hot_cast_fp16")]; + int32 var_5603 = const()[name = string("op_5603"), val = int32(-1)]; + bool input_1273_interleave_0 = const()[name = string("input_1273_interleave_0"), val = bool(false)]; + tensor input_1273_cast_fp16 = concat(axis = var_5603, interleave = input_1273_interleave_0, values = (audio_signal_cast_fp16, one_hot_cast_fp16))[name = string("input_1273_cast_fp16")]; + tensor prompt_kernel_0_weight_to_fp16 = const()[name = string("prompt_kernel_0_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(555721216)))]; + tensor prompt_kernel_0_bias_to_fp16 = const()[name = string("prompt_kernel_0_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(560439872)))]; + tensor linear_217_cast_fp16 = linear(bias = prompt_kernel_0_bias_to_fp16, weight = prompt_kernel_0_weight_to_fp16, x = input_1273_cast_fp16)[name = string("linear_217_cast_fp16")]; + tensor input_1277_cast_fp16 = relu(x = linear_217_cast_fp16)[name = string("input_1277_cast_fp16")]; + tensor prompt_kernel_2_weight_to_fp16 = const()[name = string("prompt_kernel_2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(560444032)))]; + tensor prompt_kernel_2_bias_to_fp16 = const()[name = string("prompt_kernel_2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(564638400)))]; + tensor linear_218_cast_fp16 = linear(bias = prompt_kernel_2_bias_to_fp16, weight = prompt_kernel_2_weight_to_fp16, x = input_1277_cast_fp16)[name = string("linear_218_cast_fp16")]; + string conditioned_cast_fp16_to_fp32_dtype_0 = const()[name = string("conditioned_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + tensor transpose_72_perm_0_1 = const()[name = string("transpose_72_perm_0_1"), val = tensor([0, 2, 1])]; + string var_5621_dtype_0 = const()[name = string("op_5621_dtype_0"), val = string("int32")]; + tensor var_5624_perm_0 = const()[name = string("op_5624_perm_0"), val = tensor([1, 0, 2, 3])]; + string var_5624_cast_fp16_to_fp32_dtype_0 = const()[name = string("op_5624_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + tensor var_5627_perm_0 = const()[name = string("op_5627_perm_0"), val = tensor([1, 0, 2, 3])]; + string var_5627_cast_fp16_to_fp32_dtype_0 = const()[name = string("op_5627_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + string var_5632_dtype_0 = const()[name = string("op_5632_dtype_0"), val = string("int32")]; + tensor joint_enc_weight_to_fp16 = const()[name = string("joint_enc_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(564640512)))]; + tensor joint_enc_bias_to_fp16 = const()[name = string("joint_enc_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(565951296)))]; + tensor linear_219_cast_fp16 = linear(bias = joint_enc_bias_to_fp16, weight = joint_enc_weight_to_fp16, x = linear_218_cast_fp16)[name = string("linear_219_cast_fp16")]; + string linear_219_cast_fp16_to_fp32_dtype_0 = const()[name = string("linear_219_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + tensor encoder_proj = cast(dtype = linear_219_cast_fp16_to_fp32_dtype_0, x = linear_219_cast_fp16)[name = string("cast_0")]; + tensor cache_len_out = cast(dtype = var_5632_dtype_0, x = clip_1_cast_fp16)[name = string("cast_1")]; + tensor var_5627_cast_fp16 = transpose(perm = var_5627_perm_0, x = obj_7_cast_fp16)[name = string("transpose_144")]; + tensor cache_time_out = cast(dtype = var_5627_cast_fp16_to_fp32_dtype_0, x = var_5627_cast_fp16)[name = string("cast_2")]; + tensor var_5624_cast_fp16 = transpose(perm = var_5624_perm_0, x = obj_5_cast_fp16)[name = string("transpose_145")]; + tensor cache_channel_out = cast(dtype = var_5624_cast_fp16_to_fp32_dtype_0, x = var_5624_cast_fp16)[name = string("cast_3")]; + tensor encoded_length = cast(dtype = var_5621_dtype_0, x = clip_0_cast_fp16)[name = string("cast_4")]; + tensor transpose_72_1 = transpose(perm = transpose_72_perm_0_1, x = linear_218_cast_fp16)[name = string("transpose_146")]; + tensor encoded = cast(dtype = conditioned_cast_fp16_to_fp32_dtype_0, x = transpose_72_1)[name = string("cast_5")]; + } -> (encoded, encoded_length, cache_channel_out, cache_time_out, cache_len_out, encoder_proj); +} \ No newline at end of file diff --git a/it/1120ms/encoder.mlmodelc/weights/weight.bin b/it/1120ms/encoder.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..db719d65717ba13a2abcc1fd6915682c3eeef841 --- /dev/null +++ b/it/1120ms/encoder.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:011a5342cc1e2901633a8f0468be561f50dfc289ebe51851cddb4831f9a6a23f +size 565952640 diff --git a/it/1120ms/encoder.mlpackage/Data/com.apple.CoreML/model.mlmodel b/it/1120ms/encoder.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..d7e63d6d9ce257a9bcf94ee5f3694e8fe97c90ab --- /dev/null +++ b/it/1120ms/encoder.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:14655ea0dbccc666356247f05afc87bbea77ff469344bfb084526e00c0ce6d15 +size 804512 diff --git a/it/1120ms/encoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/it/1120ms/encoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..db719d65717ba13a2abcc1fd6915682c3eeef841 --- /dev/null +++ b/it/1120ms/encoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:011a5342cc1e2901633a8f0468be561f50dfc289ebe51851cddb4831f9a6a23f +size 565952640 diff --git a/it/1120ms/encoder.mlpackage/Manifest.json b/it/1120ms/encoder.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..343d61905a62576e3ef5ee2fbcc7f1a66f2a4aa2 --- /dev/null +++ b/it/1120ms/encoder.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "2A6DB86E-A930-4F02-B70E-CA03D7F3AD84": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + }, + "712F619E-4FA1-4165-A290-6C1344B0A816": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + } + }, + "rootModelIdentifier": "2A6DB86E-A930-4F02-B70E-CA03D7F3AD84" +} diff --git a/it/1120ms/joint.mlmodelc/analytics/coremldata.bin b/it/1120ms/joint.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..ee72f6da06176879d781b1a46f441fa01048d662 --- /dev/null +++ b/it/1120ms/joint.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cae47ef3b8cb372625305f4cdd31159b5ce56be1470c94613ff946be342f6d7a +size 243 diff --git a/it/1120ms/joint.mlmodelc/coremldata.bin b/it/1120ms/joint.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..07f143747ee2f43103809647f4058203bf60dc56 --- /dev/null +++ b/it/1120ms/joint.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:215d8dd2e33da0c37f08e6b0c0a1e997a3e056f2cb9113fdcf17f8027a61216d +size 341 diff --git a/it/1120ms/joint.mlmodelc/model.mil b/it/1120ms/joint.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..bfde40ec94bf61746424d2d3e196a4fba198de2d --- /dev/null +++ b/it/1120ms/joint.mlmodelc/model.mil @@ -0,0 +1,31 @@ +program(1.3) +[buildInfo = dict({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.5.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.3.0"}})] +{ + func main(tensor decoder, tensor encoder) { + tensor input_1_perm_0 = const()[name = string("input_1_perm_0"), val = tensor([0, 2, 1])]; + string encoder_to_fp16_dtype_0 = const()[name = string("encoder_to_fp16_dtype_0"), val = string("fp16")]; + tensor input_3_perm_0 = const()[name = string("input_3_perm_0"), val = tensor([0, 2, 1])]; + string decoder_to_fp16_dtype_0 = const()[name = string("decoder_to_fp16_dtype_0"), val = string("fp16")]; + tensor module_enc_weight_to_fp16 = const()[name = string("module_enc_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))]; + tensor module_enc_bias_to_fp16 = const()[name = string("module_enc_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1310848)))]; + tensor encoder_to_fp16 = cast(dtype = encoder_to_fp16_dtype_0, x = encoder)[name = string("cast_2")]; + tensor input_1_cast_fp16 = transpose(perm = input_1_perm_0, x = encoder_to_fp16)[name = string("transpose_1")]; + tensor linear_0_cast_fp16 = linear(bias = module_enc_bias_to_fp16, weight = module_enc_weight_to_fp16, x = input_1_cast_fp16)[name = string("linear_0_cast_fp16")]; + tensor module_pred_weight_to_fp16 = const()[name = string("module_pred_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1312192)))]; + tensor module_pred_bias_to_fp16 = const()[name = string("module_pred_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2131456)))]; + tensor decoder_to_fp16 = cast(dtype = decoder_to_fp16_dtype_0, x = decoder)[name = string("cast_1")]; + tensor input_3_cast_fp16 = transpose(perm = input_3_perm_0, x = decoder_to_fp16)[name = string("transpose_0")]; + tensor linear_1_cast_fp16 = linear(bias = module_pred_bias_to_fp16, weight = module_pred_weight_to_fp16, x = input_3_cast_fp16)[name = string("linear_1_cast_fp16")]; + tensor var_23_axes_0 = const()[name = string("op_23_axes_0"), val = tensor([2])]; + tensor var_23_cast_fp16 = expand_dims(axes = var_23_axes_0, x = linear_0_cast_fp16)[name = string("op_23_cast_fp16")]; + tensor var_25_axes_0 = const()[name = string("op_25_axes_0"), val = tensor([1])]; + tensor var_25_cast_fp16 = expand_dims(axes = var_25_axes_0, x = linear_1_cast_fp16)[name = string("op_25_cast_fp16")]; + tensor input_5_cast_fp16 = add(x = var_23_cast_fp16, y = var_25_cast_fp16)[name = string("input_5_cast_fp16")]; + tensor input_7_cast_fp16 = relu(x = input_5_cast_fp16)[name = string("input_7_cast_fp16")]; + tensor module_joint_net_2_weight_to_fp16 = const()[name = string("module_joint_net_2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2132800)))]; + tensor module_joint_net_2_bias_to_fp16 = const()[name = string("module_joint_net_2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3164544)))]; + tensor linear_2_cast_fp16 = linear(bias = module_joint_net_2_bias_to_fp16, weight = module_joint_net_2_weight_to_fp16, x = input_7_cast_fp16)[name = string("linear_2_cast_fp16")]; + string linear_2_cast_fp16_to_fp32_dtype_0 = const()[name = string("linear_2_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + tensor logits = cast(dtype = linear_2_cast_fp16_to_fp32_dtype_0, x = linear_2_cast_fp16)[name = string("cast_0")]; + } -> (logits); +} \ No newline at end of file diff --git a/it/1120ms/joint.mlmodelc/weights/weight.bin b/it/1120ms/joint.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..3673a1d2a1060ecda8626c76ffacdfd89c5f00c1 --- /dev/null +++ b/it/1120ms/joint.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1740c99cbe34ebeaa0163c8421135b4586df09960ef07fe02abb2a94b5693411 +size 3166220 diff --git a/it/1120ms/joint.mlpackage/Data/com.apple.CoreML/model.mlmodel b/it/1120ms/joint.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..f7f468ddd814131e36b8af9ed7a3358576bffcf0 --- /dev/null +++ b/it/1120ms/joint.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d902ae83b8a93f0f4240a4a6939466dbd1a6b2291f1615d81d7ac26d9115bc23 +size 4486 diff --git a/it/1120ms/joint.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/it/1120ms/joint.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..3673a1d2a1060ecda8626c76ffacdfd89c5f00c1 --- /dev/null +++ b/it/1120ms/joint.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1740c99cbe34ebeaa0163c8421135b4586df09960ef07fe02abb2a94b5693411 +size 3166220 diff --git a/it/1120ms/joint.mlpackage/Manifest.json b/it/1120ms/joint.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..fea9f1ee7eee62ace96d28134fe38a74b32b40c9 --- /dev/null +++ b/it/1120ms/joint.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "6D4EAD3B-A17D-4807-8572-3E0A513C0C7E": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + }, + "8242AF76-3606-4A98-8B3D-E0BB730201A4": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + } + }, + "rootModelIdentifier": "6D4EAD3B-A17D-4807-8572-3E0A513C0C7E" +} diff --git a/it/1120ms/joint_noencproj_batched.mlmodelc/analytics/coremldata.bin b/it/1120ms/joint_noencproj_batched.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..d2ffa5c9579dd51ce490b9c8fe47e3c27a05c723 --- /dev/null +++ b/it/1120ms/joint_noencproj_batched.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:24510a0ca90eea355a0e446ba54d63004a3f2c51d30eb3e415e4af938ac1acba +size 243 diff --git a/it/1120ms/joint_noencproj_batched.mlmodelc/coremldata.bin b/it/1120ms/joint_noencproj_batched.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..03999c10ebcacf3633f026f01254ed98ae34fa25 --- /dev/null +++ b/it/1120ms/joint_noencproj_batched.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dc93bb8f241754f1359837f597d600b029d687ce0c57a4280fa586f8c8386337 +size 406 diff --git a/it/1120ms/joint_noencproj_batched.mlmodelc/model.mil b/it/1120ms/joint_noencproj_batched.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..58942374262618031d39a52f7a009b81c7f24c24 --- /dev/null +++ b/it/1120ms/joint_noencproj_batched.mlmodelc/model.mil @@ -0,0 +1,26 @@ +program(1.3) +[buildInfo = dict({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.10.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor decoder, tensor encoder_proj) { + tensor input_1_perm_0 = const()[name = string("input_1_perm_0"), val = tensor([0, 2, 1])]; + string decoder_to_fp16_dtype_0 = const()[name = string("decoder_to_fp16_dtype_0"), val = string("fp16")]; + tensor joint_module_pred_weight_to_fp16 = const()[name = string("joint_module_pred_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))]; + tensor joint_module_pred_bias_to_fp16 = const()[name = string("joint_module_pred_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(819328)))]; + tensor decoder_to_fp16 = cast(dtype = decoder_to_fp16_dtype_0, x = decoder)[name = string("cast_2")]; + tensor input_1_cast_fp16 = transpose(perm = input_1_perm_0, x = decoder_to_fp16)[name = string("transpose_0")]; + tensor linear_0_cast_fp16 = linear(bias = joint_module_pred_bias_to_fp16, weight = joint_module_pred_weight_to_fp16, x = input_1_cast_fp16)[name = string("linear_0_cast_fp16")]; + tensor var_15_axes_0 = const()[name = string("op_15_axes_0"), val = tensor([2])]; + string encoder_proj_to_fp16_dtype_0 = const()[name = string("encoder_proj_to_fp16_dtype_0"), val = string("fp16")]; + tensor encoder_proj_to_fp16 = cast(dtype = encoder_proj_to_fp16_dtype_0, x = encoder_proj)[name = string("cast_1")]; + tensor var_15_cast_fp16 = expand_dims(axes = var_15_axes_0, x = encoder_proj_to_fp16)[name = string("op_15_cast_fp16")]; + tensor var_17_axes_0 = const()[name = string("op_17_axes_0"), val = tensor([1])]; + tensor var_17_cast_fp16 = expand_dims(axes = var_17_axes_0, x = linear_0_cast_fp16)[name = string("op_17_cast_fp16")]; + tensor input_3_cast_fp16 = add(x = var_15_cast_fp16, y = var_17_cast_fp16)[name = string("input_3_cast_fp16")]; + tensor input_5_cast_fp16 = relu(x = input_3_cast_fp16)[name = string("input_5_cast_fp16")]; + tensor joint_module_joint_net_2_weight_to_fp16 = const()[name = string("joint_module_joint_net_2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(820672)))]; + tensor joint_module_joint_net_2_bias_to_fp16 = const()[name = string("joint_module_joint_net_2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1852416)))]; + tensor linear_1_cast_fp16 = linear(bias = joint_module_joint_net_2_bias_to_fp16, weight = joint_module_joint_net_2_weight_to_fp16, x = input_5_cast_fp16)[name = string("linear_1_cast_fp16")]; + string linear_1_cast_fp16_to_fp32_dtype_0 = const()[name = string("linear_1_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + tensor logits = cast(dtype = linear_1_cast_fp16_to_fp32_dtype_0, x = linear_1_cast_fp16)[name = string("cast_0")]; + } -> (logits); +} \ No newline at end of file diff --git a/it/1120ms/joint_noencproj_batched.mlmodelc/weights/weight.bin b/it/1120ms/joint_noencproj_batched.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..40ceadd152241059aa378e2ddb6cc9f649e0b59c --- /dev/null +++ b/it/1120ms/joint_noencproj_batched.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cd83e82dcfec315f28c8a8872b0d7f22e668a2c485821de86a0379ae2b3864ad +size 1854092 diff --git a/it/1120ms/joint_noencproj_batched.mlpackage/Data/com.apple.CoreML/model.mlmodel b/it/1120ms/joint_noencproj_batched.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..49f1fd816a2725ab32d473dfe600f911656f675b --- /dev/null +++ b/it/1120ms/joint_noencproj_batched.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:26a19f2e269811070e57eda978e8226ca9dcca6177b46b95f916538b5f815716 +size 3839 diff --git a/it/1120ms/joint_noencproj_batched.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/it/1120ms/joint_noencproj_batched.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..40ceadd152241059aa378e2ddb6cc9f649e0b59c --- /dev/null +++ b/it/1120ms/joint_noencproj_batched.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cd83e82dcfec315f28c8a8872b0d7f22e668a2c485821de86a0379ae2b3864ad +size 1854092 diff --git a/it/1120ms/joint_noencproj_batched.mlpackage/Manifest.json b/it/1120ms/joint_noencproj_batched.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..e3857a03022f3cec16823ed1d8a741fee56cfce4 --- /dev/null +++ b/it/1120ms/joint_noencproj_batched.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "4233CE8E-FB95-4FF9-BCD8-2A834D55C580": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + }, + "96E0F26C-90DC-49EE-B510-D0FB3FC812CC": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + } + }, + "rootModelIdentifier": "96E0F26C-90DC-49EE-B510-D0FB3FC812CC" +} diff --git a/it/1120ms/metadata.json b/it/1120ms/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..0e6e0df0226852fce0ab88bf3e12e3ef09751d73 --- /dev/null +++ b/it/1120ms/metadata.json @@ -0,0 +1,198 @@ +{ + "model": "nvidia/nemotron-asr-streaming-multilingual-0.6b", + "model_class": "nemo.collections.asr.models.rnnt_bpe_models_prompt.EncDecRNNTBPEModelWithPrompt", + "sample_rate": 16000, + "mel_features": 128, + "chunk_mel_frames": 112, + "pre_encode_cache": 9, + "total_mel_frames": 121, + "att_context_size": [ + 42, + 13 + ], + "vocab_size": 805, + "blank_idx": 805, + "vocab_pruned": true, + "vocab_pruned_original_size": 13087, + "cache_channel_shape": [ + 1, + 24, + 42, + 1024 + ], + "cache_time_shape": [ + 1, + 24, + 1024, + 8 + ], + "decoder_hidden": 640, + "decoder_layers": 2, + "encoder_dim": 1024, + "num_prompts": 128, + "prompt_dictionary": { + "af-ZA": 54, + "am-ET": 49, + "ar": 7, + "ar-AR": 7, + "auto": 101, + "ay-BO": 81, + "az-AZ": 66, + "bg": 30, + "bg-BG": 30, + "bn-IN": 36, + "cs": 22, + "cs-CZ": 22, + "da": 25, + "da-DK": 25, + "de": 9, + "de-DE": 9, + "el": 21, + "el-GR": 21, + "en": 0, + "en-GB": 1, + "en-US": 0, + "enGB": 1, + "es": 3, + "es-ES": 2, + "es-US": 3, + "esES": 2, + "et": 60, + "et-EE": 60, + "fa-IR": 38, + "fi": 26, + "fi-FI": 26, + "fr": 8, + "fr-CA": 100, + "fr-FR": 8, + "gn-PY": 82, + "gu-IN": 42, + "ha-NG": 50, + "haw-US": 97, + "he-IL": 64, + "hi": 6, + "hi-HI": 6, + "hi-IN": 6, + "hr": 29, + "hr-HR": 29, + "hu": 23, + "hu-HU": 23, + "hy-AM": 68, + "id-ID": 34, + "ig-NG": 53, + "it": 15, + "it-IT": 15, + "ja-JA": 10, + "ja-JP": 10, + "ka-GE": 67, + "km-KH": 47, + "kn-IN": 43, + "ko": 14, + "ko-KO": 14, + "ko-KR": 14, + "ku-TR": 65, + "ky-KG": 71, + "ln-CD": 58, + "lt": 31, + "lt-LT": 31, + "lv": 61, + "lv-LV": 61, + "mi-NZ": 96, + "ml-IN": 44, + "mr-IN": 41, + "ms-MY": 35, + "mt-MT": 102, + "nah-MX": 83, + "nb": 103, + "nb-NO": 103, + "ne-NP": 46, + "nl": 16, + "nl-NL": 16, + "nn": 104, + "nn-NO": 104, + "no": 27, + "no-NO": 27, + "ny-MW": 57, + "or-KE": 59, + "pl": 17, + "pl-PL": 17, + "pt": 13, + "pt-BR": 12, + "pt-PT": 13, + "qu-PE": 80, + "ro": 20, + "ro-RO": 20, + "ru": 11, + "ru-RU": 11, + "rw-RW": 55, + "si-LK": 45, + "sk": 28, + "sk-SK": 28, + "sl": 62, + "sl-SI": 62, + "sm-WS": 98, + "so-SO": 56, + "sv": 24, + "sv-SE": 24, + "sw-KE": 48, + "ta-IN": 39, + "te-IN": 40, + "tg-TJ": 70, + "th-TH": 32, + "to-TO": 99, + "tr": 18, + "tr-TR": 18, + "uk": 19, + "uk-UA": 19, + "ur-PK": 37, + "uz-UZ": 69, + "vi-VN": 33, + "yo-NG": 52, + "zh-CN": 4, + "zh-TW": 5, + "zh-ZH": 4, + "zu-ZA": 51 + }, + "default_prompt_id": 101, + "lang_tag_token_ids": [ + 1, + 63, + 115, + 167, + 226, + 227, + 259, + 276, + 328, + 353, + 368, + 462, + 481, + 499, + 518, + 542, + 571, + 602, + 603, + 612, + 624, + 646, + 647, + 667, + 689, + 699, + 720, + 727, + 747, + 748, + 750, + 751, + 752, + 756, + 774, + 787, + 788, + 801, + 802 + ] +} \ No newline at end of file diff --git a/it/1120ms/preprocessor.mlmodelc/analytics/coremldata.bin b/it/1120ms/preprocessor.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..caf6367f51e931ec1ff5cc630cadd6c50bd7d4ab --- /dev/null +++ b/it/1120ms/preprocessor.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:97be07beee7a6a28b7db85d47087d5b018ebcd1fc0b1565707141d574244bdc9 +size 243 diff --git a/it/1120ms/preprocessor.mlmodelc/coremldata.bin b/it/1120ms/preprocessor.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..0944be2be6b80940760d1f5f5f0f11ac817288bb --- /dev/null +++ b/it/1120ms/preprocessor.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f7b11e08aba46d1845d8ad3f247717e0f6fae35b21d71d52e44a69ea73587bfe +size 371 diff --git a/it/1120ms/preprocessor.mlmodelc/model.mil b/it/1120ms/preprocessor.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..0b8261362f9cbf465b530a0d2d0ee9a2b2f462cd --- /dev/null +++ b/it/1120ms/preprocessor.mlmodelc/model.mil @@ -0,0 +1,122 @@ +program(1.3) +[buildInfo = dict({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.5.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.3.0"}})] +{ + func main(tensor audio, tensor audio_length) [FlexibleShapeInformation = tuple>>, tuple, ?>>>>((("DefaultShapes", {{"audio", [1, 1]}}), ("RangeDims", {{"audio", [[1, 1], [1, 480000]]}})))] { + int32 var_9 = const()[name = string("op_9"), val = int32(1)]; + int32 var_10 = const()[name = string("op_10"), val = int32(160)]; + int32 var_12 = const()[name = string("op_12"), val = int32(0)]; + int32 var_33 = const()[name = string("op_33"), val = int32(512)]; + tensor var_34 = add(x = audio_length, y = var_33)[name = string("op_34")]; + int32 var_35 = const()[name = string("op_35"), val = int32(512)]; + tensor var_36 = sub(x = var_34, y = var_35)[name = string("op_36")]; + tensor floor_div_0 = floor_div(x = var_36, y = var_10)[name = string("floor_div_0")]; + tensor var_39 = equal(x = audio_length, y = var_12)[name = string("op_39")]; + tensor var_40 = const()[name = string("op_40"), val = tensor([0])]; + tensor mel_length = select(a = var_40, b = floor_div_0, cond = var_39)[name = string("seq_len")]; + string audio_to_fp16_dtype_0 = const()[name = string("audio_to_fp16_dtype_0"), val = string("fp16")]; + tensor audio_to_fp16 = cast(dtype = audio_to_fp16_dtype_0, x = audio)[name = string("cast_14")]; + tensor var_42_shape_cast_fp16 = shape(x = audio_to_fp16)[name = string("op_42_shape_cast_fp16")]; + int32 gather_0_axis_0 = const()[name = string("gather_0_axis_0"), val = int32(0)]; + int32 gather_0_batch_dims_0 = const()[name = string("gather_0_batch_dims_0"), val = int32(0)]; + bool gather_0_validate_indices_0 = const()[name = string("gather_0_validate_indices_0"), val = bool(false)]; + string var_42_shape_cast_fp16_to_int16_dtype_0 = const()[name = string("op_42_shape_cast_fp16_to_int16_dtype_0"), val = string("int16")]; + uint16 select_0_to_uint16 = const()[name = string("select_0_to_uint16"), val = uint16(1)]; + tensor var_42_shape_cast_fp16_to_int16 = cast(dtype = var_42_shape_cast_fp16_to_int16_dtype_0, x = var_42_shape_cast_fp16)[name = string("cast_13")]; + int16 gather_0_cast_uint16 = gather(axis = gather_0_axis_0, batch_dims = gather_0_batch_dims_0, indices = select_0_to_uint16, validate_indices = gather_0_validate_indices_0, x = var_42_shape_cast_fp16_to_int16)[name = string("gather_0_cast_uint16")]; + string gather_0_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_0_cast_uint16_to_int32_dtype_0"), val = string("int32")]; + int32 const_0 = const()[name = string("const_0"), val = int32(0)]; + int32 const_1 = const()[name = string("const_1"), val = int32(1)]; + int32 gather_0_cast_uint16_to_int32 = cast(dtype = gather_0_cast_uint16_to_int32_dtype_0, x = gather_0_cast_uint16)[name = string("cast_12")]; + tensor var_43 = range_1d(end = gather_0_cast_uint16_to_int32, start = const_0, step = const_1)[name = string("op_43")]; + tensor var_44_axes_0 = const()[name = string("op_44_axes_0"), val = tensor([0])]; + tensor var_44 = expand_dims(axes = var_44_axes_0, x = var_43)[name = string("op_44")]; + tensor var_45_axes_0 = const()[name = string("op_45_axes_0"), val = tensor([1])]; + tensor var_45 = expand_dims(axes = var_45_axes_0, x = audio_length)[name = string("op_45")]; + tensor timemask = less(x = var_44, y = var_45)[name = string("timemask")]; + tensor var_48_begin_0 = const()[name = string("op_48_begin_0"), val = tensor([0, 0])]; + tensor var_48_end_0 = const()[name = string("op_48_end_0"), val = tensor([1, 1])]; + tensor var_48_end_mask_0 = const()[name = string("op_48_end_mask_0"), val = tensor([true, false])]; + tensor var_48_squeeze_mask_0 = const()[name = string("op_48_squeeze_mask_0"), val = tensor([false, true])]; + tensor var_48_cast_fp16 = slice_by_index(begin = var_48_begin_0, end = var_48_end_0, end_mask = var_48_end_mask_0, squeeze_mask = var_48_squeeze_mask_0, x = audio_to_fp16)[name = string("op_48_cast_fp16")]; + tensor var_49_axes_0 = const()[name = string("op_49_axes_0"), val = tensor([1])]; + tensor var_49_cast_fp16 = expand_dims(axes = var_49_axes_0, x = var_48_cast_fp16)[name = string("op_49_cast_fp16")]; + tensor var_51_begin_0 = const()[name = string("op_51_begin_0"), val = tensor([0, 1])]; + tensor var_51_end_0 = const()[name = string("op_51_end_0"), val = tensor([1, 0])]; + tensor var_51_end_mask_0 = const()[name = string("op_51_end_mask_0"), val = tensor([true, true])]; + tensor var_51_cast_fp16 = slice_by_index(begin = var_51_begin_0, end = var_51_end_0, end_mask = var_51_end_mask_0, x = audio_to_fp16)[name = string("op_51_cast_fp16")]; + tensor var_53_begin_0 = const()[name = string("op_53_begin_0"), val = tensor([0, 0])]; + tensor var_53_end_0 = const()[name = string("op_53_end_0"), val = tensor([1, -1])]; + tensor var_53_end_mask_0 = const()[name = string("op_53_end_mask_0"), val = tensor([true, false])]; + tensor var_53_cast_fp16 = slice_by_index(begin = var_53_begin_0, end = var_53_end_0, end_mask = var_53_end_mask_0, x = audio_to_fp16)[name = string("op_53_cast_fp16")]; + fp16 var_54_to_fp16 = const()[name = string("op_54_to_fp16"), val = fp16(0x1.f0cp-1)]; + tensor var_55_cast_fp16 = mul(x = var_53_cast_fp16, y = var_54_to_fp16)[name = string("op_55_cast_fp16")]; + tensor var_56_cast_fp16 = sub(x = var_51_cast_fp16, y = var_55_cast_fp16)[name = string("op_56_cast_fp16")]; + bool x_3_interleave_0 = const()[name = string("x_3_interleave_0"), val = bool(false)]; + tensor x_3_cast_fp16 = concat(axis = var_9, interleave = x_3_interleave_0, values = (var_49_cast_fp16, var_56_cast_fp16))[name = string("x_3_cast_fp16")]; + tensor var_59 = logical_not(x = timemask)[name = string("op_59")]; + fp16 var_16_to_fp16 = const()[name = string("op_16_to_fp16"), val = fp16(0x0p+0)]; + tensor input_1_cast_fp16 = select(a = var_16_to_fp16, b = x_3_cast_fp16, cond = var_59)[name = string("input_1_cast_fp16")]; + tensor concat_1x = const()[name = string("concat_1x"), val = tensor([1, 1, -1])]; + tensor input_3_cast_fp16 = reshape(shape = concat_1x, x = input_1_cast_fp16)[name = string("input_3_cast_fp16")]; + tensor input_5_pad_0 = const()[name = string("input_5_pad_0"), val = tensor([0, 0, 0, 0, 256, 256])]; + string input_5_mode_0 = const()[name = string("input_5_mode_0"), val = string("constant")]; + fp16 const_3_to_fp16 = const()[name = string("const_3_to_fp16"), val = fp16(0x0p+0)]; + tensor input_5_cast_fp16 = pad(constant_val = const_3_to_fp16, mode = input_5_mode_0, pad = input_5_pad_0, x = input_3_cast_fp16)[name = string("input_5_cast_fp16")]; + tensor concat_2x = const()[name = string("concat_2x"), val = tensor([1, -1])]; + tensor input_cast_fp16 = reshape(shape = concat_2x, x = input_5_cast_fp16)[name = string("input_cast_fp16")]; + tensor expand_dims_3 = const()[name = string("expand_dims_3"), val = tensor([160])]; + tensor expand_dims_4_axes_0 = const()[name = string("expand_dims_4_axes_0"), val = tensor([1])]; + tensor expand_dims_4_cast_fp16 = expand_dims(axes = expand_dims_4_axes_0, x = input_cast_fp16)[name = string("expand_dims_4_cast_fp16")]; + string conv_0_pad_type_0 = const()[name = string("conv_0_pad_type_0"), val = string("valid")]; + tensor conv_0_pad_0 = const()[name = string("conv_0_pad_0"), val = tensor([0, 0])]; + tensor conv_0_dilations_0 = const()[name = string("conv_0_dilations_0"), val = tensor([1])]; + int32 conv_0_groups_0 = const()[name = string("conv_0_groups_0"), val = int32(1)]; + tensor expand_dims_1_to_fp16 = const()[name = string("expand_dims_1_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))]; + tensor conv_0_cast_fp16 = conv(dilations = conv_0_dilations_0, groups = conv_0_groups_0, pad = conv_0_pad_0, pad_type = conv_0_pad_type_0, strides = expand_dims_3, weight = expand_dims_1_to_fp16, x = expand_dims_4_cast_fp16)[name = string("conv_0_cast_fp16")]; + string conv_1_pad_type_0 = const()[name = string("conv_1_pad_type_0"), val = string("valid")]; + tensor conv_1_pad_0 = const()[name = string("conv_1_pad_0"), val = tensor([0, 0])]; + tensor conv_1_dilations_0 = const()[name = string("conv_1_dilations_0"), val = tensor([1])]; + int32 conv_1_groups_0 = const()[name = string("conv_1_groups_0"), val = int32(1)]; + tensor expand_dims_2_to_fp16 = const()[name = string("expand_dims_2_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(263296)))]; + tensor conv_1_cast_fp16 = conv(dilations = conv_1_dilations_0, groups = conv_1_groups_0, pad = conv_1_pad_0, pad_type = conv_1_pad_type_0, strides = expand_dims_3, weight = expand_dims_2_to_fp16, x = expand_dims_4_cast_fp16)[name = string("conv_1_cast_fp16")]; + int32 stack_0_axis_0 = const()[name = string("stack_0_axis_0"), val = int32(-1)]; + tensor stack_0_cast_fp16 = stack(axis = stack_0_axis_0, values = (conv_0_cast_fp16, conv_1_cast_fp16))[name = string("stack_0_cast_fp16")]; + fp16 var_19_promoted_to_fp16 = const()[name = string("op_19_promoted_to_fp16"), val = fp16(0x1p+1)]; + tensor var_74_cast_fp16 = pow(x = stack_0_cast_fp16, y = var_19_promoted_to_fp16)[name = string("op_74_cast_fp16")]; + tensor var_76_axes_0 = const()[name = string("op_76_axes_0"), val = tensor([-1])]; + bool var_76_keep_dims_0 = const()[name = string("op_76_keep_dims_0"), val = bool(false)]; + tensor var_76_cast_fp16 = reduce_sum(axes = var_76_axes_0, keep_dims = var_76_keep_dims_0, x = var_74_cast_fp16)[name = string("op_76_cast_fp16")]; + tensor x_11_cast_fp16 = identity(x = var_76_cast_fp16)[name = string("x_11_cast_fp16")]; + bool x_13_transpose_x_0 = const()[name = string("x_13_transpose_x_0"), val = bool(false)]; + bool x_13_transpose_y_0 = const()[name = string("x_13_transpose_y_0"), val = bool(false)]; + tensor const_4_to_fp16 = const()[name = string("const_4_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(526528)))]; + tensor x_13_cast_fp16 = matmul(transpose_x = x_13_transpose_x_0, transpose_y = x_13_transpose_y_0, x = const_4_to_fp16, y = x_11_cast_fp16)[name = string("x_13_cast_fp16")]; + fp16 var_83_to_fp16 = const()[name = string("op_83_to_fp16"), val = fp16(0x1p-24)]; + tensor var_84_cast_fp16 = add(x = x_13_cast_fp16, y = var_83_to_fp16)[name = string("op_84_cast_fp16")]; + fp32 x_epsilon_0 = const()[name = string("x_epsilon_0"), val = fp32(0x1p-149)]; + tensor x_cast_fp16 = log(epsilon = x_epsilon_0, x = var_84_cast_fp16)[name = string("x_cast_fp16")]; + tensor var_86_shape_cast_fp16 = shape(x = x_cast_fp16)[name = string("op_86_shape_cast_fp16")]; + int32 gather_5_axis_0 = const()[name = string("gather_5_axis_0"), val = int32(0)]; + int32 gather_5_batch_dims_0 = const()[name = string("gather_5_batch_dims_0"), val = int32(0)]; + bool gather_5_validate_indices_0 = const()[name = string("gather_5_validate_indices_0"), val = bool(false)]; + string var_86_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_86_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")]; + uint16 select_5_to_uint16 = const()[name = string("select_5_to_uint16"), val = uint16(2)]; + tensor var_86_shape_cast_fp16_to_uint16 = cast(dtype = var_86_shape_cast_fp16_to_uint16_dtype_0, x = var_86_shape_cast_fp16)[name = string("cast_11")]; + uint16 gather_5_cast_uint16 = gather(axis = gather_5_axis_0, batch_dims = gather_5_batch_dims_0, indices = select_5_to_uint16, validate_indices = gather_5_validate_indices_0, x = var_86_shape_cast_fp16_to_uint16)[name = string("gather_5_cast_uint16")]; + string gather_5_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_5_cast_uint16_to_int32_dtype_0"), val = string("int32")]; + int32 const_5 = const()[name = string("const_5"), val = int32(0)]; + int32 const_6 = const()[name = string("const_6"), val = int32(1)]; + int32 gather_5_cast_uint16_to_int32 = cast(dtype = gather_5_cast_uint16_to_int32_dtype_0, x = gather_5_cast_uint16)[name = string("cast_10")]; + tensor mask_1 = range_1d(end = gather_5_cast_uint16_to_int32, start = const_5, step = const_6)[name = string("mask_1")]; + tensor expand_dims_0_axes_0 = const()[name = string("expand_dims_0_axes_0"), val = tensor([0])]; + tensor expand_dims_0 = expand_dims(axes = expand_dims_0_axes_0, x = mask_1)[name = string("expand_dims_0")]; + tensor var_91_axes_0 = const()[name = string("op_91_axes_0"), val = tensor([1])]; + tensor var_91 = expand_dims(axes = var_91_axes_0, x = mel_length)[name = string("op_91")]; + tensor mask = greater_equal(x = expand_dims_0, y = var_91)[name = string("mask")]; + tensor var_93_axes_0 = const()[name = string("op_93_axes_0"), val = tensor([1])]; + tensor var_93 = expand_dims(axes = var_93_axes_0, x = mask)[name = string("op_93")]; + tensor processed_signal_cast_fp16 = select(a = var_16_to_fp16, b = x_cast_fp16, cond = var_93)[name = string("processed_signal_cast_fp16")]; + string processed_signal_cast_fp16_to_fp32_dtype_0 = const()[name = string("processed_signal_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + tensor mel = cast(dtype = processed_signal_cast_fp16_to_fp32_dtype_0, x = processed_signal_cast_fp16)[name = string("cast_9")]; + } -> (mel, mel_length); +} \ No newline at end of file diff --git a/it/1120ms/preprocessor.mlmodelc/weights/weight.bin b/it/1120ms/preprocessor.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..86dd375f6649d262d58c9cf8b89006ceab216411 --- /dev/null +++ b/it/1120ms/preprocessor.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:297514e2b211d14b0e53cb97193d679bb89ead98d28e578f3f1d049ddbcc36b3 +size 592384 diff --git a/it/1120ms/preprocessor.mlpackage/Data/com.apple.CoreML/model.mlmodel b/it/1120ms/preprocessor.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..a760b77aa823c74ebff0487b3aea510a88447fd7 --- /dev/null +++ b/it/1120ms/preprocessor.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f1df1cad6f43ce4fa090e0f9a33bb5ddf25a0aaeca2be136339878b5b9de45c0 +size 15878 diff --git a/it/1120ms/preprocessor.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/it/1120ms/preprocessor.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..86dd375f6649d262d58c9cf8b89006ceab216411 --- /dev/null +++ b/it/1120ms/preprocessor.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:297514e2b211d14b0e53cb97193d679bb89ead98d28e578f3f1d049ddbcc36b3 +size 592384 diff --git a/it/1120ms/preprocessor.mlpackage/Manifest.json b/it/1120ms/preprocessor.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..d289893c92607e581c8cdaebf26772223814cb0d --- /dev/null +++ b/it/1120ms/preprocessor.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "162F4F34-AB45-4D09-8F9D-2F76AAE79E0F": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + }, + "B5F24360-A5E1-4F41-9872-3D77094622F0": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + } + }, + "rootModelIdentifier": "B5F24360-A5E1-4F41-9872-3D77094622F0" +} diff --git a/it/1120ms/tokenizer.json b/it/1120ms/tokenizer.json new file mode 100644 index 0000000000000000000000000000000000000000..033007249c80c55ca310e0c90339004b867968f6 --- /dev/null +++ b/it/1120ms/tokenizer.json @@ -0,0 +1,808 @@ +{ + "0": "", + "1": "", + "2": "▁", + "3": ".", + "4": ",", + "5": "e", + "6": "t", + "7": "a", + "8": "s", + "9": "o", + "10": "i", + "11": "r", + "12": "l", + "13": "u", + "14": "d", + "15": "c", + "16": "h", + "17": "m", + "18": "p", + "19": "n", + "20": "g", + "21": "f", + "22": "en", + "23": "in", + "24": "on", + "25": "y", + "26": "er", + "27": "an", + "28": "w", + "29": "0", + "30": "b", + "31": "v", + "32": "2", + "33": "1", + "34": "k", + "35": "3", + "36": "5", + "37": "q", + "38": "I", + "39": "é", + "40": "4", + "41": "z", + "42": "A", + "43": "6", + "44": "j", + "45": "E", + "46": "7", + "47": "8", + "48": "9", + "49": "S", + "50": "x", + "51": "C", + "52": "P", + "53": "U", + "54": "N", + "55": "O", + "56": "à", + "57": "L", + "58": "è", + "59": "V", + "60": "R", + "61": "ó", + "62": "J", + "63": "", + "64": "st", + "65": "ch", + "66": "le", + "67": "li", + "68": "▁po", + "69": "no", + "70": "to", + "71": "me", + "72": "te", + "73": "ho", + "74": "▁pro", + "75": "ro", + "76": "▁na", + "77": "ce", + "78": "la", + "79": "ni", + "80": "ra", + "81": "ti", + "82": "lo", + "83": "ko", + "84": "po", + "85": "je", + "86": "de", + "87": "na", + "88": "mi", + "89": "ci", + "90": "▁by", + "91": "ve", + "92": "▁za", + "93": "▁A", + "94": "re", + "95": "ou", + "96": "vo", + "97": "né", + "98": "va", + "99": "mo", + "100": "ze", + "101": "ne", + "102": "ka", + "103": "ky", + "104": "ovat", + "105": "▁jak", + "106": "ny", + "107": "vi", + "108": "ent", + "109": "prav", + "110": "Z", + "111": "▁Je", + "112": "oval", + "113": "ă", + "114": "X", + "115": "", + "116": "▁for", + "117": "▁det", + "118": "▁at", + "119": "et", + "120": "▁og", + "121": "▁vi", + "122": "al", + "123": "▁de", + "124": "▁der", + "125": "or", + "126": "om", + "127": "and", + "128": "▁har", + "129": "at", + "130": "▁af", + "131": "ge", + "132": "ar", + "133": "is", + "134": "▁med", + "135": "▁be", + "136": "un", + "137": "lig", + "138": "▁man", + "139": "ig", + "140": "▁som", + "141": "▁Og", + "142": "el", + "143": "ag", + "144": "erne", + "145": "▁den", + "146": "ste", + "147": "id", + "148": "▁kan", + "149": "ske", + "150": "iv", + "151": "ion", + "152": "am", + "153": "ur", + "154": "for", + "155": "else", + "156": "▁sig", + "157": "▁ind", + "158": "ende", + "159": "▁Vi", + "160": "ation", + "161": "mme", + "162": "▁op", + "163": "▁fra", + "164": "▁alle", + "165": "▁Men", + "166": "▁var", + "167": "", + "168": "▁die", + "169": "▁und", + "170": "sch", + "171": "it", + "172": "gen", + "173": "▁W", + "174": "▁B", + "175": "▁E", + "176": "▁F", + "177": "ll", + "178": "▁es", + "179": "▁K", + "180": "ie", + "181": "au", + "182": "ich", + "183": "ck", + "184": "ten", + "185": "mal", + "186": "ein", + "187": "▁T", + "188": "▁dann", + "189": "▁mit", + "190": "ter", + "191": "tz", + "192": "▁G", + "193": "ben", + "194": "um", + "195": "us", + "196": "il", + "197": "ut", + "198": "▁ver", + "199": "ri", + "200": "ach", + "201": "ol", + "202": "▁Da", + "203": "sp", + "204": "ell", + "205": "▁was", + "206": "▁ja", + "207": "wi", + "208": "rei", + "209": "▁Ge", + "210": "und", + "211": "▁St", + "212": "▁sie", + "213": "▁Ja", + "214": "▁du", + "215": "▁Sch", + "216": "▁Ma", + "217": "▁De", + "218": "▁Sie", + "219": "▁vor", + "220": "▁Le", + "221": "▁In", + "222": "▁Ver", + "223": "▁Re", + "224": "▁Mi", + "225": "▁Ha", + "226": "", + "227": "", + "228": "ma", + "229": "ta", + "230": "se", + "231": "da", + "232": "si", + "233": "ks", + "234": "ga", + "235": "he", + "236": "mu", + "237": "tu", + "238": "ha", + "239": "ja", + "240": "gi", + "241": "gu", + "242": "ju", + "243": "est", + "244": "▁pa", + "245": "tud", + "246": "nda", + "247": "ke", + "248": "sta", + "249": "sed", + "250": "di", + "251": "▁su", + "252": "ide", + "253": "pool", + "254": "val", + "255": "▁Me", + "256": "ment", + "257": "ndus", + "258": "Q", + "259": "", + "260": "ssa", + "261": "lla", + "262": "ki", + "263": "pa", + "264": "lle", + "265": "lu", + "266": "tta", + "267": "isi", + "268": "tte", + "269": "ista", + "270": "llis", + "271": "vu", + "272": "▁voi", + "273": "utta", + "274": "iden", + "275": "parlament", + "276": "", + "277": "▁est", + "278": "▁c", + "279": "▁que", + "280": "es", + "281": "▁un", + "282": "▁pas", + "283": "▁qui", + "284": "▁il", + "285": "▁des", + "286": "▁qu", + "287": "▁par", + "288": "ant", + "289": "▁C", + "290": "tre", + "291": "ir", + "292": "elle", + "293": "eur", + "294": "▁sur", + "295": "▁con", + "296": "ement", + "297": "tion", + "298": "mp", + "299": "▁comme", + "300": "ac", + "301": "▁là", + "302": "che", + "303": "que", + "304": "ul", + "305": "▁bien", + "306": "age", + "307": "▁mon", + "308": "end", + "309": "ver", + "310": "tra", + "311": "ille", + "312": "ff", + "313": "▁ex", + "314": "▁Il", + "315": "im", + "316": "▁dire", + "317": "ance", + "318": "aire", + "319": "▁app", + "320": "onne", + "321": "mb", + "322": "man", + "323": "▁quand", + "324": "port", + "325": "form", + "326": "ture", + "327": "ù", + "328": "", + "329": "ok", + "330": "gy", + "331": "ek", + "332": "em", + "333": "▁is", + "334": "os", + "335": "ak", + "336": "ban", + "337": "ik", + "338": "▁nem", + "339": "oz", + "340": "cs", + "341": "nek", + "342": "bb", + "343": "▁ha", + "344": "ott", + "345": "▁van", + "346": "▁fel", + "347": "leg", + "348": "▁ami", + "349": "tart", + "350": "rend", + "351": "▁fog", + "352": "▁volt", + "353": "", + "354": "▁bi", + "355": "▁sa", + "356": "ru", + "357": "go", + "358": "sti", + "359": "▁pri", + "360": "ima", + "361": "nu", + "362": "▁pre", + "363": "zi", + "364": "vr", + "365": "ca", + "366": "ba", + "367": "▁raz", + "368": "", + "369": "▁di", + "370": "▁che", + "371": "▁è", + "372": "co", + "373": "▁per", + "374": "▁non", + "375": "do", + "376": "gli", + "377": "so", + "378": "amo", + "379": "sa", + "380": "ndo", + "381": "▁una", + "382": "fi", + "383": "pi", + "384": "nti", + "385": "tto", + "386": "tro", + "387": "▁fa", + "388": "chi", + "389": "▁qua", + "390": "zione", + "391": "bi", + "392": "▁del", + "393": "mente", + "394": "pe", + "395": "ssi", + "396": "▁sono", + "397": "▁questo", + "398": "nte", + "399": "tti", + "400": "tà", + "401": "▁nel", + "402": "▁anche", + "403": "sso", + "404": "▁perché", + "405": "▁più", + "406": "nta", + "407": "▁come", + "408": "cu", + "409": "▁quindi", + "410": "ggi", + "411": "nza", + "412": "sto", + "413": "ò", + "414": "▁della", + "415": "gra", + "416": "▁fare", + "417": "spe", + "418": "cco", + "419": "nde", + "420": "mento", + "421": "fe", + "422": "gio", + "423": "pu", + "424": "▁questa", + "425": "zza", + "426": "sci", + "427": "▁dei", + "428": "▁poi", + "429": "sco", + "430": "stra", + "431": "▁quel", + "432": "qui", + "433": "▁delle", + "434": "▁cosa", + "435": "▁molto", + "436": "sse", + "437": "zioni", + "438": "▁inter", + "439": "sce", + "440": "▁fatto", + "441": "▁com", + "442": "▁quello", + "443": "▁essere", + "444": "▁due", + "445": "▁abbiamo", + "446": "▁comp", + "447": "▁tutti", + "448": "ì", + "449": "▁prima", + "450": "▁parte", + "451": "▁così", + "452": "▁sempre", + "453": "▁tutto", + "454": "▁video", + "455": "▁imp", + "456": "▁cui", + "457": "▁dove", + "458": "▁Quindi", + "459": "sione", + "460": "rebbe", + "461": "scri", + "462": "", + "463": "ai", + "464": "▁ir", + "465": "as", + "466": "▁tai", + "467": "uo", + "468": "tin", + "469": "▁vis", + "470": "ly", + "471": "gal", + "472": "tar", + "473": "▁Ir", + "474": "▁turi", + "475": "▁Tai", + "476": "▁nu", + "477": "▁mes", + "478": "išk", + "479": "imas", + "480": "▁pra", + "481": "", + "482": "iem", + "483": "▁pie", + "484": "ies", + "485": "ot", + "486": "▁vien", + "487": "▁Un", + "488": "iet", + "489": "▁Ta", + "490": "dar", + "491": "sim", + "492": "gan", + "493": "▁ap", + "494": "▁nav", + "495": "▁Nu", + "496": "▁cit", + "497": "▁Ne", + "498": "▁20", + "499": "", + "500": "▁dat", + "501": "▁we", + "502": "▁En", + "503": "▁dan", + "504": "▁zo", + "505": "▁met", + "506": "▁wat", + "507": "der", + "508": "ui", + "509": "den", + "510": "op", + "511": "oor", + "512": "▁of", + "513": "ven", + "514": "acht", + "515": "▁even", + "516": "▁wil", + "517": "vol", + "518": "", + "519": "nie", + "520": "wa", + "521": "cie", + "522": "nia", + "523": "wo", + "524": "rze", + "525": "by", + "526": "za", + "527": "dy", + "528": "ry", + "529": "ego", + "530": "mie", + "531": "rz", + "532": "pie", + "533": "▁bo", + "534": "bie", + "535": "▁Po", + "536": "ski", + "537": "nego", + "538": "▁No", + "539": "▁Na", + "540": "▁Was", + "541": "▁musi", + "542": "", + "543": "-", + "544": "▁para", + "545": "▁uma", + "546": "▁pe", + "547": "▁tem", + "548": "▁gente", + "549": "▁O", + "550": "▁ele", + "551": "pre", + "552": "ria", + "553": "▁fo", + "554": "mos", + "555": "bo", + "556": "▁por", + "557": "nto", + "558": "be", + "559": "▁esse", + "560": "ente", + "561": "▁essa", + "562": "▁mas", + "563": "qua", + "564": "fica", + "565": "▁Se", + "566": "▁Por", + "567": "▁Co", + "568": "iza", + "569": "▁sua", + "570": "▁quando", + "571": "", + "572": "▁mai", + "573": "sc", + "574": "are", + "575": "▁din", + "576": "▁este", + "577": "rea", + "578": "ele", + "579": "du", + "580": "▁M", + "581": "▁fac", + "582": "lor", + "583": "▁mult", + "584": "per", + "585": "cur", + "586": "tor", + "587": "inte", + "588": "▁sau", + "589": "tat", + "590": "ori", + "591": "▁prim", + "592": "▁spun", + "593": "▁lui", + "594": "▁sub", + "595": "itate", + "596": "▁prin", + "597": "▁alt", + "598": "stru", + "599": "▁vede", + "600": "fer", + "601": "▁chiar", + "602": "", + "603": "", + "604": "ov", + "605": "ob", + "606": "▁bol", + "607": "ali", + "608": "rov", + "609": "rob", + "610": "▁spo", + "611": "osti", + "612": "", + "613": "sl", + "614": "udi", + "615": "del", + "616": "▁sem", + "617": "▁samo", + "618": "▁pred", + "619": "nost", + "620": "▁Pre", + "621": "▁prot", + "622": "▁internet", + "623": "▁film", + "624": "", + "625": "▁att", + "626": "▁inte", + "627": "▁av", + "628": "all", + "629": "era", + "630": "pp", + "631": "▁upp", + "632": "isk", + "633": "het", + "634": "▁vill", + "635": "erna", + "636": "ande", + "637": "ade", + "638": "bil", + "639": "▁min", + "640": "▁alla", + "641": "lev", + "642": "▁oss", + "643": "land", + "644": "▁Vad", + "645": "person", + "646": "", + "647": "", + "648": "vy", + "649": "ft", + "650": "lige", + "651": "ved", + "652": "'", + "653": "▁H", + "654": "▁D", + "655": "aus", + "656": "▁N", + "657": "▁Be", + "658": "mm", + "659": "ab", + "660": "▁Er", + "661": "ssen", + "662": "rie", + "663": "lei", + "664": "▁An", + "665": "rau", + "666": "▁So", + "667": "", + "668": "▁and", + "669": "▁can", + "670": "ed", + "671": "ay", + "672": "th", + "673": "ic", + "674": "hi", + "675": "▁Oh", + "676": "▁not", + "677": "ight", + "678": "ex", + "679": "▁great", + "680": "ill", + "681": "▁don", + "682": "▁problem", + "683": "▁fine", + "684": "▁month", + "685": "▁check", + "686": "▁zero", + "687": "▁first", + "688": "▁question", + "689": "", + "690": "ive", + "691": "ate", + "692": "ad", + "693": "ng", + "694": "ity", + "695": "ther", + "696": "act", + "697": "side", + "698": "\"", + "699": "", + "700": "ción", + "701": "▁Es", + "702": "res", + "703": "▁La", + "704": "dos", + "705": "▁El", + "706": "▁las", + "707": "men", + "708": "par", + "709": "rio", + "710": "enta", + "711": "▁Ca", + "712": "▁Su", + "713": "▁son", + "714": "ncia", + "715": "▁Con", + "716": "ones", + "717": "▁San", + "718": "▁persona", + "719": "▁Com", + "720": "", + "721": "cia", + "722": "▁Y", + "723": "ron", + "724": "les", + "725": "cio", + "726": "bu", + "727": "", + "728": "ré", + "729": "▁Les", + "730": "our", + "731": "▁Ce", + "732": "com", + "733": "ale", + "734": "if", + "735": "iste", + "736": "▁parti", + "737": "avec", + "738": "app", + "739": "gue", + "740": "▁grand", + "741": "Une", + "742": "È", + "743": "av", + "744": "pri", + "745": "sion", + "746": "ard", + "747": "", + "748": "", + "749": "!", + "750": "", + "751": "", + "752": "", + "753": "ene", + "754": "opp", + "755": "▁han", + "756": "", + "757": "eg", + "758": "kk", + "759": "▁god", + "760": "dde", + "761": "inn", + "762": "dig", + "763": "ord", + "764": "▁tru", + "765": "▁sei", + "766": "ller", + "767": "car", + "768": "ito", + "769": "ram", + "770": "fa", + "771": "▁mil", + "772": "▁passa", + "773": "▁casa", + "774": "", + "775": "▁Pa", + "776": "tura", + "777": "forma", + "778": "tua", + "779": "mar", + "780": "este", + "781": "fun", + "782": "gua", + "783": "▁grande", + "784": "▁nome", + "785": "▁Sua", + "786": "var", + "787": "", + "788": "", + "789": "ş", + "790": "ğ", + "791": "ya", + "792": "▁ve", + "793": "lar", + "794": "ler", + "795": "leri", + "796": "▁bu", + "797": "lan", + "798": "ara", + "799": "▁Bu", + "800": "yo", + "801": "", + "802": "", + "803": "▁t", + "804": "nh", + "805": "" +} \ No newline at end of file diff --git a/it/2240ms/decoder.mlmodelc/analytics/coremldata.bin b/it/2240ms/decoder.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..85afd8d84c262c9e1ba71c6b460a5beb4d6b94c3 --- /dev/null +++ b/it/2240ms/decoder.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4cdca6bf678463f31354072f526088e5bdf5115ae94c04e387bb35b2c7a607d6 +size 243 diff --git a/it/2240ms/decoder.mlmodelc/coremldata.bin b/it/2240ms/decoder.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..8bb44f2c4c669bd785344007b33e6273bd87aa8c --- /dev/null +++ b/it/2240ms/decoder.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3c993f8b96ce22027cd2ed42d99b7e61f93a01197bb17cadada8eb989e946dec +size 433 diff --git a/it/2240ms/decoder.mlmodelc/model.mil b/it/2240ms/decoder.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..e034376bbf9a1dff11539e03ae80e7a65ea4f393 --- /dev/null +++ b/it/2240ms/decoder.mlmodelc/model.mil @@ -0,0 +1,64 @@ +program(1.3) +[buildInfo = dict({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.5.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.3.0"}})] +{ + func main(tensor c_in, tensor h_in, tensor token, tensor token_length) { + int32 y_axis_0 = const()[name = string("y_axis_0"), val = int32(0)]; + int32 y_batch_dims_0 = const()[name = string("y_batch_dims_0"), val = int32(0)]; + bool y_validate_indices_0 = const()[name = string("y_validate_indices_0"), val = bool(false)]; + tensor module_prediction_embed_weight_to_fp16 = const()[name = string("module_prediction_embed_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))]; + string token_to_int16_dtype_0 = const()[name = string("token_to_int16_dtype_0"), val = string("int16")]; + tensor token_to_int16 = cast(dtype = token_to_int16_dtype_0, x = token)[name = string("cast_8")]; + tensor y_cast_fp16_cast_uint16 = gather(axis = y_axis_0, batch_dims = y_batch_dims_0, indices = token_to_int16, validate_indices = y_validate_indices_0, x = module_prediction_embed_weight_to_fp16)[name = string("y_cast_fp16_cast_uint16")]; + tensor input_3_perm_0 = const()[name = string("input_3_perm_0"), val = tensor([1, 0, 2])]; + int32 split_0_num_splits_0 = const()[name = string("split_0_num_splits_0"), val = int32(2)]; + int32 split_0_axis_0 = const()[name = string("split_0_axis_0"), val = int32(0)]; + string h_in_to_fp16_dtype_0 = const()[name = string("h_in_to_fp16_dtype_0"), val = string("fp16")]; + tensor h_in_to_fp16 = cast(dtype = h_in_to_fp16_dtype_0, x = h_in)[name = string("cast_7")]; + tensor split_0_cast_fp16_0, tensor split_0_cast_fp16_1 = split(axis = split_0_axis_0, num_splits = split_0_num_splits_0, x = h_in_to_fp16)[name = string("split_0_cast_fp16")]; + int32 split_1_num_splits_0 = const()[name = string("split_1_num_splits_0"), val = int32(2)]; + int32 split_1_axis_0 = const()[name = string("split_1_axis_0"), val = int32(0)]; + string c_in_to_fp16_dtype_0 = const()[name = string("c_in_to_fp16_dtype_0"), val = string("fp16")]; + tensor c_in_to_fp16 = cast(dtype = c_in_to_fp16_dtype_0, x = c_in)[name = string("cast_6")]; + tensor split_1_cast_fp16_0, tensor split_1_cast_fp16_1 = split(axis = split_1_axis_0, num_splits = split_1_num_splits_0, x = c_in_to_fp16)[name = string("split_1_cast_fp16")]; + tensor input_lstm_layer_0_lstm_h0_squeeze_axes_0 = const()[name = string("input_lstm_layer_0_lstm_h0_squeeze_axes_0"), val = tensor([0])]; + tensor input_lstm_layer_0_lstm_h0_squeeze_cast_fp16 = squeeze(axes = input_lstm_layer_0_lstm_h0_squeeze_axes_0, x = split_0_cast_fp16_0)[name = string("input_lstm_layer_0_lstm_h0_squeeze_cast_fp16")]; + tensor input_lstm_layer_0_lstm_c0_squeeze_axes_0 = const()[name = string("input_lstm_layer_0_lstm_c0_squeeze_axes_0"), val = tensor([0])]; + tensor input_lstm_layer_0_lstm_c0_squeeze_cast_fp16 = squeeze(axes = input_lstm_layer_0_lstm_c0_squeeze_axes_0, x = split_1_cast_fp16_0)[name = string("input_lstm_layer_0_lstm_c0_squeeze_cast_fp16")]; + string input_lstm_layer_0_direction_0 = const()[name = string("input_lstm_layer_0_direction_0"), val = string("forward")]; + bool input_lstm_layer_0_output_sequence_0 = const()[name = string("input_lstm_layer_0_output_sequence_0"), val = bool(true)]; + string input_lstm_layer_0_recurrent_activation_0 = const()[name = string("input_lstm_layer_0_recurrent_activation_0"), val = string("sigmoid")]; + string input_lstm_layer_0_cell_activation_0 = const()[name = string("input_lstm_layer_0_cell_activation_0"), val = string("tanh")]; + string input_lstm_layer_0_activation_0 = const()[name = string("input_lstm_layer_0_activation_0"), val = string("tanh")]; + tensor concat_1_to_fp16 = const()[name = string("concat_1_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1031808)))]; + tensor concat_2_to_fp16 = const()[name = string("concat_2_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4308672)))]; + tensor concat_0_to_fp16 = const()[name = string("concat_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7585536)))]; + tensor input_3_cast_fp16 = transpose(perm = input_3_perm_0, x = y_cast_fp16_cast_uint16)[name = string("transpose_2")]; + tensor input_lstm_layer_0_cast_fp16_0, tensor input_lstm_layer_0_cast_fp16_1, tensor input_lstm_layer_0_cast_fp16_2 = lstm(activation = input_lstm_layer_0_activation_0, bias = concat_0_to_fp16, cell_activation = input_lstm_layer_0_cell_activation_0, direction = input_lstm_layer_0_direction_0, initial_c = input_lstm_layer_0_lstm_c0_squeeze_cast_fp16, initial_h = input_lstm_layer_0_lstm_h0_squeeze_cast_fp16, output_sequence = input_lstm_layer_0_output_sequence_0, recurrent_activation = input_lstm_layer_0_recurrent_activation_0, weight_hh = concat_2_to_fp16, weight_ih = concat_1_to_fp16, x = input_3_cast_fp16)[name = string("input_lstm_layer_0_cast_fp16")]; + tensor input_lstm_h0_squeeze_axes_0 = const()[name = string("input_lstm_h0_squeeze_axes_0"), val = tensor([0])]; + tensor input_lstm_h0_squeeze_cast_fp16 = squeeze(axes = input_lstm_h0_squeeze_axes_0, x = split_0_cast_fp16_1)[name = string("input_lstm_h0_squeeze_cast_fp16")]; + tensor input_lstm_c0_squeeze_axes_0 = const()[name = string("input_lstm_c0_squeeze_axes_0"), val = tensor([0])]; + tensor input_lstm_c0_squeeze_cast_fp16 = squeeze(axes = input_lstm_c0_squeeze_axes_0, x = split_1_cast_fp16_1)[name = string("input_lstm_c0_squeeze_cast_fp16")]; + string input_direction_0 = const()[name = string("input_direction_0"), val = string("forward")]; + bool input_output_sequence_0 = const()[name = string("input_output_sequence_0"), val = bool(true)]; + string input_recurrent_activation_0 = const()[name = string("input_recurrent_activation_0"), val = string("sigmoid")]; + string input_cell_activation_0 = const()[name = string("input_cell_activation_0"), val = string("tanh")]; + string input_activation_0 = const()[name = string("input_activation_0"), val = string("tanh")]; + tensor concat_4_to_fp16 = const()[name = string("concat_4_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7590720)))]; + tensor concat_5_to_fp16 = const()[name = string("concat_5_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10867584)))]; + tensor concat_3_to_fp16 = const()[name = string("concat_3_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14144448)))]; + tensor input_cast_fp16_0, tensor input_cast_fp16_1, tensor input_cast_fp16_2 = lstm(activation = input_activation_0, bias = concat_3_to_fp16, cell_activation = input_cell_activation_0, direction = input_direction_0, initial_c = input_lstm_c0_squeeze_cast_fp16, initial_h = input_lstm_h0_squeeze_cast_fp16, output_sequence = input_output_sequence_0, recurrent_activation = input_recurrent_activation_0, weight_hh = concat_5_to_fp16, weight_ih = concat_4_to_fp16, x = input_lstm_layer_0_cast_fp16_0)[name = string("input_cast_fp16")]; + int32 obj_3_axis_0 = const()[name = string("obj_3_axis_0"), val = int32(0)]; + tensor obj_3_cast_fp16 = stack(axis = obj_3_axis_0, values = (input_lstm_layer_0_cast_fp16_1, input_cast_fp16_1))[name = string("obj_3_cast_fp16")]; + string obj_3_cast_fp16_to_fp32_dtype_0 = const()[name = string("obj_3_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + int32 obj_axis_0 = const()[name = string("obj_axis_0"), val = int32(0)]; + tensor obj_cast_fp16 = stack(axis = obj_axis_0, values = (input_lstm_layer_0_cast_fp16_2, input_cast_fp16_2))[name = string("obj_cast_fp16")]; + string obj_cast_fp16_to_fp32_dtype_0 = const()[name = string("obj_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + tensor transpose_0_perm_0 = const()[name = string("transpose_0_perm_0"), val = tensor([1, 2, 0])]; + string transpose_0_cast_fp16_to_fp32_dtype_0 = const()[name = string("transpose_0_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + tensor transpose_0_cast_fp16 = transpose(perm = transpose_0_perm_0, x = input_cast_fp16_0)[name = string("transpose_1")]; + tensor decoder_out = cast(dtype = transpose_0_cast_fp16_to_fp32_dtype_0, x = transpose_0_cast_fp16)[name = string("cast_3")]; + tensor c_out = cast(dtype = obj_cast_fp16_to_fp32_dtype_0, x = obj_cast_fp16)[name = string("cast_4")]; + tensor h_out = cast(dtype = obj_3_cast_fp16_to_fp32_dtype_0, x = obj_3_cast_fp16)[name = string("cast_5")]; + tensor token_length_tmp = identity(x = token_length)[name = string("token_length_tmp")]; + } -> (decoder_out, h_out, c_out); +} \ No newline at end of file diff --git a/it/2240ms/decoder.mlmodelc/weights/weight.bin b/it/2240ms/decoder.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..2fac0f3a92f8a80de4d92a62819f6dba98aa4983 --- /dev/null +++ b/it/2240ms/decoder.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1360d93c68c3e9c54bda4adaec860753949f3b0dc93bc98f4edc9d6f8dd5595c +size 14149632 diff --git a/it/2240ms/decoder.mlpackage/Data/com.apple.CoreML/model.mlmodel b/it/2240ms/decoder.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..567c038e1e42f382639a9ececec8bb38c22cbde0 --- /dev/null +++ b/it/2240ms/decoder.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:73bb3afa62698bc822b6d32b3731d0bc40521e03737e3139e10a768542fca1fe +size 10359 diff --git a/it/2240ms/decoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/it/2240ms/decoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..2fac0f3a92f8a80de4d92a62819f6dba98aa4983 --- /dev/null +++ b/it/2240ms/decoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1360d93c68c3e9c54bda4adaec860753949f3b0dc93bc98f4edc9d6f8dd5595c +size 14149632 diff --git a/it/2240ms/decoder.mlpackage/Manifest.json b/it/2240ms/decoder.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..8fc74b00a3e9885d54546160ecae1f6da7736d01 --- /dev/null +++ b/it/2240ms/decoder.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "7CBCED8D-FA6A-45B0-BF60-30DB0A653074": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + }, + "AFD197FC-BECC-451A-961C-C0CA05D58065": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + } + }, + "rootModelIdentifier": "AFD197FC-BECC-451A-961C-C0CA05D58065" +} diff --git a/it/2240ms/decoder_joint.mlmodelc/analytics/coremldata.bin b/it/2240ms/decoder_joint.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..6063f90a9756de97c8450a89ef53ef04317ef653 --- /dev/null +++ b/it/2240ms/decoder_joint.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:12f4bcf5114baa2b3a37b8ebeab6c519109bd857e50ec345c458b7a6c4deb20e +size 243 diff --git a/it/2240ms/decoder_joint.mlmodelc/coremldata.bin b/it/2240ms/decoder_joint.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..6f49a6f6923a6d68c50bdf11730215b1db8a2d62 --- /dev/null +++ b/it/2240ms/decoder_joint.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:53754e2eaa7e0f7435220b47b621c5f3d8c5f2da83edd46efa5950fa723ef1d9 +size 454 diff --git a/it/2240ms/decoder_joint.mlmodelc/model.mil b/it/2240ms/decoder_joint.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..9e96b62349b7d1c4bd97fe8db2d7755704041510 --- /dev/null +++ b/it/2240ms/decoder_joint.mlmodelc/model.mil @@ -0,0 +1,83 @@ +program(1.3) +[buildInfo = dict({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.5.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.3.0"}})] +{ + func main(tensor c_in, tensor encoder, tensor h_in, tensor token, tensor token_length) { + int32 y_axis_0 = const()[name = string("y_axis_0"), val = int32(0)]; + int32 y_batch_dims_0 = const()[name = string("y_batch_dims_0"), val = int32(0)]; + bool y_validate_indices_0 = const()[name = string("y_validate_indices_0"), val = bool(false)]; + tensor decoder_module_prediction_embed_weight_to_fp16 = const()[name = string("decoder_module_prediction_embed_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))]; + string token_to_int16_dtype_0 = const()[name = string("token_to_int16_dtype_0"), val = string("int16")]; + tensor token_to_int16 = cast(dtype = token_to_int16_dtype_0, x = token)[name = string("cast_9")]; + tensor y_cast_fp16_cast_uint16 = gather(axis = y_axis_0, batch_dims = y_batch_dims_0, indices = token_to_int16, validate_indices = y_validate_indices_0, x = decoder_module_prediction_embed_weight_to_fp16)[name = string("y_cast_fp16_cast_uint16")]; + tensor input_3_perm_0 = const()[name = string("input_3_perm_0"), val = tensor([1, 0, 2])]; + int32 split_0_num_splits_0 = const()[name = string("split_0_num_splits_0"), val = int32(2)]; + int32 split_0_axis_0 = const()[name = string("split_0_axis_0"), val = int32(0)]; + string h_in_to_fp16_dtype_0 = const()[name = string("h_in_to_fp16_dtype_0"), val = string("fp16")]; + tensor h_in_to_fp16 = cast(dtype = h_in_to_fp16_dtype_0, x = h_in)[name = string("cast_8")]; + tensor split_0_cast_fp16_0, tensor split_0_cast_fp16_1 = split(axis = split_0_axis_0, num_splits = split_0_num_splits_0, x = h_in_to_fp16)[name = string("split_0_cast_fp16")]; + int32 split_1_num_splits_0 = const()[name = string("split_1_num_splits_0"), val = int32(2)]; + int32 split_1_axis_0 = const()[name = string("split_1_axis_0"), val = int32(0)]; + string c_in_to_fp16_dtype_0 = const()[name = string("c_in_to_fp16_dtype_0"), val = string("fp16")]; + tensor c_in_to_fp16 = cast(dtype = c_in_to_fp16_dtype_0, x = c_in)[name = string("cast_7")]; + tensor split_1_cast_fp16_0, tensor split_1_cast_fp16_1 = split(axis = split_1_axis_0, num_splits = split_1_num_splits_0, x = c_in_to_fp16)[name = string("split_1_cast_fp16")]; + tensor input_5_lstm_layer_0_lstm_h0_squeeze_axes_0 = const()[name = string("input_5_lstm_layer_0_lstm_h0_squeeze_axes_0"), val = tensor([0])]; + tensor input_5_lstm_layer_0_lstm_h0_squeeze_cast_fp16 = squeeze(axes = input_5_lstm_layer_0_lstm_h0_squeeze_axes_0, x = split_0_cast_fp16_0)[name = string("input_5_lstm_layer_0_lstm_h0_squeeze_cast_fp16")]; + tensor input_5_lstm_layer_0_lstm_c0_squeeze_axes_0 = const()[name = string("input_5_lstm_layer_0_lstm_c0_squeeze_axes_0"), val = tensor([0])]; + tensor input_5_lstm_layer_0_lstm_c0_squeeze_cast_fp16 = squeeze(axes = input_5_lstm_layer_0_lstm_c0_squeeze_axes_0, x = split_1_cast_fp16_0)[name = string("input_5_lstm_layer_0_lstm_c0_squeeze_cast_fp16")]; + string input_5_lstm_layer_0_direction_0 = const()[name = string("input_5_lstm_layer_0_direction_0"), val = string("forward")]; + bool input_5_lstm_layer_0_output_sequence_0 = const()[name = string("input_5_lstm_layer_0_output_sequence_0"), val = bool(true)]; + string input_5_lstm_layer_0_recurrent_activation_0 = const()[name = string("input_5_lstm_layer_0_recurrent_activation_0"), val = string("sigmoid")]; + string input_5_lstm_layer_0_cell_activation_0 = const()[name = string("input_5_lstm_layer_0_cell_activation_0"), val = string("tanh")]; + string input_5_lstm_layer_0_activation_0 = const()[name = string("input_5_lstm_layer_0_activation_0"), val = string("tanh")]; + tensor concat_1_to_fp16 = const()[name = string("concat_1_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1031808)))]; + tensor concat_2_to_fp16 = const()[name = string("concat_2_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4308672)))]; + tensor concat_0_to_fp16 = const()[name = string("concat_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7585536)))]; + tensor input_3_cast_fp16 = transpose(perm = input_3_perm_0, x = y_cast_fp16_cast_uint16)[name = string("transpose_4")]; + tensor input_5_lstm_layer_0_cast_fp16_0, tensor input_5_lstm_layer_0_cast_fp16_1, tensor input_5_lstm_layer_0_cast_fp16_2 = lstm(activation = input_5_lstm_layer_0_activation_0, bias = concat_0_to_fp16, cell_activation = input_5_lstm_layer_0_cell_activation_0, direction = input_5_lstm_layer_0_direction_0, initial_c = input_5_lstm_layer_0_lstm_c0_squeeze_cast_fp16, initial_h = input_5_lstm_layer_0_lstm_h0_squeeze_cast_fp16, output_sequence = input_5_lstm_layer_0_output_sequence_0, recurrent_activation = input_5_lstm_layer_0_recurrent_activation_0, weight_hh = concat_2_to_fp16, weight_ih = concat_1_to_fp16, x = input_3_cast_fp16)[name = string("input_5_lstm_layer_0_cast_fp16")]; + tensor input_5_lstm_h0_squeeze_axes_0 = const()[name = string("input_5_lstm_h0_squeeze_axes_0"), val = tensor([0])]; + tensor input_5_lstm_h0_squeeze_cast_fp16 = squeeze(axes = input_5_lstm_h0_squeeze_axes_0, x = split_0_cast_fp16_1)[name = string("input_5_lstm_h0_squeeze_cast_fp16")]; + tensor input_5_lstm_c0_squeeze_axes_0 = const()[name = string("input_5_lstm_c0_squeeze_axes_0"), val = tensor([0])]; + tensor input_5_lstm_c0_squeeze_cast_fp16 = squeeze(axes = input_5_lstm_c0_squeeze_axes_0, x = split_1_cast_fp16_1)[name = string("input_5_lstm_c0_squeeze_cast_fp16")]; + string input_5_direction_0 = const()[name = string("input_5_direction_0"), val = string("forward")]; + bool input_5_output_sequence_0 = const()[name = string("input_5_output_sequence_0"), val = bool(true)]; + string input_5_recurrent_activation_0 = const()[name = string("input_5_recurrent_activation_0"), val = string("sigmoid")]; + string input_5_cell_activation_0 = const()[name = string("input_5_cell_activation_0"), val = string("tanh")]; + string input_5_activation_0 = const()[name = string("input_5_activation_0"), val = string("tanh")]; + tensor concat_4_to_fp16 = const()[name = string("concat_4_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7590720)))]; + tensor concat_5_to_fp16 = const()[name = string("concat_5_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10867584)))]; + tensor concat_3_to_fp16 = const()[name = string("concat_3_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14144448)))]; + tensor input_5_cast_fp16_0, tensor input_5_cast_fp16_1, tensor input_5_cast_fp16_2 = lstm(activation = input_5_activation_0, bias = concat_3_to_fp16, cell_activation = input_5_cell_activation_0, direction = input_5_direction_0, initial_c = input_5_lstm_c0_squeeze_cast_fp16, initial_h = input_5_lstm_h0_squeeze_cast_fp16, output_sequence = input_5_output_sequence_0, recurrent_activation = input_5_recurrent_activation_0, weight_hh = concat_5_to_fp16, weight_ih = concat_4_to_fp16, x = input_5_lstm_layer_0_cast_fp16_0)[name = string("input_5_cast_fp16")]; + int32 obj_3_axis_0 = const()[name = string("obj_3_axis_0"), val = int32(0)]; + tensor obj_3_cast_fp16 = stack(axis = obj_3_axis_0, values = (input_5_lstm_layer_0_cast_fp16_1, input_5_cast_fp16_1))[name = string("obj_3_cast_fp16")]; + string obj_3_cast_fp16_to_fp32_dtype_0 = const()[name = string("obj_3_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + int32 obj_axis_0 = const()[name = string("obj_axis_0"), val = int32(0)]; + tensor obj_cast_fp16 = stack(axis = obj_axis_0, values = (input_5_lstm_layer_0_cast_fp16_2, input_5_cast_fp16_2))[name = string("obj_cast_fp16")]; + string obj_cast_fp16_to_fp32_dtype_0 = const()[name = string("obj_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + tensor transpose_1_perm_0 = const()[name = string("transpose_1_perm_0"), val = tensor([1, 0, 2])]; + tensor input_7_perm_0 = const()[name = string("input_7_perm_0"), val = tensor([0, 2, 1])]; + string encoder_to_fp16_dtype_0 = const()[name = string("encoder_to_fp16_dtype_0"), val = string("fp16")]; + tensor joint_module_enc_weight_to_fp16 = const()[name = string("joint_module_enc_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14149632)))]; + tensor joint_module_enc_bias_to_fp16 = const()[name = string("joint_module_enc_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15460416)))]; + tensor encoder_to_fp16 = cast(dtype = encoder_to_fp16_dtype_0, x = encoder)[name = string("cast_4")]; + tensor input_7_cast_fp16 = transpose(perm = input_7_perm_0, x = encoder_to_fp16)[name = string("transpose_2")]; + tensor linear_0_cast_fp16 = linear(bias = joint_module_enc_bias_to_fp16, weight = joint_module_enc_weight_to_fp16, x = input_7_cast_fp16)[name = string("linear_0_cast_fp16")]; + tensor joint_module_pred_weight_to_fp16 = const()[name = string("joint_module_pred_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15461760)))]; + tensor joint_module_pred_bias_to_fp16 = const()[name = string("joint_module_pred_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16281024)))]; + tensor transpose_1_cast_fp16 = transpose(perm = transpose_1_perm_0, x = input_5_cast_fp16_0)[name = string("transpose_3")]; + tensor linear_1_cast_fp16 = linear(bias = joint_module_pred_bias_to_fp16, weight = joint_module_pred_weight_to_fp16, x = transpose_1_cast_fp16)[name = string("linear_1_cast_fp16")]; + tensor var_79_axes_0 = const()[name = string("op_79_axes_0"), val = tensor([2])]; + tensor var_79_cast_fp16 = expand_dims(axes = var_79_axes_0, x = linear_0_cast_fp16)[name = string("op_79_cast_fp16")]; + tensor var_80_axes_0 = const()[name = string("op_80_axes_0"), val = tensor([1])]; + tensor var_80_cast_fp16 = expand_dims(axes = var_80_axes_0, x = linear_1_cast_fp16)[name = string("op_80_cast_fp16")]; + tensor input_11_cast_fp16 = add(x = var_79_cast_fp16, y = var_80_cast_fp16)[name = string("input_11_cast_fp16")]; + tensor input_13_cast_fp16 = relu(x = input_11_cast_fp16)[name = string("input_13_cast_fp16")]; + tensor joint_module_joint_net_2_weight_to_fp16 = const()[name = string("joint_module_joint_net_2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16282368)))]; + tensor joint_module_joint_net_2_bias_to_fp16 = const()[name = string("joint_module_joint_net_2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17314112)))]; + tensor linear_2_cast_fp16 = linear(bias = joint_module_joint_net_2_bias_to_fp16, weight = joint_module_joint_net_2_weight_to_fp16, x = input_13_cast_fp16)[name = string("linear_2_cast_fp16")]; + string linear_2_cast_fp16_to_fp32_dtype_0 = const()[name = string("linear_2_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + tensor logits = cast(dtype = linear_2_cast_fp16_to_fp32_dtype_0, x = linear_2_cast_fp16)[name = string("cast_3")]; + tensor c_out = cast(dtype = obj_cast_fp16_to_fp32_dtype_0, x = obj_cast_fp16)[name = string("cast_5")]; + tensor h_out = cast(dtype = obj_3_cast_fp16_to_fp32_dtype_0, x = obj_3_cast_fp16)[name = string("cast_6")]; + tensor token_length_tmp = identity(x = token_length)[name = string("token_length_tmp")]; + } -> (logits, h_out, c_out); +} \ No newline at end of file diff --git a/it/2240ms/decoder_joint.mlmodelc/weights/weight.bin b/it/2240ms/decoder_joint.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..33ea2ec2f210db3873bae9b152a8fa5b13171f2e --- /dev/null +++ b/it/2240ms/decoder_joint.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e507a69196a04e30adfafc302b6a5f5f527e45c1965c65dd81d63a621cae2064 +size 17315788 diff --git a/it/2240ms/decoder_joint.mlpackage/Data/com.apple.CoreML/model.mlmodel b/it/2240ms/decoder_joint.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..cc3525ebd701acd827b72a5d6a05caf5ddff80e9 --- /dev/null +++ b/it/2240ms/decoder_joint.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5fa6a6d89c1c07ae16f162f5a3b6809b12aafe57a663f5cdb270be3dec7b1427 +size 13745 diff --git a/it/2240ms/decoder_joint.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/it/2240ms/decoder_joint.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..33ea2ec2f210db3873bae9b152a8fa5b13171f2e --- /dev/null +++ b/it/2240ms/decoder_joint.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e507a69196a04e30adfafc302b6a5f5f527e45c1965c65dd81d63a621cae2064 +size 17315788 diff --git a/it/2240ms/decoder_joint.mlpackage/Manifest.json b/it/2240ms/decoder_joint.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..48ab93415f34542ace6e74a66d563506a10f114a --- /dev/null +++ b/it/2240ms/decoder_joint.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "2B19A50C-1D16-4D97-BE3C-D9BCF35884CF": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + }, + "9CA734BC-CFD2-4F39-B068-BE69ABCAAD1F": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + } + }, + "rootModelIdentifier": "2B19A50C-1D16-4D97-BE3C-D9BCF35884CF" +} diff --git a/it/2240ms/decoder_joint_noencproj.mlmodelc/analytics/coremldata.bin b/it/2240ms/decoder_joint_noencproj.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..3573ed8dea8350501693449f8d9e59b9543d1e3b --- /dev/null +++ b/it/2240ms/decoder_joint_noencproj.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:40ec657603479e7dbf8cdb3d6368349eb8b766a52439a26a735d1fadf1b4281d +size 243 diff --git a/it/2240ms/decoder_joint_noencproj.mlmodelc/coremldata.bin b/it/2240ms/decoder_joint_noencproj.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..94d83d6a74b8a602fbbc8c932d43ab754ba51b88 --- /dev/null +++ b/it/2240ms/decoder_joint_noencproj.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c682ba0e028fb7ab6557f8ac1006febc8ec8dd81e4ef8d3a2c05d876e2dbcc8e +size 519 diff --git a/it/2240ms/decoder_joint_noencproj.mlmodelc/model.mil b/it/2240ms/decoder_joint_noencproj.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..5d5cbd528590956dead59657945f5dab997a7da9 --- /dev/null +++ b/it/2240ms/decoder_joint_noencproj.mlmodelc/model.mil @@ -0,0 +1,91 @@ +program(1.3) +[buildInfo = dict({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.10.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor c_in, tensor encoder_proj, tensor h_in, tensor token, tensor token_length) { + int32 y_batch_dims_0 = const()[name = string("y_batch_dims_0"), val = int32(0)]; + bool y_validate_indices_0 = const()[name = string("y_validate_indices_0"), val = bool(false)]; + tensor decoder_module_prediction_embed_weight_to_fp16 = const()[name = string("decoder_module_prediction_embed_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))]; + string token_to_int16_dtype_0 = const()[name = string("token_to_int16_dtype_0"), val = string("int16")]; + string cast_1_dtype_0 = const()[name = string("cast_1_dtype_0"), val = string("int32")]; + int32 greater_equal_0_y_0 = const()[name = string("greater_equal_0_y_0"), val = int32(0)]; + tensor token_to_int16 = cast(dtype = token_to_int16_dtype_0, x = token)[name = string("cast_10")]; + tensor cast_1 = cast(dtype = cast_1_dtype_0, x = token_to_int16)[name = string("cast_9")]; + tensor greater_equal_0 = greater_equal(x = cast_1, y = greater_equal_0_y_0)[name = string("greater_equal_0")]; + int32 slice_by_index_0 = const()[name = string("slice_by_index_0"), val = int32(806)]; + tensor add_2 = add(x = cast_1, y = slice_by_index_0)[name = string("add_2")]; + tensor select_0 = select(a = cast_1, b = add_2, cond = greater_equal_0)[name = string("select_0")]; + int32 y_cast_fp16_cast_uint16_axis_0 = const()[name = string("y_cast_fp16_cast_uint16_axis_0"), val = int32(0)]; + string select_0_to_int16_dtype_0 = const()[name = string("select_0_to_int16_dtype_0"), val = string("int16")]; + tensor select_0_to_int16 = cast(dtype = select_0_to_int16_dtype_0, x = select_0)[name = string("cast_8")]; + tensor y_cast_fp16_cast_uint16_cast_uint16 = gather(axis = y_cast_fp16_cast_uint16_axis_0, batch_dims = y_batch_dims_0, indices = select_0_to_int16, validate_indices = y_validate_indices_0, x = decoder_module_prediction_embed_weight_to_fp16)[name = string("y_cast_fp16_cast_uint16_cast_uint16")]; + tensor input_3_perm_0 = const()[name = string("input_3_perm_0"), val = tensor([1, 0, 2])]; + int32 split_0_num_splits_0 = const()[name = string("split_0_num_splits_0"), val = int32(2)]; + int32 split_0_axis_0 = const()[name = string("split_0_axis_0"), val = int32(0)]; + string h_in_to_fp16_dtype_0 = const()[name = string("h_in_to_fp16_dtype_0"), val = string("fp16")]; + tensor h_in_to_fp16 = cast(dtype = h_in_to_fp16_dtype_0, x = h_in)[name = string("cast_7")]; + tensor split_0_cast_fp16_0, tensor split_0_cast_fp16_1 = split(axis = split_0_axis_0, num_splits = split_0_num_splits_0, x = h_in_to_fp16)[name = string("split_0_cast_fp16")]; + int32 split_1_num_splits_0 = const()[name = string("split_1_num_splits_0"), val = int32(2)]; + int32 split_1_axis_0 = const()[name = string("split_1_axis_0"), val = int32(0)]; + string c_in_to_fp16_dtype_0 = const()[name = string("c_in_to_fp16_dtype_0"), val = string("fp16")]; + tensor c_in_to_fp16 = cast(dtype = c_in_to_fp16_dtype_0, x = c_in)[name = string("cast_6")]; + tensor split_1_cast_fp16_0, tensor split_1_cast_fp16_1 = split(axis = split_1_axis_0, num_splits = split_1_num_splits_0, x = c_in_to_fp16)[name = string("split_1_cast_fp16")]; + tensor input_5_lstm_layer_0_lstm_h0_squeeze_axes_0 = const()[name = string("input_5_lstm_layer_0_lstm_h0_squeeze_axes_0"), val = tensor([0])]; + tensor input_5_lstm_layer_0_lstm_h0_squeeze_cast_fp16 = squeeze(axes = input_5_lstm_layer_0_lstm_h0_squeeze_axes_0, x = split_0_cast_fp16_0)[name = string("input_5_lstm_layer_0_lstm_h0_squeeze_cast_fp16")]; + tensor input_5_lstm_layer_0_lstm_c0_squeeze_axes_0 = const()[name = string("input_5_lstm_layer_0_lstm_c0_squeeze_axes_0"), val = tensor([0])]; + tensor input_5_lstm_layer_0_lstm_c0_squeeze_cast_fp16 = squeeze(axes = input_5_lstm_layer_0_lstm_c0_squeeze_axes_0, x = split_1_cast_fp16_0)[name = string("input_5_lstm_layer_0_lstm_c0_squeeze_cast_fp16")]; + string input_5_lstm_layer_0_direction_0 = const()[name = string("input_5_lstm_layer_0_direction_0"), val = string("forward")]; + bool input_5_lstm_layer_0_output_sequence_0 = const()[name = string("input_5_lstm_layer_0_output_sequence_0"), val = bool(true)]; + string input_5_lstm_layer_0_recurrent_activation_0 = const()[name = string("input_5_lstm_layer_0_recurrent_activation_0"), val = string("sigmoid")]; + string input_5_lstm_layer_0_cell_activation_0 = const()[name = string("input_5_lstm_layer_0_cell_activation_0"), val = string("tanh")]; + string input_5_lstm_layer_0_activation_0 = const()[name = string("input_5_lstm_layer_0_activation_0"), val = string("tanh")]; + tensor concat_1_to_fp16 = const()[name = string("concat_1_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1031808)))]; + tensor concat_2_to_fp16 = const()[name = string("concat_2_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4308672)))]; + tensor concat_0_to_fp16 = const()[name = string("concat_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7585536)))]; + tensor input_3_cast_fp16 = transpose(perm = input_3_perm_0, x = y_cast_fp16_cast_uint16_cast_uint16)[name = string("transpose_3")]; + tensor input_5_lstm_layer_0_cast_fp16_0, tensor input_5_lstm_layer_0_cast_fp16_1, tensor input_5_lstm_layer_0_cast_fp16_2 = lstm(activation = input_5_lstm_layer_0_activation_0, bias = concat_0_to_fp16, cell_activation = input_5_lstm_layer_0_cell_activation_0, direction = input_5_lstm_layer_0_direction_0, initial_c = input_5_lstm_layer_0_lstm_c0_squeeze_cast_fp16, initial_h = input_5_lstm_layer_0_lstm_h0_squeeze_cast_fp16, output_sequence = input_5_lstm_layer_0_output_sequence_0, recurrent_activation = input_5_lstm_layer_0_recurrent_activation_0, weight_hh = concat_2_to_fp16, weight_ih = concat_1_to_fp16, x = input_3_cast_fp16)[name = string("input_5_lstm_layer_0_cast_fp16")]; + tensor input_5_lstm_h0_squeeze_axes_0 = const()[name = string("input_5_lstm_h0_squeeze_axes_0"), val = tensor([0])]; + tensor input_5_lstm_h0_squeeze_cast_fp16 = squeeze(axes = input_5_lstm_h0_squeeze_axes_0, x = split_0_cast_fp16_1)[name = string("input_5_lstm_h0_squeeze_cast_fp16")]; + tensor input_5_lstm_c0_squeeze_axes_0 = const()[name = string("input_5_lstm_c0_squeeze_axes_0"), val = tensor([0])]; + tensor input_5_lstm_c0_squeeze_cast_fp16 = squeeze(axes = input_5_lstm_c0_squeeze_axes_0, x = split_1_cast_fp16_1)[name = string("input_5_lstm_c0_squeeze_cast_fp16")]; + string input_5_direction_0 = const()[name = string("input_5_direction_0"), val = string("forward")]; + bool input_5_output_sequence_0 = const()[name = string("input_5_output_sequence_0"), val = bool(true)]; + string input_5_recurrent_activation_0 = const()[name = string("input_5_recurrent_activation_0"), val = string("sigmoid")]; + string input_5_cell_activation_0 = const()[name = string("input_5_cell_activation_0"), val = string("tanh")]; + string input_5_activation_0 = const()[name = string("input_5_activation_0"), val = string("tanh")]; + tensor concat_4_to_fp16 = const()[name = string("concat_4_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7590720)))]; + tensor concat_5_to_fp16 = const()[name = string("concat_5_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10867584)))]; + tensor concat_3_to_fp16 = const()[name = string("concat_3_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14144448)))]; + tensor input_5_cast_fp16_0, tensor input_5_cast_fp16_1, tensor input_5_cast_fp16_2 = lstm(activation = input_5_activation_0, bias = concat_3_to_fp16, cell_activation = input_5_cell_activation_0, direction = input_5_direction_0, initial_c = input_5_lstm_c0_squeeze_cast_fp16, initial_h = input_5_lstm_h0_squeeze_cast_fp16, output_sequence = input_5_output_sequence_0, recurrent_activation = input_5_recurrent_activation_0, weight_hh = concat_5_to_fp16, weight_ih = concat_4_to_fp16, x = input_5_lstm_layer_0_cast_fp16_0)[name = string("input_5_cast_fp16")]; + int32 obj_3_axis_0 = const()[name = string("obj_3_axis_0"), val = int32(0)]; + tensor obj_3_cast_fp16 = stack(axis = obj_3_axis_0, values = (input_5_lstm_layer_0_cast_fp16_1, input_5_cast_fp16_1))[name = string("obj_3_cast_fp16")]; + string obj_3_cast_fp16_to_fp32_dtype_0 = const()[name = string("obj_3_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + int32 obj_axis_0 = const()[name = string("obj_axis_0"), val = int32(0)]; + tensor obj_cast_fp16 = stack(axis = obj_axis_0, values = (input_5_lstm_layer_0_cast_fp16_2, input_5_cast_fp16_2))[name = string("obj_cast_fp16")]; + string obj_cast_fp16_to_fp32_dtype_0 = const()[name = string("obj_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + tensor transpose_1_perm_0 = const()[name = string("transpose_1_perm_0"), val = tensor([1, 0, 2])]; + tensor joint_module_pred_weight_to_fp16 = const()[name = string("joint_module_pred_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14149632)))]; + tensor joint_module_pred_bias_to_fp16 = const()[name = string("joint_module_pred_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14968896)))]; + tensor transpose_1_cast_fp16 = transpose(perm = transpose_1_perm_0, x = input_5_cast_fp16_0)[name = string("transpose_2")]; + tensor linear_0_cast_fp16 = linear(bias = joint_module_pred_bias_to_fp16, weight = joint_module_pred_weight_to_fp16, x = transpose_1_cast_fp16)[name = string("linear_0_cast_fp16")]; + tensor f_axes_0 = const()[name = string("f_axes_0"), val = tensor([2])]; + string encoder_proj_to_fp16_dtype_0 = const()[name = string("encoder_proj_to_fp16_dtype_0"), val = string("fp16")]; + tensor encoder_proj_to_fp16 = cast(dtype = encoder_proj_to_fp16_dtype_0, x = encoder_proj)[name = string("cast_3")]; + tensor f_cast_fp16 = expand_dims(axes = f_axes_0, x = encoder_proj_to_fp16)[name = string("f_cast_fp16")]; + tensor g_axes_0 = const()[name = string("g_axes_0"), val = tensor([1])]; + tensor g_cast_fp16 = expand_dims(axes = g_axes_0, x = linear_0_cast_fp16)[name = string("g_cast_fp16")]; + tensor input_9_cast_fp16 = add(x = f_cast_fp16, y = g_cast_fp16)[name = string("input_9_cast_fp16")]; + tensor input_11_cast_fp16 = relu(x = input_9_cast_fp16)[name = string("input_11_cast_fp16")]; + tensor joint_module_joint_net_2_weight_to_fp16 = const()[name = string("joint_module_joint_net_2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14970240)))]; + tensor joint_module_joint_net_2_bias_to_fp16 = const()[name = string("joint_module_joint_net_2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16001984)))]; + tensor linear_1_cast_fp16 = linear(bias = joint_module_joint_net_2_bias_to_fp16, weight = joint_module_joint_net_2_weight_to_fp16, x = input_11_cast_fp16)[name = string("linear_1_cast_fp16")]; + int32 var_83 = const()[name = string("op_83"), val = int32(-1)]; + tensor var_85_softmax_cast_fp16 = softmax(axis = var_83, x = linear_1_cast_fp16)[name = string("op_85_softmax_cast_fp16")]; + fp32 var_85_epsilon_0 = const()[name = string("op_85_epsilon_0"), val = fp32(0x1p-149)]; + tensor var_85_cast_fp16 = log(epsilon = var_85_epsilon_0, x = var_85_softmax_cast_fp16)[name = string("op_85_cast_fp16")]; + string var_85_cast_fp16_to_fp32_dtype_0 = const()[name = string("op_85_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + tensor logits = cast(dtype = var_85_cast_fp16_to_fp32_dtype_0, x = var_85_cast_fp16)[name = string("cast_2")]; + tensor c_out = cast(dtype = obj_cast_fp16_to_fp32_dtype_0, x = obj_cast_fp16)[name = string("cast_4")]; + tensor h_out = cast(dtype = obj_3_cast_fp16_to_fp32_dtype_0, x = obj_3_cast_fp16)[name = string("cast_5")]; + tensor token_length_tmp = identity(x = token_length)[name = string("token_length_tmp")]; + } -> (logits, h_out, c_out); +} \ No newline at end of file diff --git a/it/2240ms/decoder_joint_noencproj.mlmodelc/weights/weight.bin b/it/2240ms/decoder_joint_noencproj.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..80af5cec724e7e6b117fcd6a7bc8046c27c26e75 --- /dev/null +++ b/it/2240ms/decoder_joint_noencproj.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5ca6a467ebf44612a8032c6c4ddf323e35a9ffaa15c01822528fb97144ec9439 +size 16003660 diff --git a/it/2240ms/decoder_joint_noencproj.mlpackage/Data/com.apple.CoreML/model.mlmodel b/it/2240ms/decoder_joint_noencproj.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..492f62656a116fa907f67e5963520bcaa37729f8 --- /dev/null +++ b/it/2240ms/decoder_joint_noencproj.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1a46ef6436e8c1e5af6f379658d871f2a40dff3a13c946d4f8df736b19cdc45e +size 14630 diff --git a/it/2240ms/decoder_joint_noencproj.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/it/2240ms/decoder_joint_noencproj.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..80af5cec724e7e6b117fcd6a7bc8046c27c26e75 --- /dev/null +++ b/it/2240ms/decoder_joint_noencproj.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5ca6a467ebf44612a8032c6c4ddf323e35a9ffaa15c01822528fb97144ec9439 +size 16003660 diff --git a/it/2240ms/decoder_joint_noencproj.mlpackage/Manifest.json b/it/2240ms/decoder_joint_noencproj.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..2135f9be1a1ec8555bef5683841b910c941cbd7c --- /dev/null +++ b/it/2240ms/decoder_joint_noencproj.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "E92AE661-EE59-447C-B9BE-52D50BC4FC37": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + }, + "F7138F3F-8633-4C54-A3F4-D0F6E8220E39": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + } + }, + "rootModelIdentifier": "F7138F3F-8633-4C54-A3F4-D0F6E8220E39" +} diff --git a/it/2240ms/encoder.mlmodelc/analytics/coremldata.bin b/it/2240ms/encoder.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..cf222b05b3795ceca6d95b114574c50f82d5a9fc --- /dev/null +++ b/it/2240ms/encoder.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cb008a0254143d32e93073d5db48272287719969c062c9ac55514b11dc699a3f +size 243 diff --git a/it/2240ms/encoder.mlmodelc/coremldata.bin b/it/2240ms/encoder.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..86c2db47ccec63dd162fcc86c9e12d5a1321f631 --- /dev/null +++ b/it/2240ms/encoder.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:97f4d6bec344ff13f57cb7b727dc4f7c9d81c75074518715ab43502dd657b082 +size 633 diff --git a/it/2240ms/encoder.mlmodelc/model.mil b/it/2240ms/encoder.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..0b61772b118bff8a3d2b83d31f02ad558e04cde4 --- /dev/null +++ b/it/2240ms/encoder.mlmodelc/model.mil @@ -0,0 +1,4434 @@ +program(1.3) +[buildInfo = dict({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}})] +{ + func main(tensor cache_channel, tensor cache_len, tensor cache_time, tensor mel, tensor mel_length, tensor prompt_id) { + tensor value_3_perm_0 = const()[name = string("value_3_perm_0"), val = tensor([1, 0, 2, 3])]; + string cache_channel_to_fp16_dtype_0 = const()[name = string("cache_channel_to_fp16_dtype_0"), val = string("fp16")]; + tensor value_5_perm_0 = const()[name = string("value_5_perm_0"), val = tensor([1, 0, 2, 3])]; + string cache_time_to_fp16_dtype_0 = const()[name = string("cache_time_to_fp16_dtype_0"), val = string("fp16")]; + int32 var_59 = const()[name = string("op_59"), val = int32(-1)]; + int32 var_68 = const()[name = string("op_68"), val = int32(1)]; + tensor x_1_perm_0 = const()[name = string("x_1_perm_0"), val = tensor([0, 2, 1])]; + string mel_to_fp16_dtype_0 = const()[name = string("mel_to_fp16_dtype_0"), val = string("fp16")]; + tensor tensor_1_axes_0 = const()[name = string("tensor_1_axes_0"), val = tensor([1])]; + tensor mel_to_fp16 = cast(dtype = mel_to_fp16_dtype_0, x = mel)[name = string("cast_21")]; + tensor x_1_cast_fp16 = transpose(perm = x_1_perm_0, x = mel_to_fp16)[name = string("transpose_367")]; + tensor tensor_1_cast_fp16 = expand_dims(axes = tensor_1_axes_0, x = x_1_cast_fp16)[name = string("tensor_1_cast_fp16")]; + tensor expand_dims_0 = const()[name = string("expand_dims_0"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))]; + tensor var_137_axes_0 = const()[name = string("op_137_axes_0"), val = tensor([1])]; + tensor var_137 = expand_dims(axes = var_137_axes_0, x = mel_length)[name = string("op_137")]; + tensor time_mask_1 = less(x = expand_dims_0, y = var_137)[name = string("time_mask_1")]; + tensor var_139_axes_0 = const()[name = string("op_139_axes_0"), val = tensor([-1])]; + tensor var_139 = expand_dims(axes = var_139_axes_0, x = time_mask_1)[name = string("op_139")]; + tensor var_141_reps_0 = const()[name = string("op_141_reps_0"), val = tensor([1, 1, 128])]; + tensor var_141 = tile(reps = var_141_reps_0, x = var_139)[name = string("op_141")]; + tensor var_147_axes_0 = const()[name = string("op_147_axes_0"), val = tensor([1])]; + string mask_1_to_fp16_dtype_0 = const()[name = string("mask_1_to_fp16_dtype_0"), val = string("fp16")]; + tensor var_141_to_fp16 = cast(dtype = mask_1_to_fp16_dtype_0, x = var_141)[name = string("cast_20")]; + tensor var_147_cast_fp16 = expand_dims(axes = var_147_axes_0, x = var_141_to_fp16)[name = string("op_147_cast_fp16")]; + tensor input_1_cast_fp16 = mul(x = tensor_1_cast_fp16, y = var_147_cast_fp16)[name = string("input_1_cast_fp16")]; + tensor input_3_pad_0 = const()[name = string("input_3_pad_0"), val = tensor([0, 0, 0, 0, 2, 1, 2, 1])]; + string input_3_mode_0 = const()[name = string("input_3_mode_0"), val = string("constant")]; + fp16 const_9_to_fp16 = const()[name = string("const_9_to_fp16"), val = fp16(0x0p+0)]; + tensor input_3_cast_fp16 = pad(constant_val = const_9_to_fp16, mode = input_3_mode_0, pad = input_3_pad_0, x = input_1_cast_fp16)[name = string("input_3_cast_fp16")]; + string tensor_3_pad_type_0 = const()[name = string("tensor_3_pad_type_0"), val = string("valid")]; + tensor tensor_3_strides_0 = const()[name = string("tensor_3_strides_0"), val = tensor([2, 2])]; + tensor tensor_3_pad_0 = const()[name = string("tensor_3_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor tensor_3_dilations_0 = const()[name = string("tensor_3_dilations_0"), val = tensor([1, 1])]; + int32 tensor_3_groups_0 = const()[name = string("tensor_3_groups_0"), val = int32(1)]; + tensor encoder_pre_encode_conv_0_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1088))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3456))))[name = string("encoder_pre_encode_conv_0_weight_to_fp16_quantized")]; + tensor encoder_pre_encode_conv_0_bias_to_fp16 = const()[name = string("encoder_pre_encode_conv_0_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4032)))]; + tensor tensor_3_cast_fp16 = conv(bias = encoder_pre_encode_conv_0_bias_to_fp16, dilations = tensor_3_dilations_0, groups = tensor_3_groups_0, pad = tensor_3_pad_0, pad_type = tensor_3_pad_type_0, strides = tensor_3_strides_0, weight = encoder_pre_encode_conv_0_weight_to_fp16_quantized, x = input_3_cast_fp16)[name = string("tensor_3_cast_fp16")]; + string current_lengths_1_to_fp16_dtype_0 = const()[name = string("current_lengths_1_to_fp16_dtype_0"), val = string("fp16")]; + fp16 var_160_promoted_to_fp16 = const()[name = string("op_160_promoted_to_fp16"), val = fp16(0x1p+1)]; + tensor mel_length_to_fp16 = cast(dtype = current_lengths_1_to_fp16_dtype_0, x = mel_length)[name = string("cast_19")]; + tensor var_161_cast_fp16 = add(x = mel_length_to_fp16, y = var_160_promoted_to_fp16)[name = string("op_161_cast_fp16")]; + fp16 var_162_promoted_to_fp16 = const()[name = string("op_162_promoted_to_fp16"), val = fp16(0x1p+0)]; + tensor var_163_cast_fp16 = add(x = var_161_cast_fp16, y = var_162_promoted_to_fp16)[name = string("op_163_cast_fp16")]; + fp16 var_164_promoted_to_fp16 = const()[name = string("op_164_promoted_to_fp16"), val = fp16(0x1.8p+1)]; + tensor var_165_cast_fp16 = sub(x = var_163_cast_fp16, y = var_164_promoted_to_fp16)[name = string("op_165_cast_fp16")]; + fp16 var_56_promoted_to_fp16 = const()[name = string("op_56_promoted_to_fp16"), val = fp16(0x1p+1)]; + tensor floor_div_0_cast_fp16 = floor_div(x = var_165_cast_fp16, y = var_56_promoted_to_fp16)[name = string("floor_div_0_cast_fp16")]; + fp16 var_167_promoted_to_fp16 = const()[name = string("op_167_promoted_to_fp16"), val = fp16(0x1p+0)]; + tensor current_lengths_3_cast_fp16 = add(x = floor_div_0_cast_fp16, y = var_167_promoted_to_fp16)[name = string("current_lengths_3_cast_fp16")]; + string lengths_19_dtype_0 = const()[name = string("lengths_19_dtype_0"), val = string("int32")]; + tensor expand_dims_1 = const()[name = string("expand_dims_1"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4608)))]; + tensor var_176_axes_0 = const()[name = string("op_176_axes_0"), val = tensor([1])]; + tensor current_lengths_3_cast_fp16_to_int32 = cast(dtype = lengths_19_dtype_0, x = current_lengths_3_cast_fp16)[name = string("cast_18")]; + tensor var_176 = expand_dims(axes = var_176_axes_0, x = current_lengths_3_cast_fp16_to_int32)[name = string("op_176")]; + tensor time_mask_3 = less(x = expand_dims_1, y = var_176)[name = string("time_mask_3")]; + tensor var_178_axes_0 = const()[name = string("op_178_axes_0"), val = tensor([-1])]; + tensor var_178 = expand_dims(axes = var_178_axes_0, x = time_mask_3)[name = string("op_178")]; + tensor var_180_reps_0 = const()[name = string("op_180_reps_0"), val = tensor([1, 1, 65])]; + tensor var_180 = tile(reps = var_180_reps_0, x = var_178)[name = string("op_180")]; + tensor var_186_axes_0 = const()[name = string("op_186_axes_0"), val = tensor([1])]; + string mask_3_to_fp16_dtype_0 = const()[name = string("mask_3_to_fp16_dtype_0"), val = string("fp16")]; + tensor var_180_to_fp16 = cast(dtype = mask_3_to_fp16_dtype_0, x = var_180)[name = string("cast_17")]; + tensor var_186_cast_fp16 = expand_dims(axes = var_186_axes_0, x = var_180_to_fp16)[name = string("op_186_cast_fp16")]; + tensor expanded_mask_3_reps_0 = const()[name = string("expanded_mask_3_reps_0"), val = tensor([1, 256, 1, 1])]; + tensor expanded_mask_3_cast_fp16 = tile(reps = expanded_mask_3_reps_0, x = var_186_cast_fp16)[name = string("expanded_mask_3_cast_fp16")]; + tensor input_5_cast_fp16 = mul(x = tensor_3_cast_fp16, y = expanded_mask_3_cast_fp16)[name = string("input_5_cast_fp16")]; + tensor tensor_5_cast_fp16 = relu(x = input_5_cast_fp16)[name = string("tensor_5_cast_fp16")]; + tensor input_7_cast_fp16 = mul(x = tensor_5_cast_fp16, y = expanded_mask_3_cast_fp16)[name = string("input_7_cast_fp16")]; + tensor input_9_pad_0 = const()[name = string("input_9_pad_0"), val = tensor([0, 0, 0, 0, 2, 1, 2, 1])]; + string input_9_mode_0 = const()[name = string("input_9_mode_0"), val = string("constant")]; + fp16 const_23_to_fp16 = const()[name = string("const_23_to_fp16"), val = fp16(0x0p+0)]; + tensor input_9_cast_fp16 = pad(constant_val = const_23_to_fp16, mode = input_9_mode_0, pad = input_9_pad_0, x = input_7_cast_fp16)[name = string("input_9_cast_fp16")]; + string tensor_7_pad_type_0 = const()[name = string("tensor_7_pad_type_0"), val = string("valid")]; + tensor tensor_7_strides_0 = const()[name = string("tensor_7_strides_0"), val = tensor([2, 2])]; + int32 tensor_7_groups_0 = const()[name = string("tensor_7_groups_0"), val = int32(256)]; + tensor tensor_7_pad_0 = const()[name = string("tensor_7_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor tensor_7_dilations_0 = const()[name = string("tensor_7_dilations_0"), val = tensor([1, 1])]; + tensor encoder_pre_encode_conv_2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5184))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7552))))[name = string("encoder_pre_encode_conv_2_weight_to_fp16_quantized")]; + tensor encoder_pre_encode_conv_2_bias_to_fp16 = const()[name = string("encoder_pre_encode_conv_2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8128)))]; + tensor tensor_7_cast_fp16 = conv(bias = encoder_pre_encode_conv_2_bias_to_fp16, dilations = tensor_7_dilations_0, groups = tensor_7_groups_0, pad = tensor_7_pad_0, pad_type = tensor_7_pad_type_0, strides = tensor_7_strides_0, weight = encoder_pre_encode_conv_2_weight_to_fp16_quantized, x = input_9_cast_fp16)[name = string("tensor_7_cast_fp16")]; + fp16 var_208_promoted_to_fp16 = const()[name = string("op_208_promoted_to_fp16"), val = fp16(0x1p+1)]; + tensor var_209_cast_fp16 = add(x = current_lengths_3_cast_fp16, y = var_208_promoted_to_fp16)[name = string("op_209_cast_fp16")]; + fp16 var_210_promoted_to_fp16 = const()[name = string("op_210_promoted_to_fp16"), val = fp16(0x1p+0)]; + tensor var_211_cast_fp16 = add(x = var_209_cast_fp16, y = var_210_promoted_to_fp16)[name = string("op_211_cast_fp16")]; + fp16 var_212_promoted_to_fp16 = const()[name = string("op_212_promoted_to_fp16"), val = fp16(0x1.8p+1)]; + tensor var_213_cast_fp16 = sub(x = var_211_cast_fp16, y = var_212_promoted_to_fp16)[name = string("op_213_cast_fp16")]; + fp16 var_56_promoted_1_to_fp16 = const()[name = string("op_56_promoted_1_to_fp16"), val = fp16(0x1p+1)]; + tensor floor_div_1_cast_fp16 = floor_div(x = var_213_cast_fp16, y = var_56_promoted_1_to_fp16)[name = string("floor_div_1_cast_fp16")]; + fp16 var_215_promoted_to_fp16 = const()[name = string("op_215_promoted_to_fp16"), val = fp16(0x1p+0)]; + tensor current_lengths_5_cast_fp16 = add(x = floor_div_1_cast_fp16, y = var_215_promoted_to_fp16)[name = string("current_lengths_5_cast_fp16")]; + string lengths_21_dtype_0 = const()[name = string("lengths_21_dtype_0"), val = string("int32")]; + tensor expand_dims_2 = const()[name = string("expand_dims_2"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8704)))]; + tensor var_224_axes_0 = const()[name = string("op_224_axes_0"), val = tensor([1])]; + tensor current_lengths_5_cast_fp16_to_int32 = cast(dtype = lengths_21_dtype_0, x = current_lengths_5_cast_fp16)[name = string("cast_16")]; + tensor var_224 = expand_dims(axes = var_224_axes_0, x = current_lengths_5_cast_fp16_to_int32)[name = string("op_224")]; + tensor time_mask_5 = less(x = expand_dims_2, y = var_224)[name = string("time_mask_5")]; + tensor var_226_axes_0 = const()[name = string("op_226_axes_0"), val = tensor([-1])]; + tensor var_226 = expand_dims(axes = var_226_axes_0, x = time_mask_5)[name = string("op_226")]; + tensor var_228_reps_0 = const()[name = string("op_228_reps_0"), val = tensor([1, 1, 33])]; + tensor var_228 = tile(reps = var_228_reps_0, x = var_226)[name = string("op_228")]; + tensor var_234_axes_0 = const()[name = string("op_234_axes_0"), val = tensor([1])]; + string mask_5_to_fp16_dtype_0 = const()[name = string("mask_5_to_fp16_dtype_0"), val = string("fp16")]; + tensor var_228_to_fp16 = cast(dtype = mask_5_to_fp16_dtype_0, x = var_228)[name = string("cast_15")]; + tensor var_234_cast_fp16 = expand_dims(axes = var_234_axes_0, x = var_228_to_fp16)[name = string("op_234_cast_fp16")]; + tensor expanded_mask_7_reps_0 = const()[name = string("expanded_mask_7_reps_0"), val = tensor([1, 256, 1, 1])]; + tensor expanded_mask_7_cast_fp16 = tile(reps = expanded_mask_7_reps_0, x = var_234_cast_fp16)[name = string("expanded_mask_7_cast_fp16")]; + tensor input_11_cast_fp16 = mul(x = tensor_7_cast_fp16, y = expanded_mask_7_cast_fp16)[name = string("input_11_cast_fp16")]; + string tensor_9_pad_type_0 = const()[name = string("tensor_9_pad_type_0"), val = string("valid")]; + tensor tensor_9_strides_0 = const()[name = string("tensor_9_strides_0"), val = tensor([1, 1])]; + tensor tensor_9_pad_0 = const()[name = string("tensor_9_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor tensor_9_dilations_0 = const()[name = string("tensor_9_dilations_0"), val = tensor([1, 1])]; + int32 tensor_9_groups_0 = const()[name = string("tensor_9_groups_0"), val = int32(1)]; + tensor encoder_pre_encode_conv_3_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(9024))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(74624))))[name = string("encoder_pre_encode_conv_3_weight_to_fp16_quantized")]; + tensor encoder_pre_encode_conv_3_bias_to_fp16 = const()[name = string("encoder_pre_encode_conv_3_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75200)))]; + tensor tensor_9_cast_fp16 = conv(bias = encoder_pre_encode_conv_3_bias_to_fp16, dilations = tensor_9_dilations_0, groups = tensor_9_groups_0, pad = tensor_9_pad_0, pad_type = tensor_9_pad_type_0, strides = tensor_9_strides_0, weight = encoder_pre_encode_conv_3_weight_to_fp16_quantized, x = input_11_cast_fp16)[name = string("tensor_9_cast_fp16")]; + tensor input_13_cast_fp16 = mul(x = tensor_9_cast_fp16, y = expanded_mask_7_cast_fp16)[name = string("input_13_cast_fp16")]; + tensor tensor_11_cast_fp16 = relu(x = input_13_cast_fp16)[name = string("tensor_11_cast_fp16")]; + tensor input_15_cast_fp16 = mul(x = tensor_11_cast_fp16, y = expanded_mask_7_cast_fp16)[name = string("input_15_cast_fp16")]; + tensor input_17_pad_0 = const()[name = string("input_17_pad_0"), val = tensor([0, 0, 0, 0, 2, 1, 2, 1])]; + string input_17_mode_0 = const()[name = string("input_17_mode_0"), val = string("constant")]; + fp16 const_41_to_fp16 = const()[name = string("const_41_to_fp16"), val = fp16(0x0p+0)]; + tensor input_17_cast_fp16 = pad(constant_val = const_41_to_fp16, mode = input_17_mode_0, pad = input_17_pad_0, x = input_15_cast_fp16)[name = string("input_17_cast_fp16")]; + string tensor_13_pad_type_0 = const()[name = string("tensor_13_pad_type_0"), val = string("valid")]; + tensor tensor_13_strides_0 = const()[name = string("tensor_13_strides_0"), val = tensor([2, 2])]; + int32 tensor_13_groups_0 = const()[name = string("tensor_13_groups_0"), val = int32(256)]; + tensor tensor_13_pad_0 = const()[name = string("tensor_13_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor tensor_13_dilations_0 = const()[name = string("tensor_13_dilations_0"), val = tensor([1, 1])]; + tensor encoder_pre_encode_conv_5_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75776))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78144))))[name = string("encoder_pre_encode_conv_5_weight_to_fp16_quantized")]; + tensor encoder_pre_encode_conv_5_bias_to_fp16 = const()[name = string("encoder_pre_encode_conv_5_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78720)))]; + tensor tensor_13_cast_fp16 = conv(bias = encoder_pre_encode_conv_5_bias_to_fp16, dilations = tensor_13_dilations_0, groups = tensor_13_groups_0, pad = tensor_13_pad_0, pad_type = tensor_13_pad_type_0, strides = tensor_13_strides_0, weight = encoder_pre_encode_conv_5_weight_to_fp16_quantized, x = input_17_cast_fp16)[name = string("tensor_13_cast_fp16")]; + fp16 var_271_promoted_to_fp16 = const()[name = string("op_271_promoted_to_fp16"), val = fp16(0x1p+1)]; + tensor var_272_cast_fp16 = add(x = current_lengths_5_cast_fp16, y = var_271_promoted_to_fp16)[name = string("op_272_cast_fp16")]; + fp16 var_273_promoted_to_fp16 = const()[name = string("op_273_promoted_to_fp16"), val = fp16(0x1p+0)]; + tensor var_274_cast_fp16 = add(x = var_272_cast_fp16, y = var_273_promoted_to_fp16)[name = string("op_274_cast_fp16")]; + fp16 var_275_promoted_to_fp16 = const()[name = string("op_275_promoted_to_fp16"), val = fp16(0x1.8p+1)]; + tensor var_276_cast_fp16 = sub(x = var_274_cast_fp16, y = var_275_promoted_to_fp16)[name = string("op_276_cast_fp16")]; + fp16 var_56_promoted_2_to_fp16 = const()[name = string("op_56_promoted_2_to_fp16"), val = fp16(0x1p+1)]; + tensor floor_div_2_cast_fp16 = floor_div(x = var_276_cast_fp16, y = var_56_promoted_2_to_fp16)[name = string("floor_div_2_cast_fp16")]; + fp16 var_278_promoted_to_fp16 = const()[name = string("op_278_promoted_to_fp16"), val = fp16(0x1p+0)]; + tensor current_lengths_cast_fp16 = add(x = floor_div_2_cast_fp16, y = var_278_promoted_to_fp16)[name = string("current_lengths_cast_fp16")]; + string lengths_dtype_0 = const()[name = string("lengths_dtype_0"), val = string("int32")]; + tensor expand_dims_3 = const()[name = string("expand_dims_3"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(79296)))]; + tensor var_287_axes_0 = const()[name = string("op_287_axes_0"), val = tensor([1])]; + tensor current_lengths_cast_fp16_to_int32 = cast(dtype = lengths_dtype_0, x = current_lengths_cast_fp16)[name = string("cast_14")]; + tensor var_287 = expand_dims(axes = var_287_axes_0, x = current_lengths_cast_fp16_to_int32)[name = string("op_287")]; + tensor time_mask = less(x = expand_dims_3, y = var_287)[name = string("time_mask")]; + tensor var_289_axes_0 = const()[name = string("op_289_axes_0"), val = tensor([-1])]; + tensor var_289 = expand_dims(axes = var_289_axes_0, x = time_mask)[name = string("op_289")]; + tensor var_291_reps_0 = const()[name = string("op_291_reps_0"), val = tensor([1, 1, 17])]; + tensor var_291 = tile(reps = var_291_reps_0, x = var_289)[name = string("op_291")]; + tensor var_297_axes_0 = const()[name = string("op_297_axes_0"), val = tensor([1])]; + string mask_7_to_fp16_dtype_0 = const()[name = string("mask_7_to_fp16_dtype_0"), val = string("fp16")]; + tensor var_291_to_fp16 = cast(dtype = mask_7_to_fp16_dtype_0, x = var_291)[name = string("cast_13")]; + tensor var_297_cast_fp16 = expand_dims(axes = var_297_axes_0, x = var_291_to_fp16)[name = string("op_297_cast_fp16")]; + tensor expanded_mask_13_reps_0 = const()[name = string("expanded_mask_13_reps_0"), val = tensor([1, 256, 1, 1])]; + tensor expanded_mask_13_cast_fp16 = tile(reps = expanded_mask_13_reps_0, x = var_297_cast_fp16)[name = string("expanded_mask_13_cast_fp16")]; + tensor input_19_cast_fp16 = mul(x = tensor_13_cast_fp16, y = expanded_mask_13_cast_fp16)[name = string("input_19_cast_fp16")]; + string tensor_15_pad_type_0 = const()[name = string("tensor_15_pad_type_0"), val = string("valid")]; + tensor tensor_15_strides_0 = const()[name = string("tensor_15_strides_0"), val = tensor([1, 1])]; + tensor tensor_15_pad_0 = const()[name = string("tensor_15_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor tensor_15_dilations_0 = const()[name = string("tensor_15_dilations_0"), val = tensor([1, 1])]; + int32 tensor_15_groups_0 = const()[name = string("tensor_15_groups_0"), val = int32(1)]; + tensor encoder_pre_encode_conv_6_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(79488))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(145088))))[name = string("encoder_pre_encode_conv_6_weight_to_fp16_quantized")]; + tensor encoder_pre_encode_conv_6_bias_to_fp16 = const()[name = string("encoder_pre_encode_conv_6_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(145664)))]; + tensor tensor_15_cast_fp16 = conv(bias = encoder_pre_encode_conv_6_bias_to_fp16, dilations = tensor_15_dilations_0, groups = tensor_15_groups_0, pad = tensor_15_pad_0, pad_type = tensor_15_pad_type_0, strides = tensor_15_strides_0, weight = encoder_pre_encode_conv_6_weight_to_fp16_quantized, x = input_19_cast_fp16)[name = string("tensor_15_cast_fp16")]; + tensor input_21_cast_fp16 = mul(x = tensor_15_cast_fp16, y = expanded_mask_13_cast_fp16)[name = string("input_21_cast_fp16")]; + tensor tensor_cast_fp16 = relu(x = input_21_cast_fp16)[name = string("tensor_cast_fp16")]; + tensor x_3_cast_fp16 = mul(x = tensor_cast_fp16, y = expanded_mask_13_cast_fp16)[name = string("x_3_cast_fp16")]; + tensor var_331_perm_0 = const()[name = string("op_331_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_332 = const()[name = string("op_332"), val = tensor([1, 30, -1])]; + tensor var_331_cast_fp16 = transpose(perm = var_331_perm_0, x = x_3_cast_fp16)[name = string("transpose_366")]; + tensor input_23_cast_fp16 = reshape(shape = var_332, x = var_331_cast_fp16)[name = string("input_23_cast_fp16")]; + tensor encoder_pre_encode_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(146240))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4602752))))[name = string("encoder_pre_encode_out_weight_to_fp16_quantized")]; + tensor encoder_pre_encode_out_bias_to_fp16 = const()[name = string("encoder_pre_encode_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4604864)))]; + tensor linear_0_cast_fp16 = linear(bias = encoder_pre_encode_out_bias_to_fp16, weight = encoder_pre_encode_out_weight_to_fp16_quantized, x = input_23_cast_fp16)[name = string("linear_0_cast_fp16")]; + tensor var_342_begin_0 = const()[name = string("op_342_begin_0"), val = tensor([0, 2, 0])]; + tensor var_342_end_0 = const()[name = string("op_342_end_0"), val = tensor([1, 30, 1024])]; + tensor var_342_end_mask_0 = const()[name = string("op_342_end_mask_0"), val = tensor([true, true, true])]; + tensor var_342_cast_fp16 = slice_by_index(begin = var_342_begin_0, end = var_342_end_0, end_mask = var_342_end_mask_0, x = linear_0_cast_fp16)[name = string("op_342_cast_fp16")]; + int32 var_344 = const()[name = string("op_344"), val = int32(2)]; + tensor var_345 = sub(x = current_lengths_cast_fp16_to_int32, y = var_344)[name = string("op_345")]; + string var_345_promoted_to_fp16_dtype_0 = const()[name = string("op_345_promoted_to_fp16_dtype_0"), val = string("fp16")]; + fp16 var_62_promoted_to_fp16 = const()[name = string("op_62_promoted_to_fp16"), val = fp16(0x0p+0)]; + fp16 const_61_to_fp16 = const()[name = string("const_61_to_fp16"), val = fp16(inf)]; + tensor var_345_to_fp16 = cast(dtype = var_345_promoted_to_fp16_dtype_0, x = var_345)[name = string("cast_12")]; + tensor clip_0_cast_fp16 = clip(alpha = var_62_promoted_to_fp16, beta = const_61_to_fp16, x = var_345_to_fp16)[name = string("clip_0_cast_fp16")]; + tensor max_audio_length_1 = const()[name = string("max_audio_length_1"), val = tensor([28])]; + fp16 var_361_promoted_to_fp16 = const()[name = string("op_361_promoted_to_fp16"), val = fp16(0x1.5p+5)]; + tensor padding_length_cast_fp16 = add(x = clip_0_cast_fp16, y = var_361_promoted_to_fp16)[name = string("padding_length_cast_fp16")]; + int32 const_63 = const()[name = string("const_63"), val = int32(-1)]; + tensor var_363 = mul(x = cache_len, y = const_63)[name = string("op_363")]; + int32 var_364 = const()[name = string("op_364"), val = int32(42)]; + tensor offset = add(x = var_363, y = var_364)[name = string("offset")]; + tensor var_404_axes_0 = const()[name = string("op_404_axes_0"), val = tensor([-1])]; + tensor var_404_cast_fp16 = expand_dims(axes = var_404_axes_0, x = padding_length_cast_fp16)[name = string("op_404_cast_fp16")]; + tensor var_403_promoted_to_fp16 = const()[name = string("op_403_promoted_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4606976)))]; + tensor pad_mask_1_cast_fp16 = less(x = var_403_promoted_to_fp16, y = var_404_cast_fp16)[name = string("pad_mask_1_cast_fp16")]; + tensor expand_dims_5 = const()[name = string("expand_dims_5"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4607232)))]; + tensor var_410_axes_0 = const()[name = string("op_410_axes_0"), val = tensor([-1])]; + tensor var_410 = expand_dims(axes = var_410_axes_0, x = offset)[name = string("op_410")]; + tensor pad_mask_off = greater_equal(x = expand_dims_5, y = var_410)[name = string("pad_mask_off")]; + tensor pad_mask_3 = logical_and(x = pad_mask_off, y = pad_mask_1_cast_fp16)[name = string("pad_mask_3")]; + tensor var_413_axes_0 = const()[name = string("op_413_axes_0"), val = tensor([1])]; + tensor var_413 = expand_dims(axes = var_413_axes_0, x = pad_mask_3)[name = string("op_413")]; + tensor var_414 = const()[name = string("op_414"), val = tensor([1, 70, 1])]; + tensor pad_mask_for_att_mask_1 = tile(reps = var_414, x = var_413)[name = string("pad_mask_for_att_mask_1")]; + tensor var_416_perm_0 = const()[name = string("op_416_perm_0"), val = tensor([0, 2, 1])]; + tensor var_416 = transpose(perm = var_416_perm_0, x = pad_mask_for_att_mask_1)[name = string("transpose_365")]; + tensor pad_mask_for_att_mask = logical_and(x = pad_mask_for_att_mask_1, y = var_416)[name = string("pad_mask_for_att_mask")]; + tensor const_71 = const()[name = string("const_71"), val = tensor([[[true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true]]])]; + tensor att_mask_9 = logical_and(x = pad_mask_for_att_mask, y = const_71)[name = string("att_mask_9")]; + tensor att_mask = logical_not(x = att_mask_9)[name = string("att_mask")]; + tensor pad_mask_5 = logical_not(x = pad_mask_3)[name = string("pad_mask_5")]; + tensor pad_mask_begin_0 = const()[name = string("pad_mask_begin_0"), val = tensor([0, 42])]; + tensor pad_mask_end_0 = const()[name = string("pad_mask_end_0"), val = tensor([1, 70])]; + tensor pad_mask_end_mask_0 = const()[name = string("pad_mask_end_mask_0"), val = tensor([true, true])]; + tensor pad_mask = slice_by_index(begin = pad_mask_begin_0, end = pad_mask_end_0, end_mask = pad_mask_end_mask_0, x = pad_mask_5)[name = string("pad_mask")]; + tensor mask_9_begin_0 = const()[name = string("mask_9_begin_0"), val = tensor([0, 42, 0])]; + tensor mask_9_end_0 = const()[name = string("mask_9_end_0"), val = tensor([1, 70, 70])]; + tensor mask_9_end_mask_0 = const()[name = string("mask_9_end_mask_0"), val = tensor([true, true, true])]; + tensor mask_9 = slice_by_index(begin = mask_9_begin_0, end = mask_9_end_0, end_mask = mask_9_end_mask_0, x = att_mask)[name = string("mask_9")]; + tensor cache_1_begin_0 = const()[name = string("cache_1_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor cache_1_end_0 = const()[name = string("cache_1_end_0"), val = tensor([1, 1, 42, 1024])]; + tensor cache_1_end_mask_0 = const()[name = string("cache_1_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_1_squeeze_mask_0 = const()[name = string("cache_1_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_channel_to_fp16 = cast(dtype = cache_channel_to_fp16_dtype_0, x = cache_channel)[name = string("cast_11")]; + tensor value_3_cast_fp16 = transpose(perm = value_3_perm_0, x = cache_channel_to_fp16)[name = string("transpose_364")]; + tensor cache_1_cast_fp16 = slice_by_index(begin = cache_1_begin_0, end = cache_1_end_0, end_mask = cache_1_end_mask_0, squeeze_mask = cache_1_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_1_cast_fp16")]; + tensor cache_3_begin_0 = const()[name = string("cache_3_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor cache_3_end_0 = const()[name = string("cache_3_end_0"), val = tensor([1, 1, 1024, 8])]; + tensor cache_3_end_mask_0 = const()[name = string("cache_3_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_3_squeeze_mask_0 = const()[name = string("cache_3_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_time_to_fp16 = cast(dtype = cache_time_to_fp16_dtype_0, x = cache_time)[name = string("cast_10")]; + tensor value_5_cast_fp16 = transpose(perm = value_5_perm_0, x = cache_time_to_fp16)[name = string("transpose_363")]; + tensor cache_3_cast_fp16 = slice_by_index(begin = cache_3_begin_0, end = cache_3_end_0, end_mask = cache_3_end_mask_0, squeeze_mask = cache_3_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_3_cast_fp16")]; + tensor input_27_axes_0 = const()[name = string("input_27_axes_0"), val = tensor([-1])]; + tensor encoder_layers_0_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_0_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4607616)))]; + tensor encoder_layers_0_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_0_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4609728)))]; + fp16 var_42_to_fp16 = const()[name = string("op_42_to_fp16"), val = fp16(0x1.5p-17)]; + tensor input_27_cast_fp16 = layer_norm(axes = input_27_axes_0, beta = encoder_layers_0_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_0_norm_feed_forward1_weight_to_fp16, x = var_342_cast_fp16)[name = string("input_27_cast_fp16")]; + tensor encoder_layers_0_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4611840))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8806208))))[name = string("encoder_layers_0_feed_forward1_linear1_weight_to_fp16_quantized")]; + tensor encoder_layers_0_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_0_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8814464)))]; + tensor linear_1_cast_fp16 = linear(bias = encoder_layers_0_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_0_feed_forward1_linear1_weight_to_fp16_quantized, x = input_27_cast_fp16)[name = string("linear_1_cast_fp16")]; + tensor input_31_cast_fp16 = silu(x = linear_1_cast_fp16)[name = string("input_31_cast_fp16")]; + tensor encoder_layers_0_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8822720))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13017088))))[name = string("encoder_layers_0_feed_forward1_linear2_weight_to_fp16_quantized")]; + tensor encoder_layers_0_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_0_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13019200)))]; + tensor linear_2_cast_fp16 = linear(bias = encoder_layers_0_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_0_feed_forward1_linear2_weight_to_fp16_quantized, x = input_31_cast_fp16)[name = string("linear_2_cast_fp16")]; + fp16 var_455_to_fp16 = const()[name = string("op_455_to_fp16"), val = fp16(0x1p-1)]; + tensor var_456_cast_fp16 = mul(x = linear_2_cast_fp16, y = var_455_to_fp16)[name = string("op_456_cast_fp16")]; + tensor input_37_cast_fp16 = add(x = var_342_cast_fp16, y = var_456_cast_fp16)[name = string("input_37_cast_fp16")]; + tensor key_1_axes_0 = const()[name = string("key_1_axes_0"), val = tensor([-1])]; + tensor encoder_layers_0_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_0_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13021312)))]; + tensor encoder_layers_0_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_0_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13023424)))]; + tensor key_1_cast_fp16 = layer_norm(axes = key_1_axes_0, beta = encoder_layers_0_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_0_norm_self_att_weight_to_fp16, x = input_37_cast_fp16)[name = string("key_1_cast_fp16")]; + bool input_39_interleave_0 = const()[name = string("input_39_interleave_0"), val = bool(false)]; + tensor input_39_cast_fp16 = concat(axis = var_68, interleave = input_39_interleave_0, values = (cache_1_cast_fp16, key_1_cast_fp16))[name = string("input_39_cast_fp16")]; + tensor var_478_begin_0 = const()[name = string("op_478_begin_0"), val = tensor([0, 28, 0])]; + tensor var_478_end_0 = const()[name = string("op_478_end_0"), val = tensor([1, 42, 1024])]; + tensor var_478_end_mask_0 = const()[name = string("op_478_end_mask_0"), val = tensor([true, true, true])]; + tensor var_478_cast_fp16 = slice_by_index(begin = var_478_begin_0, end = var_478_end_0, end_mask = var_478_end_mask_0, x = cache_1_cast_fp16)[name = string("op_478_cast_fp16")]; + bool var_484_interleave_0 = const()[name = string("op_484_interleave_0"), val = bool(false)]; + tensor var_484_cast_fp16 = concat(axis = var_68, interleave = var_484_interleave_0, values = (var_478_cast_fp16, key_1_cast_fp16))[name = string("op_484_cast_fp16")]; + tensor encoder_layers_0_self_attn_linear_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13025536))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14074176))))[name = string("encoder_layers_0_self_attn_linear_q_weight_to_fp16_quantized")]; + tensor encoder_layers_0_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_0_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14076288)))]; + tensor linear_3_cast_fp16 = linear(bias = encoder_layers_0_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_0_self_attn_linear_q_weight_to_fp16_quantized, x = key_1_cast_fp16)[name = string("linear_3_cast_fp16")]; + tensor var_489 = const()[name = string("op_489"), val = tensor([1, -1, 8, 128])]; + tensor q_1_cast_fp16 = reshape(shape = var_489, x = linear_3_cast_fp16)[name = string("q_1_cast_fp16")]; + tensor encoder_layers_0_self_attn_linear_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14078400))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15127040))))[name = string("encoder_layers_0_self_attn_linear_k_weight_to_fp16_quantized")]; + tensor encoder_layers_0_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_0_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15129152)))]; + tensor linear_4_cast_fp16 = linear(bias = encoder_layers_0_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_0_self_attn_linear_k_weight_to_fp16_quantized, x = input_39_cast_fp16)[name = string("linear_4_cast_fp16")]; + tensor var_494 = const()[name = string("op_494"), val = tensor([1, -1, 8, 128])]; + tensor k_1_cast_fp16 = reshape(shape = var_494, x = linear_4_cast_fp16)[name = string("k_1_cast_fp16")]; + tensor encoder_layers_0_self_attn_linear_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15131264))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16179904))))[name = string("encoder_layers_0_self_attn_linear_v_weight_to_fp16_quantized")]; + tensor encoder_layers_0_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_0_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16182016)))]; + tensor linear_5_cast_fp16 = linear(bias = encoder_layers_0_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_0_self_attn_linear_v_weight_to_fp16_quantized, x = input_39_cast_fp16)[name = string("linear_5_cast_fp16")]; + tensor var_499 = const()[name = string("op_499"), val = tensor([1, -1, 8, 128])]; + tensor v_1_cast_fp16 = reshape(shape = var_499, x = linear_5_cast_fp16)[name = string("v_1_cast_fp16")]; + tensor value_9_perm_0 = const()[name = string("value_9_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_0_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_0_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16184128)))]; + tensor var_512_cast_fp16 = add(x = q_1_cast_fp16, y = encoder_layers_0_self_attn_pos_bias_u_to_fp16)[name = string("op_512_cast_fp16")]; + tensor encoder_layers_0_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_0_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16186240)))]; + tensor var_514_cast_fp16 = add(x = q_1_cast_fp16, y = encoder_layers_0_self_attn_pos_bias_v_to_fp16)[name = string("op_514_cast_fp16")]; + tensor q_with_bias_v_1_perm_0 = const()[name = string("q_with_bias_v_1_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_7_transpose_x_0 = const()[name = string("x_7_transpose_x_0"), val = bool(false)]; + bool x_7_transpose_y_0 = const()[name = string("x_7_transpose_y_0"), val = bool(false)]; + tensor op_516_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16188352))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16330752))))[name = string("op_516_to_fp16_quantized")]; + tensor q_with_bias_v_1_cast_fp16 = transpose(perm = q_with_bias_v_1_perm_0, x = var_514_cast_fp16)[name = string("transpose_362")]; + tensor x_7_cast_fp16 = matmul(transpose_x = x_7_transpose_x_0, transpose_y = x_7_transpose_y_0, x = q_with_bias_v_1_cast_fp16, y = op_516_to_fp16_quantized)[name = string("x_7_cast_fp16")]; + tensor x_9_pad_0 = const()[name = string("x_9_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_9_mode_0 = const()[name = string("x_9_mode_0"), val = string("constant")]; + fp16 const_79_to_fp16 = const()[name = string("const_79_to_fp16"), val = fp16(0x0p+0)]; + tensor x_9_cast_fp16 = pad(constant_val = const_79_to_fp16, mode = x_9_mode_0, pad = x_9_pad_0, x = x_7_cast_fp16)[name = string("x_9_cast_fp16")]; + tensor var_524 = const()[name = string("op_524"), val = tensor([1, 8, -1, 28])]; + tensor x_11_cast_fp16 = reshape(shape = var_524, x = x_9_cast_fp16)[name = string("x_11_cast_fp16")]; + tensor var_528_begin_0 = const()[name = string("op_528_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_528_end_0 = const()[name = string("op_528_end_0"), val = tensor([1, 8, 140, 28])]; + tensor var_528_end_mask_0 = const()[name = string("op_528_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_528_cast_fp16 = slice_by_index(begin = var_528_begin_0, end = var_528_end_0, end_mask = var_528_end_mask_0, x = x_11_cast_fp16)[name = string("op_528_cast_fp16")]; + tensor var_529 = const()[name = string("op_529"), val = tensor([1, 8, 28, 139])]; + tensor matrix_bd_1_cast_fp16 = reshape(shape = var_529, x = var_528_cast_fp16)[name = string("matrix_bd_1_cast_fp16")]; + bool matrix_ac_1_transpose_x_0 = const()[name = string("matrix_ac_1_transpose_x_0"), val = bool(false)]; + bool matrix_ac_1_transpose_y_0 = const()[name = string("matrix_ac_1_transpose_y_0"), val = bool(false)]; + tensor transpose_96_perm_0 = const()[name = string("transpose_96_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_97_perm_0 = const()[name = string("transpose_97_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_97 = transpose(perm = transpose_97_perm_0, x = k_1_cast_fp16)[name = string("transpose_360")]; + tensor transpose_96 = transpose(perm = transpose_96_perm_0, x = var_512_cast_fp16)[name = string("transpose_361")]; + tensor matrix_ac_1_cast_fp16 = matmul(transpose_x = matrix_ac_1_transpose_x_0, transpose_y = matrix_ac_1_transpose_y_0, x = transpose_96, y = transpose_97)[name = string("matrix_ac_1_cast_fp16")]; + tensor matrix_bd_3_begin_0 = const()[name = string("matrix_bd_3_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_3_end_0 = const()[name = string("matrix_bd_3_end_0"), val = tensor([1, 8, 28, 70])]; + tensor matrix_bd_3_end_mask_0 = const()[name = string("matrix_bd_3_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_3_cast_fp16 = slice_by_index(begin = matrix_bd_3_begin_0, end = matrix_bd_3_end_0, end_mask = matrix_bd_3_end_mask_0, x = matrix_bd_1_cast_fp16)[name = string("matrix_bd_3_cast_fp16")]; + tensor var_538_cast_fp16 = add(x = matrix_ac_1_cast_fp16, y = matrix_bd_3_cast_fp16)[name = string("op_538_cast_fp16")]; + fp16 _inversed_scores_1_y_0_to_fp16 = const()[name = string("_inversed_scores_1_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_1_cast_fp16 = mul(x = var_538_cast_fp16, y = _inversed_scores_1_y_0_to_fp16)[name = string("_inversed_scores_1_cast_fp16")]; + tensor mask_11_axes_0 = const()[name = string("mask_11_axes_0"), val = tensor([1])]; + tensor mask_11 = expand_dims(axes = mask_11_axes_0, x = mask_9)[name = string("mask_11")]; + fp16 var_45_to_fp16 = const()[name = string("op_45_to_fp16"), val = fp16(-0x1.388p+13)]; + tensor scores_3_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_1_cast_fp16, cond = mask_11)[name = string("scores_3_cast_fp16")]; + tensor var_544_cast_fp16 = softmax(axis = var_59, x = scores_3_cast_fp16)[name = string("op_544_cast_fp16")]; + fp16 var_44_to_fp16 = const()[name = string("op_44_to_fp16"), val = fp16(0x0p+0)]; + tensor input_41_cast_fp16 = select(a = var_44_to_fp16, b = var_544_cast_fp16, cond = mask_11)[name = string("input_41_cast_fp16")]; + bool x_13_transpose_x_0 = const()[name = string("x_13_transpose_x_0"), val = bool(false)]; + bool x_13_transpose_y_0 = const()[name = string("x_13_transpose_y_0"), val = bool(false)]; + tensor value_9_cast_fp16 = transpose(perm = value_9_perm_0, x = v_1_cast_fp16)[name = string("transpose_359")]; + tensor x_13_cast_fp16 = matmul(transpose_x = x_13_transpose_x_0, transpose_y = x_13_transpose_y_0, x = input_41_cast_fp16, y = value_9_cast_fp16)[name = string("x_13_cast_fp16")]; + tensor var_548_perm_0 = const()[name = string("op_548_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_549 = const()[name = string("op_549"), val = tensor([1, -1, 1024])]; + tensor var_548_cast_fp16 = transpose(perm = var_548_perm_0, x = x_13_cast_fp16)[name = string("transpose_358")]; + tensor input_43_cast_fp16 = reshape(shape = var_549, x = var_548_cast_fp16)[name = string("input_43_cast_fp16")]; + tensor encoder_layers_0_self_attn_linear_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16331136))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17379776))))[name = string("encoder_layers_0_self_attn_linear_out_weight_to_fp16_quantized")]; + tensor encoder_layers_0_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_0_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17381888)))]; + tensor linear_7_cast_fp16 = linear(bias = encoder_layers_0_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_0_self_attn_linear_out_weight_to_fp16_quantized, x = input_43_cast_fp16)[name = string("linear_7_cast_fp16")]; + tensor input_47_cast_fp16 = add(x = input_37_cast_fp16, y = linear_7_cast_fp16)[name = string("input_47_cast_fp16")]; + tensor x_17_axes_0 = const()[name = string("x_17_axes_0"), val = tensor([-1])]; + tensor encoder_layers_0_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_0_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17384000)))]; + tensor encoder_layers_0_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_0_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17386112)))]; + tensor x_17_cast_fp16 = layer_norm(axes = x_17_axes_0, beta = encoder_layers_0_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_0_norm_conv_weight_to_fp16, x = input_47_cast_fp16)[name = string("x_17_cast_fp16")]; + tensor input_49_perm_0 = const()[name = string("input_49_perm_0"), val = tensor([0, 2, 1])]; + string input_51_pad_type_0 = const()[name = string("input_51_pad_type_0"), val = string("valid")]; + tensor input_51_strides_0 = const()[name = string("input_51_strides_0"), val = tensor([1])]; + tensor input_51_pad_0 = const()[name = string("input_51_pad_0"), val = tensor([0, 0])]; + tensor input_51_dilations_0 = const()[name = string("input_51_dilations_0"), val = tensor([1])]; + int32 input_51_groups_0 = const()[name = string("input_51_groups_0"), val = int32(1)]; + tensor encoder_layers_0_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17388224))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19485440))))[name = string("encoder_layers_0_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_49_cast_fp16 = transpose(perm = input_49_perm_0, x = x_17_cast_fp16)[name = string("transpose_357")]; + tensor input_51_cast_fp16 = conv(dilations = input_51_dilations_0, groups = input_51_groups_0, pad = input_51_pad_0, pad_type = input_51_pad_type_0, strides = input_51_strides_0, weight = encoder_layers_0_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_49_cast_fp16)[name = string("input_51_cast_fp16")]; + int32 x_19_split_num_splits_0 = const()[name = string("x_19_split_num_splits_0"), val = int32(2)]; + int32 x_19_split_axis_0 = const()[name = string("x_19_split_axis_0"), val = int32(1)]; + tensor x_19_split_cast_fp16_0, tensor x_19_split_cast_fp16_1 = split(axis = x_19_split_axis_0, num_splits = x_19_split_num_splits_0, x = input_51_cast_fp16)[name = string("x_19_split_cast_fp16")]; + tensor x_19_split_1_sigmoid_cast_fp16 = sigmoid(x = x_19_split_cast_fp16_1)[name = string("x_19_split_1_sigmoid_cast_fp16")]; + tensor x_19_cast_fp16 = mul(x = x_19_split_cast_fp16_0, y = x_19_split_1_sigmoid_cast_fp16)[name = string("x_19_cast_fp16")]; + tensor var_575_axes_0 = const()[name = string("op_575_axes_0"), val = tensor([1])]; + tensor var_575 = expand_dims(axes = var_575_axes_0, x = pad_mask)[name = string("op_575")]; + tensor input_53_cast_fp16 = select(a = var_44_to_fp16, b = x_19_cast_fp16, cond = var_575)[name = string("input_53_cast_fp16")]; + bool new_x_3_interleave_0 = const()[name = string("new_x_3_interleave_0"), val = bool(false)]; + tensor new_x_3_cast_fp16 = concat(axis = var_59, interleave = new_x_3_interleave_0, values = (cache_3_cast_fp16, input_53_cast_fp16))[name = string("new_x_3_cast_fp16")]; + tensor var_588_begin_0 = const()[name = string("op_588_begin_0"), val = tensor([0, 0, 28])]; + tensor var_588_end_0 = const()[name = string("op_588_end_0"), val = tensor([1, 1024, 36])]; + tensor var_588_end_mask_0 = const()[name = string("op_588_end_mask_0"), val = tensor([true, true, true])]; + tensor var_588_cast_fp16 = slice_by_index(begin = var_588_begin_0, end = var_588_end_0, end_mask = var_588_end_mask_0, x = new_x_3_cast_fp16)[name = string("op_588_cast_fp16")]; + string x_21_pad_type_0 = const()[name = string("x_21_pad_type_0"), val = string("valid")]; + int32 x_21_groups_0 = const()[name = string("x_21_groups_0"), val = int32(1024)]; + tensor x_21_strides_0 = const()[name = string("x_21_strides_0"), val = tensor([1])]; + tensor x_21_pad_0 = const()[name = string("x_21_pad_0"), val = tensor([0, 0])]; + tensor x_21_dilations_0 = const()[name = string("x_21_dilations_0"), val = tensor([1])]; + tensor encoder_layers_0_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19489600))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19498880))))[name = string("encoder_layers_0_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_21_cast_fp16 = conv(dilations = x_21_dilations_0, groups = x_21_groups_0, pad = x_21_pad_0, pad_type = x_21_pad_type_0, strides = x_21_strides_0, weight = encoder_layers_0_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_3_cast_fp16)[name = string("x_21_cast_fp16")]; + tensor input_55_perm_0 = const()[name = string("input_55_perm_0"), val = tensor([0, 2, 1])]; + tensor x_23_axes_0 = const()[name = string("x_23_axes_0"), val = tensor([-1])]; + tensor encoder_layers_0_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_0_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19500992)))]; + tensor encoder_layers_0_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_0_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19503104)))]; + tensor input_55_cast_fp16 = transpose(perm = input_55_perm_0, x = x_21_cast_fp16)[name = string("transpose_356")]; + tensor x_23_cast_fp16 = layer_norm(axes = x_23_axes_0, beta = encoder_layers_0_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_0_conv_batch_norm_weight_to_fp16, x = input_55_cast_fp16)[name = string("x_23_cast_fp16")]; + tensor input_57_perm_0 = const()[name = string("input_57_perm_0"), val = tensor([0, 2, 1])]; + tensor input_57_cast_fp16 = transpose(perm = input_57_perm_0, x = x_23_cast_fp16)[name = string("transpose_355")]; + tensor input_59_cast_fp16 = silu(x = input_57_cast_fp16)[name = string("input_59_cast_fp16")]; + string x_25_pad_type_0 = const()[name = string("x_25_pad_type_0"), val = string("valid")]; + tensor x_25_strides_0 = const()[name = string("x_25_strides_0"), val = tensor([1])]; + tensor x_25_pad_0 = const()[name = string("x_25_pad_0"), val = tensor([0, 0])]; + tensor x_25_dilations_0 = const()[name = string("x_25_dilations_0"), val = tensor([1])]; + int32 x_25_groups_0 = const()[name = string("x_25_groups_0"), val = int32(1)]; + tensor encoder_layers_0_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19505216))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20553856))))[name = string("encoder_layers_0_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_25_cast_fp16 = conv(dilations = x_25_dilations_0, groups = x_25_groups_0, pad = x_25_pad_0, pad_type = x_25_pad_type_0, strides = x_25_strides_0, weight = encoder_layers_0_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_59_cast_fp16)[name = string("x_25_cast_fp16")]; + tensor input_61_perm_0 = const()[name = string("input_61_perm_0"), val = tensor([0, 2, 1])]; + tensor input_61_cast_fp16 = transpose(perm = input_61_perm_0, x = x_25_cast_fp16)[name = string("transpose_354")]; + tensor input_63_cast_fp16 = add(x = input_47_cast_fp16, y = input_61_cast_fp16)[name = string("input_63_cast_fp16")]; + tensor input_65_axes_0 = const()[name = string("input_65_axes_0"), val = tensor([-1])]; + tensor encoder_layers_0_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_0_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20555968)))]; + tensor encoder_layers_0_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_0_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20558080)))]; + tensor input_65_cast_fp16 = layer_norm(axes = input_65_axes_0, beta = encoder_layers_0_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_0_norm_feed_forward2_weight_to_fp16, x = input_63_cast_fp16)[name = string("input_65_cast_fp16")]; + tensor encoder_layers_0_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20560192))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24754560))))[name = string("encoder_layers_0_feed_forward2_linear1_weight_to_fp16_quantized")]; + tensor encoder_layers_0_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_0_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24762816)))]; + tensor linear_8_cast_fp16 = linear(bias = encoder_layers_0_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_0_feed_forward2_linear1_weight_to_fp16_quantized, x = input_65_cast_fp16)[name = string("linear_8_cast_fp16")]; + tensor input_69_cast_fp16 = silu(x = linear_8_cast_fp16)[name = string("input_69_cast_fp16")]; + tensor encoder_layers_0_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24771072))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28965440))))[name = string("encoder_layers_0_feed_forward2_linear2_weight_to_fp16_quantized")]; + tensor encoder_layers_0_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_0_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28967552)))]; + tensor linear_9_cast_fp16 = linear(bias = encoder_layers_0_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_0_feed_forward2_linear2_weight_to_fp16_quantized, x = input_69_cast_fp16)[name = string("linear_9_cast_fp16")]; + fp16 var_631_to_fp16 = const()[name = string("op_631_to_fp16"), val = fp16(0x1p-1)]; + tensor var_632_cast_fp16 = mul(x = linear_9_cast_fp16, y = var_631_to_fp16)[name = string("op_632_cast_fp16")]; + tensor input_75_cast_fp16 = add(x = input_63_cast_fp16, y = var_632_cast_fp16)[name = string("input_75_cast_fp16")]; + tensor input_77_axes_0 = const()[name = string("input_77_axes_0"), val = tensor([-1])]; + tensor encoder_layers_0_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_0_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28969664)))]; + tensor encoder_layers_0_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_0_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28971776)))]; + tensor input_77_cast_fp16 = layer_norm(axes = input_77_axes_0, beta = encoder_layers_0_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_0_norm_out_weight_to_fp16, x = input_75_cast_fp16)[name = string("input_77_cast_fp16")]; + tensor cache_5_begin_0 = const()[name = string("cache_5_begin_0"), val = tensor([1, 0, 0, 0])]; + tensor cache_5_end_0 = const()[name = string("cache_5_end_0"), val = tensor([2, 1, 42, 1024])]; + tensor cache_5_end_mask_0 = const()[name = string("cache_5_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_5_squeeze_mask_0 = const()[name = string("cache_5_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_5_cast_fp16 = slice_by_index(begin = cache_5_begin_0, end = cache_5_end_0, end_mask = cache_5_end_mask_0, squeeze_mask = cache_5_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_5_cast_fp16")]; + tensor cache_7_begin_0 = const()[name = string("cache_7_begin_0"), val = tensor([1, 0, 0, 0])]; + tensor cache_7_end_0 = const()[name = string("cache_7_end_0"), val = tensor([2, 1, 1024, 8])]; + tensor cache_7_end_mask_0 = const()[name = string("cache_7_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_7_squeeze_mask_0 = const()[name = string("cache_7_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_7_cast_fp16 = slice_by_index(begin = cache_7_begin_0, end = cache_7_end_0, end_mask = cache_7_end_mask_0, squeeze_mask = cache_7_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_7_cast_fp16")]; + tensor input_79_axes_0 = const()[name = string("input_79_axes_0"), val = tensor([-1])]; + tensor encoder_layers_1_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_1_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28973888)))]; + tensor encoder_layers_1_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_1_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28976000)))]; + tensor input_79_cast_fp16 = layer_norm(axes = input_79_axes_0, beta = encoder_layers_1_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_1_norm_feed_forward1_weight_to_fp16, x = input_77_cast_fp16)[name = string("input_79_cast_fp16")]; + tensor encoder_layers_1_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28978112))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33172480))))[name = string("encoder_layers_1_feed_forward1_linear1_weight_to_fp16_quantized")]; + tensor encoder_layers_1_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_1_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33180736)))]; + tensor linear_10_cast_fp16 = linear(bias = encoder_layers_1_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_1_feed_forward1_linear1_weight_to_fp16_quantized, x = input_79_cast_fp16)[name = string("linear_10_cast_fp16")]; + tensor input_83_cast_fp16 = silu(x = linear_10_cast_fp16)[name = string("input_83_cast_fp16")]; + tensor encoder_layers_1_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33188992))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37383360))))[name = string("encoder_layers_1_feed_forward1_linear2_weight_to_fp16_quantized")]; + tensor encoder_layers_1_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_1_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37385472)))]; + tensor linear_11_cast_fp16 = linear(bias = encoder_layers_1_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_1_feed_forward1_linear2_weight_to_fp16_quantized, x = input_83_cast_fp16)[name = string("linear_11_cast_fp16")]; + fp16 var_668_to_fp16 = const()[name = string("op_668_to_fp16"), val = fp16(0x1p-1)]; + tensor var_669_cast_fp16 = mul(x = linear_11_cast_fp16, y = var_668_to_fp16)[name = string("op_669_cast_fp16")]; + tensor input_89_cast_fp16 = add(x = input_77_cast_fp16, y = var_669_cast_fp16)[name = string("input_89_cast_fp16")]; + tensor key_3_axes_0 = const()[name = string("key_3_axes_0"), val = tensor([-1])]; + tensor encoder_layers_1_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_1_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37387584)))]; + tensor encoder_layers_1_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_1_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37389696)))]; + tensor key_3_cast_fp16 = layer_norm(axes = key_3_axes_0, beta = encoder_layers_1_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_1_norm_self_att_weight_to_fp16, x = input_89_cast_fp16)[name = string("key_3_cast_fp16")]; + bool input_91_interleave_0 = const()[name = string("input_91_interleave_0"), val = bool(false)]; + tensor input_91_cast_fp16 = concat(axis = var_68, interleave = input_91_interleave_0, values = (cache_5_cast_fp16, key_3_cast_fp16))[name = string("input_91_cast_fp16")]; + tensor var_691_begin_0 = const()[name = string("op_691_begin_0"), val = tensor([0, 28, 0])]; + tensor var_691_end_0 = const()[name = string("op_691_end_0"), val = tensor([1, 42, 1024])]; + tensor var_691_end_mask_0 = const()[name = string("op_691_end_mask_0"), val = tensor([true, true, true])]; + tensor var_691_cast_fp16 = slice_by_index(begin = var_691_begin_0, end = var_691_end_0, end_mask = var_691_end_mask_0, x = cache_5_cast_fp16)[name = string("op_691_cast_fp16")]; + bool var_697_interleave_0 = const()[name = string("op_697_interleave_0"), val = bool(false)]; + tensor var_697_cast_fp16 = concat(axis = var_68, interleave = var_697_interleave_0, values = (var_691_cast_fp16, key_3_cast_fp16))[name = string("op_697_cast_fp16")]; + tensor encoder_layers_1_self_attn_linear_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37391808))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38440448))))[name = string("encoder_layers_1_self_attn_linear_q_weight_to_fp16_quantized")]; + tensor encoder_layers_1_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_1_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38442560)))]; + tensor linear_12_cast_fp16 = linear(bias = encoder_layers_1_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_1_self_attn_linear_q_weight_to_fp16_quantized, x = key_3_cast_fp16)[name = string("linear_12_cast_fp16")]; + tensor var_702 = const()[name = string("op_702"), val = tensor([1, -1, 8, 128])]; + tensor q_7_cast_fp16 = reshape(shape = var_702, x = linear_12_cast_fp16)[name = string("q_7_cast_fp16")]; + tensor encoder_layers_1_self_attn_linear_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38444672))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39493312))))[name = string("encoder_layers_1_self_attn_linear_k_weight_to_fp16_quantized")]; + tensor encoder_layers_1_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_1_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39495424)))]; + tensor linear_13_cast_fp16 = linear(bias = encoder_layers_1_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_1_self_attn_linear_k_weight_to_fp16_quantized, x = input_91_cast_fp16)[name = string("linear_13_cast_fp16")]; + tensor var_707 = const()[name = string("op_707"), val = tensor([1, -1, 8, 128])]; + tensor k_5_cast_fp16 = reshape(shape = var_707, x = linear_13_cast_fp16)[name = string("k_5_cast_fp16")]; + tensor encoder_layers_1_self_attn_linear_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39497536))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40546176))))[name = string("encoder_layers_1_self_attn_linear_v_weight_to_fp16_quantized")]; + tensor encoder_layers_1_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_1_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40548288)))]; + tensor linear_14_cast_fp16 = linear(bias = encoder_layers_1_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_1_self_attn_linear_v_weight_to_fp16_quantized, x = input_91_cast_fp16)[name = string("linear_14_cast_fp16")]; + tensor var_712 = const()[name = string("op_712"), val = tensor([1, -1, 8, 128])]; + tensor v_3_cast_fp16 = reshape(shape = var_712, x = linear_14_cast_fp16)[name = string("v_3_cast_fp16")]; + tensor value_11_perm_0 = const()[name = string("value_11_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_1_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_1_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40550400)))]; + tensor var_725_cast_fp16 = add(x = q_7_cast_fp16, y = encoder_layers_1_self_attn_pos_bias_u_to_fp16)[name = string("op_725_cast_fp16")]; + tensor encoder_layers_1_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_1_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40552512)))]; + tensor var_727_cast_fp16 = add(x = q_7_cast_fp16, y = encoder_layers_1_self_attn_pos_bias_v_to_fp16)[name = string("op_727_cast_fp16")]; + tensor q_with_bias_v_3_perm_0 = const()[name = string("q_with_bias_v_3_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_33_transpose_x_0 = const()[name = string("x_33_transpose_x_0"), val = bool(false)]; + bool x_33_transpose_y_0 = const()[name = string("x_33_transpose_y_0"), val = bool(false)]; + tensor op_729_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40554624))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40697024))))[name = string("op_729_to_fp16_quantized")]; + tensor q_with_bias_v_3_cast_fp16 = transpose(perm = q_with_bias_v_3_perm_0, x = var_727_cast_fp16)[name = string("transpose_353")]; + tensor x_33_cast_fp16 = matmul(transpose_x = x_33_transpose_x_0, transpose_y = x_33_transpose_y_0, x = q_with_bias_v_3_cast_fp16, y = op_729_to_fp16_quantized)[name = string("x_33_cast_fp16")]; + tensor x_35_pad_0 = const()[name = string("x_35_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_35_mode_0 = const()[name = string("x_35_mode_0"), val = string("constant")]; + fp16 const_92_to_fp16 = const()[name = string("const_92_to_fp16"), val = fp16(0x0p+0)]; + tensor x_35_cast_fp16 = pad(constant_val = const_92_to_fp16, mode = x_35_mode_0, pad = x_35_pad_0, x = x_33_cast_fp16)[name = string("x_35_cast_fp16")]; + tensor var_737 = const()[name = string("op_737"), val = tensor([1, 8, -1, 28])]; + tensor x_37_cast_fp16 = reshape(shape = var_737, x = x_35_cast_fp16)[name = string("x_37_cast_fp16")]; + tensor var_741_begin_0 = const()[name = string("op_741_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_741_end_0 = const()[name = string("op_741_end_0"), val = tensor([1, 8, 140, 28])]; + tensor var_741_end_mask_0 = const()[name = string("op_741_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_741_cast_fp16 = slice_by_index(begin = var_741_begin_0, end = var_741_end_0, end_mask = var_741_end_mask_0, x = x_37_cast_fp16)[name = string("op_741_cast_fp16")]; + tensor var_742 = const()[name = string("op_742"), val = tensor([1, 8, 28, 139])]; + tensor matrix_bd_5_cast_fp16 = reshape(shape = var_742, x = var_741_cast_fp16)[name = string("matrix_bd_5_cast_fp16")]; + bool matrix_ac_3_transpose_x_0 = const()[name = string("matrix_ac_3_transpose_x_0"), val = bool(false)]; + bool matrix_ac_3_transpose_y_0 = const()[name = string("matrix_ac_3_transpose_y_0"), val = bool(false)]; + tensor transpose_98_perm_0 = const()[name = string("transpose_98_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_99_perm_0 = const()[name = string("transpose_99_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_99 = transpose(perm = transpose_99_perm_0, x = k_5_cast_fp16)[name = string("transpose_351")]; + tensor transpose_98 = transpose(perm = transpose_98_perm_0, x = var_725_cast_fp16)[name = string("transpose_352")]; + tensor matrix_ac_3_cast_fp16 = matmul(transpose_x = matrix_ac_3_transpose_x_0, transpose_y = matrix_ac_3_transpose_y_0, x = transpose_98, y = transpose_99)[name = string("matrix_ac_3_cast_fp16")]; + tensor matrix_bd_7_begin_0 = const()[name = string("matrix_bd_7_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_7_end_0 = const()[name = string("matrix_bd_7_end_0"), val = tensor([1, 8, 28, 70])]; + tensor matrix_bd_7_end_mask_0 = const()[name = string("matrix_bd_7_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_7_cast_fp16 = slice_by_index(begin = matrix_bd_7_begin_0, end = matrix_bd_7_end_0, end_mask = matrix_bd_7_end_mask_0, x = matrix_bd_5_cast_fp16)[name = string("matrix_bd_7_cast_fp16")]; + tensor var_751_cast_fp16 = add(x = matrix_ac_3_cast_fp16, y = matrix_bd_7_cast_fp16)[name = string("op_751_cast_fp16")]; + fp16 _inversed_scores_5_y_0_to_fp16 = const()[name = string("_inversed_scores_5_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_5_cast_fp16 = mul(x = var_751_cast_fp16, y = _inversed_scores_5_y_0_to_fp16)[name = string("_inversed_scores_5_cast_fp16")]; + tensor scores_7_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_5_cast_fp16, cond = mask_11)[name = string("scores_7_cast_fp16")]; + tensor var_757_cast_fp16 = softmax(axis = var_59, x = scores_7_cast_fp16)[name = string("op_757_cast_fp16")]; + tensor input_93_cast_fp16 = select(a = var_44_to_fp16, b = var_757_cast_fp16, cond = mask_11)[name = string("input_93_cast_fp16")]; + bool x_39_transpose_x_0 = const()[name = string("x_39_transpose_x_0"), val = bool(false)]; + bool x_39_transpose_y_0 = const()[name = string("x_39_transpose_y_0"), val = bool(false)]; + tensor value_11_cast_fp16 = transpose(perm = value_11_perm_0, x = v_3_cast_fp16)[name = string("transpose_350")]; + tensor x_39_cast_fp16 = matmul(transpose_x = x_39_transpose_x_0, transpose_y = x_39_transpose_y_0, x = input_93_cast_fp16, y = value_11_cast_fp16)[name = string("x_39_cast_fp16")]; + tensor var_761_perm_0 = const()[name = string("op_761_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_762 = const()[name = string("op_762"), val = tensor([1, -1, 1024])]; + tensor var_761_cast_fp16 = transpose(perm = var_761_perm_0, x = x_39_cast_fp16)[name = string("transpose_349")]; + tensor input_95_cast_fp16 = reshape(shape = var_762, x = var_761_cast_fp16)[name = string("input_95_cast_fp16")]; + tensor encoder_layers_1_self_attn_linear_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40697408))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41746048))))[name = string("encoder_layers_1_self_attn_linear_out_weight_to_fp16_quantized")]; + tensor encoder_layers_1_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_1_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41748160)))]; + tensor linear_16_cast_fp16 = linear(bias = encoder_layers_1_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_1_self_attn_linear_out_weight_to_fp16_quantized, x = input_95_cast_fp16)[name = string("linear_16_cast_fp16")]; + tensor input_99_cast_fp16 = add(x = input_89_cast_fp16, y = linear_16_cast_fp16)[name = string("input_99_cast_fp16")]; + tensor x_43_axes_0 = const()[name = string("x_43_axes_0"), val = tensor([-1])]; + tensor encoder_layers_1_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_1_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41750272)))]; + tensor encoder_layers_1_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_1_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41752384)))]; + tensor x_43_cast_fp16 = layer_norm(axes = x_43_axes_0, beta = encoder_layers_1_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_1_norm_conv_weight_to_fp16, x = input_99_cast_fp16)[name = string("x_43_cast_fp16")]; + tensor input_101_perm_0 = const()[name = string("input_101_perm_0"), val = tensor([0, 2, 1])]; + string input_103_pad_type_0 = const()[name = string("input_103_pad_type_0"), val = string("valid")]; + tensor input_103_strides_0 = const()[name = string("input_103_strides_0"), val = tensor([1])]; + tensor input_103_pad_0 = const()[name = string("input_103_pad_0"), val = tensor([0, 0])]; + tensor input_103_dilations_0 = const()[name = string("input_103_dilations_0"), val = tensor([1])]; + int32 input_103_groups_0 = const()[name = string("input_103_groups_0"), val = int32(1)]; + tensor encoder_layers_1_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41754496))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43851712))))[name = string("encoder_layers_1_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_101_cast_fp16 = transpose(perm = input_101_perm_0, x = x_43_cast_fp16)[name = string("transpose_348")]; + tensor input_103_cast_fp16 = conv(dilations = input_103_dilations_0, groups = input_103_groups_0, pad = input_103_pad_0, pad_type = input_103_pad_type_0, strides = input_103_strides_0, weight = encoder_layers_1_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_101_cast_fp16)[name = string("input_103_cast_fp16")]; + int32 x_45_split_num_splits_0 = const()[name = string("x_45_split_num_splits_0"), val = int32(2)]; + int32 x_45_split_axis_0 = const()[name = string("x_45_split_axis_0"), val = int32(1)]; + tensor x_45_split_cast_fp16_0, tensor x_45_split_cast_fp16_1 = split(axis = x_45_split_axis_0, num_splits = x_45_split_num_splits_0, x = input_103_cast_fp16)[name = string("x_45_split_cast_fp16")]; + tensor x_45_split_1_sigmoid_cast_fp16 = sigmoid(x = x_45_split_cast_fp16_1)[name = string("x_45_split_1_sigmoid_cast_fp16")]; + tensor x_45_cast_fp16 = mul(x = x_45_split_cast_fp16_0, y = x_45_split_1_sigmoid_cast_fp16)[name = string("x_45_cast_fp16")]; + tensor input_105_cast_fp16 = select(a = var_44_to_fp16, b = x_45_cast_fp16, cond = var_575)[name = string("input_105_cast_fp16")]; + bool new_x_7_interleave_0 = const()[name = string("new_x_7_interleave_0"), val = bool(false)]; + tensor new_x_7_cast_fp16 = concat(axis = var_59, interleave = new_x_7_interleave_0, values = (cache_7_cast_fp16, input_105_cast_fp16))[name = string("new_x_7_cast_fp16")]; + tensor var_801_begin_0 = const()[name = string("op_801_begin_0"), val = tensor([0, 0, 28])]; + tensor var_801_end_0 = const()[name = string("op_801_end_0"), val = tensor([1, 1024, 36])]; + tensor var_801_end_mask_0 = const()[name = string("op_801_end_mask_0"), val = tensor([true, true, true])]; + tensor var_801_cast_fp16 = slice_by_index(begin = var_801_begin_0, end = var_801_end_0, end_mask = var_801_end_mask_0, x = new_x_7_cast_fp16)[name = string("op_801_cast_fp16")]; + string x_47_pad_type_0 = const()[name = string("x_47_pad_type_0"), val = string("valid")]; + int32 x_47_groups_0 = const()[name = string("x_47_groups_0"), val = int32(1024)]; + tensor x_47_strides_0 = const()[name = string("x_47_strides_0"), val = tensor([1])]; + tensor x_47_pad_0 = const()[name = string("x_47_pad_0"), val = tensor([0, 0])]; + tensor x_47_dilations_0 = const()[name = string("x_47_dilations_0"), val = tensor([1])]; + tensor encoder_layers_1_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43855872))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43865152))))[name = string("encoder_layers_1_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_47_cast_fp16 = conv(dilations = x_47_dilations_0, groups = x_47_groups_0, pad = x_47_pad_0, pad_type = x_47_pad_type_0, strides = x_47_strides_0, weight = encoder_layers_1_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_7_cast_fp16)[name = string("x_47_cast_fp16")]; + tensor input_107_perm_0 = const()[name = string("input_107_perm_0"), val = tensor([0, 2, 1])]; + tensor x_49_axes_0 = const()[name = string("x_49_axes_0"), val = tensor([-1])]; + tensor encoder_layers_1_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_1_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43867264)))]; + tensor encoder_layers_1_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_1_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43869376)))]; + tensor input_107_cast_fp16 = transpose(perm = input_107_perm_0, x = x_47_cast_fp16)[name = string("transpose_347")]; + tensor x_49_cast_fp16 = layer_norm(axes = x_49_axes_0, beta = encoder_layers_1_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_1_conv_batch_norm_weight_to_fp16, x = input_107_cast_fp16)[name = string("x_49_cast_fp16")]; + tensor input_109_perm_0 = const()[name = string("input_109_perm_0"), val = tensor([0, 2, 1])]; + tensor input_109_cast_fp16 = transpose(perm = input_109_perm_0, x = x_49_cast_fp16)[name = string("transpose_346")]; + tensor input_111_cast_fp16 = silu(x = input_109_cast_fp16)[name = string("input_111_cast_fp16")]; + string x_51_pad_type_0 = const()[name = string("x_51_pad_type_0"), val = string("valid")]; + tensor x_51_strides_0 = const()[name = string("x_51_strides_0"), val = tensor([1])]; + tensor x_51_pad_0 = const()[name = string("x_51_pad_0"), val = tensor([0, 0])]; + tensor x_51_dilations_0 = const()[name = string("x_51_dilations_0"), val = tensor([1])]; + int32 x_51_groups_0 = const()[name = string("x_51_groups_0"), val = int32(1)]; + tensor encoder_layers_1_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43871488))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44920128))))[name = string("encoder_layers_1_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_51_cast_fp16 = conv(dilations = x_51_dilations_0, groups = x_51_groups_0, pad = x_51_pad_0, pad_type = x_51_pad_type_0, strides = x_51_strides_0, weight = encoder_layers_1_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_111_cast_fp16)[name = string("x_51_cast_fp16")]; + tensor input_113_perm_0 = const()[name = string("input_113_perm_0"), val = tensor([0, 2, 1])]; + tensor input_113_cast_fp16 = transpose(perm = input_113_perm_0, x = x_51_cast_fp16)[name = string("transpose_345")]; + tensor input_115_cast_fp16 = add(x = input_99_cast_fp16, y = input_113_cast_fp16)[name = string("input_115_cast_fp16")]; + tensor input_117_axes_0 = const()[name = string("input_117_axes_0"), val = tensor([-1])]; + tensor encoder_layers_1_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_1_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44922240)))]; + tensor encoder_layers_1_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_1_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44924352)))]; + tensor input_117_cast_fp16 = layer_norm(axes = input_117_axes_0, beta = encoder_layers_1_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_1_norm_feed_forward2_weight_to_fp16, x = input_115_cast_fp16)[name = string("input_117_cast_fp16")]; + tensor encoder_layers_1_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44926464))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49120832))))[name = string("encoder_layers_1_feed_forward2_linear1_weight_to_fp16_quantized")]; + tensor encoder_layers_1_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_1_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49129088)))]; + tensor linear_17_cast_fp16 = linear(bias = encoder_layers_1_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_1_feed_forward2_linear1_weight_to_fp16_quantized, x = input_117_cast_fp16)[name = string("linear_17_cast_fp16")]; + tensor input_121_cast_fp16 = silu(x = linear_17_cast_fp16)[name = string("input_121_cast_fp16")]; + tensor encoder_layers_1_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49137344))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53331712))))[name = string("encoder_layers_1_feed_forward2_linear2_weight_to_fp16_quantized")]; + tensor encoder_layers_1_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_1_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53333824)))]; + tensor linear_18_cast_fp16 = linear(bias = encoder_layers_1_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_1_feed_forward2_linear2_weight_to_fp16_quantized, x = input_121_cast_fp16)[name = string("linear_18_cast_fp16")]; + fp16 var_844_to_fp16 = const()[name = string("op_844_to_fp16"), val = fp16(0x1p-1)]; + tensor var_845_cast_fp16 = mul(x = linear_18_cast_fp16, y = var_844_to_fp16)[name = string("op_845_cast_fp16")]; + tensor input_127_cast_fp16 = add(x = input_115_cast_fp16, y = var_845_cast_fp16)[name = string("input_127_cast_fp16")]; + tensor input_129_axes_0 = const()[name = string("input_129_axes_0"), val = tensor([-1])]; + tensor encoder_layers_1_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_1_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53335936)))]; + tensor encoder_layers_1_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_1_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53338048)))]; + tensor input_129_cast_fp16 = layer_norm(axes = input_129_axes_0, beta = encoder_layers_1_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_1_norm_out_weight_to_fp16, x = input_127_cast_fp16)[name = string("input_129_cast_fp16")]; + tensor cache_9_begin_0 = const()[name = string("cache_9_begin_0"), val = tensor([2, 0, 0, 0])]; + tensor cache_9_end_0 = const()[name = string("cache_9_end_0"), val = tensor([3, 1, 42, 1024])]; + tensor cache_9_end_mask_0 = const()[name = string("cache_9_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_9_squeeze_mask_0 = const()[name = string("cache_9_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_9_cast_fp16 = slice_by_index(begin = cache_9_begin_0, end = cache_9_end_0, end_mask = cache_9_end_mask_0, squeeze_mask = cache_9_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_9_cast_fp16")]; + tensor cache_11_begin_0 = const()[name = string("cache_11_begin_0"), val = tensor([2, 0, 0, 0])]; + tensor cache_11_end_0 = const()[name = string("cache_11_end_0"), val = tensor([3, 1, 1024, 8])]; + tensor cache_11_end_mask_0 = const()[name = string("cache_11_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_11_squeeze_mask_0 = const()[name = string("cache_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_11_cast_fp16 = slice_by_index(begin = cache_11_begin_0, end = cache_11_end_0, end_mask = cache_11_end_mask_0, squeeze_mask = cache_11_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_11_cast_fp16")]; + tensor input_131_axes_0 = const()[name = string("input_131_axes_0"), val = tensor([-1])]; + tensor encoder_layers_2_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_2_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53340160)))]; + tensor encoder_layers_2_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_2_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53342272)))]; + tensor input_131_cast_fp16 = layer_norm(axes = input_131_axes_0, beta = encoder_layers_2_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_2_norm_feed_forward1_weight_to_fp16, x = input_129_cast_fp16)[name = string("input_131_cast_fp16")]; + tensor encoder_layers_2_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53344384))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57538752))))[name = string("encoder_layers_2_feed_forward1_linear1_weight_to_fp16_quantized")]; + tensor encoder_layers_2_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_2_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57547008)))]; + tensor linear_19_cast_fp16 = linear(bias = encoder_layers_2_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_2_feed_forward1_linear1_weight_to_fp16_quantized, x = input_131_cast_fp16)[name = string("linear_19_cast_fp16")]; + tensor input_135_cast_fp16 = silu(x = linear_19_cast_fp16)[name = string("input_135_cast_fp16")]; + tensor encoder_layers_2_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57555264))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61749632))))[name = string("encoder_layers_2_feed_forward1_linear2_weight_to_fp16_quantized")]; + tensor encoder_layers_2_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_2_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61751744)))]; + tensor linear_20_cast_fp16 = linear(bias = encoder_layers_2_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_2_feed_forward1_linear2_weight_to_fp16_quantized, x = input_135_cast_fp16)[name = string("linear_20_cast_fp16")]; + fp16 var_881_to_fp16 = const()[name = string("op_881_to_fp16"), val = fp16(0x1p-1)]; + tensor var_882_cast_fp16 = mul(x = linear_20_cast_fp16, y = var_881_to_fp16)[name = string("op_882_cast_fp16")]; + tensor input_141_cast_fp16 = add(x = input_129_cast_fp16, y = var_882_cast_fp16)[name = string("input_141_cast_fp16")]; + tensor key_5_axes_0 = const()[name = string("key_5_axes_0"), val = tensor([-1])]; + tensor encoder_layers_2_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_2_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61753856)))]; + tensor encoder_layers_2_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_2_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61755968)))]; + tensor key_5_cast_fp16 = layer_norm(axes = key_5_axes_0, beta = encoder_layers_2_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_2_norm_self_att_weight_to_fp16, x = input_141_cast_fp16)[name = string("key_5_cast_fp16")]; + bool input_143_interleave_0 = const()[name = string("input_143_interleave_0"), val = bool(false)]; + tensor input_143_cast_fp16 = concat(axis = var_68, interleave = input_143_interleave_0, values = (cache_9_cast_fp16, key_5_cast_fp16))[name = string("input_143_cast_fp16")]; + tensor var_904_begin_0 = const()[name = string("op_904_begin_0"), val = tensor([0, 28, 0])]; + tensor var_904_end_0 = const()[name = string("op_904_end_0"), val = tensor([1, 42, 1024])]; + tensor var_904_end_mask_0 = const()[name = string("op_904_end_mask_0"), val = tensor([true, true, true])]; + tensor var_904_cast_fp16 = slice_by_index(begin = var_904_begin_0, end = var_904_end_0, end_mask = var_904_end_mask_0, x = cache_9_cast_fp16)[name = string("op_904_cast_fp16")]; + bool var_910_interleave_0 = const()[name = string("op_910_interleave_0"), val = bool(false)]; + tensor var_910_cast_fp16 = concat(axis = var_68, interleave = var_910_interleave_0, values = (var_904_cast_fp16, key_5_cast_fp16))[name = string("op_910_cast_fp16")]; + tensor encoder_layers_2_self_attn_linear_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61758080))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62806720))))[name = string("encoder_layers_2_self_attn_linear_q_weight_to_fp16_quantized")]; + tensor encoder_layers_2_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_2_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62808832)))]; + tensor linear_21_cast_fp16 = linear(bias = encoder_layers_2_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_2_self_attn_linear_q_weight_to_fp16_quantized, x = key_5_cast_fp16)[name = string("linear_21_cast_fp16")]; + tensor var_915 = const()[name = string("op_915"), val = tensor([1, -1, 8, 128])]; + tensor q_13_cast_fp16 = reshape(shape = var_915, x = linear_21_cast_fp16)[name = string("q_13_cast_fp16")]; + tensor encoder_layers_2_self_attn_linear_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62810944))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(63859584))))[name = string("encoder_layers_2_self_attn_linear_k_weight_to_fp16_quantized")]; + tensor encoder_layers_2_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_2_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(63861696)))]; + tensor linear_22_cast_fp16 = linear(bias = encoder_layers_2_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_2_self_attn_linear_k_weight_to_fp16_quantized, x = input_143_cast_fp16)[name = string("linear_22_cast_fp16")]; + tensor var_920 = const()[name = string("op_920"), val = tensor([1, -1, 8, 128])]; + tensor k_9_cast_fp16 = reshape(shape = var_920, x = linear_22_cast_fp16)[name = string("k_9_cast_fp16")]; + tensor encoder_layers_2_self_attn_linear_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(63863808))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64912448))))[name = string("encoder_layers_2_self_attn_linear_v_weight_to_fp16_quantized")]; + tensor encoder_layers_2_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_2_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64914560)))]; + tensor linear_23_cast_fp16 = linear(bias = encoder_layers_2_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_2_self_attn_linear_v_weight_to_fp16_quantized, x = input_143_cast_fp16)[name = string("linear_23_cast_fp16")]; + tensor var_925 = const()[name = string("op_925"), val = tensor([1, -1, 8, 128])]; + tensor v_5_cast_fp16 = reshape(shape = var_925, x = linear_23_cast_fp16)[name = string("v_5_cast_fp16")]; + tensor value_13_perm_0 = const()[name = string("value_13_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_2_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_2_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64916672)))]; + tensor var_938_cast_fp16 = add(x = q_13_cast_fp16, y = encoder_layers_2_self_attn_pos_bias_u_to_fp16)[name = string("op_938_cast_fp16")]; + tensor encoder_layers_2_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_2_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64918784)))]; + tensor var_940_cast_fp16 = add(x = q_13_cast_fp16, y = encoder_layers_2_self_attn_pos_bias_v_to_fp16)[name = string("op_940_cast_fp16")]; + tensor q_with_bias_v_5_perm_0 = const()[name = string("q_with_bias_v_5_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_59_transpose_x_0 = const()[name = string("x_59_transpose_x_0"), val = bool(false)]; + bool x_59_transpose_y_0 = const()[name = string("x_59_transpose_y_0"), val = bool(false)]; + tensor op_942_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64920896))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65063296))))[name = string("op_942_to_fp16_quantized")]; + tensor q_with_bias_v_5_cast_fp16 = transpose(perm = q_with_bias_v_5_perm_0, x = var_940_cast_fp16)[name = string("transpose_344")]; + tensor x_59_cast_fp16 = matmul(transpose_x = x_59_transpose_x_0, transpose_y = x_59_transpose_y_0, x = q_with_bias_v_5_cast_fp16, y = op_942_to_fp16_quantized)[name = string("x_59_cast_fp16")]; + tensor x_61_pad_0 = const()[name = string("x_61_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_61_mode_0 = const()[name = string("x_61_mode_0"), val = string("constant")]; + fp16 const_105_to_fp16 = const()[name = string("const_105_to_fp16"), val = fp16(0x0p+0)]; + tensor x_61_cast_fp16 = pad(constant_val = const_105_to_fp16, mode = x_61_mode_0, pad = x_61_pad_0, x = x_59_cast_fp16)[name = string("x_61_cast_fp16")]; + tensor var_950 = const()[name = string("op_950"), val = tensor([1, 8, -1, 28])]; + tensor x_63_cast_fp16 = reshape(shape = var_950, x = x_61_cast_fp16)[name = string("x_63_cast_fp16")]; + tensor var_954_begin_0 = const()[name = string("op_954_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_954_end_0 = const()[name = string("op_954_end_0"), val = tensor([1, 8, 140, 28])]; + tensor var_954_end_mask_0 = const()[name = string("op_954_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_954_cast_fp16 = slice_by_index(begin = var_954_begin_0, end = var_954_end_0, end_mask = var_954_end_mask_0, x = x_63_cast_fp16)[name = string("op_954_cast_fp16")]; + tensor var_955 = const()[name = string("op_955"), val = tensor([1, 8, 28, 139])]; + tensor matrix_bd_9_cast_fp16 = reshape(shape = var_955, x = var_954_cast_fp16)[name = string("matrix_bd_9_cast_fp16")]; + bool matrix_ac_5_transpose_x_0 = const()[name = string("matrix_ac_5_transpose_x_0"), val = bool(false)]; + bool matrix_ac_5_transpose_y_0 = const()[name = string("matrix_ac_5_transpose_y_0"), val = bool(false)]; + tensor transpose_100_perm_0 = const()[name = string("transpose_100_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_101_perm_0 = const()[name = string("transpose_101_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_101 = transpose(perm = transpose_101_perm_0, x = k_9_cast_fp16)[name = string("transpose_342")]; + tensor transpose_100 = transpose(perm = transpose_100_perm_0, x = var_938_cast_fp16)[name = string("transpose_343")]; + tensor matrix_ac_5_cast_fp16 = matmul(transpose_x = matrix_ac_5_transpose_x_0, transpose_y = matrix_ac_5_transpose_y_0, x = transpose_100, y = transpose_101)[name = string("matrix_ac_5_cast_fp16")]; + tensor matrix_bd_11_begin_0 = const()[name = string("matrix_bd_11_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_11_end_0 = const()[name = string("matrix_bd_11_end_0"), val = tensor([1, 8, 28, 70])]; + tensor matrix_bd_11_end_mask_0 = const()[name = string("matrix_bd_11_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_11_cast_fp16 = slice_by_index(begin = matrix_bd_11_begin_0, end = matrix_bd_11_end_0, end_mask = matrix_bd_11_end_mask_0, x = matrix_bd_9_cast_fp16)[name = string("matrix_bd_11_cast_fp16")]; + tensor var_964_cast_fp16 = add(x = matrix_ac_5_cast_fp16, y = matrix_bd_11_cast_fp16)[name = string("op_964_cast_fp16")]; + fp16 _inversed_scores_9_y_0_to_fp16 = const()[name = string("_inversed_scores_9_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_9_cast_fp16 = mul(x = var_964_cast_fp16, y = _inversed_scores_9_y_0_to_fp16)[name = string("_inversed_scores_9_cast_fp16")]; + tensor scores_11_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_9_cast_fp16, cond = mask_11)[name = string("scores_11_cast_fp16")]; + tensor var_970_cast_fp16 = softmax(axis = var_59, x = scores_11_cast_fp16)[name = string("op_970_cast_fp16")]; + tensor input_145_cast_fp16 = select(a = var_44_to_fp16, b = var_970_cast_fp16, cond = mask_11)[name = string("input_145_cast_fp16")]; + bool x_65_transpose_x_0 = const()[name = string("x_65_transpose_x_0"), val = bool(false)]; + bool x_65_transpose_y_0 = const()[name = string("x_65_transpose_y_0"), val = bool(false)]; + tensor value_13_cast_fp16 = transpose(perm = value_13_perm_0, x = v_5_cast_fp16)[name = string("transpose_341")]; + tensor x_65_cast_fp16 = matmul(transpose_x = x_65_transpose_x_0, transpose_y = x_65_transpose_y_0, x = input_145_cast_fp16, y = value_13_cast_fp16)[name = string("x_65_cast_fp16")]; + tensor var_974_perm_0 = const()[name = string("op_974_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_975 = const()[name = string("op_975"), val = tensor([1, -1, 1024])]; + tensor var_974_cast_fp16 = transpose(perm = var_974_perm_0, x = x_65_cast_fp16)[name = string("transpose_340")]; + tensor input_147_cast_fp16 = reshape(shape = var_975, x = var_974_cast_fp16)[name = string("input_147_cast_fp16")]; + tensor encoder_layers_2_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65063680))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65850176))))[name = string("encoder_layers_2_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_2_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_2_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65850368)))]; + tensor linear_25_cast_fp16 = linear(bias = encoder_layers_2_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_2_self_attn_linear_out_weight_to_fp16_palettized, x = input_147_cast_fp16)[name = string("linear_25_cast_fp16")]; + tensor input_151_cast_fp16 = add(x = input_141_cast_fp16, y = linear_25_cast_fp16)[name = string("input_151_cast_fp16")]; + tensor x_69_axes_0 = const()[name = string("x_69_axes_0"), val = tensor([-1])]; + tensor encoder_layers_2_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_2_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65852480)))]; + tensor encoder_layers_2_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_2_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65854592)))]; + tensor x_69_cast_fp16 = layer_norm(axes = x_69_axes_0, beta = encoder_layers_2_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_2_norm_conv_weight_to_fp16, x = input_151_cast_fp16)[name = string("x_69_cast_fp16")]; + tensor input_153_perm_0 = const()[name = string("input_153_perm_0"), val = tensor([0, 2, 1])]; + string input_155_pad_type_0 = const()[name = string("input_155_pad_type_0"), val = string("valid")]; + tensor input_155_strides_0 = const()[name = string("input_155_strides_0"), val = tensor([1])]; + tensor input_155_pad_0 = const()[name = string("input_155_pad_0"), val = tensor([0, 0])]; + tensor input_155_dilations_0 = const()[name = string("input_155_dilations_0"), val = tensor([1])]; + int32 input_155_groups_0 = const()[name = string("input_155_groups_0"), val = int32(1)]; + tensor encoder_layers_2_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65856704))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67953920))))[name = string("encoder_layers_2_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_153_cast_fp16 = transpose(perm = input_153_perm_0, x = x_69_cast_fp16)[name = string("transpose_339")]; + tensor input_155_cast_fp16 = conv(dilations = input_155_dilations_0, groups = input_155_groups_0, pad = input_155_pad_0, pad_type = input_155_pad_type_0, strides = input_155_strides_0, weight = encoder_layers_2_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_153_cast_fp16)[name = string("input_155_cast_fp16")]; + int32 x_71_split_num_splits_0 = const()[name = string("x_71_split_num_splits_0"), val = int32(2)]; + int32 x_71_split_axis_0 = const()[name = string("x_71_split_axis_0"), val = int32(1)]; + tensor x_71_split_cast_fp16_0, tensor x_71_split_cast_fp16_1 = split(axis = x_71_split_axis_0, num_splits = x_71_split_num_splits_0, x = input_155_cast_fp16)[name = string("x_71_split_cast_fp16")]; + tensor x_71_split_1_sigmoid_cast_fp16 = sigmoid(x = x_71_split_cast_fp16_1)[name = string("x_71_split_1_sigmoid_cast_fp16")]; + tensor x_71_cast_fp16 = mul(x = x_71_split_cast_fp16_0, y = x_71_split_1_sigmoid_cast_fp16)[name = string("x_71_cast_fp16")]; + tensor input_157_cast_fp16 = select(a = var_44_to_fp16, b = x_71_cast_fp16, cond = var_575)[name = string("input_157_cast_fp16")]; + bool new_x_11_interleave_0 = const()[name = string("new_x_11_interleave_0"), val = bool(false)]; + tensor new_x_11_cast_fp16 = concat(axis = var_59, interleave = new_x_11_interleave_0, values = (cache_11_cast_fp16, input_157_cast_fp16))[name = string("new_x_11_cast_fp16")]; + tensor var_1014_begin_0 = const()[name = string("op_1014_begin_0"), val = tensor([0, 0, 28])]; + tensor var_1014_end_0 = const()[name = string("op_1014_end_0"), val = tensor([1, 1024, 36])]; + tensor var_1014_end_mask_0 = const()[name = string("op_1014_end_mask_0"), val = tensor([true, true, true])]; + tensor var_1014_cast_fp16 = slice_by_index(begin = var_1014_begin_0, end = var_1014_end_0, end_mask = var_1014_end_mask_0, x = new_x_11_cast_fp16)[name = string("op_1014_cast_fp16")]; + string x_73_pad_type_0 = const()[name = string("x_73_pad_type_0"), val = string("valid")]; + int32 x_73_groups_0 = const()[name = string("x_73_groups_0"), val = int32(1024)]; + tensor x_73_strides_0 = const()[name = string("x_73_strides_0"), val = tensor([1])]; + tensor x_73_pad_0 = const()[name = string("x_73_pad_0"), val = tensor([0, 0])]; + tensor x_73_dilations_0 = const()[name = string("x_73_dilations_0"), val = tensor([1])]; + tensor encoder_layers_2_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67958080))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67967360))))[name = string("encoder_layers_2_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_73_cast_fp16 = conv(dilations = x_73_dilations_0, groups = x_73_groups_0, pad = x_73_pad_0, pad_type = x_73_pad_type_0, strides = x_73_strides_0, weight = encoder_layers_2_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_11_cast_fp16)[name = string("x_73_cast_fp16")]; + tensor input_159_perm_0 = const()[name = string("input_159_perm_0"), val = tensor([0, 2, 1])]; + tensor x_75_axes_0 = const()[name = string("x_75_axes_0"), val = tensor([-1])]; + tensor encoder_layers_2_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_2_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67969472)))]; + tensor encoder_layers_2_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_2_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67971584)))]; + tensor input_159_cast_fp16 = transpose(perm = input_159_perm_0, x = x_73_cast_fp16)[name = string("transpose_338")]; + tensor x_75_cast_fp16 = layer_norm(axes = x_75_axes_0, beta = encoder_layers_2_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_2_conv_batch_norm_weight_to_fp16, x = input_159_cast_fp16)[name = string("x_75_cast_fp16")]; + tensor input_161_perm_0 = const()[name = string("input_161_perm_0"), val = tensor([0, 2, 1])]; + tensor input_161_cast_fp16 = transpose(perm = input_161_perm_0, x = x_75_cast_fp16)[name = string("transpose_337")]; + tensor input_163_cast_fp16 = silu(x = input_161_cast_fp16)[name = string("input_163_cast_fp16")]; + string x_77_pad_type_0 = const()[name = string("x_77_pad_type_0"), val = string("valid")]; + tensor x_77_strides_0 = const()[name = string("x_77_strides_0"), val = tensor([1])]; + tensor x_77_pad_0 = const()[name = string("x_77_pad_0"), val = tensor([0, 0])]; + tensor x_77_dilations_0 = const()[name = string("x_77_dilations_0"), val = tensor([1])]; + int32 x_77_groups_0 = const()[name = string("x_77_groups_0"), val = int32(1)]; + tensor encoder_layers_2_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67973696))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69022336))))[name = string("encoder_layers_2_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_77_cast_fp16 = conv(dilations = x_77_dilations_0, groups = x_77_groups_0, pad = x_77_pad_0, pad_type = x_77_pad_type_0, strides = x_77_strides_0, weight = encoder_layers_2_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_163_cast_fp16)[name = string("x_77_cast_fp16")]; + tensor input_165_perm_0 = const()[name = string("input_165_perm_0"), val = tensor([0, 2, 1])]; + tensor input_165_cast_fp16 = transpose(perm = input_165_perm_0, x = x_77_cast_fp16)[name = string("transpose_336")]; + tensor input_167_cast_fp16 = add(x = input_151_cast_fp16, y = input_165_cast_fp16)[name = string("input_167_cast_fp16")]; + tensor input_169_axes_0 = const()[name = string("input_169_axes_0"), val = tensor([-1])]; + tensor encoder_layers_2_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_2_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69024448)))]; + tensor encoder_layers_2_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_2_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69026560)))]; + tensor input_169_cast_fp16 = layer_norm(axes = input_169_axes_0, beta = encoder_layers_2_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_2_norm_feed_forward2_weight_to_fp16, x = input_167_cast_fp16)[name = string("input_169_cast_fp16")]; + tensor encoder_layers_2_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69028672))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72174464))))[name = string("encoder_layers_2_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_2_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_2_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72174656)))]; + tensor linear_26_cast_fp16 = linear(bias = encoder_layers_2_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_2_feed_forward2_linear1_weight_to_fp16_palettized, x = input_169_cast_fp16)[name = string("linear_26_cast_fp16")]; + tensor input_173_cast_fp16 = silu(x = linear_26_cast_fp16)[name = string("input_173_cast_fp16")]; + tensor encoder_layers_2_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72182912))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75328704))))[name = string("encoder_layers_2_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_2_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_2_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75328896)))]; + tensor linear_27_cast_fp16 = linear(bias = encoder_layers_2_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_2_feed_forward2_linear2_weight_to_fp16_palettized, x = input_173_cast_fp16)[name = string("linear_27_cast_fp16")]; + fp16 var_1057_to_fp16 = const()[name = string("op_1057_to_fp16"), val = fp16(0x1p-1)]; + tensor var_1058_cast_fp16 = mul(x = linear_27_cast_fp16, y = var_1057_to_fp16)[name = string("op_1058_cast_fp16")]; + tensor input_179_cast_fp16 = add(x = input_167_cast_fp16, y = var_1058_cast_fp16)[name = string("input_179_cast_fp16")]; + tensor input_181_axes_0 = const()[name = string("input_181_axes_0"), val = tensor([-1])]; + tensor encoder_layers_2_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_2_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75331008)))]; + tensor encoder_layers_2_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_2_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75333120)))]; + tensor input_181_cast_fp16 = layer_norm(axes = input_181_axes_0, beta = encoder_layers_2_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_2_norm_out_weight_to_fp16, x = input_179_cast_fp16)[name = string("input_181_cast_fp16")]; + tensor cache_13_begin_0 = const()[name = string("cache_13_begin_0"), val = tensor([3, 0, 0, 0])]; + tensor cache_13_end_0 = const()[name = string("cache_13_end_0"), val = tensor([4, 1, 42, 1024])]; + tensor cache_13_end_mask_0 = const()[name = string("cache_13_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_13_squeeze_mask_0 = const()[name = string("cache_13_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_13_cast_fp16 = slice_by_index(begin = cache_13_begin_0, end = cache_13_end_0, end_mask = cache_13_end_mask_0, squeeze_mask = cache_13_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_13_cast_fp16")]; + tensor cache_15_begin_0 = const()[name = string("cache_15_begin_0"), val = tensor([3, 0, 0, 0])]; + tensor cache_15_end_0 = const()[name = string("cache_15_end_0"), val = tensor([4, 1, 1024, 8])]; + tensor cache_15_end_mask_0 = const()[name = string("cache_15_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_15_squeeze_mask_0 = const()[name = string("cache_15_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_15_cast_fp16 = slice_by_index(begin = cache_15_begin_0, end = cache_15_end_0, end_mask = cache_15_end_mask_0, squeeze_mask = cache_15_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_15_cast_fp16")]; + tensor input_183_axes_0 = const()[name = string("input_183_axes_0"), val = tensor([-1])]; + tensor encoder_layers_3_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_3_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75335232)))]; + tensor encoder_layers_3_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_3_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75337344)))]; + tensor input_183_cast_fp16 = layer_norm(axes = input_183_axes_0, beta = encoder_layers_3_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_3_norm_feed_forward1_weight_to_fp16, x = input_181_cast_fp16)[name = string("input_183_cast_fp16")]; + tensor encoder_layers_3_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75339456))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78485248))))[name = string("encoder_layers_3_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_3_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_3_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78485440)))]; + tensor linear_28_cast_fp16 = linear(bias = encoder_layers_3_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_3_feed_forward1_linear1_weight_to_fp16_palettized, x = input_183_cast_fp16)[name = string("linear_28_cast_fp16")]; + tensor input_187_cast_fp16 = silu(x = linear_28_cast_fp16)[name = string("input_187_cast_fp16")]; + tensor encoder_layers_3_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78493696))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81639488))))[name = string("encoder_layers_3_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_3_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_3_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81639680)))]; + tensor linear_29_cast_fp16 = linear(bias = encoder_layers_3_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_3_feed_forward1_linear2_weight_to_fp16_palettized, x = input_187_cast_fp16)[name = string("linear_29_cast_fp16")]; + fp16 var_1094_to_fp16 = const()[name = string("op_1094_to_fp16"), val = fp16(0x1p-1)]; + tensor var_1095_cast_fp16 = mul(x = linear_29_cast_fp16, y = var_1094_to_fp16)[name = string("op_1095_cast_fp16")]; + tensor input_193_cast_fp16 = add(x = input_181_cast_fp16, y = var_1095_cast_fp16)[name = string("input_193_cast_fp16")]; + tensor key_7_axes_0 = const()[name = string("key_7_axes_0"), val = tensor([-1])]; + tensor encoder_layers_3_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_3_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81641792)))]; + tensor encoder_layers_3_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_3_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81643904)))]; + tensor key_7_cast_fp16 = layer_norm(axes = key_7_axes_0, beta = encoder_layers_3_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_3_norm_self_att_weight_to_fp16, x = input_193_cast_fp16)[name = string("key_7_cast_fp16")]; + bool input_195_interleave_0 = const()[name = string("input_195_interleave_0"), val = bool(false)]; + tensor input_195_cast_fp16 = concat(axis = var_68, interleave = input_195_interleave_0, values = (cache_13_cast_fp16, key_7_cast_fp16))[name = string("input_195_cast_fp16")]; + tensor var_1117_begin_0 = const()[name = string("op_1117_begin_0"), val = tensor([0, 28, 0])]; + tensor var_1117_end_0 = const()[name = string("op_1117_end_0"), val = tensor([1, 42, 1024])]; + tensor var_1117_end_mask_0 = const()[name = string("op_1117_end_mask_0"), val = tensor([true, true, true])]; + tensor var_1117_cast_fp16 = slice_by_index(begin = var_1117_begin_0, end = var_1117_end_0, end_mask = var_1117_end_mask_0, x = cache_13_cast_fp16)[name = string("op_1117_cast_fp16")]; + bool var_1123_interleave_0 = const()[name = string("op_1123_interleave_0"), val = bool(false)]; + tensor var_1123_cast_fp16 = concat(axis = var_68, interleave = var_1123_interleave_0, values = (var_1117_cast_fp16, key_7_cast_fp16))[name = string("op_1123_cast_fp16")]; + tensor encoder_layers_3_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81646016))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82432512))))[name = string("encoder_layers_3_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_3_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_3_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82432704)))]; + tensor linear_30_cast_fp16 = linear(bias = encoder_layers_3_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_3_self_attn_linear_q_weight_to_fp16_palettized, x = key_7_cast_fp16)[name = string("linear_30_cast_fp16")]; + tensor var_1128 = const()[name = string("op_1128"), val = tensor([1, -1, 8, 128])]; + tensor q_19_cast_fp16 = reshape(shape = var_1128, x = linear_30_cast_fp16)[name = string("q_19_cast_fp16")]; + tensor encoder_layers_3_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82434816))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83221312))))[name = string("encoder_layers_3_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_3_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_3_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83221504)))]; + tensor linear_31_cast_fp16 = linear(bias = encoder_layers_3_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_3_self_attn_linear_k_weight_to_fp16_palettized, x = input_195_cast_fp16)[name = string("linear_31_cast_fp16")]; + tensor var_1133 = const()[name = string("op_1133"), val = tensor([1, -1, 8, 128])]; + tensor k_13_cast_fp16 = reshape(shape = var_1133, x = linear_31_cast_fp16)[name = string("k_13_cast_fp16")]; + tensor encoder_layers_3_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83223616))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84010112))))[name = string("encoder_layers_3_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_3_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_3_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84010304)))]; + tensor linear_32_cast_fp16 = linear(bias = encoder_layers_3_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_3_self_attn_linear_v_weight_to_fp16_palettized, x = input_195_cast_fp16)[name = string("linear_32_cast_fp16")]; + tensor var_1138 = const()[name = string("op_1138"), val = tensor([1, -1, 8, 128])]; + tensor v_7_cast_fp16 = reshape(shape = var_1138, x = linear_32_cast_fp16)[name = string("v_7_cast_fp16")]; + tensor value_15_perm_0 = const()[name = string("value_15_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_3_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_3_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84012416)))]; + tensor var_1151_cast_fp16 = add(x = q_19_cast_fp16, y = encoder_layers_3_self_attn_pos_bias_u_to_fp16)[name = string("op_1151_cast_fp16")]; + tensor encoder_layers_3_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_3_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84014528)))]; + tensor var_1153_cast_fp16 = add(x = q_19_cast_fp16, y = encoder_layers_3_self_attn_pos_bias_v_to_fp16)[name = string("op_1153_cast_fp16")]; + tensor q_with_bias_v_7_perm_0 = const()[name = string("q_with_bias_v_7_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_85_transpose_x_0 = const()[name = string("x_85_transpose_x_0"), val = bool(false)]; + bool x_85_transpose_y_0 = const()[name = string("x_85_transpose_y_0"), val = bool(false)]; + tensor op_1155_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84016640))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84159040))))[name = string("op_1155_to_fp16_quantized")]; + tensor q_with_bias_v_7_cast_fp16 = transpose(perm = q_with_bias_v_7_perm_0, x = var_1153_cast_fp16)[name = string("transpose_335")]; + tensor x_85_cast_fp16 = matmul(transpose_x = x_85_transpose_x_0, transpose_y = x_85_transpose_y_0, x = q_with_bias_v_7_cast_fp16, y = op_1155_to_fp16_quantized)[name = string("x_85_cast_fp16")]; + tensor x_87_pad_0 = const()[name = string("x_87_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_87_mode_0 = const()[name = string("x_87_mode_0"), val = string("constant")]; + fp16 const_118_to_fp16 = const()[name = string("const_118_to_fp16"), val = fp16(0x0p+0)]; + tensor x_87_cast_fp16 = pad(constant_val = const_118_to_fp16, mode = x_87_mode_0, pad = x_87_pad_0, x = x_85_cast_fp16)[name = string("x_87_cast_fp16")]; + tensor var_1163 = const()[name = string("op_1163"), val = tensor([1, 8, -1, 28])]; + tensor x_89_cast_fp16 = reshape(shape = var_1163, x = x_87_cast_fp16)[name = string("x_89_cast_fp16")]; + tensor var_1167_begin_0 = const()[name = string("op_1167_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_1167_end_0 = const()[name = string("op_1167_end_0"), val = tensor([1, 8, 140, 28])]; + tensor var_1167_end_mask_0 = const()[name = string("op_1167_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_1167_cast_fp16 = slice_by_index(begin = var_1167_begin_0, end = var_1167_end_0, end_mask = var_1167_end_mask_0, x = x_89_cast_fp16)[name = string("op_1167_cast_fp16")]; + tensor var_1168 = const()[name = string("op_1168"), val = tensor([1, 8, 28, 139])]; + tensor matrix_bd_13_cast_fp16 = reshape(shape = var_1168, x = var_1167_cast_fp16)[name = string("matrix_bd_13_cast_fp16")]; + bool matrix_ac_7_transpose_x_0 = const()[name = string("matrix_ac_7_transpose_x_0"), val = bool(false)]; + bool matrix_ac_7_transpose_y_0 = const()[name = string("matrix_ac_7_transpose_y_0"), val = bool(false)]; + tensor transpose_102_perm_0 = const()[name = string("transpose_102_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_103_perm_0 = const()[name = string("transpose_103_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_103 = transpose(perm = transpose_103_perm_0, x = k_13_cast_fp16)[name = string("transpose_333")]; + tensor transpose_102 = transpose(perm = transpose_102_perm_0, x = var_1151_cast_fp16)[name = string("transpose_334")]; + tensor matrix_ac_7_cast_fp16 = matmul(transpose_x = matrix_ac_7_transpose_x_0, transpose_y = matrix_ac_7_transpose_y_0, x = transpose_102, y = transpose_103)[name = string("matrix_ac_7_cast_fp16")]; + tensor matrix_bd_15_begin_0 = const()[name = string("matrix_bd_15_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_15_end_0 = const()[name = string("matrix_bd_15_end_0"), val = tensor([1, 8, 28, 70])]; + tensor matrix_bd_15_end_mask_0 = const()[name = string("matrix_bd_15_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_15_cast_fp16 = slice_by_index(begin = matrix_bd_15_begin_0, end = matrix_bd_15_end_0, end_mask = matrix_bd_15_end_mask_0, x = matrix_bd_13_cast_fp16)[name = string("matrix_bd_15_cast_fp16")]; + tensor var_1177_cast_fp16 = add(x = matrix_ac_7_cast_fp16, y = matrix_bd_15_cast_fp16)[name = string("op_1177_cast_fp16")]; + fp16 _inversed_scores_13_y_0_to_fp16 = const()[name = string("_inversed_scores_13_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_13_cast_fp16 = mul(x = var_1177_cast_fp16, y = _inversed_scores_13_y_0_to_fp16)[name = string("_inversed_scores_13_cast_fp16")]; + tensor scores_15_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_13_cast_fp16, cond = mask_11)[name = string("scores_15_cast_fp16")]; + tensor var_1183_cast_fp16 = softmax(axis = var_59, x = scores_15_cast_fp16)[name = string("op_1183_cast_fp16")]; + tensor input_197_cast_fp16 = select(a = var_44_to_fp16, b = var_1183_cast_fp16, cond = mask_11)[name = string("input_197_cast_fp16")]; + bool x_91_transpose_x_0 = const()[name = string("x_91_transpose_x_0"), val = bool(false)]; + bool x_91_transpose_y_0 = const()[name = string("x_91_transpose_y_0"), val = bool(false)]; + tensor value_15_cast_fp16 = transpose(perm = value_15_perm_0, x = v_7_cast_fp16)[name = string("transpose_332")]; + tensor x_91_cast_fp16 = matmul(transpose_x = x_91_transpose_x_0, transpose_y = x_91_transpose_y_0, x = input_197_cast_fp16, y = value_15_cast_fp16)[name = string("x_91_cast_fp16")]; + tensor var_1187_perm_0 = const()[name = string("op_1187_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1188 = const()[name = string("op_1188"), val = tensor([1, -1, 1024])]; + tensor var_1187_cast_fp16 = transpose(perm = var_1187_perm_0, x = x_91_cast_fp16)[name = string("transpose_331")]; + tensor input_199_cast_fp16 = reshape(shape = var_1188, x = var_1187_cast_fp16)[name = string("input_199_cast_fp16")]; + tensor encoder_layers_3_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84159424))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84945920))))[name = string("encoder_layers_3_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_3_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_3_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84946112)))]; + tensor linear_34_cast_fp16 = linear(bias = encoder_layers_3_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_3_self_attn_linear_out_weight_to_fp16_palettized, x = input_199_cast_fp16)[name = string("linear_34_cast_fp16")]; + tensor input_203_cast_fp16 = add(x = input_193_cast_fp16, y = linear_34_cast_fp16)[name = string("input_203_cast_fp16")]; + tensor x_95_axes_0 = const()[name = string("x_95_axes_0"), val = tensor([-1])]; + tensor encoder_layers_3_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_3_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84948224)))]; + tensor encoder_layers_3_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_3_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84950336)))]; + tensor x_95_cast_fp16 = layer_norm(axes = x_95_axes_0, beta = encoder_layers_3_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_3_norm_conv_weight_to_fp16, x = input_203_cast_fp16)[name = string("x_95_cast_fp16")]; + tensor input_205_perm_0 = const()[name = string("input_205_perm_0"), val = tensor([0, 2, 1])]; + string input_207_pad_type_0 = const()[name = string("input_207_pad_type_0"), val = string("valid")]; + tensor input_207_strides_0 = const()[name = string("input_207_strides_0"), val = tensor([1])]; + tensor input_207_pad_0 = const()[name = string("input_207_pad_0"), val = tensor([0, 0])]; + tensor input_207_dilations_0 = const()[name = string("input_207_dilations_0"), val = tensor([1])]; + int32 input_207_groups_0 = const()[name = string("input_207_groups_0"), val = int32(1)]; + tensor encoder_layers_3_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84952448))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87049664))))[name = string("encoder_layers_3_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_205_cast_fp16 = transpose(perm = input_205_perm_0, x = x_95_cast_fp16)[name = string("transpose_330")]; + tensor input_207_cast_fp16 = conv(dilations = input_207_dilations_0, groups = input_207_groups_0, pad = input_207_pad_0, pad_type = input_207_pad_type_0, strides = input_207_strides_0, weight = encoder_layers_3_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_205_cast_fp16)[name = string("input_207_cast_fp16")]; + int32 x_97_split_num_splits_0 = const()[name = string("x_97_split_num_splits_0"), val = int32(2)]; + int32 x_97_split_axis_0 = const()[name = string("x_97_split_axis_0"), val = int32(1)]; + tensor x_97_split_cast_fp16_0, tensor x_97_split_cast_fp16_1 = split(axis = x_97_split_axis_0, num_splits = x_97_split_num_splits_0, x = input_207_cast_fp16)[name = string("x_97_split_cast_fp16")]; + tensor x_97_split_1_sigmoid_cast_fp16 = sigmoid(x = x_97_split_cast_fp16_1)[name = string("x_97_split_1_sigmoid_cast_fp16")]; + tensor x_97_cast_fp16 = mul(x = x_97_split_cast_fp16_0, y = x_97_split_1_sigmoid_cast_fp16)[name = string("x_97_cast_fp16")]; + tensor input_209_cast_fp16 = select(a = var_44_to_fp16, b = x_97_cast_fp16, cond = var_575)[name = string("input_209_cast_fp16")]; + bool new_x_15_interleave_0 = const()[name = string("new_x_15_interleave_0"), val = bool(false)]; + tensor new_x_15_cast_fp16 = concat(axis = var_59, interleave = new_x_15_interleave_0, values = (cache_15_cast_fp16, input_209_cast_fp16))[name = string("new_x_15_cast_fp16")]; + tensor var_1227_begin_0 = const()[name = string("op_1227_begin_0"), val = tensor([0, 0, 28])]; + tensor var_1227_end_0 = const()[name = string("op_1227_end_0"), val = tensor([1, 1024, 36])]; + tensor var_1227_end_mask_0 = const()[name = string("op_1227_end_mask_0"), val = tensor([true, true, true])]; + tensor var_1227_cast_fp16 = slice_by_index(begin = var_1227_begin_0, end = var_1227_end_0, end_mask = var_1227_end_mask_0, x = new_x_15_cast_fp16)[name = string("op_1227_cast_fp16")]; + string x_99_pad_type_0 = const()[name = string("x_99_pad_type_0"), val = string("valid")]; + int32 x_99_groups_0 = const()[name = string("x_99_groups_0"), val = int32(1024)]; + tensor x_99_strides_0 = const()[name = string("x_99_strides_0"), val = tensor([1])]; + tensor x_99_pad_0 = const()[name = string("x_99_pad_0"), val = tensor([0, 0])]; + tensor x_99_dilations_0 = const()[name = string("x_99_dilations_0"), val = tensor([1])]; + tensor encoder_layers_3_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87053824))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87063104))))[name = string("encoder_layers_3_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_99_cast_fp16 = conv(dilations = x_99_dilations_0, groups = x_99_groups_0, pad = x_99_pad_0, pad_type = x_99_pad_type_0, strides = x_99_strides_0, weight = encoder_layers_3_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_15_cast_fp16)[name = string("x_99_cast_fp16")]; + tensor input_211_perm_0 = const()[name = string("input_211_perm_0"), val = tensor([0, 2, 1])]; + tensor x_101_axes_0 = const()[name = string("x_101_axes_0"), val = tensor([-1])]; + tensor encoder_layers_3_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_3_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87065216)))]; + tensor encoder_layers_3_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_3_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87067328)))]; + tensor input_211_cast_fp16 = transpose(perm = input_211_perm_0, x = x_99_cast_fp16)[name = string("transpose_329")]; + tensor x_101_cast_fp16 = layer_norm(axes = x_101_axes_0, beta = encoder_layers_3_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_3_conv_batch_norm_weight_to_fp16, x = input_211_cast_fp16)[name = string("x_101_cast_fp16")]; + tensor input_213_perm_0 = const()[name = string("input_213_perm_0"), val = tensor([0, 2, 1])]; + tensor input_213_cast_fp16 = transpose(perm = input_213_perm_0, x = x_101_cast_fp16)[name = string("transpose_328")]; + tensor input_215_cast_fp16 = silu(x = input_213_cast_fp16)[name = string("input_215_cast_fp16")]; + string x_103_pad_type_0 = const()[name = string("x_103_pad_type_0"), val = string("valid")]; + tensor x_103_strides_0 = const()[name = string("x_103_strides_0"), val = tensor([1])]; + tensor x_103_pad_0 = const()[name = string("x_103_pad_0"), val = tensor([0, 0])]; + tensor x_103_dilations_0 = const()[name = string("x_103_dilations_0"), val = tensor([1])]; + int32 x_103_groups_0 = const()[name = string("x_103_groups_0"), val = int32(1)]; + tensor encoder_layers_3_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87069440))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88118080))))[name = string("encoder_layers_3_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_103_cast_fp16 = conv(dilations = x_103_dilations_0, groups = x_103_groups_0, pad = x_103_pad_0, pad_type = x_103_pad_type_0, strides = x_103_strides_0, weight = encoder_layers_3_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_215_cast_fp16)[name = string("x_103_cast_fp16")]; + tensor input_217_perm_0 = const()[name = string("input_217_perm_0"), val = tensor([0, 2, 1])]; + tensor input_217_cast_fp16 = transpose(perm = input_217_perm_0, x = x_103_cast_fp16)[name = string("transpose_327")]; + tensor input_219_cast_fp16 = add(x = input_203_cast_fp16, y = input_217_cast_fp16)[name = string("input_219_cast_fp16")]; + tensor input_221_axes_0 = const()[name = string("input_221_axes_0"), val = tensor([-1])]; + tensor encoder_layers_3_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_3_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88120192)))]; + tensor encoder_layers_3_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_3_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88122304)))]; + tensor input_221_cast_fp16 = layer_norm(axes = input_221_axes_0, beta = encoder_layers_3_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_3_norm_feed_forward2_weight_to_fp16, x = input_219_cast_fp16)[name = string("input_221_cast_fp16")]; + tensor encoder_layers_3_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88124416))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(91270208))))[name = string("encoder_layers_3_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_3_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_3_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(91270400)))]; + tensor linear_35_cast_fp16 = linear(bias = encoder_layers_3_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_3_feed_forward2_linear1_weight_to_fp16_palettized, x = input_221_cast_fp16)[name = string("linear_35_cast_fp16")]; + tensor input_225_cast_fp16 = silu(x = linear_35_cast_fp16)[name = string("input_225_cast_fp16")]; + tensor encoder_layers_3_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(91278656))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94424448))))[name = string("encoder_layers_3_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_3_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_3_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94424640)))]; + tensor linear_36_cast_fp16 = linear(bias = encoder_layers_3_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_3_feed_forward2_linear2_weight_to_fp16_palettized, x = input_225_cast_fp16)[name = string("linear_36_cast_fp16")]; + fp16 var_1270_to_fp16 = const()[name = string("op_1270_to_fp16"), val = fp16(0x1p-1)]; + tensor var_1271_cast_fp16 = mul(x = linear_36_cast_fp16, y = var_1270_to_fp16)[name = string("op_1271_cast_fp16")]; + tensor input_231_cast_fp16 = add(x = input_219_cast_fp16, y = var_1271_cast_fp16)[name = string("input_231_cast_fp16")]; + tensor input_233_axes_0 = const()[name = string("input_233_axes_0"), val = tensor([-1])]; + tensor encoder_layers_3_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_3_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94426752)))]; + tensor encoder_layers_3_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_3_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94428864)))]; + tensor input_233_cast_fp16 = layer_norm(axes = input_233_axes_0, beta = encoder_layers_3_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_3_norm_out_weight_to_fp16, x = input_231_cast_fp16)[name = string("input_233_cast_fp16")]; + tensor cache_17_begin_0 = const()[name = string("cache_17_begin_0"), val = tensor([4, 0, 0, 0])]; + tensor cache_17_end_0 = const()[name = string("cache_17_end_0"), val = tensor([5, 1, 42, 1024])]; + tensor cache_17_end_mask_0 = const()[name = string("cache_17_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_17_squeeze_mask_0 = const()[name = string("cache_17_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_17_cast_fp16 = slice_by_index(begin = cache_17_begin_0, end = cache_17_end_0, end_mask = cache_17_end_mask_0, squeeze_mask = cache_17_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_17_cast_fp16")]; + tensor cache_19_begin_0 = const()[name = string("cache_19_begin_0"), val = tensor([4, 0, 0, 0])]; + tensor cache_19_end_0 = const()[name = string("cache_19_end_0"), val = tensor([5, 1, 1024, 8])]; + tensor cache_19_end_mask_0 = const()[name = string("cache_19_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_19_squeeze_mask_0 = const()[name = string("cache_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_19_cast_fp16 = slice_by_index(begin = cache_19_begin_0, end = cache_19_end_0, end_mask = cache_19_end_mask_0, squeeze_mask = cache_19_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_19_cast_fp16")]; + tensor input_235_axes_0 = const()[name = string("input_235_axes_0"), val = tensor([-1])]; + tensor encoder_layers_4_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_4_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94430976)))]; + tensor encoder_layers_4_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_4_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94433088)))]; + tensor input_235_cast_fp16 = layer_norm(axes = input_235_axes_0, beta = encoder_layers_4_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_4_norm_feed_forward1_weight_to_fp16, x = input_233_cast_fp16)[name = string("input_235_cast_fp16")]; + tensor encoder_layers_4_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94435200))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97580992))))[name = string("encoder_layers_4_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_4_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_4_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97581184)))]; + tensor linear_37_cast_fp16 = linear(bias = encoder_layers_4_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_4_feed_forward1_linear1_weight_to_fp16_palettized, x = input_235_cast_fp16)[name = string("linear_37_cast_fp16")]; + tensor input_239_cast_fp16 = silu(x = linear_37_cast_fp16)[name = string("input_239_cast_fp16")]; + tensor encoder_layers_4_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97589440))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100735232))))[name = string("encoder_layers_4_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_4_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_4_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100735424)))]; + tensor linear_38_cast_fp16 = linear(bias = encoder_layers_4_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_4_feed_forward1_linear2_weight_to_fp16_palettized, x = input_239_cast_fp16)[name = string("linear_38_cast_fp16")]; + fp16 var_1307_to_fp16 = const()[name = string("op_1307_to_fp16"), val = fp16(0x1p-1)]; + tensor var_1308_cast_fp16 = mul(x = linear_38_cast_fp16, y = var_1307_to_fp16)[name = string("op_1308_cast_fp16")]; + tensor input_245_cast_fp16 = add(x = input_233_cast_fp16, y = var_1308_cast_fp16)[name = string("input_245_cast_fp16")]; + tensor key_9_axes_0 = const()[name = string("key_9_axes_0"), val = tensor([-1])]; + tensor encoder_layers_4_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_4_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100737536)))]; + tensor encoder_layers_4_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_4_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100739648)))]; + tensor key_9_cast_fp16 = layer_norm(axes = key_9_axes_0, beta = encoder_layers_4_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_4_norm_self_att_weight_to_fp16, x = input_245_cast_fp16)[name = string("key_9_cast_fp16")]; + bool input_247_interleave_0 = const()[name = string("input_247_interleave_0"), val = bool(false)]; + tensor input_247_cast_fp16 = concat(axis = var_68, interleave = input_247_interleave_0, values = (cache_17_cast_fp16, key_9_cast_fp16))[name = string("input_247_cast_fp16")]; + tensor var_1330_begin_0 = const()[name = string("op_1330_begin_0"), val = tensor([0, 28, 0])]; + tensor var_1330_end_0 = const()[name = string("op_1330_end_0"), val = tensor([1, 42, 1024])]; + tensor var_1330_end_mask_0 = const()[name = string("op_1330_end_mask_0"), val = tensor([true, true, true])]; + tensor var_1330_cast_fp16 = slice_by_index(begin = var_1330_begin_0, end = var_1330_end_0, end_mask = var_1330_end_mask_0, x = cache_17_cast_fp16)[name = string("op_1330_cast_fp16")]; + bool var_1336_interleave_0 = const()[name = string("op_1336_interleave_0"), val = bool(false)]; + tensor var_1336_cast_fp16 = concat(axis = var_68, interleave = var_1336_interleave_0, values = (var_1330_cast_fp16, key_9_cast_fp16))[name = string("op_1336_cast_fp16")]; + tensor encoder_layers_4_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100741760))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(101528256))))[name = string("encoder_layers_4_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_4_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_4_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(101528448)))]; + tensor linear_39_cast_fp16 = linear(bias = encoder_layers_4_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_4_self_attn_linear_q_weight_to_fp16_palettized, x = key_9_cast_fp16)[name = string("linear_39_cast_fp16")]; + tensor var_1341 = const()[name = string("op_1341"), val = tensor([1, -1, 8, 128])]; + tensor q_25_cast_fp16 = reshape(shape = var_1341, x = linear_39_cast_fp16)[name = string("q_25_cast_fp16")]; + tensor encoder_layers_4_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(101530560))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102317056))))[name = string("encoder_layers_4_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_4_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_4_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102317248)))]; + tensor linear_40_cast_fp16 = linear(bias = encoder_layers_4_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_4_self_attn_linear_k_weight_to_fp16_palettized, x = input_247_cast_fp16)[name = string("linear_40_cast_fp16")]; + tensor var_1346 = const()[name = string("op_1346"), val = tensor([1, -1, 8, 128])]; + tensor k_17_cast_fp16 = reshape(shape = var_1346, x = linear_40_cast_fp16)[name = string("k_17_cast_fp16")]; + tensor encoder_layers_4_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102319360))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103105856))))[name = string("encoder_layers_4_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_4_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_4_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103106048)))]; + tensor linear_41_cast_fp16 = linear(bias = encoder_layers_4_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_4_self_attn_linear_v_weight_to_fp16_palettized, x = input_247_cast_fp16)[name = string("linear_41_cast_fp16")]; + tensor var_1351 = const()[name = string("op_1351"), val = tensor([1, -1, 8, 128])]; + tensor v_9_cast_fp16 = reshape(shape = var_1351, x = linear_41_cast_fp16)[name = string("v_9_cast_fp16")]; + tensor value_17_perm_0 = const()[name = string("value_17_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_4_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_4_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103108160)))]; + tensor var_1364_cast_fp16 = add(x = q_25_cast_fp16, y = encoder_layers_4_self_attn_pos_bias_u_to_fp16)[name = string("op_1364_cast_fp16")]; + tensor encoder_layers_4_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_4_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103110272)))]; + tensor var_1366_cast_fp16 = add(x = q_25_cast_fp16, y = encoder_layers_4_self_attn_pos_bias_v_to_fp16)[name = string("op_1366_cast_fp16")]; + tensor q_with_bias_v_9_perm_0 = const()[name = string("q_with_bias_v_9_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_111_transpose_x_0 = const()[name = string("x_111_transpose_x_0"), val = bool(false)]; + bool x_111_transpose_y_0 = const()[name = string("x_111_transpose_y_0"), val = bool(false)]; + tensor op_1368_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103112384))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103254784))))[name = string("op_1368_to_fp16_quantized")]; + tensor q_with_bias_v_9_cast_fp16 = transpose(perm = q_with_bias_v_9_perm_0, x = var_1366_cast_fp16)[name = string("transpose_326")]; + tensor x_111_cast_fp16 = matmul(transpose_x = x_111_transpose_x_0, transpose_y = x_111_transpose_y_0, x = q_with_bias_v_9_cast_fp16, y = op_1368_to_fp16_quantized)[name = string("x_111_cast_fp16")]; + tensor x_113_pad_0 = const()[name = string("x_113_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_113_mode_0 = const()[name = string("x_113_mode_0"), val = string("constant")]; + fp16 const_131_to_fp16 = const()[name = string("const_131_to_fp16"), val = fp16(0x0p+0)]; + tensor x_113_cast_fp16 = pad(constant_val = const_131_to_fp16, mode = x_113_mode_0, pad = x_113_pad_0, x = x_111_cast_fp16)[name = string("x_113_cast_fp16")]; + tensor var_1376 = const()[name = string("op_1376"), val = tensor([1, 8, -1, 28])]; + tensor x_115_cast_fp16 = reshape(shape = var_1376, x = x_113_cast_fp16)[name = string("x_115_cast_fp16")]; + tensor var_1380_begin_0 = const()[name = string("op_1380_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_1380_end_0 = const()[name = string("op_1380_end_0"), val = tensor([1, 8, 140, 28])]; + tensor var_1380_end_mask_0 = const()[name = string("op_1380_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_1380_cast_fp16 = slice_by_index(begin = var_1380_begin_0, end = var_1380_end_0, end_mask = var_1380_end_mask_0, x = x_115_cast_fp16)[name = string("op_1380_cast_fp16")]; + tensor var_1381 = const()[name = string("op_1381"), val = tensor([1, 8, 28, 139])]; + tensor matrix_bd_17_cast_fp16 = reshape(shape = var_1381, x = var_1380_cast_fp16)[name = string("matrix_bd_17_cast_fp16")]; + bool matrix_ac_9_transpose_x_0 = const()[name = string("matrix_ac_9_transpose_x_0"), val = bool(false)]; + bool matrix_ac_9_transpose_y_0 = const()[name = string("matrix_ac_9_transpose_y_0"), val = bool(false)]; + tensor transpose_104_perm_0 = const()[name = string("transpose_104_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_105_perm_0 = const()[name = string("transpose_105_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_105 = transpose(perm = transpose_105_perm_0, x = k_17_cast_fp16)[name = string("transpose_324")]; + tensor transpose_104 = transpose(perm = transpose_104_perm_0, x = var_1364_cast_fp16)[name = string("transpose_325")]; + tensor matrix_ac_9_cast_fp16 = matmul(transpose_x = matrix_ac_9_transpose_x_0, transpose_y = matrix_ac_9_transpose_y_0, x = transpose_104, y = transpose_105)[name = string("matrix_ac_9_cast_fp16")]; + tensor matrix_bd_19_begin_0 = const()[name = string("matrix_bd_19_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_19_end_0 = const()[name = string("matrix_bd_19_end_0"), val = tensor([1, 8, 28, 70])]; + tensor matrix_bd_19_end_mask_0 = const()[name = string("matrix_bd_19_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_19_cast_fp16 = slice_by_index(begin = matrix_bd_19_begin_0, end = matrix_bd_19_end_0, end_mask = matrix_bd_19_end_mask_0, x = matrix_bd_17_cast_fp16)[name = string("matrix_bd_19_cast_fp16")]; + tensor var_1390_cast_fp16 = add(x = matrix_ac_9_cast_fp16, y = matrix_bd_19_cast_fp16)[name = string("op_1390_cast_fp16")]; + fp16 _inversed_scores_17_y_0_to_fp16 = const()[name = string("_inversed_scores_17_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_17_cast_fp16 = mul(x = var_1390_cast_fp16, y = _inversed_scores_17_y_0_to_fp16)[name = string("_inversed_scores_17_cast_fp16")]; + tensor scores_19_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_17_cast_fp16, cond = mask_11)[name = string("scores_19_cast_fp16")]; + tensor var_1396_cast_fp16 = softmax(axis = var_59, x = scores_19_cast_fp16)[name = string("op_1396_cast_fp16")]; + tensor input_249_cast_fp16 = select(a = var_44_to_fp16, b = var_1396_cast_fp16, cond = mask_11)[name = string("input_249_cast_fp16")]; + bool x_117_transpose_x_0 = const()[name = string("x_117_transpose_x_0"), val = bool(false)]; + bool x_117_transpose_y_0 = const()[name = string("x_117_transpose_y_0"), val = bool(false)]; + tensor value_17_cast_fp16 = transpose(perm = value_17_perm_0, x = v_9_cast_fp16)[name = string("transpose_323")]; + tensor x_117_cast_fp16 = matmul(transpose_x = x_117_transpose_x_0, transpose_y = x_117_transpose_y_0, x = input_249_cast_fp16, y = value_17_cast_fp16)[name = string("x_117_cast_fp16")]; + tensor var_1400_perm_0 = const()[name = string("op_1400_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1401 = const()[name = string("op_1401"), val = tensor([1, -1, 1024])]; + tensor var_1400_cast_fp16 = transpose(perm = var_1400_perm_0, x = x_117_cast_fp16)[name = string("transpose_322")]; + tensor input_251_cast_fp16 = reshape(shape = var_1401, x = var_1400_cast_fp16)[name = string("input_251_cast_fp16")]; + tensor encoder_layers_4_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103255168))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104041664))))[name = string("encoder_layers_4_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_4_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_4_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104041856)))]; + tensor linear_43_cast_fp16 = linear(bias = encoder_layers_4_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_4_self_attn_linear_out_weight_to_fp16_palettized, x = input_251_cast_fp16)[name = string("linear_43_cast_fp16")]; + tensor input_255_cast_fp16 = add(x = input_245_cast_fp16, y = linear_43_cast_fp16)[name = string("input_255_cast_fp16")]; + tensor x_121_axes_0 = const()[name = string("x_121_axes_0"), val = tensor([-1])]; + tensor encoder_layers_4_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_4_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104043968)))]; + tensor encoder_layers_4_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_4_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104046080)))]; + tensor x_121_cast_fp16 = layer_norm(axes = x_121_axes_0, beta = encoder_layers_4_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_4_norm_conv_weight_to_fp16, x = input_255_cast_fp16)[name = string("x_121_cast_fp16")]; + tensor input_257_perm_0 = const()[name = string("input_257_perm_0"), val = tensor([0, 2, 1])]; + string input_259_pad_type_0 = const()[name = string("input_259_pad_type_0"), val = string("valid")]; + tensor input_259_strides_0 = const()[name = string("input_259_strides_0"), val = tensor([1])]; + tensor input_259_pad_0 = const()[name = string("input_259_pad_0"), val = tensor([0, 0])]; + tensor input_259_dilations_0 = const()[name = string("input_259_dilations_0"), val = tensor([1])]; + int32 input_259_groups_0 = const()[name = string("input_259_groups_0"), val = int32(1)]; + tensor encoder_layers_4_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104048192))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106145408))))[name = string("encoder_layers_4_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_257_cast_fp16 = transpose(perm = input_257_perm_0, x = x_121_cast_fp16)[name = string("transpose_321")]; + tensor input_259_cast_fp16 = conv(dilations = input_259_dilations_0, groups = input_259_groups_0, pad = input_259_pad_0, pad_type = input_259_pad_type_0, strides = input_259_strides_0, weight = encoder_layers_4_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_257_cast_fp16)[name = string("input_259_cast_fp16")]; + int32 x_123_split_num_splits_0 = const()[name = string("x_123_split_num_splits_0"), val = int32(2)]; + int32 x_123_split_axis_0 = const()[name = string("x_123_split_axis_0"), val = int32(1)]; + tensor x_123_split_cast_fp16_0, tensor x_123_split_cast_fp16_1 = split(axis = x_123_split_axis_0, num_splits = x_123_split_num_splits_0, x = input_259_cast_fp16)[name = string("x_123_split_cast_fp16")]; + tensor x_123_split_1_sigmoid_cast_fp16 = sigmoid(x = x_123_split_cast_fp16_1)[name = string("x_123_split_1_sigmoid_cast_fp16")]; + tensor x_123_cast_fp16 = mul(x = x_123_split_cast_fp16_0, y = x_123_split_1_sigmoid_cast_fp16)[name = string("x_123_cast_fp16")]; + tensor input_261_cast_fp16 = select(a = var_44_to_fp16, b = x_123_cast_fp16, cond = var_575)[name = string("input_261_cast_fp16")]; + bool new_x_19_interleave_0 = const()[name = string("new_x_19_interleave_0"), val = bool(false)]; + tensor new_x_19_cast_fp16 = concat(axis = var_59, interleave = new_x_19_interleave_0, values = (cache_19_cast_fp16, input_261_cast_fp16))[name = string("new_x_19_cast_fp16")]; + tensor var_1440_begin_0 = const()[name = string("op_1440_begin_0"), val = tensor([0, 0, 28])]; + tensor var_1440_end_0 = const()[name = string("op_1440_end_0"), val = tensor([1, 1024, 36])]; + tensor var_1440_end_mask_0 = const()[name = string("op_1440_end_mask_0"), val = tensor([true, true, true])]; + tensor var_1440_cast_fp16 = slice_by_index(begin = var_1440_begin_0, end = var_1440_end_0, end_mask = var_1440_end_mask_0, x = new_x_19_cast_fp16)[name = string("op_1440_cast_fp16")]; + string x_125_pad_type_0 = const()[name = string("x_125_pad_type_0"), val = string("valid")]; + int32 x_125_groups_0 = const()[name = string("x_125_groups_0"), val = int32(1024)]; + tensor x_125_strides_0 = const()[name = string("x_125_strides_0"), val = tensor([1])]; + tensor x_125_pad_0 = const()[name = string("x_125_pad_0"), val = tensor([0, 0])]; + tensor x_125_dilations_0 = const()[name = string("x_125_dilations_0"), val = tensor([1])]; + tensor encoder_layers_4_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106149568))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106158848))))[name = string("encoder_layers_4_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_125_cast_fp16 = conv(dilations = x_125_dilations_0, groups = x_125_groups_0, pad = x_125_pad_0, pad_type = x_125_pad_type_0, strides = x_125_strides_0, weight = encoder_layers_4_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_19_cast_fp16)[name = string("x_125_cast_fp16")]; + tensor input_263_perm_0 = const()[name = string("input_263_perm_0"), val = tensor([0, 2, 1])]; + tensor x_127_axes_0 = const()[name = string("x_127_axes_0"), val = tensor([-1])]; + tensor encoder_layers_4_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_4_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106160960)))]; + tensor encoder_layers_4_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_4_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106163072)))]; + tensor input_263_cast_fp16 = transpose(perm = input_263_perm_0, x = x_125_cast_fp16)[name = string("transpose_320")]; + tensor x_127_cast_fp16 = layer_norm(axes = x_127_axes_0, beta = encoder_layers_4_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_4_conv_batch_norm_weight_to_fp16, x = input_263_cast_fp16)[name = string("x_127_cast_fp16")]; + tensor input_265_perm_0 = const()[name = string("input_265_perm_0"), val = tensor([0, 2, 1])]; + tensor input_265_cast_fp16 = transpose(perm = input_265_perm_0, x = x_127_cast_fp16)[name = string("transpose_319")]; + tensor input_267_cast_fp16 = silu(x = input_265_cast_fp16)[name = string("input_267_cast_fp16")]; + string x_129_pad_type_0 = const()[name = string("x_129_pad_type_0"), val = string("valid")]; + tensor x_129_strides_0 = const()[name = string("x_129_strides_0"), val = tensor([1])]; + tensor x_129_pad_0 = const()[name = string("x_129_pad_0"), val = tensor([0, 0])]; + tensor x_129_dilations_0 = const()[name = string("x_129_dilations_0"), val = tensor([1])]; + int32 x_129_groups_0 = const()[name = string("x_129_groups_0"), val = int32(1)]; + tensor encoder_layers_4_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106165184))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107213824))))[name = string("encoder_layers_4_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_129_cast_fp16 = conv(dilations = x_129_dilations_0, groups = x_129_groups_0, pad = x_129_pad_0, pad_type = x_129_pad_type_0, strides = x_129_strides_0, weight = encoder_layers_4_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_267_cast_fp16)[name = string("x_129_cast_fp16")]; + tensor input_269_perm_0 = const()[name = string("input_269_perm_0"), val = tensor([0, 2, 1])]; + tensor input_269_cast_fp16 = transpose(perm = input_269_perm_0, x = x_129_cast_fp16)[name = string("transpose_318")]; + tensor input_271_cast_fp16 = add(x = input_255_cast_fp16, y = input_269_cast_fp16)[name = string("input_271_cast_fp16")]; + tensor input_273_axes_0 = const()[name = string("input_273_axes_0"), val = tensor([-1])]; + tensor encoder_layers_4_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_4_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107215936)))]; + tensor encoder_layers_4_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_4_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107218048)))]; + tensor input_273_cast_fp16 = layer_norm(axes = input_273_axes_0, beta = encoder_layers_4_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_4_norm_feed_forward2_weight_to_fp16, x = input_271_cast_fp16)[name = string("input_273_cast_fp16")]; + tensor encoder_layers_4_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107220160))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110365952))))[name = string("encoder_layers_4_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_4_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_4_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110366144)))]; + tensor linear_44_cast_fp16 = linear(bias = encoder_layers_4_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_4_feed_forward2_linear1_weight_to_fp16_palettized, x = input_273_cast_fp16)[name = string("linear_44_cast_fp16")]; + tensor input_277_cast_fp16 = silu(x = linear_44_cast_fp16)[name = string("input_277_cast_fp16")]; + tensor encoder_layers_4_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110374400))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113520192))))[name = string("encoder_layers_4_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_4_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_4_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113520384)))]; + tensor linear_45_cast_fp16 = linear(bias = encoder_layers_4_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_4_feed_forward2_linear2_weight_to_fp16_palettized, x = input_277_cast_fp16)[name = string("linear_45_cast_fp16")]; + fp16 var_1483_to_fp16 = const()[name = string("op_1483_to_fp16"), val = fp16(0x1p-1)]; + tensor var_1484_cast_fp16 = mul(x = linear_45_cast_fp16, y = var_1483_to_fp16)[name = string("op_1484_cast_fp16")]; + tensor input_283_cast_fp16 = add(x = input_271_cast_fp16, y = var_1484_cast_fp16)[name = string("input_283_cast_fp16")]; + tensor input_285_axes_0 = const()[name = string("input_285_axes_0"), val = tensor([-1])]; + tensor encoder_layers_4_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_4_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113522496)))]; + tensor encoder_layers_4_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_4_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113524608)))]; + tensor input_285_cast_fp16 = layer_norm(axes = input_285_axes_0, beta = encoder_layers_4_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_4_norm_out_weight_to_fp16, x = input_283_cast_fp16)[name = string("input_285_cast_fp16")]; + tensor cache_21_begin_0 = const()[name = string("cache_21_begin_0"), val = tensor([5, 0, 0, 0])]; + tensor cache_21_end_0 = const()[name = string("cache_21_end_0"), val = tensor([6, 1, 42, 1024])]; + tensor cache_21_end_mask_0 = const()[name = string("cache_21_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_21_squeeze_mask_0 = const()[name = string("cache_21_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_21_cast_fp16 = slice_by_index(begin = cache_21_begin_0, end = cache_21_end_0, end_mask = cache_21_end_mask_0, squeeze_mask = cache_21_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_21_cast_fp16")]; + tensor cache_23_begin_0 = const()[name = string("cache_23_begin_0"), val = tensor([5, 0, 0, 0])]; + tensor cache_23_end_0 = const()[name = string("cache_23_end_0"), val = tensor([6, 1, 1024, 8])]; + tensor cache_23_end_mask_0 = const()[name = string("cache_23_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_23_squeeze_mask_0 = const()[name = string("cache_23_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_23_cast_fp16 = slice_by_index(begin = cache_23_begin_0, end = cache_23_end_0, end_mask = cache_23_end_mask_0, squeeze_mask = cache_23_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_23_cast_fp16")]; + tensor input_287_axes_0 = const()[name = string("input_287_axes_0"), val = tensor([-1])]; + tensor encoder_layers_5_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_5_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113526720)))]; + tensor encoder_layers_5_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_5_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113528832)))]; + tensor input_287_cast_fp16 = layer_norm(axes = input_287_axes_0, beta = encoder_layers_5_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_5_norm_feed_forward1_weight_to_fp16, x = input_285_cast_fp16)[name = string("input_287_cast_fp16")]; + tensor encoder_layers_5_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113530944))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116676736))))[name = string("encoder_layers_5_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_5_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_5_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116676928)))]; + tensor linear_46_cast_fp16 = linear(bias = encoder_layers_5_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_5_feed_forward1_linear1_weight_to_fp16_palettized, x = input_287_cast_fp16)[name = string("linear_46_cast_fp16")]; + tensor input_291_cast_fp16 = silu(x = linear_46_cast_fp16)[name = string("input_291_cast_fp16")]; + tensor encoder_layers_5_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116685184))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119830976))))[name = string("encoder_layers_5_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_5_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_5_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119831168)))]; + tensor linear_47_cast_fp16 = linear(bias = encoder_layers_5_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_5_feed_forward1_linear2_weight_to_fp16_palettized, x = input_291_cast_fp16)[name = string("linear_47_cast_fp16")]; + fp16 var_1520_to_fp16 = const()[name = string("op_1520_to_fp16"), val = fp16(0x1p-1)]; + tensor var_1521_cast_fp16 = mul(x = linear_47_cast_fp16, y = var_1520_to_fp16)[name = string("op_1521_cast_fp16")]; + tensor input_297_cast_fp16 = add(x = input_285_cast_fp16, y = var_1521_cast_fp16)[name = string("input_297_cast_fp16")]; + tensor key_11_axes_0 = const()[name = string("key_11_axes_0"), val = tensor([-1])]; + tensor encoder_layers_5_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_5_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119833280)))]; + tensor encoder_layers_5_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_5_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119835392)))]; + tensor key_11_cast_fp16 = layer_norm(axes = key_11_axes_0, beta = encoder_layers_5_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_5_norm_self_att_weight_to_fp16, x = input_297_cast_fp16)[name = string("key_11_cast_fp16")]; + bool input_299_interleave_0 = const()[name = string("input_299_interleave_0"), val = bool(false)]; + tensor input_299_cast_fp16 = concat(axis = var_68, interleave = input_299_interleave_0, values = (cache_21_cast_fp16, key_11_cast_fp16))[name = string("input_299_cast_fp16")]; + tensor var_1543_begin_0 = const()[name = string("op_1543_begin_0"), val = tensor([0, 28, 0])]; + tensor var_1543_end_0 = const()[name = string("op_1543_end_0"), val = tensor([1, 42, 1024])]; + tensor var_1543_end_mask_0 = const()[name = string("op_1543_end_mask_0"), val = tensor([true, true, true])]; + tensor var_1543_cast_fp16 = slice_by_index(begin = var_1543_begin_0, end = var_1543_end_0, end_mask = var_1543_end_mask_0, x = cache_21_cast_fp16)[name = string("op_1543_cast_fp16")]; + bool var_1549_interleave_0 = const()[name = string("op_1549_interleave_0"), val = bool(false)]; + tensor var_1549_cast_fp16 = concat(axis = var_68, interleave = var_1549_interleave_0, values = (var_1543_cast_fp16, key_11_cast_fp16))[name = string("op_1549_cast_fp16")]; + tensor encoder_layers_5_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119837504))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120624000))))[name = string("encoder_layers_5_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_5_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_5_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120624192)))]; + tensor linear_48_cast_fp16 = linear(bias = encoder_layers_5_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_5_self_attn_linear_q_weight_to_fp16_palettized, x = key_11_cast_fp16)[name = string("linear_48_cast_fp16")]; + tensor var_1554 = const()[name = string("op_1554"), val = tensor([1, -1, 8, 128])]; + tensor q_31_cast_fp16 = reshape(shape = var_1554, x = linear_48_cast_fp16)[name = string("q_31_cast_fp16")]; + tensor encoder_layers_5_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120626304))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121412800))))[name = string("encoder_layers_5_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_5_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_5_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121412992)))]; + tensor linear_49_cast_fp16 = linear(bias = encoder_layers_5_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_5_self_attn_linear_k_weight_to_fp16_palettized, x = input_299_cast_fp16)[name = string("linear_49_cast_fp16")]; + tensor var_1559 = const()[name = string("op_1559"), val = tensor([1, -1, 8, 128])]; + tensor k_21_cast_fp16 = reshape(shape = var_1559, x = linear_49_cast_fp16)[name = string("k_21_cast_fp16")]; + tensor encoder_layers_5_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121415104))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122201600))))[name = string("encoder_layers_5_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_5_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_5_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122201792)))]; + tensor linear_50_cast_fp16 = linear(bias = encoder_layers_5_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_5_self_attn_linear_v_weight_to_fp16_palettized, x = input_299_cast_fp16)[name = string("linear_50_cast_fp16")]; + tensor var_1564 = const()[name = string("op_1564"), val = tensor([1, -1, 8, 128])]; + tensor v_11_cast_fp16 = reshape(shape = var_1564, x = linear_50_cast_fp16)[name = string("v_11_cast_fp16")]; + tensor value_19_perm_0 = const()[name = string("value_19_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_5_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_5_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122203904)))]; + tensor var_1577_cast_fp16 = add(x = q_31_cast_fp16, y = encoder_layers_5_self_attn_pos_bias_u_to_fp16)[name = string("op_1577_cast_fp16")]; + tensor encoder_layers_5_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_5_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122206016)))]; + tensor var_1579_cast_fp16 = add(x = q_31_cast_fp16, y = encoder_layers_5_self_attn_pos_bias_v_to_fp16)[name = string("op_1579_cast_fp16")]; + tensor q_with_bias_v_11_perm_0 = const()[name = string("q_with_bias_v_11_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_137_transpose_x_0 = const()[name = string("x_137_transpose_x_0"), val = bool(false)]; + bool x_137_transpose_y_0 = const()[name = string("x_137_transpose_y_0"), val = bool(false)]; + tensor op_1581_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122208128))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122350528))))[name = string("op_1581_to_fp16_quantized")]; + tensor q_with_bias_v_11_cast_fp16 = transpose(perm = q_with_bias_v_11_perm_0, x = var_1579_cast_fp16)[name = string("transpose_317")]; + tensor x_137_cast_fp16 = matmul(transpose_x = x_137_transpose_x_0, transpose_y = x_137_transpose_y_0, x = q_with_bias_v_11_cast_fp16, y = op_1581_to_fp16_quantized)[name = string("x_137_cast_fp16")]; + tensor x_139_pad_0 = const()[name = string("x_139_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_139_mode_0 = const()[name = string("x_139_mode_0"), val = string("constant")]; + fp16 const_144_to_fp16 = const()[name = string("const_144_to_fp16"), val = fp16(0x0p+0)]; + tensor x_139_cast_fp16 = pad(constant_val = const_144_to_fp16, mode = x_139_mode_0, pad = x_139_pad_0, x = x_137_cast_fp16)[name = string("x_139_cast_fp16")]; + tensor var_1589 = const()[name = string("op_1589"), val = tensor([1, 8, -1, 28])]; + tensor x_141_cast_fp16 = reshape(shape = var_1589, x = x_139_cast_fp16)[name = string("x_141_cast_fp16")]; + tensor var_1593_begin_0 = const()[name = string("op_1593_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_1593_end_0 = const()[name = string("op_1593_end_0"), val = tensor([1, 8, 140, 28])]; + tensor var_1593_end_mask_0 = const()[name = string("op_1593_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_1593_cast_fp16 = slice_by_index(begin = var_1593_begin_0, end = var_1593_end_0, end_mask = var_1593_end_mask_0, x = x_141_cast_fp16)[name = string("op_1593_cast_fp16")]; + tensor var_1594 = const()[name = string("op_1594"), val = tensor([1, 8, 28, 139])]; + tensor matrix_bd_21_cast_fp16 = reshape(shape = var_1594, x = var_1593_cast_fp16)[name = string("matrix_bd_21_cast_fp16")]; + bool matrix_ac_11_transpose_x_0 = const()[name = string("matrix_ac_11_transpose_x_0"), val = bool(false)]; + bool matrix_ac_11_transpose_y_0 = const()[name = string("matrix_ac_11_transpose_y_0"), val = bool(false)]; + tensor transpose_106_perm_0 = const()[name = string("transpose_106_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_107_perm_0 = const()[name = string("transpose_107_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_107 = transpose(perm = transpose_107_perm_0, x = k_21_cast_fp16)[name = string("transpose_315")]; + tensor transpose_106 = transpose(perm = transpose_106_perm_0, x = var_1577_cast_fp16)[name = string("transpose_316")]; + tensor matrix_ac_11_cast_fp16 = matmul(transpose_x = matrix_ac_11_transpose_x_0, transpose_y = matrix_ac_11_transpose_y_0, x = transpose_106, y = transpose_107)[name = string("matrix_ac_11_cast_fp16")]; + tensor matrix_bd_23_begin_0 = const()[name = string("matrix_bd_23_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_23_end_0 = const()[name = string("matrix_bd_23_end_0"), val = tensor([1, 8, 28, 70])]; + tensor matrix_bd_23_end_mask_0 = const()[name = string("matrix_bd_23_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_23_cast_fp16 = slice_by_index(begin = matrix_bd_23_begin_0, end = matrix_bd_23_end_0, end_mask = matrix_bd_23_end_mask_0, x = matrix_bd_21_cast_fp16)[name = string("matrix_bd_23_cast_fp16")]; + tensor var_1603_cast_fp16 = add(x = matrix_ac_11_cast_fp16, y = matrix_bd_23_cast_fp16)[name = string("op_1603_cast_fp16")]; + fp16 _inversed_scores_21_y_0_to_fp16 = const()[name = string("_inversed_scores_21_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_21_cast_fp16 = mul(x = var_1603_cast_fp16, y = _inversed_scores_21_y_0_to_fp16)[name = string("_inversed_scores_21_cast_fp16")]; + tensor scores_23_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_21_cast_fp16, cond = mask_11)[name = string("scores_23_cast_fp16")]; + tensor var_1609_cast_fp16 = softmax(axis = var_59, x = scores_23_cast_fp16)[name = string("op_1609_cast_fp16")]; + tensor input_301_cast_fp16 = select(a = var_44_to_fp16, b = var_1609_cast_fp16, cond = mask_11)[name = string("input_301_cast_fp16")]; + bool x_143_transpose_x_0 = const()[name = string("x_143_transpose_x_0"), val = bool(false)]; + bool x_143_transpose_y_0 = const()[name = string("x_143_transpose_y_0"), val = bool(false)]; + tensor value_19_cast_fp16 = transpose(perm = value_19_perm_0, x = v_11_cast_fp16)[name = string("transpose_314")]; + tensor x_143_cast_fp16 = matmul(transpose_x = x_143_transpose_x_0, transpose_y = x_143_transpose_y_0, x = input_301_cast_fp16, y = value_19_cast_fp16)[name = string("x_143_cast_fp16")]; + tensor var_1613_perm_0 = const()[name = string("op_1613_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1614 = const()[name = string("op_1614"), val = tensor([1, -1, 1024])]; + tensor var_1613_cast_fp16 = transpose(perm = var_1613_perm_0, x = x_143_cast_fp16)[name = string("transpose_313")]; + tensor input_303_cast_fp16 = reshape(shape = var_1614, x = var_1613_cast_fp16)[name = string("input_303_cast_fp16")]; + tensor encoder_layers_5_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122350912))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123137408))))[name = string("encoder_layers_5_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_5_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_5_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123137600)))]; + tensor linear_52_cast_fp16 = linear(bias = encoder_layers_5_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_5_self_attn_linear_out_weight_to_fp16_palettized, x = input_303_cast_fp16)[name = string("linear_52_cast_fp16")]; + tensor input_307_cast_fp16 = add(x = input_297_cast_fp16, y = linear_52_cast_fp16)[name = string("input_307_cast_fp16")]; + tensor x_147_axes_0 = const()[name = string("x_147_axes_0"), val = tensor([-1])]; + tensor encoder_layers_5_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_5_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123139712)))]; + tensor encoder_layers_5_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_5_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123141824)))]; + tensor x_147_cast_fp16 = layer_norm(axes = x_147_axes_0, beta = encoder_layers_5_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_5_norm_conv_weight_to_fp16, x = input_307_cast_fp16)[name = string("x_147_cast_fp16")]; + tensor input_309_perm_0 = const()[name = string("input_309_perm_0"), val = tensor([0, 2, 1])]; + string input_311_pad_type_0 = const()[name = string("input_311_pad_type_0"), val = string("valid")]; + tensor input_311_strides_0 = const()[name = string("input_311_strides_0"), val = tensor([1])]; + tensor input_311_pad_0 = const()[name = string("input_311_pad_0"), val = tensor([0, 0])]; + tensor input_311_dilations_0 = const()[name = string("input_311_dilations_0"), val = tensor([1])]; + int32 input_311_groups_0 = const()[name = string("input_311_groups_0"), val = int32(1)]; + tensor encoder_layers_5_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123143936))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125241152))))[name = string("encoder_layers_5_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_309_cast_fp16 = transpose(perm = input_309_perm_0, x = x_147_cast_fp16)[name = string("transpose_312")]; + tensor input_311_cast_fp16 = conv(dilations = input_311_dilations_0, groups = input_311_groups_0, pad = input_311_pad_0, pad_type = input_311_pad_type_0, strides = input_311_strides_0, weight = encoder_layers_5_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_309_cast_fp16)[name = string("input_311_cast_fp16")]; + int32 x_149_split_num_splits_0 = const()[name = string("x_149_split_num_splits_0"), val = int32(2)]; + int32 x_149_split_axis_0 = const()[name = string("x_149_split_axis_0"), val = int32(1)]; + tensor x_149_split_cast_fp16_0, tensor x_149_split_cast_fp16_1 = split(axis = x_149_split_axis_0, num_splits = x_149_split_num_splits_0, x = input_311_cast_fp16)[name = string("x_149_split_cast_fp16")]; + tensor x_149_split_1_sigmoid_cast_fp16 = sigmoid(x = x_149_split_cast_fp16_1)[name = string("x_149_split_1_sigmoid_cast_fp16")]; + tensor x_149_cast_fp16 = mul(x = x_149_split_cast_fp16_0, y = x_149_split_1_sigmoid_cast_fp16)[name = string("x_149_cast_fp16")]; + tensor input_313_cast_fp16 = select(a = var_44_to_fp16, b = x_149_cast_fp16, cond = var_575)[name = string("input_313_cast_fp16")]; + bool new_x_23_interleave_0 = const()[name = string("new_x_23_interleave_0"), val = bool(false)]; + tensor new_x_23_cast_fp16 = concat(axis = var_59, interleave = new_x_23_interleave_0, values = (cache_23_cast_fp16, input_313_cast_fp16))[name = string("new_x_23_cast_fp16")]; + tensor var_1653_begin_0 = const()[name = string("op_1653_begin_0"), val = tensor([0, 0, 28])]; + tensor var_1653_end_0 = const()[name = string("op_1653_end_0"), val = tensor([1, 1024, 36])]; + tensor var_1653_end_mask_0 = const()[name = string("op_1653_end_mask_0"), val = tensor([true, true, true])]; + tensor var_1653_cast_fp16 = slice_by_index(begin = var_1653_begin_0, end = var_1653_end_0, end_mask = var_1653_end_mask_0, x = new_x_23_cast_fp16)[name = string("op_1653_cast_fp16")]; + string x_151_pad_type_0 = const()[name = string("x_151_pad_type_0"), val = string("valid")]; + int32 x_151_groups_0 = const()[name = string("x_151_groups_0"), val = int32(1024)]; + tensor x_151_strides_0 = const()[name = string("x_151_strides_0"), val = tensor([1])]; + tensor x_151_pad_0 = const()[name = string("x_151_pad_0"), val = tensor([0, 0])]; + tensor x_151_dilations_0 = const()[name = string("x_151_dilations_0"), val = tensor([1])]; + tensor encoder_layers_5_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125245312))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125254592))))[name = string("encoder_layers_5_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_151_cast_fp16 = conv(dilations = x_151_dilations_0, groups = x_151_groups_0, pad = x_151_pad_0, pad_type = x_151_pad_type_0, strides = x_151_strides_0, weight = encoder_layers_5_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_23_cast_fp16)[name = string("x_151_cast_fp16")]; + tensor input_315_perm_0 = const()[name = string("input_315_perm_0"), val = tensor([0, 2, 1])]; + tensor x_153_axes_0 = const()[name = string("x_153_axes_0"), val = tensor([-1])]; + tensor encoder_layers_5_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_5_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125256704)))]; + tensor encoder_layers_5_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_5_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125258816)))]; + tensor input_315_cast_fp16 = transpose(perm = input_315_perm_0, x = x_151_cast_fp16)[name = string("transpose_311")]; + tensor x_153_cast_fp16 = layer_norm(axes = x_153_axes_0, beta = encoder_layers_5_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_5_conv_batch_norm_weight_to_fp16, x = input_315_cast_fp16)[name = string("x_153_cast_fp16")]; + tensor input_317_perm_0 = const()[name = string("input_317_perm_0"), val = tensor([0, 2, 1])]; + tensor input_317_cast_fp16 = transpose(perm = input_317_perm_0, x = x_153_cast_fp16)[name = string("transpose_310")]; + tensor input_319_cast_fp16 = silu(x = input_317_cast_fp16)[name = string("input_319_cast_fp16")]; + string x_155_pad_type_0 = const()[name = string("x_155_pad_type_0"), val = string("valid")]; + tensor x_155_strides_0 = const()[name = string("x_155_strides_0"), val = tensor([1])]; + tensor x_155_pad_0 = const()[name = string("x_155_pad_0"), val = tensor([0, 0])]; + tensor x_155_dilations_0 = const()[name = string("x_155_dilations_0"), val = tensor([1])]; + int32 x_155_groups_0 = const()[name = string("x_155_groups_0"), val = int32(1)]; + tensor encoder_layers_5_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125260928))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126309568))))[name = string("encoder_layers_5_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_155_cast_fp16 = conv(dilations = x_155_dilations_0, groups = x_155_groups_0, pad = x_155_pad_0, pad_type = x_155_pad_type_0, strides = x_155_strides_0, weight = encoder_layers_5_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_319_cast_fp16)[name = string("x_155_cast_fp16")]; + tensor input_321_perm_0 = const()[name = string("input_321_perm_0"), val = tensor([0, 2, 1])]; + tensor input_321_cast_fp16 = transpose(perm = input_321_perm_0, x = x_155_cast_fp16)[name = string("transpose_309")]; + tensor input_323_cast_fp16 = add(x = input_307_cast_fp16, y = input_321_cast_fp16)[name = string("input_323_cast_fp16")]; + tensor input_325_axes_0 = const()[name = string("input_325_axes_0"), val = tensor([-1])]; + tensor encoder_layers_5_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_5_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126311680)))]; + tensor encoder_layers_5_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_5_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126313792)))]; + tensor input_325_cast_fp16 = layer_norm(axes = input_325_axes_0, beta = encoder_layers_5_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_5_norm_feed_forward2_weight_to_fp16, x = input_323_cast_fp16)[name = string("input_325_cast_fp16")]; + tensor encoder_layers_5_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126315904))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(129461696))))[name = string("encoder_layers_5_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_5_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_5_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(129461888)))]; + tensor linear_53_cast_fp16 = linear(bias = encoder_layers_5_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_5_feed_forward2_linear1_weight_to_fp16_palettized, x = input_325_cast_fp16)[name = string("linear_53_cast_fp16")]; + tensor input_329_cast_fp16 = silu(x = linear_53_cast_fp16)[name = string("input_329_cast_fp16")]; + tensor encoder_layers_5_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(129470144))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132615936))))[name = string("encoder_layers_5_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_5_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_5_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132616128)))]; + tensor linear_54_cast_fp16 = linear(bias = encoder_layers_5_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_5_feed_forward2_linear2_weight_to_fp16_palettized, x = input_329_cast_fp16)[name = string("linear_54_cast_fp16")]; + fp16 var_1696_to_fp16 = const()[name = string("op_1696_to_fp16"), val = fp16(0x1p-1)]; + tensor var_1697_cast_fp16 = mul(x = linear_54_cast_fp16, y = var_1696_to_fp16)[name = string("op_1697_cast_fp16")]; + tensor input_335_cast_fp16 = add(x = input_323_cast_fp16, y = var_1697_cast_fp16)[name = string("input_335_cast_fp16")]; + tensor input_337_axes_0 = const()[name = string("input_337_axes_0"), val = tensor([-1])]; + tensor encoder_layers_5_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_5_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132618240)))]; + tensor encoder_layers_5_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_5_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132620352)))]; + tensor input_337_cast_fp16 = layer_norm(axes = input_337_axes_0, beta = encoder_layers_5_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_5_norm_out_weight_to_fp16, x = input_335_cast_fp16)[name = string("input_337_cast_fp16")]; + tensor cache_25_begin_0 = const()[name = string("cache_25_begin_0"), val = tensor([6, 0, 0, 0])]; + tensor cache_25_end_0 = const()[name = string("cache_25_end_0"), val = tensor([7, 1, 42, 1024])]; + tensor cache_25_end_mask_0 = const()[name = string("cache_25_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_25_squeeze_mask_0 = const()[name = string("cache_25_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_25_cast_fp16 = slice_by_index(begin = cache_25_begin_0, end = cache_25_end_0, end_mask = cache_25_end_mask_0, squeeze_mask = cache_25_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_25_cast_fp16")]; + tensor cache_27_begin_0 = const()[name = string("cache_27_begin_0"), val = tensor([6, 0, 0, 0])]; + tensor cache_27_end_0 = const()[name = string("cache_27_end_0"), val = tensor([7, 1, 1024, 8])]; + tensor cache_27_end_mask_0 = const()[name = string("cache_27_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_27_squeeze_mask_0 = const()[name = string("cache_27_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_27_cast_fp16 = slice_by_index(begin = cache_27_begin_0, end = cache_27_end_0, end_mask = cache_27_end_mask_0, squeeze_mask = cache_27_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_27_cast_fp16")]; + tensor input_339_axes_0 = const()[name = string("input_339_axes_0"), val = tensor([-1])]; + tensor encoder_layers_6_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_6_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132622464)))]; + tensor encoder_layers_6_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_6_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132624576)))]; + tensor input_339_cast_fp16 = layer_norm(axes = input_339_axes_0, beta = encoder_layers_6_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_6_norm_feed_forward1_weight_to_fp16, x = input_337_cast_fp16)[name = string("input_339_cast_fp16")]; + tensor encoder_layers_6_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132626688))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135772480))))[name = string("encoder_layers_6_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_6_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_6_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135772672)))]; + tensor linear_55_cast_fp16 = linear(bias = encoder_layers_6_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_6_feed_forward1_linear1_weight_to_fp16_palettized, x = input_339_cast_fp16)[name = string("linear_55_cast_fp16")]; + tensor input_343_cast_fp16 = silu(x = linear_55_cast_fp16)[name = string("input_343_cast_fp16")]; + tensor encoder_layers_6_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135780928))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138926720))))[name = string("encoder_layers_6_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_6_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_6_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138926912)))]; + tensor linear_56_cast_fp16 = linear(bias = encoder_layers_6_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_6_feed_forward1_linear2_weight_to_fp16_palettized, x = input_343_cast_fp16)[name = string("linear_56_cast_fp16")]; + fp16 var_1733_to_fp16 = const()[name = string("op_1733_to_fp16"), val = fp16(0x1p-1)]; + tensor var_1734_cast_fp16 = mul(x = linear_56_cast_fp16, y = var_1733_to_fp16)[name = string("op_1734_cast_fp16")]; + tensor input_349_cast_fp16 = add(x = input_337_cast_fp16, y = var_1734_cast_fp16)[name = string("input_349_cast_fp16")]; + tensor key_13_axes_0 = const()[name = string("key_13_axes_0"), val = tensor([-1])]; + tensor encoder_layers_6_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_6_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138929024)))]; + tensor encoder_layers_6_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_6_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138931136)))]; + tensor key_13_cast_fp16 = layer_norm(axes = key_13_axes_0, beta = encoder_layers_6_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_6_norm_self_att_weight_to_fp16, x = input_349_cast_fp16)[name = string("key_13_cast_fp16")]; + bool input_351_interleave_0 = const()[name = string("input_351_interleave_0"), val = bool(false)]; + tensor input_351_cast_fp16 = concat(axis = var_68, interleave = input_351_interleave_0, values = (cache_25_cast_fp16, key_13_cast_fp16))[name = string("input_351_cast_fp16")]; + tensor var_1756_begin_0 = const()[name = string("op_1756_begin_0"), val = tensor([0, 28, 0])]; + tensor var_1756_end_0 = const()[name = string("op_1756_end_0"), val = tensor([1, 42, 1024])]; + tensor var_1756_end_mask_0 = const()[name = string("op_1756_end_mask_0"), val = tensor([true, true, true])]; + tensor var_1756_cast_fp16 = slice_by_index(begin = var_1756_begin_0, end = var_1756_end_0, end_mask = var_1756_end_mask_0, x = cache_25_cast_fp16)[name = string("op_1756_cast_fp16")]; + bool var_1762_interleave_0 = const()[name = string("op_1762_interleave_0"), val = bool(false)]; + tensor var_1762_cast_fp16 = concat(axis = var_68, interleave = var_1762_interleave_0, values = (var_1756_cast_fp16, key_13_cast_fp16))[name = string("op_1762_cast_fp16")]; + tensor encoder_layers_6_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138933248))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(139719744))))[name = string("encoder_layers_6_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_6_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_6_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(139719936)))]; + tensor linear_57_cast_fp16 = linear(bias = encoder_layers_6_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_6_self_attn_linear_q_weight_to_fp16_palettized, x = key_13_cast_fp16)[name = string("linear_57_cast_fp16")]; + tensor var_1767 = const()[name = string("op_1767"), val = tensor([1, -1, 8, 128])]; + tensor q_37_cast_fp16 = reshape(shape = var_1767, x = linear_57_cast_fp16)[name = string("q_37_cast_fp16")]; + tensor encoder_layers_6_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(139722048))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140508544))))[name = string("encoder_layers_6_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_6_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_6_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140508736)))]; + tensor linear_58_cast_fp16 = linear(bias = encoder_layers_6_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_6_self_attn_linear_k_weight_to_fp16_palettized, x = input_351_cast_fp16)[name = string("linear_58_cast_fp16")]; + tensor var_1772 = const()[name = string("op_1772"), val = tensor([1, -1, 8, 128])]; + tensor k_25_cast_fp16 = reshape(shape = var_1772, x = linear_58_cast_fp16)[name = string("k_25_cast_fp16")]; + tensor encoder_layers_6_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140510848))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141297344))))[name = string("encoder_layers_6_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_6_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_6_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141297536)))]; + tensor linear_59_cast_fp16 = linear(bias = encoder_layers_6_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_6_self_attn_linear_v_weight_to_fp16_palettized, x = input_351_cast_fp16)[name = string("linear_59_cast_fp16")]; + tensor var_1777 = const()[name = string("op_1777"), val = tensor([1, -1, 8, 128])]; + tensor v_13_cast_fp16 = reshape(shape = var_1777, x = linear_59_cast_fp16)[name = string("v_13_cast_fp16")]; + tensor value_21_perm_0 = const()[name = string("value_21_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_6_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_6_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141299648)))]; + tensor var_1790_cast_fp16 = add(x = q_37_cast_fp16, y = encoder_layers_6_self_attn_pos_bias_u_to_fp16)[name = string("op_1790_cast_fp16")]; + tensor encoder_layers_6_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_6_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141301760)))]; + tensor var_1792_cast_fp16 = add(x = q_37_cast_fp16, y = encoder_layers_6_self_attn_pos_bias_v_to_fp16)[name = string("op_1792_cast_fp16")]; + tensor q_with_bias_v_13_perm_0 = const()[name = string("q_with_bias_v_13_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_163_transpose_x_0 = const()[name = string("x_163_transpose_x_0"), val = bool(false)]; + bool x_163_transpose_y_0 = const()[name = string("x_163_transpose_y_0"), val = bool(false)]; + tensor op_1794_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141303872))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141446272))))[name = string("op_1794_to_fp16_quantized")]; + tensor q_with_bias_v_13_cast_fp16 = transpose(perm = q_with_bias_v_13_perm_0, x = var_1792_cast_fp16)[name = string("transpose_308")]; + tensor x_163_cast_fp16 = matmul(transpose_x = x_163_transpose_x_0, transpose_y = x_163_transpose_y_0, x = q_with_bias_v_13_cast_fp16, y = op_1794_to_fp16_quantized)[name = string("x_163_cast_fp16")]; + tensor x_165_pad_0 = const()[name = string("x_165_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_165_mode_0 = const()[name = string("x_165_mode_0"), val = string("constant")]; + fp16 const_157_to_fp16 = const()[name = string("const_157_to_fp16"), val = fp16(0x0p+0)]; + tensor x_165_cast_fp16 = pad(constant_val = const_157_to_fp16, mode = x_165_mode_0, pad = x_165_pad_0, x = x_163_cast_fp16)[name = string("x_165_cast_fp16")]; + tensor var_1802 = const()[name = string("op_1802"), val = tensor([1, 8, -1, 28])]; + tensor x_167_cast_fp16 = reshape(shape = var_1802, x = x_165_cast_fp16)[name = string("x_167_cast_fp16")]; + tensor var_1806_begin_0 = const()[name = string("op_1806_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_1806_end_0 = const()[name = string("op_1806_end_0"), val = tensor([1, 8, 140, 28])]; + tensor var_1806_end_mask_0 = const()[name = string("op_1806_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_1806_cast_fp16 = slice_by_index(begin = var_1806_begin_0, end = var_1806_end_0, end_mask = var_1806_end_mask_0, x = x_167_cast_fp16)[name = string("op_1806_cast_fp16")]; + tensor var_1807 = const()[name = string("op_1807"), val = tensor([1, 8, 28, 139])]; + tensor matrix_bd_25_cast_fp16 = reshape(shape = var_1807, x = var_1806_cast_fp16)[name = string("matrix_bd_25_cast_fp16")]; + bool matrix_ac_13_transpose_x_0 = const()[name = string("matrix_ac_13_transpose_x_0"), val = bool(false)]; + bool matrix_ac_13_transpose_y_0 = const()[name = string("matrix_ac_13_transpose_y_0"), val = bool(false)]; + tensor transpose_108_perm_0 = const()[name = string("transpose_108_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_109_perm_0 = const()[name = string("transpose_109_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_109 = transpose(perm = transpose_109_perm_0, x = k_25_cast_fp16)[name = string("transpose_306")]; + tensor transpose_108 = transpose(perm = transpose_108_perm_0, x = var_1790_cast_fp16)[name = string("transpose_307")]; + tensor matrix_ac_13_cast_fp16 = matmul(transpose_x = matrix_ac_13_transpose_x_0, transpose_y = matrix_ac_13_transpose_y_0, x = transpose_108, y = transpose_109)[name = string("matrix_ac_13_cast_fp16")]; + tensor matrix_bd_27_begin_0 = const()[name = string("matrix_bd_27_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_27_end_0 = const()[name = string("matrix_bd_27_end_0"), val = tensor([1, 8, 28, 70])]; + tensor matrix_bd_27_end_mask_0 = const()[name = string("matrix_bd_27_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_27_cast_fp16 = slice_by_index(begin = matrix_bd_27_begin_0, end = matrix_bd_27_end_0, end_mask = matrix_bd_27_end_mask_0, x = matrix_bd_25_cast_fp16)[name = string("matrix_bd_27_cast_fp16")]; + tensor var_1816_cast_fp16 = add(x = matrix_ac_13_cast_fp16, y = matrix_bd_27_cast_fp16)[name = string("op_1816_cast_fp16")]; + fp16 _inversed_scores_25_y_0_to_fp16 = const()[name = string("_inversed_scores_25_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_25_cast_fp16 = mul(x = var_1816_cast_fp16, y = _inversed_scores_25_y_0_to_fp16)[name = string("_inversed_scores_25_cast_fp16")]; + tensor scores_27_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_25_cast_fp16, cond = mask_11)[name = string("scores_27_cast_fp16")]; + tensor var_1822_cast_fp16 = softmax(axis = var_59, x = scores_27_cast_fp16)[name = string("op_1822_cast_fp16")]; + tensor input_353_cast_fp16 = select(a = var_44_to_fp16, b = var_1822_cast_fp16, cond = mask_11)[name = string("input_353_cast_fp16")]; + bool x_169_transpose_x_0 = const()[name = string("x_169_transpose_x_0"), val = bool(false)]; + bool x_169_transpose_y_0 = const()[name = string("x_169_transpose_y_0"), val = bool(false)]; + tensor value_21_cast_fp16 = transpose(perm = value_21_perm_0, x = v_13_cast_fp16)[name = string("transpose_305")]; + tensor x_169_cast_fp16 = matmul(transpose_x = x_169_transpose_x_0, transpose_y = x_169_transpose_y_0, x = input_353_cast_fp16, y = value_21_cast_fp16)[name = string("x_169_cast_fp16")]; + tensor var_1826_perm_0 = const()[name = string("op_1826_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1827 = const()[name = string("op_1827"), val = tensor([1, -1, 1024])]; + tensor var_1826_cast_fp16 = transpose(perm = var_1826_perm_0, x = x_169_cast_fp16)[name = string("transpose_304")]; + tensor input_355_cast_fp16 = reshape(shape = var_1827, x = var_1826_cast_fp16)[name = string("input_355_cast_fp16")]; + tensor encoder_layers_6_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141446656))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142233152))))[name = string("encoder_layers_6_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_6_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_6_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142233344)))]; + tensor linear_61_cast_fp16 = linear(bias = encoder_layers_6_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_6_self_attn_linear_out_weight_to_fp16_palettized, x = input_355_cast_fp16)[name = string("linear_61_cast_fp16")]; + tensor input_359_cast_fp16 = add(x = input_349_cast_fp16, y = linear_61_cast_fp16)[name = string("input_359_cast_fp16")]; + tensor x_173_axes_0 = const()[name = string("x_173_axes_0"), val = tensor([-1])]; + tensor encoder_layers_6_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_6_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142235456)))]; + tensor encoder_layers_6_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_6_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142237568)))]; + tensor x_173_cast_fp16 = layer_norm(axes = x_173_axes_0, beta = encoder_layers_6_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_6_norm_conv_weight_to_fp16, x = input_359_cast_fp16)[name = string("x_173_cast_fp16")]; + tensor input_361_perm_0 = const()[name = string("input_361_perm_0"), val = tensor([0, 2, 1])]; + string input_363_pad_type_0 = const()[name = string("input_363_pad_type_0"), val = string("valid")]; + tensor input_363_strides_0 = const()[name = string("input_363_strides_0"), val = tensor([1])]; + tensor input_363_pad_0 = const()[name = string("input_363_pad_0"), val = tensor([0, 0])]; + tensor input_363_dilations_0 = const()[name = string("input_363_dilations_0"), val = tensor([1])]; + int32 input_363_groups_0 = const()[name = string("input_363_groups_0"), val = int32(1)]; + tensor encoder_layers_6_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142239680))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144336896))))[name = string("encoder_layers_6_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_361_cast_fp16 = transpose(perm = input_361_perm_0, x = x_173_cast_fp16)[name = string("transpose_303")]; + tensor input_363_cast_fp16 = conv(dilations = input_363_dilations_0, groups = input_363_groups_0, pad = input_363_pad_0, pad_type = input_363_pad_type_0, strides = input_363_strides_0, weight = encoder_layers_6_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_361_cast_fp16)[name = string("input_363_cast_fp16")]; + int32 x_175_split_num_splits_0 = const()[name = string("x_175_split_num_splits_0"), val = int32(2)]; + int32 x_175_split_axis_0 = const()[name = string("x_175_split_axis_0"), val = int32(1)]; + tensor x_175_split_cast_fp16_0, tensor x_175_split_cast_fp16_1 = split(axis = x_175_split_axis_0, num_splits = x_175_split_num_splits_0, x = input_363_cast_fp16)[name = string("x_175_split_cast_fp16")]; + tensor x_175_split_1_sigmoid_cast_fp16 = sigmoid(x = x_175_split_cast_fp16_1)[name = string("x_175_split_1_sigmoid_cast_fp16")]; + tensor x_175_cast_fp16 = mul(x = x_175_split_cast_fp16_0, y = x_175_split_1_sigmoid_cast_fp16)[name = string("x_175_cast_fp16")]; + tensor input_365_cast_fp16 = select(a = var_44_to_fp16, b = x_175_cast_fp16, cond = var_575)[name = string("input_365_cast_fp16")]; + bool new_x_27_interleave_0 = const()[name = string("new_x_27_interleave_0"), val = bool(false)]; + tensor new_x_27_cast_fp16 = concat(axis = var_59, interleave = new_x_27_interleave_0, values = (cache_27_cast_fp16, input_365_cast_fp16))[name = string("new_x_27_cast_fp16")]; + tensor var_1866_begin_0 = const()[name = string("op_1866_begin_0"), val = tensor([0, 0, 28])]; + tensor var_1866_end_0 = const()[name = string("op_1866_end_0"), val = tensor([1, 1024, 36])]; + tensor var_1866_end_mask_0 = const()[name = string("op_1866_end_mask_0"), val = tensor([true, true, true])]; + tensor var_1866_cast_fp16 = slice_by_index(begin = var_1866_begin_0, end = var_1866_end_0, end_mask = var_1866_end_mask_0, x = new_x_27_cast_fp16)[name = string("op_1866_cast_fp16")]; + string x_177_pad_type_0 = const()[name = string("x_177_pad_type_0"), val = string("valid")]; + int32 x_177_groups_0 = const()[name = string("x_177_groups_0"), val = int32(1024)]; + tensor x_177_strides_0 = const()[name = string("x_177_strides_0"), val = tensor([1])]; + tensor x_177_pad_0 = const()[name = string("x_177_pad_0"), val = tensor([0, 0])]; + tensor x_177_dilations_0 = const()[name = string("x_177_dilations_0"), val = tensor([1])]; + tensor encoder_layers_6_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144341056))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144350336))))[name = string("encoder_layers_6_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_177_cast_fp16 = conv(dilations = x_177_dilations_0, groups = x_177_groups_0, pad = x_177_pad_0, pad_type = x_177_pad_type_0, strides = x_177_strides_0, weight = encoder_layers_6_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_27_cast_fp16)[name = string("x_177_cast_fp16")]; + tensor input_367_perm_0 = const()[name = string("input_367_perm_0"), val = tensor([0, 2, 1])]; + tensor x_179_axes_0 = const()[name = string("x_179_axes_0"), val = tensor([-1])]; + tensor encoder_layers_6_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_6_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144352448)))]; + tensor encoder_layers_6_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_6_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144354560)))]; + tensor input_367_cast_fp16 = transpose(perm = input_367_perm_0, x = x_177_cast_fp16)[name = string("transpose_302")]; + tensor x_179_cast_fp16 = layer_norm(axes = x_179_axes_0, beta = encoder_layers_6_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_6_conv_batch_norm_weight_to_fp16, x = input_367_cast_fp16)[name = string("x_179_cast_fp16")]; + tensor input_369_perm_0 = const()[name = string("input_369_perm_0"), val = tensor([0, 2, 1])]; + tensor input_369_cast_fp16 = transpose(perm = input_369_perm_0, x = x_179_cast_fp16)[name = string("transpose_301")]; + tensor input_371_cast_fp16 = silu(x = input_369_cast_fp16)[name = string("input_371_cast_fp16")]; + string x_181_pad_type_0 = const()[name = string("x_181_pad_type_0"), val = string("valid")]; + tensor x_181_strides_0 = const()[name = string("x_181_strides_0"), val = tensor([1])]; + tensor x_181_pad_0 = const()[name = string("x_181_pad_0"), val = tensor([0, 0])]; + tensor x_181_dilations_0 = const()[name = string("x_181_dilations_0"), val = tensor([1])]; + int32 x_181_groups_0 = const()[name = string("x_181_groups_0"), val = int32(1)]; + tensor encoder_layers_6_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144356672))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(145405312))))[name = string("encoder_layers_6_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_181_cast_fp16 = conv(dilations = x_181_dilations_0, groups = x_181_groups_0, pad = x_181_pad_0, pad_type = x_181_pad_type_0, strides = x_181_strides_0, weight = encoder_layers_6_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_371_cast_fp16)[name = string("x_181_cast_fp16")]; + tensor input_373_perm_0 = const()[name = string("input_373_perm_0"), val = tensor([0, 2, 1])]; + tensor input_373_cast_fp16 = transpose(perm = input_373_perm_0, x = x_181_cast_fp16)[name = string("transpose_300")]; + tensor input_375_cast_fp16 = add(x = input_359_cast_fp16, y = input_373_cast_fp16)[name = string("input_375_cast_fp16")]; + tensor input_377_axes_0 = const()[name = string("input_377_axes_0"), val = tensor([-1])]; + tensor encoder_layers_6_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_6_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(145407424)))]; + tensor encoder_layers_6_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_6_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(145409536)))]; + tensor input_377_cast_fp16 = layer_norm(axes = input_377_axes_0, beta = encoder_layers_6_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_6_norm_feed_forward2_weight_to_fp16, x = input_375_cast_fp16)[name = string("input_377_cast_fp16")]; + tensor encoder_layers_6_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(145411648))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(148557440))))[name = string("encoder_layers_6_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_6_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_6_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(148557632)))]; + tensor linear_62_cast_fp16 = linear(bias = encoder_layers_6_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_6_feed_forward2_linear1_weight_to_fp16_palettized, x = input_377_cast_fp16)[name = string("linear_62_cast_fp16")]; + tensor input_381_cast_fp16 = silu(x = linear_62_cast_fp16)[name = string("input_381_cast_fp16")]; + tensor encoder_layers_6_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(148565888))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151711680))))[name = string("encoder_layers_6_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_6_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_6_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151711872)))]; + tensor linear_63_cast_fp16 = linear(bias = encoder_layers_6_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_6_feed_forward2_linear2_weight_to_fp16_palettized, x = input_381_cast_fp16)[name = string("linear_63_cast_fp16")]; + fp16 var_1909_to_fp16 = const()[name = string("op_1909_to_fp16"), val = fp16(0x1p-1)]; + tensor var_1910_cast_fp16 = mul(x = linear_63_cast_fp16, y = var_1909_to_fp16)[name = string("op_1910_cast_fp16")]; + tensor input_387_cast_fp16 = add(x = input_375_cast_fp16, y = var_1910_cast_fp16)[name = string("input_387_cast_fp16")]; + tensor input_389_axes_0 = const()[name = string("input_389_axes_0"), val = tensor([-1])]; + tensor encoder_layers_6_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_6_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151713984)))]; + tensor encoder_layers_6_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_6_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151716096)))]; + tensor input_389_cast_fp16 = layer_norm(axes = input_389_axes_0, beta = encoder_layers_6_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_6_norm_out_weight_to_fp16, x = input_387_cast_fp16)[name = string("input_389_cast_fp16")]; + tensor cache_29_begin_0 = const()[name = string("cache_29_begin_0"), val = tensor([7, 0, 0, 0])]; + tensor cache_29_end_0 = const()[name = string("cache_29_end_0"), val = tensor([8, 1, 42, 1024])]; + tensor cache_29_end_mask_0 = const()[name = string("cache_29_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_29_squeeze_mask_0 = const()[name = string("cache_29_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_29_cast_fp16 = slice_by_index(begin = cache_29_begin_0, end = cache_29_end_0, end_mask = cache_29_end_mask_0, squeeze_mask = cache_29_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_29_cast_fp16")]; + tensor cache_31_begin_0 = const()[name = string("cache_31_begin_0"), val = tensor([7, 0, 0, 0])]; + tensor cache_31_end_0 = const()[name = string("cache_31_end_0"), val = tensor([8, 1, 1024, 8])]; + tensor cache_31_end_mask_0 = const()[name = string("cache_31_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_31_squeeze_mask_0 = const()[name = string("cache_31_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_31_cast_fp16 = slice_by_index(begin = cache_31_begin_0, end = cache_31_end_0, end_mask = cache_31_end_mask_0, squeeze_mask = cache_31_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_31_cast_fp16")]; + tensor input_391_axes_0 = const()[name = string("input_391_axes_0"), val = tensor([-1])]; + tensor encoder_layers_7_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_7_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151718208)))]; + tensor encoder_layers_7_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_7_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151720320)))]; + tensor input_391_cast_fp16 = layer_norm(axes = input_391_axes_0, beta = encoder_layers_7_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_7_norm_feed_forward1_weight_to_fp16, x = input_389_cast_fp16)[name = string("input_391_cast_fp16")]; + tensor encoder_layers_7_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151722432))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(154868224))))[name = string("encoder_layers_7_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_7_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_7_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(154868416)))]; + tensor linear_64_cast_fp16 = linear(bias = encoder_layers_7_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_7_feed_forward1_linear1_weight_to_fp16_palettized, x = input_391_cast_fp16)[name = string("linear_64_cast_fp16")]; + tensor input_395_cast_fp16 = silu(x = linear_64_cast_fp16)[name = string("input_395_cast_fp16")]; + tensor encoder_layers_7_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(154876672))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158022464))))[name = string("encoder_layers_7_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_7_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_7_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158022656)))]; + tensor linear_65_cast_fp16 = linear(bias = encoder_layers_7_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_7_feed_forward1_linear2_weight_to_fp16_palettized, x = input_395_cast_fp16)[name = string("linear_65_cast_fp16")]; + fp16 var_1946_to_fp16 = const()[name = string("op_1946_to_fp16"), val = fp16(0x1p-1)]; + tensor var_1947_cast_fp16 = mul(x = linear_65_cast_fp16, y = var_1946_to_fp16)[name = string("op_1947_cast_fp16")]; + tensor input_401_cast_fp16 = add(x = input_389_cast_fp16, y = var_1947_cast_fp16)[name = string("input_401_cast_fp16")]; + tensor key_15_axes_0 = const()[name = string("key_15_axes_0"), val = tensor([-1])]; + tensor encoder_layers_7_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_7_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158024768)))]; + tensor encoder_layers_7_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_7_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158026880)))]; + tensor key_15_cast_fp16 = layer_norm(axes = key_15_axes_0, beta = encoder_layers_7_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_7_norm_self_att_weight_to_fp16, x = input_401_cast_fp16)[name = string("key_15_cast_fp16")]; + bool input_403_interleave_0 = const()[name = string("input_403_interleave_0"), val = bool(false)]; + tensor input_403_cast_fp16 = concat(axis = var_68, interleave = input_403_interleave_0, values = (cache_29_cast_fp16, key_15_cast_fp16))[name = string("input_403_cast_fp16")]; + tensor var_1969_begin_0 = const()[name = string("op_1969_begin_0"), val = tensor([0, 28, 0])]; + tensor var_1969_end_0 = const()[name = string("op_1969_end_0"), val = tensor([1, 42, 1024])]; + tensor var_1969_end_mask_0 = const()[name = string("op_1969_end_mask_0"), val = tensor([true, true, true])]; + tensor var_1969_cast_fp16 = slice_by_index(begin = var_1969_begin_0, end = var_1969_end_0, end_mask = var_1969_end_mask_0, x = cache_29_cast_fp16)[name = string("op_1969_cast_fp16")]; + bool var_1975_interleave_0 = const()[name = string("op_1975_interleave_0"), val = bool(false)]; + tensor var_1975_cast_fp16 = concat(axis = var_68, interleave = var_1975_interleave_0, values = (var_1969_cast_fp16, key_15_cast_fp16))[name = string("op_1975_cast_fp16")]; + tensor encoder_layers_7_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158028992))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158815488))))[name = string("encoder_layers_7_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_7_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_7_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158815680)))]; + tensor linear_66_cast_fp16 = linear(bias = encoder_layers_7_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_7_self_attn_linear_q_weight_to_fp16_palettized, x = key_15_cast_fp16)[name = string("linear_66_cast_fp16")]; + tensor var_1980 = const()[name = string("op_1980"), val = tensor([1, -1, 8, 128])]; + tensor q_43_cast_fp16 = reshape(shape = var_1980, x = linear_66_cast_fp16)[name = string("q_43_cast_fp16")]; + tensor encoder_layers_7_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158817792))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(159604288))))[name = string("encoder_layers_7_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_7_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_7_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(159604480)))]; + tensor linear_67_cast_fp16 = linear(bias = encoder_layers_7_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_7_self_attn_linear_k_weight_to_fp16_palettized, x = input_403_cast_fp16)[name = string("linear_67_cast_fp16")]; + tensor var_1985 = const()[name = string("op_1985"), val = tensor([1, -1, 8, 128])]; + tensor k_29_cast_fp16 = reshape(shape = var_1985, x = linear_67_cast_fp16)[name = string("k_29_cast_fp16")]; + tensor encoder_layers_7_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(159606592))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160393088))))[name = string("encoder_layers_7_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_7_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_7_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160393280)))]; + tensor linear_68_cast_fp16 = linear(bias = encoder_layers_7_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_7_self_attn_linear_v_weight_to_fp16_palettized, x = input_403_cast_fp16)[name = string("linear_68_cast_fp16")]; + tensor var_1990 = const()[name = string("op_1990"), val = tensor([1, -1, 8, 128])]; + tensor v_15_cast_fp16 = reshape(shape = var_1990, x = linear_68_cast_fp16)[name = string("v_15_cast_fp16")]; + tensor value_23_perm_0 = const()[name = string("value_23_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_7_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_7_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160395392)))]; + tensor var_2003_cast_fp16 = add(x = q_43_cast_fp16, y = encoder_layers_7_self_attn_pos_bias_u_to_fp16)[name = string("op_2003_cast_fp16")]; + tensor encoder_layers_7_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_7_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160397504)))]; + tensor var_2005_cast_fp16 = add(x = q_43_cast_fp16, y = encoder_layers_7_self_attn_pos_bias_v_to_fp16)[name = string("op_2005_cast_fp16")]; + tensor q_with_bias_v_15_perm_0 = const()[name = string("q_with_bias_v_15_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_189_transpose_x_0 = const()[name = string("x_189_transpose_x_0"), val = bool(false)]; + bool x_189_transpose_y_0 = const()[name = string("x_189_transpose_y_0"), val = bool(false)]; + tensor op_2007_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160399616))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160542016))))[name = string("op_2007_to_fp16_quantized")]; + tensor q_with_bias_v_15_cast_fp16 = transpose(perm = q_with_bias_v_15_perm_0, x = var_2005_cast_fp16)[name = string("transpose_299")]; + tensor x_189_cast_fp16 = matmul(transpose_x = x_189_transpose_x_0, transpose_y = x_189_transpose_y_0, x = q_with_bias_v_15_cast_fp16, y = op_2007_to_fp16_quantized)[name = string("x_189_cast_fp16")]; + tensor x_191_pad_0 = const()[name = string("x_191_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_191_mode_0 = const()[name = string("x_191_mode_0"), val = string("constant")]; + fp16 const_170_to_fp16 = const()[name = string("const_170_to_fp16"), val = fp16(0x0p+0)]; + tensor x_191_cast_fp16 = pad(constant_val = const_170_to_fp16, mode = x_191_mode_0, pad = x_191_pad_0, x = x_189_cast_fp16)[name = string("x_191_cast_fp16")]; + tensor var_2015 = const()[name = string("op_2015"), val = tensor([1, 8, -1, 28])]; + tensor x_193_cast_fp16 = reshape(shape = var_2015, x = x_191_cast_fp16)[name = string("x_193_cast_fp16")]; + tensor var_2019_begin_0 = const()[name = string("op_2019_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_2019_end_0 = const()[name = string("op_2019_end_0"), val = tensor([1, 8, 140, 28])]; + tensor var_2019_end_mask_0 = const()[name = string("op_2019_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_2019_cast_fp16 = slice_by_index(begin = var_2019_begin_0, end = var_2019_end_0, end_mask = var_2019_end_mask_0, x = x_193_cast_fp16)[name = string("op_2019_cast_fp16")]; + tensor var_2020 = const()[name = string("op_2020"), val = tensor([1, 8, 28, 139])]; + tensor matrix_bd_29_cast_fp16 = reshape(shape = var_2020, x = var_2019_cast_fp16)[name = string("matrix_bd_29_cast_fp16")]; + bool matrix_ac_15_transpose_x_0 = const()[name = string("matrix_ac_15_transpose_x_0"), val = bool(false)]; + bool matrix_ac_15_transpose_y_0 = const()[name = string("matrix_ac_15_transpose_y_0"), val = bool(false)]; + tensor transpose_110_perm_0 = const()[name = string("transpose_110_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_111_perm_0 = const()[name = string("transpose_111_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_111 = transpose(perm = transpose_111_perm_0, x = k_29_cast_fp16)[name = string("transpose_297")]; + tensor transpose_110 = transpose(perm = transpose_110_perm_0, x = var_2003_cast_fp16)[name = string("transpose_298")]; + tensor matrix_ac_15_cast_fp16 = matmul(transpose_x = matrix_ac_15_transpose_x_0, transpose_y = matrix_ac_15_transpose_y_0, x = transpose_110, y = transpose_111)[name = string("matrix_ac_15_cast_fp16")]; + tensor matrix_bd_31_begin_0 = const()[name = string("matrix_bd_31_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_31_end_0 = const()[name = string("matrix_bd_31_end_0"), val = tensor([1, 8, 28, 70])]; + tensor matrix_bd_31_end_mask_0 = const()[name = string("matrix_bd_31_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_31_cast_fp16 = slice_by_index(begin = matrix_bd_31_begin_0, end = matrix_bd_31_end_0, end_mask = matrix_bd_31_end_mask_0, x = matrix_bd_29_cast_fp16)[name = string("matrix_bd_31_cast_fp16")]; + tensor var_2029_cast_fp16 = add(x = matrix_ac_15_cast_fp16, y = matrix_bd_31_cast_fp16)[name = string("op_2029_cast_fp16")]; + fp16 _inversed_scores_29_y_0_to_fp16 = const()[name = string("_inversed_scores_29_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_29_cast_fp16 = mul(x = var_2029_cast_fp16, y = _inversed_scores_29_y_0_to_fp16)[name = string("_inversed_scores_29_cast_fp16")]; + tensor scores_31_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_29_cast_fp16, cond = mask_11)[name = string("scores_31_cast_fp16")]; + tensor var_2035_cast_fp16 = softmax(axis = var_59, x = scores_31_cast_fp16)[name = string("op_2035_cast_fp16")]; + tensor input_405_cast_fp16 = select(a = var_44_to_fp16, b = var_2035_cast_fp16, cond = mask_11)[name = string("input_405_cast_fp16")]; + bool x_195_transpose_x_0 = const()[name = string("x_195_transpose_x_0"), val = bool(false)]; + bool x_195_transpose_y_0 = const()[name = string("x_195_transpose_y_0"), val = bool(false)]; + tensor value_23_cast_fp16 = transpose(perm = value_23_perm_0, x = v_15_cast_fp16)[name = string("transpose_296")]; + tensor x_195_cast_fp16 = matmul(transpose_x = x_195_transpose_x_0, transpose_y = x_195_transpose_y_0, x = input_405_cast_fp16, y = value_23_cast_fp16)[name = string("x_195_cast_fp16")]; + tensor var_2039_perm_0 = const()[name = string("op_2039_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_2040 = const()[name = string("op_2040"), val = tensor([1, -1, 1024])]; + tensor var_2039_cast_fp16 = transpose(perm = var_2039_perm_0, x = x_195_cast_fp16)[name = string("transpose_295")]; + tensor input_407_cast_fp16 = reshape(shape = var_2040, x = var_2039_cast_fp16)[name = string("input_407_cast_fp16")]; + tensor encoder_layers_7_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160542400))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161328896))))[name = string("encoder_layers_7_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_7_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_7_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161329088)))]; + tensor linear_70_cast_fp16 = linear(bias = encoder_layers_7_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_7_self_attn_linear_out_weight_to_fp16_palettized, x = input_407_cast_fp16)[name = string("linear_70_cast_fp16")]; + tensor input_411_cast_fp16 = add(x = input_401_cast_fp16, y = linear_70_cast_fp16)[name = string("input_411_cast_fp16")]; + tensor x_199_axes_0 = const()[name = string("x_199_axes_0"), val = tensor([-1])]; + tensor encoder_layers_7_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_7_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161331200)))]; + tensor encoder_layers_7_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_7_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161333312)))]; + tensor x_199_cast_fp16 = layer_norm(axes = x_199_axes_0, beta = encoder_layers_7_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_7_norm_conv_weight_to_fp16, x = input_411_cast_fp16)[name = string("x_199_cast_fp16")]; + tensor input_413_perm_0 = const()[name = string("input_413_perm_0"), val = tensor([0, 2, 1])]; + string input_415_pad_type_0 = const()[name = string("input_415_pad_type_0"), val = string("valid")]; + tensor input_415_strides_0 = const()[name = string("input_415_strides_0"), val = tensor([1])]; + tensor input_415_pad_0 = const()[name = string("input_415_pad_0"), val = tensor([0, 0])]; + tensor input_415_dilations_0 = const()[name = string("input_415_dilations_0"), val = tensor([1])]; + int32 input_415_groups_0 = const()[name = string("input_415_groups_0"), val = int32(1)]; + tensor encoder_layers_7_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161335424))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163432640))))[name = string("encoder_layers_7_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_413_cast_fp16 = transpose(perm = input_413_perm_0, x = x_199_cast_fp16)[name = string("transpose_294")]; + tensor input_415_cast_fp16 = conv(dilations = input_415_dilations_0, groups = input_415_groups_0, pad = input_415_pad_0, pad_type = input_415_pad_type_0, strides = input_415_strides_0, weight = encoder_layers_7_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_413_cast_fp16)[name = string("input_415_cast_fp16")]; + int32 x_201_split_num_splits_0 = const()[name = string("x_201_split_num_splits_0"), val = int32(2)]; + int32 x_201_split_axis_0 = const()[name = string("x_201_split_axis_0"), val = int32(1)]; + tensor x_201_split_cast_fp16_0, tensor x_201_split_cast_fp16_1 = split(axis = x_201_split_axis_0, num_splits = x_201_split_num_splits_0, x = input_415_cast_fp16)[name = string("x_201_split_cast_fp16")]; + tensor x_201_split_1_sigmoid_cast_fp16 = sigmoid(x = x_201_split_cast_fp16_1)[name = string("x_201_split_1_sigmoid_cast_fp16")]; + tensor x_201_cast_fp16 = mul(x = x_201_split_cast_fp16_0, y = x_201_split_1_sigmoid_cast_fp16)[name = string("x_201_cast_fp16")]; + tensor input_417_cast_fp16 = select(a = var_44_to_fp16, b = x_201_cast_fp16, cond = var_575)[name = string("input_417_cast_fp16")]; + bool new_x_31_interleave_0 = const()[name = string("new_x_31_interleave_0"), val = bool(false)]; + tensor new_x_31_cast_fp16 = concat(axis = var_59, interleave = new_x_31_interleave_0, values = (cache_31_cast_fp16, input_417_cast_fp16))[name = string("new_x_31_cast_fp16")]; + tensor var_2079_begin_0 = const()[name = string("op_2079_begin_0"), val = tensor([0, 0, 28])]; + tensor var_2079_end_0 = const()[name = string("op_2079_end_0"), val = tensor([1, 1024, 36])]; + tensor var_2079_end_mask_0 = const()[name = string("op_2079_end_mask_0"), val = tensor([true, true, true])]; + tensor var_2079_cast_fp16 = slice_by_index(begin = var_2079_begin_0, end = var_2079_end_0, end_mask = var_2079_end_mask_0, x = new_x_31_cast_fp16)[name = string("op_2079_cast_fp16")]; + string x_203_pad_type_0 = const()[name = string("x_203_pad_type_0"), val = string("valid")]; + int32 x_203_groups_0 = const()[name = string("x_203_groups_0"), val = int32(1024)]; + tensor x_203_strides_0 = const()[name = string("x_203_strides_0"), val = tensor([1])]; + tensor x_203_pad_0 = const()[name = string("x_203_pad_0"), val = tensor([0, 0])]; + tensor x_203_dilations_0 = const()[name = string("x_203_dilations_0"), val = tensor([1])]; + tensor encoder_layers_7_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163436800))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163446080))))[name = string("encoder_layers_7_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_203_cast_fp16 = conv(dilations = x_203_dilations_0, groups = x_203_groups_0, pad = x_203_pad_0, pad_type = x_203_pad_type_0, strides = x_203_strides_0, weight = encoder_layers_7_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_31_cast_fp16)[name = string("x_203_cast_fp16")]; + tensor input_419_perm_0 = const()[name = string("input_419_perm_0"), val = tensor([0, 2, 1])]; + tensor x_205_axes_0 = const()[name = string("x_205_axes_0"), val = tensor([-1])]; + tensor encoder_layers_7_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_7_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163448192)))]; + tensor encoder_layers_7_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_7_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163450304)))]; + tensor input_419_cast_fp16 = transpose(perm = input_419_perm_0, x = x_203_cast_fp16)[name = string("transpose_293")]; + tensor x_205_cast_fp16 = layer_norm(axes = x_205_axes_0, beta = encoder_layers_7_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_7_conv_batch_norm_weight_to_fp16, x = input_419_cast_fp16)[name = string("x_205_cast_fp16")]; + tensor input_421_perm_0 = const()[name = string("input_421_perm_0"), val = tensor([0, 2, 1])]; + tensor input_421_cast_fp16 = transpose(perm = input_421_perm_0, x = x_205_cast_fp16)[name = string("transpose_292")]; + tensor input_423_cast_fp16 = silu(x = input_421_cast_fp16)[name = string("input_423_cast_fp16")]; + string x_207_pad_type_0 = const()[name = string("x_207_pad_type_0"), val = string("valid")]; + tensor x_207_strides_0 = const()[name = string("x_207_strides_0"), val = tensor([1])]; + tensor x_207_pad_0 = const()[name = string("x_207_pad_0"), val = tensor([0, 0])]; + tensor x_207_dilations_0 = const()[name = string("x_207_dilations_0"), val = tensor([1])]; + int32 x_207_groups_0 = const()[name = string("x_207_groups_0"), val = int32(1)]; + tensor encoder_layers_7_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163452416))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(164501056))))[name = string("encoder_layers_7_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_207_cast_fp16 = conv(dilations = x_207_dilations_0, groups = x_207_groups_0, pad = x_207_pad_0, pad_type = x_207_pad_type_0, strides = x_207_strides_0, weight = encoder_layers_7_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_423_cast_fp16)[name = string("x_207_cast_fp16")]; + tensor input_425_perm_0 = const()[name = string("input_425_perm_0"), val = tensor([0, 2, 1])]; + tensor input_425_cast_fp16 = transpose(perm = input_425_perm_0, x = x_207_cast_fp16)[name = string("transpose_291")]; + tensor input_427_cast_fp16 = add(x = input_411_cast_fp16, y = input_425_cast_fp16)[name = string("input_427_cast_fp16")]; + tensor input_429_axes_0 = const()[name = string("input_429_axes_0"), val = tensor([-1])]; + tensor encoder_layers_7_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_7_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(164503168)))]; + tensor encoder_layers_7_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_7_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(164505280)))]; + tensor input_429_cast_fp16 = layer_norm(axes = input_429_axes_0, beta = encoder_layers_7_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_7_norm_feed_forward2_weight_to_fp16, x = input_427_cast_fp16)[name = string("input_429_cast_fp16")]; + tensor encoder_layers_7_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(164507392))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167653184))))[name = string("encoder_layers_7_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_7_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_7_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167653376)))]; + tensor linear_71_cast_fp16 = linear(bias = encoder_layers_7_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_7_feed_forward2_linear1_weight_to_fp16_palettized, x = input_429_cast_fp16)[name = string("linear_71_cast_fp16")]; + tensor input_433_cast_fp16 = silu(x = linear_71_cast_fp16)[name = string("input_433_cast_fp16")]; + tensor encoder_layers_7_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167661632))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170807424))))[name = string("encoder_layers_7_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_7_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_7_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170807616)))]; + tensor linear_72_cast_fp16 = linear(bias = encoder_layers_7_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_7_feed_forward2_linear2_weight_to_fp16_palettized, x = input_433_cast_fp16)[name = string("linear_72_cast_fp16")]; + fp16 var_2122_to_fp16 = const()[name = string("op_2122_to_fp16"), val = fp16(0x1p-1)]; + tensor var_2123_cast_fp16 = mul(x = linear_72_cast_fp16, y = var_2122_to_fp16)[name = string("op_2123_cast_fp16")]; + tensor input_439_cast_fp16 = add(x = input_427_cast_fp16, y = var_2123_cast_fp16)[name = string("input_439_cast_fp16")]; + tensor input_441_axes_0 = const()[name = string("input_441_axes_0"), val = tensor([-1])]; + tensor encoder_layers_7_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_7_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170809728)))]; + tensor encoder_layers_7_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_7_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170811840)))]; + tensor input_441_cast_fp16 = layer_norm(axes = input_441_axes_0, beta = encoder_layers_7_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_7_norm_out_weight_to_fp16, x = input_439_cast_fp16)[name = string("input_441_cast_fp16")]; + tensor cache_33_begin_0 = const()[name = string("cache_33_begin_0"), val = tensor([8, 0, 0, 0])]; + tensor cache_33_end_0 = const()[name = string("cache_33_end_0"), val = tensor([9, 1, 42, 1024])]; + tensor cache_33_end_mask_0 = const()[name = string("cache_33_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_33_squeeze_mask_0 = const()[name = string("cache_33_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_33_cast_fp16 = slice_by_index(begin = cache_33_begin_0, end = cache_33_end_0, end_mask = cache_33_end_mask_0, squeeze_mask = cache_33_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_33_cast_fp16")]; + tensor cache_35_begin_0 = const()[name = string("cache_35_begin_0"), val = tensor([8, 0, 0, 0])]; + tensor cache_35_end_0 = const()[name = string("cache_35_end_0"), val = tensor([9, 1, 1024, 8])]; + tensor cache_35_end_mask_0 = const()[name = string("cache_35_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_35_squeeze_mask_0 = const()[name = string("cache_35_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_35_cast_fp16 = slice_by_index(begin = cache_35_begin_0, end = cache_35_end_0, end_mask = cache_35_end_mask_0, squeeze_mask = cache_35_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_35_cast_fp16")]; + tensor input_443_axes_0 = const()[name = string("input_443_axes_0"), val = tensor([-1])]; + tensor encoder_layers_8_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_8_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170813952)))]; + tensor encoder_layers_8_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_8_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170816064)))]; + tensor input_443_cast_fp16 = layer_norm(axes = input_443_axes_0, beta = encoder_layers_8_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_8_norm_feed_forward1_weight_to_fp16, x = input_441_cast_fp16)[name = string("input_443_cast_fp16")]; + tensor encoder_layers_8_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170818176))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(173963968))))[name = string("encoder_layers_8_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_8_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_8_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(173964160)))]; + tensor linear_73_cast_fp16 = linear(bias = encoder_layers_8_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_8_feed_forward1_linear1_weight_to_fp16_palettized, x = input_443_cast_fp16)[name = string("linear_73_cast_fp16")]; + tensor input_447_cast_fp16 = silu(x = linear_73_cast_fp16)[name = string("input_447_cast_fp16")]; + tensor encoder_layers_8_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(173972416))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(177118208))))[name = string("encoder_layers_8_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_8_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_8_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(177118400)))]; + tensor linear_74_cast_fp16 = linear(bias = encoder_layers_8_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_8_feed_forward1_linear2_weight_to_fp16_palettized, x = input_447_cast_fp16)[name = string("linear_74_cast_fp16")]; + fp16 var_2159_to_fp16 = const()[name = string("op_2159_to_fp16"), val = fp16(0x1p-1)]; + tensor var_2160_cast_fp16 = mul(x = linear_74_cast_fp16, y = var_2159_to_fp16)[name = string("op_2160_cast_fp16")]; + tensor input_453_cast_fp16 = add(x = input_441_cast_fp16, y = var_2160_cast_fp16)[name = string("input_453_cast_fp16")]; + tensor key_17_axes_0 = const()[name = string("key_17_axes_0"), val = tensor([-1])]; + tensor encoder_layers_8_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_8_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(177120512)))]; + tensor encoder_layers_8_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_8_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(177122624)))]; + tensor key_17_cast_fp16 = layer_norm(axes = key_17_axes_0, beta = encoder_layers_8_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_8_norm_self_att_weight_to_fp16, x = input_453_cast_fp16)[name = string("key_17_cast_fp16")]; + bool input_455_interleave_0 = const()[name = string("input_455_interleave_0"), val = bool(false)]; + tensor input_455_cast_fp16 = concat(axis = var_68, interleave = input_455_interleave_0, values = (cache_33_cast_fp16, key_17_cast_fp16))[name = string("input_455_cast_fp16")]; + tensor var_2182_begin_0 = const()[name = string("op_2182_begin_0"), val = tensor([0, 28, 0])]; + tensor var_2182_end_0 = const()[name = string("op_2182_end_0"), val = tensor([1, 42, 1024])]; + tensor var_2182_end_mask_0 = const()[name = string("op_2182_end_mask_0"), val = tensor([true, true, true])]; + tensor var_2182_cast_fp16 = slice_by_index(begin = var_2182_begin_0, end = var_2182_end_0, end_mask = var_2182_end_mask_0, x = cache_33_cast_fp16)[name = string("op_2182_cast_fp16")]; + bool var_2188_interleave_0 = const()[name = string("op_2188_interleave_0"), val = bool(false)]; + tensor var_2188_cast_fp16 = concat(axis = var_68, interleave = var_2188_interleave_0, values = (var_2182_cast_fp16, key_17_cast_fp16))[name = string("op_2188_cast_fp16")]; + tensor encoder_layers_8_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(177124736))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(177911232))))[name = string("encoder_layers_8_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_8_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_8_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(177911424)))]; + tensor linear_75_cast_fp16 = linear(bias = encoder_layers_8_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_8_self_attn_linear_q_weight_to_fp16_palettized, x = key_17_cast_fp16)[name = string("linear_75_cast_fp16")]; + tensor var_2193 = const()[name = string("op_2193"), val = tensor([1, -1, 8, 128])]; + tensor q_49_cast_fp16 = reshape(shape = var_2193, x = linear_75_cast_fp16)[name = string("q_49_cast_fp16")]; + tensor encoder_layers_8_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(177913536))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(178700032))))[name = string("encoder_layers_8_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_8_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_8_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(178700224)))]; + tensor linear_76_cast_fp16 = linear(bias = encoder_layers_8_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_8_self_attn_linear_k_weight_to_fp16_palettized, x = input_455_cast_fp16)[name = string("linear_76_cast_fp16")]; + tensor var_2198 = const()[name = string("op_2198"), val = tensor([1, -1, 8, 128])]; + tensor k_33_cast_fp16 = reshape(shape = var_2198, x = linear_76_cast_fp16)[name = string("k_33_cast_fp16")]; + tensor encoder_layers_8_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(178702336))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179488832))))[name = string("encoder_layers_8_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_8_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_8_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179489024)))]; + tensor linear_77_cast_fp16 = linear(bias = encoder_layers_8_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_8_self_attn_linear_v_weight_to_fp16_palettized, x = input_455_cast_fp16)[name = string("linear_77_cast_fp16")]; + tensor var_2203 = const()[name = string("op_2203"), val = tensor([1, -1, 8, 128])]; + tensor v_17_cast_fp16 = reshape(shape = var_2203, x = linear_77_cast_fp16)[name = string("v_17_cast_fp16")]; + tensor value_25_perm_0 = const()[name = string("value_25_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_8_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_8_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179491136)))]; + tensor var_2216_cast_fp16 = add(x = q_49_cast_fp16, y = encoder_layers_8_self_attn_pos_bias_u_to_fp16)[name = string("op_2216_cast_fp16")]; + tensor encoder_layers_8_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_8_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179493248)))]; + tensor var_2218_cast_fp16 = add(x = q_49_cast_fp16, y = encoder_layers_8_self_attn_pos_bias_v_to_fp16)[name = string("op_2218_cast_fp16")]; + tensor q_with_bias_v_17_perm_0 = const()[name = string("q_with_bias_v_17_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_215_transpose_x_0 = const()[name = string("x_215_transpose_x_0"), val = bool(false)]; + bool x_215_transpose_y_0 = const()[name = string("x_215_transpose_y_0"), val = bool(false)]; + tensor op_2220_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179495360))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179637760))))[name = string("op_2220_to_fp16_quantized")]; + tensor q_with_bias_v_17_cast_fp16 = transpose(perm = q_with_bias_v_17_perm_0, x = var_2218_cast_fp16)[name = string("transpose_290")]; + tensor x_215_cast_fp16 = matmul(transpose_x = x_215_transpose_x_0, transpose_y = x_215_transpose_y_0, x = q_with_bias_v_17_cast_fp16, y = op_2220_to_fp16_quantized)[name = string("x_215_cast_fp16")]; + tensor x_217_pad_0 = const()[name = string("x_217_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_217_mode_0 = const()[name = string("x_217_mode_0"), val = string("constant")]; + fp16 const_183_to_fp16 = const()[name = string("const_183_to_fp16"), val = fp16(0x0p+0)]; + tensor x_217_cast_fp16 = pad(constant_val = const_183_to_fp16, mode = x_217_mode_0, pad = x_217_pad_0, x = x_215_cast_fp16)[name = string("x_217_cast_fp16")]; + tensor var_2228 = const()[name = string("op_2228"), val = tensor([1, 8, -1, 28])]; + tensor x_219_cast_fp16 = reshape(shape = var_2228, x = x_217_cast_fp16)[name = string("x_219_cast_fp16")]; + tensor var_2232_begin_0 = const()[name = string("op_2232_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_2232_end_0 = const()[name = string("op_2232_end_0"), val = tensor([1, 8, 140, 28])]; + tensor var_2232_end_mask_0 = const()[name = string("op_2232_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_2232_cast_fp16 = slice_by_index(begin = var_2232_begin_0, end = var_2232_end_0, end_mask = var_2232_end_mask_0, x = x_219_cast_fp16)[name = string("op_2232_cast_fp16")]; + tensor var_2233 = const()[name = string("op_2233"), val = tensor([1, 8, 28, 139])]; + tensor matrix_bd_33_cast_fp16 = reshape(shape = var_2233, x = var_2232_cast_fp16)[name = string("matrix_bd_33_cast_fp16")]; + bool matrix_ac_17_transpose_x_0 = const()[name = string("matrix_ac_17_transpose_x_0"), val = bool(false)]; + bool matrix_ac_17_transpose_y_0 = const()[name = string("matrix_ac_17_transpose_y_0"), val = bool(false)]; + tensor transpose_112_perm_0 = const()[name = string("transpose_112_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_113_perm_0 = const()[name = string("transpose_113_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_113 = transpose(perm = transpose_113_perm_0, x = k_33_cast_fp16)[name = string("transpose_288")]; + tensor transpose_112 = transpose(perm = transpose_112_perm_0, x = var_2216_cast_fp16)[name = string("transpose_289")]; + tensor matrix_ac_17_cast_fp16 = matmul(transpose_x = matrix_ac_17_transpose_x_0, transpose_y = matrix_ac_17_transpose_y_0, x = transpose_112, y = transpose_113)[name = string("matrix_ac_17_cast_fp16")]; + tensor matrix_bd_35_begin_0 = const()[name = string("matrix_bd_35_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_35_end_0 = const()[name = string("matrix_bd_35_end_0"), val = tensor([1, 8, 28, 70])]; + tensor matrix_bd_35_end_mask_0 = const()[name = string("matrix_bd_35_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_35_cast_fp16 = slice_by_index(begin = matrix_bd_35_begin_0, end = matrix_bd_35_end_0, end_mask = matrix_bd_35_end_mask_0, x = matrix_bd_33_cast_fp16)[name = string("matrix_bd_35_cast_fp16")]; + tensor var_2242_cast_fp16 = add(x = matrix_ac_17_cast_fp16, y = matrix_bd_35_cast_fp16)[name = string("op_2242_cast_fp16")]; + fp16 _inversed_scores_33_y_0_to_fp16 = const()[name = string("_inversed_scores_33_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_33_cast_fp16 = mul(x = var_2242_cast_fp16, y = _inversed_scores_33_y_0_to_fp16)[name = string("_inversed_scores_33_cast_fp16")]; + tensor scores_35_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_33_cast_fp16, cond = mask_11)[name = string("scores_35_cast_fp16")]; + tensor var_2248_cast_fp16 = softmax(axis = var_59, x = scores_35_cast_fp16)[name = string("op_2248_cast_fp16")]; + tensor input_457_cast_fp16 = select(a = var_44_to_fp16, b = var_2248_cast_fp16, cond = mask_11)[name = string("input_457_cast_fp16")]; + bool x_221_transpose_x_0 = const()[name = string("x_221_transpose_x_0"), val = bool(false)]; + bool x_221_transpose_y_0 = const()[name = string("x_221_transpose_y_0"), val = bool(false)]; + tensor value_25_cast_fp16 = transpose(perm = value_25_perm_0, x = v_17_cast_fp16)[name = string("transpose_287")]; + tensor x_221_cast_fp16 = matmul(transpose_x = x_221_transpose_x_0, transpose_y = x_221_transpose_y_0, x = input_457_cast_fp16, y = value_25_cast_fp16)[name = string("x_221_cast_fp16")]; + tensor var_2252_perm_0 = const()[name = string("op_2252_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_2253 = const()[name = string("op_2253"), val = tensor([1, -1, 1024])]; + tensor var_2252_cast_fp16 = transpose(perm = var_2252_perm_0, x = x_221_cast_fp16)[name = string("transpose_286")]; + tensor input_459_cast_fp16 = reshape(shape = var_2253, x = var_2252_cast_fp16)[name = string("input_459_cast_fp16")]; + tensor encoder_layers_8_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179638144))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(180424640))))[name = string("encoder_layers_8_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_8_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_8_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(180424832)))]; + tensor linear_79_cast_fp16 = linear(bias = encoder_layers_8_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_8_self_attn_linear_out_weight_to_fp16_palettized, x = input_459_cast_fp16)[name = string("linear_79_cast_fp16")]; + tensor input_463_cast_fp16 = add(x = input_453_cast_fp16, y = linear_79_cast_fp16)[name = string("input_463_cast_fp16")]; + tensor x_225_axes_0 = const()[name = string("x_225_axes_0"), val = tensor([-1])]; + tensor encoder_layers_8_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_8_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(180426944)))]; + tensor encoder_layers_8_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_8_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(180429056)))]; + tensor x_225_cast_fp16 = layer_norm(axes = x_225_axes_0, beta = encoder_layers_8_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_8_norm_conv_weight_to_fp16, x = input_463_cast_fp16)[name = string("x_225_cast_fp16")]; + tensor input_465_perm_0 = const()[name = string("input_465_perm_0"), val = tensor([0, 2, 1])]; + string input_467_pad_type_0 = const()[name = string("input_467_pad_type_0"), val = string("valid")]; + tensor input_467_strides_0 = const()[name = string("input_467_strides_0"), val = tensor([1])]; + tensor input_467_pad_0 = const()[name = string("input_467_pad_0"), val = tensor([0, 0])]; + tensor input_467_dilations_0 = const()[name = string("input_467_dilations_0"), val = tensor([1])]; + int32 input_467_groups_0 = const()[name = string("input_467_groups_0"), val = int32(1)]; + tensor encoder_layers_8_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(180431168))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182528384))))[name = string("encoder_layers_8_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_465_cast_fp16 = transpose(perm = input_465_perm_0, x = x_225_cast_fp16)[name = string("transpose_285")]; + tensor input_467_cast_fp16 = conv(dilations = input_467_dilations_0, groups = input_467_groups_0, pad = input_467_pad_0, pad_type = input_467_pad_type_0, strides = input_467_strides_0, weight = encoder_layers_8_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_465_cast_fp16)[name = string("input_467_cast_fp16")]; + int32 x_227_split_num_splits_0 = const()[name = string("x_227_split_num_splits_0"), val = int32(2)]; + int32 x_227_split_axis_0 = const()[name = string("x_227_split_axis_0"), val = int32(1)]; + tensor x_227_split_cast_fp16_0, tensor x_227_split_cast_fp16_1 = split(axis = x_227_split_axis_0, num_splits = x_227_split_num_splits_0, x = input_467_cast_fp16)[name = string("x_227_split_cast_fp16")]; + tensor x_227_split_1_sigmoid_cast_fp16 = sigmoid(x = x_227_split_cast_fp16_1)[name = string("x_227_split_1_sigmoid_cast_fp16")]; + tensor x_227_cast_fp16 = mul(x = x_227_split_cast_fp16_0, y = x_227_split_1_sigmoid_cast_fp16)[name = string("x_227_cast_fp16")]; + tensor input_469_cast_fp16 = select(a = var_44_to_fp16, b = x_227_cast_fp16, cond = var_575)[name = string("input_469_cast_fp16")]; + bool new_x_35_interleave_0 = const()[name = string("new_x_35_interleave_0"), val = bool(false)]; + tensor new_x_35_cast_fp16 = concat(axis = var_59, interleave = new_x_35_interleave_0, values = (cache_35_cast_fp16, input_469_cast_fp16))[name = string("new_x_35_cast_fp16")]; + tensor var_2292_begin_0 = const()[name = string("op_2292_begin_0"), val = tensor([0, 0, 28])]; + tensor var_2292_end_0 = const()[name = string("op_2292_end_0"), val = tensor([1, 1024, 36])]; + tensor var_2292_end_mask_0 = const()[name = string("op_2292_end_mask_0"), val = tensor([true, true, true])]; + tensor var_2292_cast_fp16 = slice_by_index(begin = var_2292_begin_0, end = var_2292_end_0, end_mask = var_2292_end_mask_0, x = new_x_35_cast_fp16)[name = string("op_2292_cast_fp16")]; + string x_229_pad_type_0 = const()[name = string("x_229_pad_type_0"), val = string("valid")]; + int32 x_229_groups_0 = const()[name = string("x_229_groups_0"), val = int32(1024)]; + tensor x_229_strides_0 = const()[name = string("x_229_strides_0"), val = tensor([1])]; + tensor x_229_pad_0 = const()[name = string("x_229_pad_0"), val = tensor([0, 0])]; + tensor x_229_dilations_0 = const()[name = string("x_229_dilations_0"), val = tensor([1])]; + tensor encoder_layers_8_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182532544))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182541824))))[name = string("encoder_layers_8_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_229_cast_fp16 = conv(dilations = x_229_dilations_0, groups = x_229_groups_0, pad = x_229_pad_0, pad_type = x_229_pad_type_0, strides = x_229_strides_0, weight = encoder_layers_8_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_35_cast_fp16)[name = string("x_229_cast_fp16")]; + tensor input_471_perm_0 = const()[name = string("input_471_perm_0"), val = tensor([0, 2, 1])]; + tensor x_231_axes_0 = const()[name = string("x_231_axes_0"), val = tensor([-1])]; + tensor encoder_layers_8_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_8_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182543936)))]; + tensor encoder_layers_8_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_8_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182546048)))]; + tensor input_471_cast_fp16 = transpose(perm = input_471_perm_0, x = x_229_cast_fp16)[name = string("transpose_284")]; + tensor x_231_cast_fp16 = layer_norm(axes = x_231_axes_0, beta = encoder_layers_8_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_8_conv_batch_norm_weight_to_fp16, x = input_471_cast_fp16)[name = string("x_231_cast_fp16")]; + tensor input_473_perm_0 = const()[name = string("input_473_perm_0"), val = tensor([0, 2, 1])]; + tensor input_473_cast_fp16 = transpose(perm = input_473_perm_0, x = x_231_cast_fp16)[name = string("transpose_283")]; + tensor input_475_cast_fp16 = silu(x = input_473_cast_fp16)[name = string("input_475_cast_fp16")]; + string x_233_pad_type_0 = const()[name = string("x_233_pad_type_0"), val = string("valid")]; + tensor x_233_strides_0 = const()[name = string("x_233_strides_0"), val = tensor([1])]; + tensor x_233_pad_0 = const()[name = string("x_233_pad_0"), val = tensor([0, 0])]; + tensor x_233_dilations_0 = const()[name = string("x_233_dilations_0"), val = tensor([1])]; + int32 x_233_groups_0 = const()[name = string("x_233_groups_0"), val = int32(1)]; + tensor encoder_layers_8_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182548160))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183596800))))[name = string("encoder_layers_8_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_233_cast_fp16 = conv(dilations = x_233_dilations_0, groups = x_233_groups_0, pad = x_233_pad_0, pad_type = x_233_pad_type_0, strides = x_233_strides_0, weight = encoder_layers_8_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_475_cast_fp16)[name = string("x_233_cast_fp16")]; + tensor input_477_perm_0 = const()[name = string("input_477_perm_0"), val = tensor([0, 2, 1])]; + tensor input_477_cast_fp16 = transpose(perm = input_477_perm_0, x = x_233_cast_fp16)[name = string("transpose_282")]; + tensor input_479_cast_fp16 = add(x = input_463_cast_fp16, y = input_477_cast_fp16)[name = string("input_479_cast_fp16")]; + tensor input_481_axes_0 = const()[name = string("input_481_axes_0"), val = tensor([-1])]; + tensor encoder_layers_8_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_8_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183598912)))]; + tensor encoder_layers_8_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_8_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183601024)))]; + tensor input_481_cast_fp16 = layer_norm(axes = input_481_axes_0, beta = encoder_layers_8_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_8_norm_feed_forward2_weight_to_fp16, x = input_479_cast_fp16)[name = string("input_481_cast_fp16")]; + tensor encoder_layers_8_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183603136))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186748928))))[name = string("encoder_layers_8_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_8_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_8_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186749120)))]; + tensor linear_80_cast_fp16 = linear(bias = encoder_layers_8_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_8_feed_forward2_linear1_weight_to_fp16_palettized, x = input_481_cast_fp16)[name = string("linear_80_cast_fp16")]; + tensor input_485_cast_fp16 = silu(x = linear_80_cast_fp16)[name = string("input_485_cast_fp16")]; + tensor encoder_layers_8_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186757376))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(189903168))))[name = string("encoder_layers_8_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_8_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_8_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(189903360)))]; + tensor linear_81_cast_fp16 = linear(bias = encoder_layers_8_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_8_feed_forward2_linear2_weight_to_fp16_palettized, x = input_485_cast_fp16)[name = string("linear_81_cast_fp16")]; + fp16 var_2335_to_fp16 = const()[name = string("op_2335_to_fp16"), val = fp16(0x1p-1)]; + tensor var_2336_cast_fp16 = mul(x = linear_81_cast_fp16, y = var_2335_to_fp16)[name = string("op_2336_cast_fp16")]; + tensor input_491_cast_fp16 = add(x = input_479_cast_fp16, y = var_2336_cast_fp16)[name = string("input_491_cast_fp16")]; + tensor input_493_axes_0 = const()[name = string("input_493_axes_0"), val = tensor([-1])]; + tensor encoder_layers_8_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_8_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(189905472)))]; + tensor encoder_layers_8_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_8_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(189907584)))]; + tensor input_493_cast_fp16 = layer_norm(axes = input_493_axes_0, beta = encoder_layers_8_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_8_norm_out_weight_to_fp16, x = input_491_cast_fp16)[name = string("input_493_cast_fp16")]; + tensor cache_37_begin_0 = const()[name = string("cache_37_begin_0"), val = tensor([9, 0, 0, 0])]; + tensor cache_37_end_0 = const()[name = string("cache_37_end_0"), val = tensor([10, 1, 42, 1024])]; + tensor cache_37_end_mask_0 = const()[name = string("cache_37_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_37_squeeze_mask_0 = const()[name = string("cache_37_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_37_cast_fp16 = slice_by_index(begin = cache_37_begin_0, end = cache_37_end_0, end_mask = cache_37_end_mask_0, squeeze_mask = cache_37_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_37_cast_fp16")]; + tensor cache_39_begin_0 = const()[name = string("cache_39_begin_0"), val = tensor([9, 0, 0, 0])]; + tensor cache_39_end_0 = const()[name = string("cache_39_end_0"), val = tensor([10, 1, 1024, 8])]; + tensor cache_39_end_mask_0 = const()[name = string("cache_39_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_39_squeeze_mask_0 = const()[name = string("cache_39_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_39_cast_fp16 = slice_by_index(begin = cache_39_begin_0, end = cache_39_end_0, end_mask = cache_39_end_mask_0, squeeze_mask = cache_39_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_39_cast_fp16")]; + tensor input_495_axes_0 = const()[name = string("input_495_axes_0"), val = tensor([-1])]; + tensor encoder_layers_9_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_9_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(189909696)))]; + tensor encoder_layers_9_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_9_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(189911808)))]; + tensor input_495_cast_fp16 = layer_norm(axes = input_495_axes_0, beta = encoder_layers_9_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_9_norm_feed_forward1_weight_to_fp16, x = input_493_cast_fp16)[name = string("input_495_cast_fp16")]; + tensor encoder_layers_9_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(189913920))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(193059712))))[name = string("encoder_layers_9_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_9_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_9_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(193059904)))]; + tensor linear_82_cast_fp16 = linear(bias = encoder_layers_9_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_9_feed_forward1_linear1_weight_to_fp16_palettized, x = input_495_cast_fp16)[name = string("linear_82_cast_fp16")]; + tensor input_499_cast_fp16 = silu(x = linear_82_cast_fp16)[name = string("input_499_cast_fp16")]; + tensor encoder_layers_9_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(193068160))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196213952))))[name = string("encoder_layers_9_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_9_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_9_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196214144)))]; + tensor linear_83_cast_fp16 = linear(bias = encoder_layers_9_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_9_feed_forward1_linear2_weight_to_fp16_palettized, x = input_499_cast_fp16)[name = string("linear_83_cast_fp16")]; + fp16 var_2372_to_fp16 = const()[name = string("op_2372_to_fp16"), val = fp16(0x1p-1)]; + tensor var_2373_cast_fp16 = mul(x = linear_83_cast_fp16, y = var_2372_to_fp16)[name = string("op_2373_cast_fp16")]; + tensor input_505_cast_fp16 = add(x = input_493_cast_fp16, y = var_2373_cast_fp16)[name = string("input_505_cast_fp16")]; + tensor key_19_axes_0 = const()[name = string("key_19_axes_0"), val = tensor([-1])]; + tensor encoder_layers_9_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_9_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196216256)))]; + tensor encoder_layers_9_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_9_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196218368)))]; + tensor key_19_cast_fp16 = layer_norm(axes = key_19_axes_0, beta = encoder_layers_9_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_9_norm_self_att_weight_to_fp16, x = input_505_cast_fp16)[name = string("key_19_cast_fp16")]; + bool input_507_interleave_0 = const()[name = string("input_507_interleave_0"), val = bool(false)]; + tensor input_507_cast_fp16 = concat(axis = var_68, interleave = input_507_interleave_0, values = (cache_37_cast_fp16, key_19_cast_fp16))[name = string("input_507_cast_fp16")]; + tensor var_2395_begin_0 = const()[name = string("op_2395_begin_0"), val = tensor([0, 28, 0])]; + tensor var_2395_end_0 = const()[name = string("op_2395_end_0"), val = tensor([1, 42, 1024])]; + tensor var_2395_end_mask_0 = const()[name = string("op_2395_end_mask_0"), val = tensor([true, true, true])]; + tensor var_2395_cast_fp16 = slice_by_index(begin = var_2395_begin_0, end = var_2395_end_0, end_mask = var_2395_end_mask_0, x = cache_37_cast_fp16)[name = string("op_2395_cast_fp16")]; + bool var_2401_interleave_0 = const()[name = string("op_2401_interleave_0"), val = bool(false)]; + tensor var_2401_cast_fp16 = concat(axis = var_68, interleave = var_2401_interleave_0, values = (var_2395_cast_fp16, key_19_cast_fp16))[name = string("op_2401_cast_fp16")]; + tensor encoder_layers_9_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196220480))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197006976))))[name = string("encoder_layers_9_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_9_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_9_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197007168)))]; + tensor linear_84_cast_fp16 = linear(bias = encoder_layers_9_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_9_self_attn_linear_q_weight_to_fp16_palettized, x = key_19_cast_fp16)[name = string("linear_84_cast_fp16")]; + tensor var_2406 = const()[name = string("op_2406"), val = tensor([1, -1, 8, 128])]; + tensor q_55_cast_fp16 = reshape(shape = var_2406, x = linear_84_cast_fp16)[name = string("q_55_cast_fp16")]; + tensor encoder_layers_9_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197009280))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197795776))))[name = string("encoder_layers_9_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_9_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_9_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197795968)))]; + tensor linear_85_cast_fp16 = linear(bias = encoder_layers_9_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_9_self_attn_linear_k_weight_to_fp16_palettized, x = input_507_cast_fp16)[name = string("linear_85_cast_fp16")]; + tensor var_2411 = const()[name = string("op_2411"), val = tensor([1, -1, 8, 128])]; + tensor k_37_cast_fp16 = reshape(shape = var_2411, x = linear_85_cast_fp16)[name = string("k_37_cast_fp16")]; + tensor encoder_layers_9_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197798080))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198584576))))[name = string("encoder_layers_9_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_9_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_9_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198584768)))]; + tensor linear_86_cast_fp16 = linear(bias = encoder_layers_9_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_9_self_attn_linear_v_weight_to_fp16_palettized, x = input_507_cast_fp16)[name = string("linear_86_cast_fp16")]; + tensor var_2416 = const()[name = string("op_2416"), val = tensor([1, -1, 8, 128])]; + tensor v_19_cast_fp16 = reshape(shape = var_2416, x = linear_86_cast_fp16)[name = string("v_19_cast_fp16")]; + tensor value_27_perm_0 = const()[name = string("value_27_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_9_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_9_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198586880)))]; + tensor var_2429_cast_fp16 = add(x = q_55_cast_fp16, y = encoder_layers_9_self_attn_pos_bias_u_to_fp16)[name = string("op_2429_cast_fp16")]; + tensor encoder_layers_9_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_9_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198588992)))]; + tensor var_2431_cast_fp16 = add(x = q_55_cast_fp16, y = encoder_layers_9_self_attn_pos_bias_v_to_fp16)[name = string("op_2431_cast_fp16")]; + tensor q_with_bias_v_19_perm_0 = const()[name = string("q_with_bias_v_19_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_241_transpose_x_0 = const()[name = string("x_241_transpose_x_0"), val = bool(false)]; + bool x_241_transpose_y_0 = const()[name = string("x_241_transpose_y_0"), val = bool(false)]; + tensor op_2433_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198591104))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198733504))))[name = string("op_2433_to_fp16_quantized")]; + tensor q_with_bias_v_19_cast_fp16 = transpose(perm = q_with_bias_v_19_perm_0, x = var_2431_cast_fp16)[name = string("transpose_281")]; + tensor x_241_cast_fp16 = matmul(transpose_x = x_241_transpose_x_0, transpose_y = x_241_transpose_y_0, x = q_with_bias_v_19_cast_fp16, y = op_2433_to_fp16_quantized)[name = string("x_241_cast_fp16")]; + tensor x_243_pad_0 = const()[name = string("x_243_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_243_mode_0 = const()[name = string("x_243_mode_0"), val = string("constant")]; + fp16 const_196_to_fp16 = const()[name = string("const_196_to_fp16"), val = fp16(0x0p+0)]; + tensor x_243_cast_fp16 = pad(constant_val = const_196_to_fp16, mode = x_243_mode_0, pad = x_243_pad_0, x = x_241_cast_fp16)[name = string("x_243_cast_fp16")]; + tensor var_2441 = const()[name = string("op_2441"), val = tensor([1, 8, -1, 28])]; + tensor x_245_cast_fp16 = reshape(shape = var_2441, x = x_243_cast_fp16)[name = string("x_245_cast_fp16")]; + tensor var_2445_begin_0 = const()[name = string("op_2445_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_2445_end_0 = const()[name = string("op_2445_end_0"), val = tensor([1, 8, 140, 28])]; + tensor var_2445_end_mask_0 = const()[name = string("op_2445_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_2445_cast_fp16 = slice_by_index(begin = var_2445_begin_0, end = var_2445_end_0, end_mask = var_2445_end_mask_0, x = x_245_cast_fp16)[name = string("op_2445_cast_fp16")]; + tensor var_2446 = const()[name = string("op_2446"), val = tensor([1, 8, 28, 139])]; + tensor matrix_bd_37_cast_fp16 = reshape(shape = var_2446, x = var_2445_cast_fp16)[name = string("matrix_bd_37_cast_fp16")]; + bool matrix_ac_19_transpose_x_0 = const()[name = string("matrix_ac_19_transpose_x_0"), val = bool(false)]; + bool matrix_ac_19_transpose_y_0 = const()[name = string("matrix_ac_19_transpose_y_0"), val = bool(false)]; + tensor transpose_114_perm_0 = const()[name = string("transpose_114_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_115_perm_0 = const()[name = string("transpose_115_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_115 = transpose(perm = transpose_115_perm_0, x = k_37_cast_fp16)[name = string("transpose_279")]; + tensor transpose_114 = transpose(perm = transpose_114_perm_0, x = var_2429_cast_fp16)[name = string("transpose_280")]; + tensor matrix_ac_19_cast_fp16 = matmul(transpose_x = matrix_ac_19_transpose_x_0, transpose_y = matrix_ac_19_transpose_y_0, x = transpose_114, y = transpose_115)[name = string("matrix_ac_19_cast_fp16")]; + tensor matrix_bd_39_begin_0 = const()[name = string("matrix_bd_39_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_39_end_0 = const()[name = string("matrix_bd_39_end_0"), val = tensor([1, 8, 28, 70])]; + tensor matrix_bd_39_end_mask_0 = const()[name = string("matrix_bd_39_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_39_cast_fp16 = slice_by_index(begin = matrix_bd_39_begin_0, end = matrix_bd_39_end_0, end_mask = matrix_bd_39_end_mask_0, x = matrix_bd_37_cast_fp16)[name = string("matrix_bd_39_cast_fp16")]; + tensor var_2455_cast_fp16 = add(x = matrix_ac_19_cast_fp16, y = matrix_bd_39_cast_fp16)[name = string("op_2455_cast_fp16")]; + fp16 _inversed_scores_37_y_0_to_fp16 = const()[name = string("_inversed_scores_37_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_37_cast_fp16 = mul(x = var_2455_cast_fp16, y = _inversed_scores_37_y_0_to_fp16)[name = string("_inversed_scores_37_cast_fp16")]; + tensor scores_39_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_37_cast_fp16, cond = mask_11)[name = string("scores_39_cast_fp16")]; + tensor var_2461_cast_fp16 = softmax(axis = var_59, x = scores_39_cast_fp16)[name = string("op_2461_cast_fp16")]; + tensor input_509_cast_fp16 = select(a = var_44_to_fp16, b = var_2461_cast_fp16, cond = mask_11)[name = string("input_509_cast_fp16")]; + bool x_247_transpose_x_0 = const()[name = string("x_247_transpose_x_0"), val = bool(false)]; + bool x_247_transpose_y_0 = const()[name = string("x_247_transpose_y_0"), val = bool(false)]; + tensor value_27_cast_fp16 = transpose(perm = value_27_perm_0, x = v_19_cast_fp16)[name = string("transpose_278")]; + tensor x_247_cast_fp16 = matmul(transpose_x = x_247_transpose_x_0, transpose_y = x_247_transpose_y_0, x = input_509_cast_fp16, y = value_27_cast_fp16)[name = string("x_247_cast_fp16")]; + tensor var_2465_perm_0 = const()[name = string("op_2465_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_2466 = const()[name = string("op_2466"), val = tensor([1, -1, 1024])]; + tensor var_2465_cast_fp16 = transpose(perm = var_2465_perm_0, x = x_247_cast_fp16)[name = string("transpose_277")]; + tensor input_511_cast_fp16 = reshape(shape = var_2466, x = var_2465_cast_fp16)[name = string("input_511_cast_fp16")]; + tensor encoder_layers_9_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198733888))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199520384))))[name = string("encoder_layers_9_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_9_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_9_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199520576)))]; + tensor linear_88_cast_fp16 = linear(bias = encoder_layers_9_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_9_self_attn_linear_out_weight_to_fp16_palettized, x = input_511_cast_fp16)[name = string("linear_88_cast_fp16")]; + tensor input_515_cast_fp16 = add(x = input_505_cast_fp16, y = linear_88_cast_fp16)[name = string("input_515_cast_fp16")]; + tensor x_251_axes_0 = const()[name = string("x_251_axes_0"), val = tensor([-1])]; + tensor encoder_layers_9_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_9_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199522688)))]; + tensor encoder_layers_9_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_9_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199524800)))]; + tensor x_251_cast_fp16 = layer_norm(axes = x_251_axes_0, beta = encoder_layers_9_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_9_norm_conv_weight_to_fp16, x = input_515_cast_fp16)[name = string("x_251_cast_fp16")]; + tensor input_517_perm_0 = const()[name = string("input_517_perm_0"), val = tensor([0, 2, 1])]; + string input_519_pad_type_0 = const()[name = string("input_519_pad_type_0"), val = string("valid")]; + tensor input_519_strides_0 = const()[name = string("input_519_strides_0"), val = tensor([1])]; + tensor input_519_pad_0 = const()[name = string("input_519_pad_0"), val = tensor([0, 0])]; + tensor input_519_dilations_0 = const()[name = string("input_519_dilations_0"), val = tensor([1])]; + int32 input_519_groups_0 = const()[name = string("input_519_groups_0"), val = int32(1)]; + tensor encoder_layers_9_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199526912))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(201624128))))[name = string("encoder_layers_9_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_517_cast_fp16 = transpose(perm = input_517_perm_0, x = x_251_cast_fp16)[name = string("transpose_276")]; + tensor input_519_cast_fp16 = conv(dilations = input_519_dilations_0, groups = input_519_groups_0, pad = input_519_pad_0, pad_type = input_519_pad_type_0, strides = input_519_strides_0, weight = encoder_layers_9_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_517_cast_fp16)[name = string("input_519_cast_fp16")]; + int32 x_253_split_num_splits_0 = const()[name = string("x_253_split_num_splits_0"), val = int32(2)]; + int32 x_253_split_axis_0 = const()[name = string("x_253_split_axis_0"), val = int32(1)]; + tensor x_253_split_cast_fp16_0, tensor x_253_split_cast_fp16_1 = split(axis = x_253_split_axis_0, num_splits = x_253_split_num_splits_0, x = input_519_cast_fp16)[name = string("x_253_split_cast_fp16")]; + tensor x_253_split_1_sigmoid_cast_fp16 = sigmoid(x = x_253_split_cast_fp16_1)[name = string("x_253_split_1_sigmoid_cast_fp16")]; + tensor x_253_cast_fp16 = mul(x = x_253_split_cast_fp16_0, y = x_253_split_1_sigmoid_cast_fp16)[name = string("x_253_cast_fp16")]; + tensor input_521_cast_fp16 = select(a = var_44_to_fp16, b = x_253_cast_fp16, cond = var_575)[name = string("input_521_cast_fp16")]; + bool new_x_39_interleave_0 = const()[name = string("new_x_39_interleave_0"), val = bool(false)]; + tensor new_x_39_cast_fp16 = concat(axis = var_59, interleave = new_x_39_interleave_0, values = (cache_39_cast_fp16, input_521_cast_fp16))[name = string("new_x_39_cast_fp16")]; + tensor var_2505_begin_0 = const()[name = string("op_2505_begin_0"), val = tensor([0, 0, 28])]; + tensor var_2505_end_0 = const()[name = string("op_2505_end_0"), val = tensor([1, 1024, 36])]; + tensor var_2505_end_mask_0 = const()[name = string("op_2505_end_mask_0"), val = tensor([true, true, true])]; + tensor var_2505_cast_fp16 = slice_by_index(begin = var_2505_begin_0, end = var_2505_end_0, end_mask = var_2505_end_mask_0, x = new_x_39_cast_fp16)[name = string("op_2505_cast_fp16")]; + string x_255_pad_type_0 = const()[name = string("x_255_pad_type_0"), val = string("valid")]; + int32 x_255_groups_0 = const()[name = string("x_255_groups_0"), val = int32(1024)]; + tensor x_255_strides_0 = const()[name = string("x_255_strides_0"), val = tensor([1])]; + tensor x_255_pad_0 = const()[name = string("x_255_pad_0"), val = tensor([0, 0])]; + tensor x_255_dilations_0 = const()[name = string("x_255_dilations_0"), val = tensor([1])]; + tensor encoder_layers_9_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(201628288))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(201637568))))[name = string("encoder_layers_9_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_255_cast_fp16 = conv(dilations = x_255_dilations_0, groups = x_255_groups_0, pad = x_255_pad_0, pad_type = x_255_pad_type_0, strides = x_255_strides_0, weight = encoder_layers_9_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_39_cast_fp16)[name = string("x_255_cast_fp16")]; + tensor input_523_perm_0 = const()[name = string("input_523_perm_0"), val = tensor([0, 2, 1])]; + tensor x_257_axes_0 = const()[name = string("x_257_axes_0"), val = tensor([-1])]; + tensor encoder_layers_9_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_9_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(201639680)))]; + tensor encoder_layers_9_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_9_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(201641792)))]; + tensor input_523_cast_fp16 = transpose(perm = input_523_perm_0, x = x_255_cast_fp16)[name = string("transpose_275")]; + tensor x_257_cast_fp16 = layer_norm(axes = x_257_axes_0, beta = encoder_layers_9_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_9_conv_batch_norm_weight_to_fp16, x = input_523_cast_fp16)[name = string("x_257_cast_fp16")]; + tensor input_525_perm_0 = const()[name = string("input_525_perm_0"), val = tensor([0, 2, 1])]; + tensor input_525_cast_fp16 = transpose(perm = input_525_perm_0, x = x_257_cast_fp16)[name = string("transpose_274")]; + tensor input_527_cast_fp16 = silu(x = input_525_cast_fp16)[name = string("input_527_cast_fp16")]; + string x_259_pad_type_0 = const()[name = string("x_259_pad_type_0"), val = string("valid")]; + tensor x_259_strides_0 = const()[name = string("x_259_strides_0"), val = tensor([1])]; + tensor x_259_pad_0 = const()[name = string("x_259_pad_0"), val = tensor([0, 0])]; + tensor x_259_dilations_0 = const()[name = string("x_259_dilations_0"), val = tensor([1])]; + int32 x_259_groups_0 = const()[name = string("x_259_groups_0"), val = int32(1)]; + tensor encoder_layers_9_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(201643904))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202692544))))[name = string("encoder_layers_9_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_259_cast_fp16 = conv(dilations = x_259_dilations_0, groups = x_259_groups_0, pad = x_259_pad_0, pad_type = x_259_pad_type_0, strides = x_259_strides_0, weight = encoder_layers_9_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_527_cast_fp16)[name = string("x_259_cast_fp16")]; + tensor input_529_perm_0 = const()[name = string("input_529_perm_0"), val = tensor([0, 2, 1])]; + tensor input_529_cast_fp16 = transpose(perm = input_529_perm_0, x = x_259_cast_fp16)[name = string("transpose_273")]; + tensor input_531_cast_fp16 = add(x = input_515_cast_fp16, y = input_529_cast_fp16)[name = string("input_531_cast_fp16")]; + tensor input_533_axes_0 = const()[name = string("input_533_axes_0"), val = tensor([-1])]; + tensor encoder_layers_9_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_9_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202694656)))]; + tensor encoder_layers_9_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_9_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202696768)))]; + tensor input_533_cast_fp16 = layer_norm(axes = input_533_axes_0, beta = encoder_layers_9_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_9_norm_feed_forward2_weight_to_fp16, x = input_531_cast_fp16)[name = string("input_533_cast_fp16")]; + tensor encoder_layers_9_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202698880))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(205844672))))[name = string("encoder_layers_9_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_9_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_9_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(205844864)))]; + tensor linear_89_cast_fp16 = linear(bias = encoder_layers_9_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_9_feed_forward2_linear1_weight_to_fp16_palettized, x = input_533_cast_fp16)[name = string("linear_89_cast_fp16")]; + tensor input_537_cast_fp16 = silu(x = linear_89_cast_fp16)[name = string("input_537_cast_fp16")]; + tensor encoder_layers_9_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(205853120))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(208998912))))[name = string("encoder_layers_9_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_9_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_9_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(208999104)))]; + tensor linear_90_cast_fp16 = linear(bias = encoder_layers_9_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_9_feed_forward2_linear2_weight_to_fp16_palettized, x = input_537_cast_fp16)[name = string("linear_90_cast_fp16")]; + fp16 var_2548_to_fp16 = const()[name = string("op_2548_to_fp16"), val = fp16(0x1p-1)]; + tensor var_2549_cast_fp16 = mul(x = linear_90_cast_fp16, y = var_2548_to_fp16)[name = string("op_2549_cast_fp16")]; + tensor input_543_cast_fp16 = add(x = input_531_cast_fp16, y = var_2549_cast_fp16)[name = string("input_543_cast_fp16")]; + tensor input_545_axes_0 = const()[name = string("input_545_axes_0"), val = tensor([-1])]; + tensor encoder_layers_9_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_9_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(209001216)))]; + tensor encoder_layers_9_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_9_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(209003328)))]; + tensor input_545_cast_fp16 = layer_norm(axes = input_545_axes_0, beta = encoder_layers_9_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_9_norm_out_weight_to_fp16, x = input_543_cast_fp16)[name = string("input_545_cast_fp16")]; + tensor cache_41_begin_0 = const()[name = string("cache_41_begin_0"), val = tensor([10, 0, 0, 0])]; + tensor cache_41_end_0 = const()[name = string("cache_41_end_0"), val = tensor([11, 1, 42, 1024])]; + tensor cache_41_end_mask_0 = const()[name = string("cache_41_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_41_squeeze_mask_0 = const()[name = string("cache_41_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_41_cast_fp16 = slice_by_index(begin = cache_41_begin_0, end = cache_41_end_0, end_mask = cache_41_end_mask_0, squeeze_mask = cache_41_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_41_cast_fp16")]; + tensor cache_43_begin_0 = const()[name = string("cache_43_begin_0"), val = tensor([10, 0, 0, 0])]; + tensor cache_43_end_0 = const()[name = string("cache_43_end_0"), val = tensor([11, 1, 1024, 8])]; + tensor cache_43_end_mask_0 = const()[name = string("cache_43_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_43_squeeze_mask_0 = const()[name = string("cache_43_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_43_cast_fp16 = slice_by_index(begin = cache_43_begin_0, end = cache_43_end_0, end_mask = cache_43_end_mask_0, squeeze_mask = cache_43_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_43_cast_fp16")]; + tensor input_547_axes_0 = const()[name = string("input_547_axes_0"), val = tensor([-1])]; + tensor encoder_layers_10_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_10_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(209005440)))]; + tensor encoder_layers_10_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_10_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(209007552)))]; + tensor input_547_cast_fp16 = layer_norm(axes = input_547_axes_0, beta = encoder_layers_10_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_10_norm_feed_forward1_weight_to_fp16, x = input_545_cast_fp16)[name = string("input_547_cast_fp16")]; + tensor encoder_layers_10_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(209009664))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212155456))))[name = string("encoder_layers_10_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_10_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_10_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212155648)))]; + tensor linear_91_cast_fp16 = linear(bias = encoder_layers_10_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_10_feed_forward1_linear1_weight_to_fp16_palettized, x = input_547_cast_fp16)[name = string("linear_91_cast_fp16")]; + tensor input_551_cast_fp16 = silu(x = linear_91_cast_fp16)[name = string("input_551_cast_fp16")]; + tensor encoder_layers_10_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212163904))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215309696))))[name = string("encoder_layers_10_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_10_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_10_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215309888)))]; + tensor linear_92_cast_fp16 = linear(bias = encoder_layers_10_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_10_feed_forward1_linear2_weight_to_fp16_palettized, x = input_551_cast_fp16)[name = string("linear_92_cast_fp16")]; + fp16 var_2585_to_fp16 = const()[name = string("op_2585_to_fp16"), val = fp16(0x1p-1)]; + tensor var_2586_cast_fp16 = mul(x = linear_92_cast_fp16, y = var_2585_to_fp16)[name = string("op_2586_cast_fp16")]; + tensor input_557_cast_fp16 = add(x = input_545_cast_fp16, y = var_2586_cast_fp16)[name = string("input_557_cast_fp16")]; + tensor key_21_axes_0 = const()[name = string("key_21_axes_0"), val = tensor([-1])]; + tensor encoder_layers_10_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_10_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215312000)))]; + tensor encoder_layers_10_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_10_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215314112)))]; + tensor key_21_cast_fp16 = layer_norm(axes = key_21_axes_0, beta = encoder_layers_10_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_10_norm_self_att_weight_to_fp16, x = input_557_cast_fp16)[name = string("key_21_cast_fp16")]; + bool input_559_interleave_0 = const()[name = string("input_559_interleave_0"), val = bool(false)]; + tensor input_559_cast_fp16 = concat(axis = var_68, interleave = input_559_interleave_0, values = (cache_41_cast_fp16, key_21_cast_fp16))[name = string("input_559_cast_fp16")]; + tensor var_2608_begin_0 = const()[name = string("op_2608_begin_0"), val = tensor([0, 28, 0])]; + tensor var_2608_end_0 = const()[name = string("op_2608_end_0"), val = tensor([1, 42, 1024])]; + tensor var_2608_end_mask_0 = const()[name = string("op_2608_end_mask_0"), val = tensor([true, true, true])]; + tensor var_2608_cast_fp16 = slice_by_index(begin = var_2608_begin_0, end = var_2608_end_0, end_mask = var_2608_end_mask_0, x = cache_41_cast_fp16)[name = string("op_2608_cast_fp16")]; + bool var_2614_interleave_0 = const()[name = string("op_2614_interleave_0"), val = bool(false)]; + tensor var_2614_cast_fp16 = concat(axis = var_68, interleave = var_2614_interleave_0, values = (var_2608_cast_fp16, key_21_cast_fp16))[name = string("op_2614_cast_fp16")]; + tensor encoder_layers_10_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215316224))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216102720))))[name = string("encoder_layers_10_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_10_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_10_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216102912)))]; + tensor linear_93_cast_fp16 = linear(bias = encoder_layers_10_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_10_self_attn_linear_q_weight_to_fp16_palettized, x = key_21_cast_fp16)[name = string("linear_93_cast_fp16")]; + tensor var_2619 = const()[name = string("op_2619"), val = tensor([1, -1, 8, 128])]; + tensor q_61_cast_fp16 = reshape(shape = var_2619, x = linear_93_cast_fp16)[name = string("q_61_cast_fp16")]; + tensor encoder_layers_10_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216105024))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216891520))))[name = string("encoder_layers_10_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_10_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_10_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216891712)))]; + tensor linear_94_cast_fp16 = linear(bias = encoder_layers_10_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_10_self_attn_linear_k_weight_to_fp16_palettized, x = input_559_cast_fp16)[name = string("linear_94_cast_fp16")]; + tensor var_2624 = const()[name = string("op_2624"), val = tensor([1, -1, 8, 128])]; + tensor k_41_cast_fp16 = reshape(shape = var_2624, x = linear_94_cast_fp16)[name = string("k_41_cast_fp16")]; + tensor encoder_layers_10_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216893824))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217680320))))[name = string("encoder_layers_10_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_10_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_10_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217680512)))]; + tensor linear_95_cast_fp16 = linear(bias = encoder_layers_10_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_10_self_attn_linear_v_weight_to_fp16_palettized, x = input_559_cast_fp16)[name = string("linear_95_cast_fp16")]; + tensor var_2629 = const()[name = string("op_2629"), val = tensor([1, -1, 8, 128])]; + tensor v_21_cast_fp16 = reshape(shape = var_2629, x = linear_95_cast_fp16)[name = string("v_21_cast_fp16")]; + tensor value_29_perm_0 = const()[name = string("value_29_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_10_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_10_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217682624)))]; + tensor var_2642_cast_fp16 = add(x = q_61_cast_fp16, y = encoder_layers_10_self_attn_pos_bias_u_to_fp16)[name = string("op_2642_cast_fp16")]; + tensor encoder_layers_10_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_10_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217684736)))]; + tensor var_2644_cast_fp16 = add(x = q_61_cast_fp16, y = encoder_layers_10_self_attn_pos_bias_v_to_fp16)[name = string("op_2644_cast_fp16")]; + tensor q_with_bias_v_21_perm_0 = const()[name = string("q_with_bias_v_21_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_267_transpose_x_0 = const()[name = string("x_267_transpose_x_0"), val = bool(false)]; + bool x_267_transpose_y_0 = const()[name = string("x_267_transpose_y_0"), val = bool(false)]; + tensor op_2646_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217686848))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217829248))))[name = string("op_2646_to_fp16_quantized")]; + tensor q_with_bias_v_21_cast_fp16 = transpose(perm = q_with_bias_v_21_perm_0, x = var_2644_cast_fp16)[name = string("transpose_272")]; + tensor x_267_cast_fp16 = matmul(transpose_x = x_267_transpose_x_0, transpose_y = x_267_transpose_y_0, x = q_with_bias_v_21_cast_fp16, y = op_2646_to_fp16_quantized)[name = string("x_267_cast_fp16")]; + tensor x_269_pad_0 = const()[name = string("x_269_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_269_mode_0 = const()[name = string("x_269_mode_0"), val = string("constant")]; + fp16 const_209_to_fp16 = const()[name = string("const_209_to_fp16"), val = fp16(0x0p+0)]; + tensor x_269_cast_fp16 = pad(constant_val = const_209_to_fp16, mode = x_269_mode_0, pad = x_269_pad_0, x = x_267_cast_fp16)[name = string("x_269_cast_fp16")]; + tensor var_2654 = const()[name = string("op_2654"), val = tensor([1, 8, -1, 28])]; + tensor x_271_cast_fp16 = reshape(shape = var_2654, x = x_269_cast_fp16)[name = string("x_271_cast_fp16")]; + tensor var_2658_begin_0 = const()[name = string("op_2658_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_2658_end_0 = const()[name = string("op_2658_end_0"), val = tensor([1, 8, 140, 28])]; + tensor var_2658_end_mask_0 = const()[name = string("op_2658_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_2658_cast_fp16 = slice_by_index(begin = var_2658_begin_0, end = var_2658_end_0, end_mask = var_2658_end_mask_0, x = x_271_cast_fp16)[name = string("op_2658_cast_fp16")]; + tensor var_2659 = const()[name = string("op_2659"), val = tensor([1, 8, 28, 139])]; + tensor matrix_bd_41_cast_fp16 = reshape(shape = var_2659, x = var_2658_cast_fp16)[name = string("matrix_bd_41_cast_fp16")]; + bool matrix_ac_21_transpose_x_0 = const()[name = string("matrix_ac_21_transpose_x_0"), val = bool(false)]; + bool matrix_ac_21_transpose_y_0 = const()[name = string("matrix_ac_21_transpose_y_0"), val = bool(false)]; + tensor transpose_116_perm_0 = const()[name = string("transpose_116_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_117_perm_0 = const()[name = string("transpose_117_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_117 = transpose(perm = transpose_117_perm_0, x = k_41_cast_fp16)[name = string("transpose_270")]; + tensor transpose_116 = transpose(perm = transpose_116_perm_0, x = var_2642_cast_fp16)[name = string("transpose_271")]; + tensor matrix_ac_21_cast_fp16 = matmul(transpose_x = matrix_ac_21_transpose_x_0, transpose_y = matrix_ac_21_transpose_y_0, x = transpose_116, y = transpose_117)[name = string("matrix_ac_21_cast_fp16")]; + tensor matrix_bd_43_begin_0 = const()[name = string("matrix_bd_43_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_43_end_0 = const()[name = string("matrix_bd_43_end_0"), val = tensor([1, 8, 28, 70])]; + tensor matrix_bd_43_end_mask_0 = const()[name = string("matrix_bd_43_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_43_cast_fp16 = slice_by_index(begin = matrix_bd_43_begin_0, end = matrix_bd_43_end_0, end_mask = matrix_bd_43_end_mask_0, x = matrix_bd_41_cast_fp16)[name = string("matrix_bd_43_cast_fp16")]; + tensor var_2668_cast_fp16 = add(x = matrix_ac_21_cast_fp16, y = matrix_bd_43_cast_fp16)[name = string("op_2668_cast_fp16")]; + fp16 _inversed_scores_41_y_0_to_fp16 = const()[name = string("_inversed_scores_41_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_41_cast_fp16 = mul(x = var_2668_cast_fp16, y = _inversed_scores_41_y_0_to_fp16)[name = string("_inversed_scores_41_cast_fp16")]; + tensor scores_43_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_41_cast_fp16, cond = mask_11)[name = string("scores_43_cast_fp16")]; + tensor var_2674_cast_fp16 = softmax(axis = var_59, x = scores_43_cast_fp16)[name = string("op_2674_cast_fp16")]; + tensor input_561_cast_fp16 = select(a = var_44_to_fp16, b = var_2674_cast_fp16, cond = mask_11)[name = string("input_561_cast_fp16")]; + bool x_273_transpose_x_0 = const()[name = string("x_273_transpose_x_0"), val = bool(false)]; + bool x_273_transpose_y_0 = const()[name = string("x_273_transpose_y_0"), val = bool(false)]; + tensor value_29_cast_fp16 = transpose(perm = value_29_perm_0, x = v_21_cast_fp16)[name = string("transpose_269")]; + tensor x_273_cast_fp16 = matmul(transpose_x = x_273_transpose_x_0, transpose_y = x_273_transpose_y_0, x = input_561_cast_fp16, y = value_29_cast_fp16)[name = string("x_273_cast_fp16")]; + tensor var_2678_perm_0 = const()[name = string("op_2678_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_2679 = const()[name = string("op_2679"), val = tensor([1, -1, 1024])]; + tensor var_2678_cast_fp16 = transpose(perm = var_2678_perm_0, x = x_273_cast_fp16)[name = string("transpose_268")]; + tensor input_563_cast_fp16 = reshape(shape = var_2679, x = var_2678_cast_fp16)[name = string("input_563_cast_fp16")]; + tensor encoder_layers_10_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217829632))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218616128))))[name = string("encoder_layers_10_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_10_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_10_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218616320)))]; + tensor linear_97_cast_fp16 = linear(bias = encoder_layers_10_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_10_self_attn_linear_out_weight_to_fp16_palettized, x = input_563_cast_fp16)[name = string("linear_97_cast_fp16")]; + tensor input_567_cast_fp16 = add(x = input_557_cast_fp16, y = linear_97_cast_fp16)[name = string("input_567_cast_fp16")]; + tensor x_277_axes_0 = const()[name = string("x_277_axes_0"), val = tensor([-1])]; + tensor encoder_layers_10_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_10_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218618432)))]; + tensor encoder_layers_10_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_10_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218620544)))]; + tensor x_277_cast_fp16 = layer_norm(axes = x_277_axes_0, beta = encoder_layers_10_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_10_norm_conv_weight_to_fp16, x = input_567_cast_fp16)[name = string("x_277_cast_fp16")]; + tensor input_569_perm_0 = const()[name = string("input_569_perm_0"), val = tensor([0, 2, 1])]; + string input_571_pad_type_0 = const()[name = string("input_571_pad_type_0"), val = string("valid")]; + tensor input_571_strides_0 = const()[name = string("input_571_strides_0"), val = tensor([1])]; + tensor input_571_pad_0 = const()[name = string("input_571_pad_0"), val = tensor([0, 0])]; + tensor input_571_dilations_0 = const()[name = string("input_571_dilations_0"), val = tensor([1])]; + int32 input_571_groups_0 = const()[name = string("input_571_groups_0"), val = int32(1)]; + tensor encoder_layers_10_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218622656))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220719872))))[name = string("encoder_layers_10_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_569_cast_fp16 = transpose(perm = input_569_perm_0, x = x_277_cast_fp16)[name = string("transpose_267")]; + tensor input_571_cast_fp16 = conv(dilations = input_571_dilations_0, groups = input_571_groups_0, pad = input_571_pad_0, pad_type = input_571_pad_type_0, strides = input_571_strides_0, weight = encoder_layers_10_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_569_cast_fp16)[name = string("input_571_cast_fp16")]; + int32 x_279_split_num_splits_0 = const()[name = string("x_279_split_num_splits_0"), val = int32(2)]; + int32 x_279_split_axis_0 = const()[name = string("x_279_split_axis_0"), val = int32(1)]; + tensor x_279_split_cast_fp16_0, tensor x_279_split_cast_fp16_1 = split(axis = x_279_split_axis_0, num_splits = x_279_split_num_splits_0, x = input_571_cast_fp16)[name = string("x_279_split_cast_fp16")]; + tensor x_279_split_1_sigmoid_cast_fp16 = sigmoid(x = x_279_split_cast_fp16_1)[name = string("x_279_split_1_sigmoid_cast_fp16")]; + tensor x_279_cast_fp16 = mul(x = x_279_split_cast_fp16_0, y = x_279_split_1_sigmoid_cast_fp16)[name = string("x_279_cast_fp16")]; + tensor input_573_cast_fp16 = select(a = var_44_to_fp16, b = x_279_cast_fp16, cond = var_575)[name = string("input_573_cast_fp16")]; + bool new_x_43_interleave_0 = const()[name = string("new_x_43_interleave_0"), val = bool(false)]; + tensor new_x_43_cast_fp16 = concat(axis = var_59, interleave = new_x_43_interleave_0, values = (cache_43_cast_fp16, input_573_cast_fp16))[name = string("new_x_43_cast_fp16")]; + tensor var_2718_begin_0 = const()[name = string("op_2718_begin_0"), val = tensor([0, 0, 28])]; + tensor var_2718_end_0 = const()[name = string("op_2718_end_0"), val = tensor([1, 1024, 36])]; + tensor var_2718_end_mask_0 = const()[name = string("op_2718_end_mask_0"), val = tensor([true, true, true])]; + tensor var_2718_cast_fp16 = slice_by_index(begin = var_2718_begin_0, end = var_2718_end_0, end_mask = var_2718_end_mask_0, x = new_x_43_cast_fp16)[name = string("op_2718_cast_fp16")]; + string x_281_pad_type_0 = const()[name = string("x_281_pad_type_0"), val = string("valid")]; + int32 x_281_groups_0 = const()[name = string("x_281_groups_0"), val = int32(1024)]; + tensor x_281_strides_0 = const()[name = string("x_281_strides_0"), val = tensor([1])]; + tensor x_281_pad_0 = const()[name = string("x_281_pad_0"), val = tensor([0, 0])]; + tensor x_281_dilations_0 = const()[name = string("x_281_dilations_0"), val = tensor([1])]; + tensor encoder_layers_10_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220724032))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220733312))))[name = string("encoder_layers_10_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_281_cast_fp16 = conv(dilations = x_281_dilations_0, groups = x_281_groups_0, pad = x_281_pad_0, pad_type = x_281_pad_type_0, strides = x_281_strides_0, weight = encoder_layers_10_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_43_cast_fp16)[name = string("x_281_cast_fp16")]; + tensor input_575_perm_0 = const()[name = string("input_575_perm_0"), val = tensor([0, 2, 1])]; + tensor x_283_axes_0 = const()[name = string("x_283_axes_0"), val = tensor([-1])]; + tensor encoder_layers_10_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_10_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220735424)))]; + tensor encoder_layers_10_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_10_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220737536)))]; + tensor input_575_cast_fp16 = transpose(perm = input_575_perm_0, x = x_281_cast_fp16)[name = string("transpose_266")]; + tensor x_283_cast_fp16 = layer_norm(axes = x_283_axes_0, beta = encoder_layers_10_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_10_conv_batch_norm_weight_to_fp16, x = input_575_cast_fp16)[name = string("x_283_cast_fp16")]; + tensor input_577_perm_0 = const()[name = string("input_577_perm_0"), val = tensor([0, 2, 1])]; + tensor input_577_cast_fp16 = transpose(perm = input_577_perm_0, x = x_283_cast_fp16)[name = string("transpose_265")]; + tensor input_579_cast_fp16 = silu(x = input_577_cast_fp16)[name = string("input_579_cast_fp16")]; + string x_285_pad_type_0 = const()[name = string("x_285_pad_type_0"), val = string("valid")]; + tensor x_285_strides_0 = const()[name = string("x_285_strides_0"), val = tensor([1])]; + tensor x_285_pad_0 = const()[name = string("x_285_pad_0"), val = tensor([0, 0])]; + tensor x_285_dilations_0 = const()[name = string("x_285_dilations_0"), val = tensor([1])]; + int32 x_285_groups_0 = const()[name = string("x_285_groups_0"), val = int32(1)]; + tensor encoder_layers_10_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220739648))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221788288))))[name = string("encoder_layers_10_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_285_cast_fp16 = conv(dilations = x_285_dilations_0, groups = x_285_groups_0, pad = x_285_pad_0, pad_type = x_285_pad_type_0, strides = x_285_strides_0, weight = encoder_layers_10_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_579_cast_fp16)[name = string("x_285_cast_fp16")]; + tensor input_581_perm_0 = const()[name = string("input_581_perm_0"), val = tensor([0, 2, 1])]; + tensor input_581_cast_fp16 = transpose(perm = input_581_perm_0, x = x_285_cast_fp16)[name = string("transpose_264")]; + tensor input_583_cast_fp16 = add(x = input_567_cast_fp16, y = input_581_cast_fp16)[name = string("input_583_cast_fp16")]; + tensor input_585_axes_0 = const()[name = string("input_585_axes_0"), val = tensor([-1])]; + tensor encoder_layers_10_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_10_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221790400)))]; + tensor encoder_layers_10_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_10_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221792512)))]; + tensor input_585_cast_fp16 = layer_norm(axes = input_585_axes_0, beta = encoder_layers_10_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_10_norm_feed_forward2_weight_to_fp16, x = input_583_cast_fp16)[name = string("input_585_cast_fp16")]; + tensor encoder_layers_10_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221794624))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(224940416))))[name = string("encoder_layers_10_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_10_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_10_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(224940608)))]; + tensor linear_98_cast_fp16 = linear(bias = encoder_layers_10_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_10_feed_forward2_linear1_weight_to_fp16_palettized, x = input_585_cast_fp16)[name = string("linear_98_cast_fp16")]; + tensor input_589_cast_fp16 = silu(x = linear_98_cast_fp16)[name = string("input_589_cast_fp16")]; + tensor encoder_layers_10_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(224948864))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(228094656))))[name = string("encoder_layers_10_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_10_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_10_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(228094848)))]; + tensor linear_99_cast_fp16 = linear(bias = encoder_layers_10_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_10_feed_forward2_linear2_weight_to_fp16_palettized, x = input_589_cast_fp16)[name = string("linear_99_cast_fp16")]; + fp16 var_2761_to_fp16 = const()[name = string("op_2761_to_fp16"), val = fp16(0x1p-1)]; + tensor var_2762_cast_fp16 = mul(x = linear_99_cast_fp16, y = var_2761_to_fp16)[name = string("op_2762_cast_fp16")]; + tensor input_595_cast_fp16 = add(x = input_583_cast_fp16, y = var_2762_cast_fp16)[name = string("input_595_cast_fp16")]; + tensor input_597_axes_0 = const()[name = string("input_597_axes_0"), val = tensor([-1])]; + tensor encoder_layers_10_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_10_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(228096960)))]; + tensor encoder_layers_10_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_10_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(228099072)))]; + tensor input_597_cast_fp16 = layer_norm(axes = input_597_axes_0, beta = encoder_layers_10_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_10_norm_out_weight_to_fp16, x = input_595_cast_fp16)[name = string("input_597_cast_fp16")]; + tensor cache_45_begin_0 = const()[name = string("cache_45_begin_0"), val = tensor([11, 0, 0, 0])]; + tensor cache_45_end_0 = const()[name = string("cache_45_end_0"), val = tensor([12, 1, 42, 1024])]; + tensor cache_45_end_mask_0 = const()[name = string("cache_45_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_45_squeeze_mask_0 = const()[name = string("cache_45_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_45_cast_fp16 = slice_by_index(begin = cache_45_begin_0, end = cache_45_end_0, end_mask = cache_45_end_mask_0, squeeze_mask = cache_45_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_45_cast_fp16")]; + tensor cache_47_begin_0 = const()[name = string("cache_47_begin_0"), val = tensor([11, 0, 0, 0])]; + tensor cache_47_end_0 = const()[name = string("cache_47_end_0"), val = tensor([12, 1, 1024, 8])]; + tensor cache_47_end_mask_0 = const()[name = string("cache_47_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_47_squeeze_mask_0 = const()[name = string("cache_47_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_47_cast_fp16 = slice_by_index(begin = cache_47_begin_0, end = cache_47_end_0, end_mask = cache_47_end_mask_0, squeeze_mask = cache_47_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_47_cast_fp16")]; + tensor input_599_axes_0 = const()[name = string("input_599_axes_0"), val = tensor([-1])]; + tensor encoder_layers_11_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_11_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(228101184)))]; + tensor encoder_layers_11_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_11_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(228103296)))]; + tensor input_599_cast_fp16 = layer_norm(axes = input_599_axes_0, beta = encoder_layers_11_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_11_norm_feed_forward1_weight_to_fp16, x = input_597_cast_fp16)[name = string("input_599_cast_fp16")]; + tensor encoder_layers_11_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(228105408))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(231251200))))[name = string("encoder_layers_11_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_11_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_11_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(231251392)))]; + tensor linear_100_cast_fp16 = linear(bias = encoder_layers_11_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_11_feed_forward1_linear1_weight_to_fp16_palettized, x = input_599_cast_fp16)[name = string("linear_100_cast_fp16")]; + tensor input_603_cast_fp16 = silu(x = linear_100_cast_fp16)[name = string("input_603_cast_fp16")]; + tensor encoder_layers_11_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(231259648))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234405440))))[name = string("encoder_layers_11_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_11_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_11_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234405632)))]; + tensor linear_101_cast_fp16 = linear(bias = encoder_layers_11_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_11_feed_forward1_linear2_weight_to_fp16_palettized, x = input_603_cast_fp16)[name = string("linear_101_cast_fp16")]; + fp16 var_2798_to_fp16 = const()[name = string("op_2798_to_fp16"), val = fp16(0x1p-1)]; + tensor var_2799_cast_fp16 = mul(x = linear_101_cast_fp16, y = var_2798_to_fp16)[name = string("op_2799_cast_fp16")]; + tensor input_609_cast_fp16 = add(x = input_597_cast_fp16, y = var_2799_cast_fp16)[name = string("input_609_cast_fp16")]; + tensor key_23_axes_0 = const()[name = string("key_23_axes_0"), val = tensor([-1])]; + tensor encoder_layers_11_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_11_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234407744)))]; + tensor encoder_layers_11_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_11_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234409856)))]; + tensor key_23_cast_fp16 = layer_norm(axes = key_23_axes_0, beta = encoder_layers_11_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_11_norm_self_att_weight_to_fp16, x = input_609_cast_fp16)[name = string("key_23_cast_fp16")]; + bool input_611_interleave_0 = const()[name = string("input_611_interleave_0"), val = bool(false)]; + tensor input_611_cast_fp16 = concat(axis = var_68, interleave = input_611_interleave_0, values = (cache_45_cast_fp16, key_23_cast_fp16))[name = string("input_611_cast_fp16")]; + tensor var_2821_begin_0 = const()[name = string("op_2821_begin_0"), val = tensor([0, 28, 0])]; + tensor var_2821_end_0 = const()[name = string("op_2821_end_0"), val = tensor([1, 42, 1024])]; + tensor var_2821_end_mask_0 = const()[name = string("op_2821_end_mask_0"), val = tensor([true, true, true])]; + tensor var_2821_cast_fp16 = slice_by_index(begin = var_2821_begin_0, end = var_2821_end_0, end_mask = var_2821_end_mask_0, x = cache_45_cast_fp16)[name = string("op_2821_cast_fp16")]; + bool var_2827_interleave_0 = const()[name = string("op_2827_interleave_0"), val = bool(false)]; + tensor var_2827_cast_fp16 = concat(axis = var_68, interleave = var_2827_interleave_0, values = (var_2821_cast_fp16, key_23_cast_fp16))[name = string("op_2827_cast_fp16")]; + tensor encoder_layers_11_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234411968))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(235198464))))[name = string("encoder_layers_11_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_11_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_11_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(235198656)))]; + tensor linear_102_cast_fp16 = linear(bias = encoder_layers_11_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_11_self_attn_linear_q_weight_to_fp16_palettized, x = key_23_cast_fp16)[name = string("linear_102_cast_fp16")]; + tensor var_2832 = const()[name = string("op_2832"), val = tensor([1, -1, 8, 128])]; + tensor q_67_cast_fp16 = reshape(shape = var_2832, x = linear_102_cast_fp16)[name = string("q_67_cast_fp16")]; + tensor encoder_layers_11_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(235200768))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(235987264))))[name = string("encoder_layers_11_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_11_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_11_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(235987456)))]; + tensor linear_103_cast_fp16 = linear(bias = encoder_layers_11_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_11_self_attn_linear_k_weight_to_fp16_palettized, x = input_611_cast_fp16)[name = string("linear_103_cast_fp16")]; + tensor var_2837 = const()[name = string("op_2837"), val = tensor([1, -1, 8, 128])]; + tensor k_45_cast_fp16 = reshape(shape = var_2837, x = linear_103_cast_fp16)[name = string("k_45_cast_fp16")]; + tensor encoder_layers_11_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(235989568))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236776064))))[name = string("encoder_layers_11_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_11_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_11_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236776256)))]; + tensor linear_104_cast_fp16 = linear(bias = encoder_layers_11_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_11_self_attn_linear_v_weight_to_fp16_palettized, x = input_611_cast_fp16)[name = string("linear_104_cast_fp16")]; + tensor var_2842 = const()[name = string("op_2842"), val = tensor([1, -1, 8, 128])]; + tensor v_23_cast_fp16 = reshape(shape = var_2842, x = linear_104_cast_fp16)[name = string("v_23_cast_fp16")]; + tensor value_31_perm_0 = const()[name = string("value_31_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_11_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_11_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236778368)))]; + tensor var_2855_cast_fp16 = add(x = q_67_cast_fp16, y = encoder_layers_11_self_attn_pos_bias_u_to_fp16)[name = string("op_2855_cast_fp16")]; + tensor encoder_layers_11_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_11_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236780480)))]; + tensor var_2857_cast_fp16 = add(x = q_67_cast_fp16, y = encoder_layers_11_self_attn_pos_bias_v_to_fp16)[name = string("op_2857_cast_fp16")]; + tensor q_with_bias_v_23_perm_0 = const()[name = string("q_with_bias_v_23_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_293_transpose_x_0 = const()[name = string("x_293_transpose_x_0"), val = bool(false)]; + bool x_293_transpose_y_0 = const()[name = string("x_293_transpose_y_0"), val = bool(false)]; + tensor op_2859_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236782592))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236924992))))[name = string("op_2859_to_fp16_quantized")]; + tensor q_with_bias_v_23_cast_fp16 = transpose(perm = q_with_bias_v_23_perm_0, x = var_2857_cast_fp16)[name = string("transpose_263")]; + tensor x_293_cast_fp16 = matmul(transpose_x = x_293_transpose_x_0, transpose_y = x_293_transpose_y_0, x = q_with_bias_v_23_cast_fp16, y = op_2859_to_fp16_quantized)[name = string("x_293_cast_fp16")]; + tensor x_295_pad_0 = const()[name = string("x_295_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_295_mode_0 = const()[name = string("x_295_mode_0"), val = string("constant")]; + fp16 const_222_to_fp16 = const()[name = string("const_222_to_fp16"), val = fp16(0x0p+0)]; + tensor x_295_cast_fp16 = pad(constant_val = const_222_to_fp16, mode = x_295_mode_0, pad = x_295_pad_0, x = x_293_cast_fp16)[name = string("x_295_cast_fp16")]; + tensor var_2867 = const()[name = string("op_2867"), val = tensor([1, 8, -1, 28])]; + tensor x_297_cast_fp16 = reshape(shape = var_2867, x = x_295_cast_fp16)[name = string("x_297_cast_fp16")]; + tensor var_2871_begin_0 = const()[name = string("op_2871_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_2871_end_0 = const()[name = string("op_2871_end_0"), val = tensor([1, 8, 140, 28])]; + tensor var_2871_end_mask_0 = const()[name = string("op_2871_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_2871_cast_fp16 = slice_by_index(begin = var_2871_begin_0, end = var_2871_end_0, end_mask = var_2871_end_mask_0, x = x_297_cast_fp16)[name = string("op_2871_cast_fp16")]; + tensor var_2872 = const()[name = string("op_2872"), val = tensor([1, 8, 28, 139])]; + tensor matrix_bd_45_cast_fp16 = reshape(shape = var_2872, x = var_2871_cast_fp16)[name = string("matrix_bd_45_cast_fp16")]; + bool matrix_ac_23_transpose_x_0 = const()[name = string("matrix_ac_23_transpose_x_0"), val = bool(false)]; + bool matrix_ac_23_transpose_y_0 = const()[name = string("matrix_ac_23_transpose_y_0"), val = bool(false)]; + tensor transpose_118_perm_0 = const()[name = string("transpose_118_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_119_perm_0 = const()[name = string("transpose_119_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_119 = transpose(perm = transpose_119_perm_0, x = k_45_cast_fp16)[name = string("transpose_261")]; + tensor transpose_118 = transpose(perm = transpose_118_perm_0, x = var_2855_cast_fp16)[name = string("transpose_262")]; + tensor matrix_ac_23_cast_fp16 = matmul(transpose_x = matrix_ac_23_transpose_x_0, transpose_y = matrix_ac_23_transpose_y_0, x = transpose_118, y = transpose_119)[name = string("matrix_ac_23_cast_fp16")]; + tensor matrix_bd_47_begin_0 = const()[name = string("matrix_bd_47_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_47_end_0 = const()[name = string("matrix_bd_47_end_0"), val = tensor([1, 8, 28, 70])]; + tensor matrix_bd_47_end_mask_0 = const()[name = string("matrix_bd_47_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_47_cast_fp16 = slice_by_index(begin = matrix_bd_47_begin_0, end = matrix_bd_47_end_0, end_mask = matrix_bd_47_end_mask_0, x = matrix_bd_45_cast_fp16)[name = string("matrix_bd_47_cast_fp16")]; + tensor var_2881_cast_fp16 = add(x = matrix_ac_23_cast_fp16, y = matrix_bd_47_cast_fp16)[name = string("op_2881_cast_fp16")]; + fp16 _inversed_scores_45_y_0_to_fp16 = const()[name = string("_inversed_scores_45_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_45_cast_fp16 = mul(x = var_2881_cast_fp16, y = _inversed_scores_45_y_0_to_fp16)[name = string("_inversed_scores_45_cast_fp16")]; + tensor scores_47_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_45_cast_fp16, cond = mask_11)[name = string("scores_47_cast_fp16")]; + tensor var_2887_cast_fp16 = softmax(axis = var_59, x = scores_47_cast_fp16)[name = string("op_2887_cast_fp16")]; + tensor input_613_cast_fp16 = select(a = var_44_to_fp16, b = var_2887_cast_fp16, cond = mask_11)[name = string("input_613_cast_fp16")]; + bool x_299_transpose_x_0 = const()[name = string("x_299_transpose_x_0"), val = bool(false)]; + bool x_299_transpose_y_0 = const()[name = string("x_299_transpose_y_0"), val = bool(false)]; + tensor value_31_cast_fp16 = transpose(perm = value_31_perm_0, x = v_23_cast_fp16)[name = string("transpose_260")]; + tensor x_299_cast_fp16 = matmul(transpose_x = x_299_transpose_x_0, transpose_y = x_299_transpose_y_0, x = input_613_cast_fp16, y = value_31_cast_fp16)[name = string("x_299_cast_fp16")]; + tensor var_2891_perm_0 = const()[name = string("op_2891_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_2892 = const()[name = string("op_2892"), val = tensor([1, -1, 1024])]; + tensor var_2891_cast_fp16 = transpose(perm = var_2891_perm_0, x = x_299_cast_fp16)[name = string("transpose_259")]; + tensor input_615_cast_fp16 = reshape(shape = var_2892, x = var_2891_cast_fp16)[name = string("input_615_cast_fp16")]; + tensor encoder_layers_11_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236925376))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(237711872))))[name = string("encoder_layers_11_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_11_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_11_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(237712064)))]; + tensor linear_106_cast_fp16 = linear(bias = encoder_layers_11_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_11_self_attn_linear_out_weight_to_fp16_palettized, x = input_615_cast_fp16)[name = string("linear_106_cast_fp16")]; + tensor input_619_cast_fp16 = add(x = input_609_cast_fp16, y = linear_106_cast_fp16)[name = string("input_619_cast_fp16")]; + tensor x_303_axes_0 = const()[name = string("x_303_axes_0"), val = tensor([-1])]; + tensor encoder_layers_11_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_11_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(237714176)))]; + tensor encoder_layers_11_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_11_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(237716288)))]; + tensor x_303_cast_fp16 = layer_norm(axes = x_303_axes_0, beta = encoder_layers_11_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_11_norm_conv_weight_to_fp16, x = input_619_cast_fp16)[name = string("x_303_cast_fp16")]; + tensor input_621_perm_0 = const()[name = string("input_621_perm_0"), val = tensor([0, 2, 1])]; + string input_623_pad_type_0 = const()[name = string("input_623_pad_type_0"), val = string("valid")]; + tensor input_623_strides_0 = const()[name = string("input_623_strides_0"), val = tensor([1])]; + tensor input_623_pad_0 = const()[name = string("input_623_pad_0"), val = tensor([0, 0])]; + tensor input_623_dilations_0 = const()[name = string("input_623_dilations_0"), val = tensor([1])]; + int32 input_623_groups_0 = const()[name = string("input_623_groups_0"), val = int32(1)]; + tensor encoder_layers_11_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(237718400))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(239815616))))[name = string("encoder_layers_11_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_621_cast_fp16 = transpose(perm = input_621_perm_0, x = x_303_cast_fp16)[name = string("transpose_258")]; + tensor input_623_cast_fp16 = conv(dilations = input_623_dilations_0, groups = input_623_groups_0, pad = input_623_pad_0, pad_type = input_623_pad_type_0, strides = input_623_strides_0, weight = encoder_layers_11_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_621_cast_fp16)[name = string("input_623_cast_fp16")]; + int32 x_305_split_num_splits_0 = const()[name = string("x_305_split_num_splits_0"), val = int32(2)]; + int32 x_305_split_axis_0 = const()[name = string("x_305_split_axis_0"), val = int32(1)]; + tensor x_305_split_cast_fp16_0, tensor x_305_split_cast_fp16_1 = split(axis = x_305_split_axis_0, num_splits = x_305_split_num_splits_0, x = input_623_cast_fp16)[name = string("x_305_split_cast_fp16")]; + tensor x_305_split_1_sigmoid_cast_fp16 = sigmoid(x = x_305_split_cast_fp16_1)[name = string("x_305_split_1_sigmoid_cast_fp16")]; + tensor x_305_cast_fp16 = mul(x = x_305_split_cast_fp16_0, y = x_305_split_1_sigmoid_cast_fp16)[name = string("x_305_cast_fp16")]; + tensor input_625_cast_fp16 = select(a = var_44_to_fp16, b = x_305_cast_fp16, cond = var_575)[name = string("input_625_cast_fp16")]; + bool new_x_47_interleave_0 = const()[name = string("new_x_47_interleave_0"), val = bool(false)]; + tensor new_x_47_cast_fp16 = concat(axis = var_59, interleave = new_x_47_interleave_0, values = (cache_47_cast_fp16, input_625_cast_fp16))[name = string("new_x_47_cast_fp16")]; + tensor var_2931_begin_0 = const()[name = string("op_2931_begin_0"), val = tensor([0, 0, 28])]; + tensor var_2931_end_0 = const()[name = string("op_2931_end_0"), val = tensor([1, 1024, 36])]; + tensor var_2931_end_mask_0 = const()[name = string("op_2931_end_mask_0"), val = tensor([true, true, true])]; + tensor var_2931_cast_fp16 = slice_by_index(begin = var_2931_begin_0, end = var_2931_end_0, end_mask = var_2931_end_mask_0, x = new_x_47_cast_fp16)[name = string("op_2931_cast_fp16")]; + string x_307_pad_type_0 = const()[name = string("x_307_pad_type_0"), val = string("valid")]; + int32 x_307_groups_0 = const()[name = string("x_307_groups_0"), val = int32(1024)]; + tensor x_307_strides_0 = const()[name = string("x_307_strides_0"), val = tensor([1])]; + tensor x_307_pad_0 = const()[name = string("x_307_pad_0"), val = tensor([0, 0])]; + tensor x_307_dilations_0 = const()[name = string("x_307_dilations_0"), val = tensor([1])]; + tensor encoder_layers_11_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(239819776))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(239829056))))[name = string("encoder_layers_11_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_307_cast_fp16 = conv(dilations = x_307_dilations_0, groups = x_307_groups_0, pad = x_307_pad_0, pad_type = x_307_pad_type_0, strides = x_307_strides_0, weight = encoder_layers_11_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_47_cast_fp16)[name = string("x_307_cast_fp16")]; + tensor input_627_perm_0 = const()[name = string("input_627_perm_0"), val = tensor([0, 2, 1])]; + tensor x_309_axes_0 = const()[name = string("x_309_axes_0"), val = tensor([-1])]; + tensor encoder_layers_11_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_11_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(239831168)))]; + tensor encoder_layers_11_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_11_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(239833280)))]; + tensor input_627_cast_fp16 = transpose(perm = input_627_perm_0, x = x_307_cast_fp16)[name = string("transpose_257")]; + tensor x_309_cast_fp16 = layer_norm(axes = x_309_axes_0, beta = encoder_layers_11_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_11_conv_batch_norm_weight_to_fp16, x = input_627_cast_fp16)[name = string("x_309_cast_fp16")]; + tensor input_629_perm_0 = const()[name = string("input_629_perm_0"), val = tensor([0, 2, 1])]; + tensor input_629_cast_fp16 = transpose(perm = input_629_perm_0, x = x_309_cast_fp16)[name = string("transpose_256")]; + tensor input_631_cast_fp16 = silu(x = input_629_cast_fp16)[name = string("input_631_cast_fp16")]; + string x_311_pad_type_0 = const()[name = string("x_311_pad_type_0"), val = string("valid")]; + tensor x_311_strides_0 = const()[name = string("x_311_strides_0"), val = tensor([1])]; + tensor x_311_pad_0 = const()[name = string("x_311_pad_0"), val = tensor([0, 0])]; + tensor x_311_dilations_0 = const()[name = string("x_311_dilations_0"), val = tensor([1])]; + int32 x_311_groups_0 = const()[name = string("x_311_groups_0"), val = int32(1)]; + tensor encoder_layers_11_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(239835392))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240884032))))[name = string("encoder_layers_11_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_311_cast_fp16 = conv(dilations = x_311_dilations_0, groups = x_311_groups_0, pad = x_311_pad_0, pad_type = x_311_pad_type_0, strides = x_311_strides_0, weight = encoder_layers_11_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_631_cast_fp16)[name = string("x_311_cast_fp16")]; + tensor input_633_perm_0 = const()[name = string("input_633_perm_0"), val = tensor([0, 2, 1])]; + tensor input_633_cast_fp16 = transpose(perm = input_633_perm_0, x = x_311_cast_fp16)[name = string("transpose_255")]; + tensor input_635_cast_fp16 = add(x = input_619_cast_fp16, y = input_633_cast_fp16)[name = string("input_635_cast_fp16")]; + tensor input_637_axes_0 = const()[name = string("input_637_axes_0"), val = tensor([-1])]; + tensor encoder_layers_11_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_11_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240886144)))]; + tensor encoder_layers_11_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_11_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240888256)))]; + tensor input_637_cast_fp16 = layer_norm(axes = input_637_axes_0, beta = encoder_layers_11_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_11_norm_feed_forward2_weight_to_fp16, x = input_635_cast_fp16)[name = string("input_637_cast_fp16")]; + tensor encoder_layers_11_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240890368))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(244036160))))[name = string("encoder_layers_11_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_11_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_11_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(244036352)))]; + tensor linear_107_cast_fp16 = linear(bias = encoder_layers_11_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_11_feed_forward2_linear1_weight_to_fp16_palettized, x = input_637_cast_fp16)[name = string("linear_107_cast_fp16")]; + tensor input_641_cast_fp16 = silu(x = linear_107_cast_fp16)[name = string("input_641_cast_fp16")]; + tensor encoder_layers_11_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(244044608))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(247190400))))[name = string("encoder_layers_11_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_11_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_11_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(247190592)))]; + tensor linear_108_cast_fp16 = linear(bias = encoder_layers_11_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_11_feed_forward2_linear2_weight_to_fp16_palettized, x = input_641_cast_fp16)[name = string("linear_108_cast_fp16")]; + fp16 var_2974_to_fp16 = const()[name = string("op_2974_to_fp16"), val = fp16(0x1p-1)]; + tensor var_2975_cast_fp16 = mul(x = linear_108_cast_fp16, y = var_2974_to_fp16)[name = string("op_2975_cast_fp16")]; + tensor input_647_cast_fp16 = add(x = input_635_cast_fp16, y = var_2975_cast_fp16)[name = string("input_647_cast_fp16")]; + tensor input_649_axes_0 = const()[name = string("input_649_axes_0"), val = tensor([-1])]; + tensor encoder_layers_11_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_11_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(247192704)))]; + tensor encoder_layers_11_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_11_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(247194816)))]; + tensor input_649_cast_fp16 = layer_norm(axes = input_649_axes_0, beta = encoder_layers_11_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_11_norm_out_weight_to_fp16, x = input_647_cast_fp16)[name = string("input_649_cast_fp16")]; + tensor cache_49_begin_0 = const()[name = string("cache_49_begin_0"), val = tensor([12, 0, 0, 0])]; + tensor cache_49_end_0 = const()[name = string("cache_49_end_0"), val = tensor([13, 1, 42, 1024])]; + tensor cache_49_end_mask_0 = const()[name = string("cache_49_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_49_squeeze_mask_0 = const()[name = string("cache_49_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_49_cast_fp16 = slice_by_index(begin = cache_49_begin_0, end = cache_49_end_0, end_mask = cache_49_end_mask_0, squeeze_mask = cache_49_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_49_cast_fp16")]; + tensor cache_51_begin_0 = const()[name = string("cache_51_begin_0"), val = tensor([12, 0, 0, 0])]; + tensor cache_51_end_0 = const()[name = string("cache_51_end_0"), val = tensor([13, 1, 1024, 8])]; + tensor cache_51_end_mask_0 = const()[name = string("cache_51_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_51_squeeze_mask_0 = const()[name = string("cache_51_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_51_cast_fp16 = slice_by_index(begin = cache_51_begin_0, end = cache_51_end_0, end_mask = cache_51_end_mask_0, squeeze_mask = cache_51_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_51_cast_fp16")]; + tensor input_651_axes_0 = const()[name = string("input_651_axes_0"), val = tensor([-1])]; + tensor encoder_layers_12_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_12_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(247196928)))]; + tensor encoder_layers_12_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_12_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(247199040)))]; + tensor input_651_cast_fp16 = layer_norm(axes = input_651_axes_0, beta = encoder_layers_12_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_12_norm_feed_forward1_weight_to_fp16, x = input_649_cast_fp16)[name = string("input_651_cast_fp16")]; + tensor encoder_layers_12_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(247201152))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(250346944))))[name = string("encoder_layers_12_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_12_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_12_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(250347136)))]; + tensor linear_109_cast_fp16 = linear(bias = encoder_layers_12_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_12_feed_forward1_linear1_weight_to_fp16_palettized, x = input_651_cast_fp16)[name = string("linear_109_cast_fp16")]; + tensor input_655_cast_fp16 = silu(x = linear_109_cast_fp16)[name = string("input_655_cast_fp16")]; + tensor encoder_layers_12_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(250355392))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253501184))))[name = string("encoder_layers_12_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_12_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_12_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253501376)))]; + tensor linear_110_cast_fp16 = linear(bias = encoder_layers_12_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_12_feed_forward1_linear2_weight_to_fp16_palettized, x = input_655_cast_fp16)[name = string("linear_110_cast_fp16")]; + fp16 var_3011_to_fp16 = const()[name = string("op_3011_to_fp16"), val = fp16(0x1p-1)]; + tensor var_3012_cast_fp16 = mul(x = linear_110_cast_fp16, y = var_3011_to_fp16)[name = string("op_3012_cast_fp16")]; + tensor input_661_cast_fp16 = add(x = input_649_cast_fp16, y = var_3012_cast_fp16)[name = string("input_661_cast_fp16")]; + tensor key_25_axes_0 = const()[name = string("key_25_axes_0"), val = tensor([-1])]; + tensor encoder_layers_12_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_12_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253503488)))]; + tensor encoder_layers_12_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_12_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253505600)))]; + tensor key_25_cast_fp16 = layer_norm(axes = key_25_axes_0, beta = encoder_layers_12_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_12_norm_self_att_weight_to_fp16, x = input_661_cast_fp16)[name = string("key_25_cast_fp16")]; + bool input_663_interleave_0 = const()[name = string("input_663_interleave_0"), val = bool(false)]; + tensor input_663_cast_fp16 = concat(axis = var_68, interleave = input_663_interleave_0, values = (cache_49_cast_fp16, key_25_cast_fp16))[name = string("input_663_cast_fp16")]; + tensor var_3034_begin_0 = const()[name = string("op_3034_begin_0"), val = tensor([0, 28, 0])]; + tensor var_3034_end_0 = const()[name = string("op_3034_end_0"), val = tensor([1, 42, 1024])]; + tensor var_3034_end_mask_0 = const()[name = string("op_3034_end_mask_0"), val = tensor([true, true, true])]; + tensor var_3034_cast_fp16 = slice_by_index(begin = var_3034_begin_0, end = var_3034_end_0, end_mask = var_3034_end_mask_0, x = cache_49_cast_fp16)[name = string("op_3034_cast_fp16")]; + bool var_3040_interleave_0 = const()[name = string("op_3040_interleave_0"), val = bool(false)]; + tensor var_3040_cast_fp16 = concat(axis = var_68, interleave = var_3040_interleave_0, values = (var_3034_cast_fp16, key_25_cast_fp16))[name = string("op_3040_cast_fp16")]; + tensor encoder_layers_12_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253507712))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(254294208))))[name = string("encoder_layers_12_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_12_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_12_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(254294400)))]; + tensor linear_111_cast_fp16 = linear(bias = encoder_layers_12_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_12_self_attn_linear_q_weight_to_fp16_palettized, x = key_25_cast_fp16)[name = string("linear_111_cast_fp16")]; + tensor var_3045 = const()[name = string("op_3045"), val = tensor([1, -1, 8, 128])]; + tensor q_73_cast_fp16 = reshape(shape = var_3045, x = linear_111_cast_fp16)[name = string("q_73_cast_fp16")]; + tensor encoder_layers_12_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(254296512))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255083008))))[name = string("encoder_layers_12_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_12_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_12_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255083200)))]; + tensor linear_112_cast_fp16 = linear(bias = encoder_layers_12_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_12_self_attn_linear_k_weight_to_fp16_palettized, x = input_663_cast_fp16)[name = string("linear_112_cast_fp16")]; + tensor var_3050 = const()[name = string("op_3050"), val = tensor([1, -1, 8, 128])]; + tensor k_49_cast_fp16 = reshape(shape = var_3050, x = linear_112_cast_fp16)[name = string("k_49_cast_fp16")]; + tensor encoder_layers_12_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255085312))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255871808))))[name = string("encoder_layers_12_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_12_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_12_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255872000)))]; + tensor linear_113_cast_fp16 = linear(bias = encoder_layers_12_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_12_self_attn_linear_v_weight_to_fp16_palettized, x = input_663_cast_fp16)[name = string("linear_113_cast_fp16")]; + tensor var_3055 = const()[name = string("op_3055"), val = tensor([1, -1, 8, 128])]; + tensor v_25_cast_fp16 = reshape(shape = var_3055, x = linear_113_cast_fp16)[name = string("v_25_cast_fp16")]; + tensor value_33_perm_0 = const()[name = string("value_33_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_12_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_12_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255874112)))]; + tensor var_3068_cast_fp16 = add(x = q_73_cast_fp16, y = encoder_layers_12_self_attn_pos_bias_u_to_fp16)[name = string("op_3068_cast_fp16")]; + tensor encoder_layers_12_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_12_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255876224)))]; + tensor var_3070_cast_fp16 = add(x = q_73_cast_fp16, y = encoder_layers_12_self_attn_pos_bias_v_to_fp16)[name = string("op_3070_cast_fp16")]; + tensor q_with_bias_v_25_perm_0 = const()[name = string("q_with_bias_v_25_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_319_transpose_x_0 = const()[name = string("x_319_transpose_x_0"), val = bool(false)]; + bool x_319_transpose_y_0 = const()[name = string("x_319_transpose_y_0"), val = bool(false)]; + tensor op_3072_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255878336))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256020736))))[name = string("op_3072_to_fp16_quantized")]; + tensor q_with_bias_v_25_cast_fp16 = transpose(perm = q_with_bias_v_25_perm_0, x = var_3070_cast_fp16)[name = string("transpose_254")]; + tensor x_319_cast_fp16 = matmul(transpose_x = x_319_transpose_x_0, transpose_y = x_319_transpose_y_0, x = q_with_bias_v_25_cast_fp16, y = op_3072_to_fp16_quantized)[name = string("x_319_cast_fp16")]; + tensor x_321_pad_0 = const()[name = string("x_321_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_321_mode_0 = const()[name = string("x_321_mode_0"), val = string("constant")]; + fp16 const_235_to_fp16 = const()[name = string("const_235_to_fp16"), val = fp16(0x0p+0)]; + tensor x_321_cast_fp16 = pad(constant_val = const_235_to_fp16, mode = x_321_mode_0, pad = x_321_pad_0, x = x_319_cast_fp16)[name = string("x_321_cast_fp16")]; + tensor var_3080 = const()[name = string("op_3080"), val = tensor([1, 8, -1, 28])]; + tensor x_323_cast_fp16 = reshape(shape = var_3080, x = x_321_cast_fp16)[name = string("x_323_cast_fp16")]; + tensor var_3084_begin_0 = const()[name = string("op_3084_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_3084_end_0 = const()[name = string("op_3084_end_0"), val = tensor([1, 8, 140, 28])]; + tensor var_3084_end_mask_0 = const()[name = string("op_3084_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_3084_cast_fp16 = slice_by_index(begin = var_3084_begin_0, end = var_3084_end_0, end_mask = var_3084_end_mask_0, x = x_323_cast_fp16)[name = string("op_3084_cast_fp16")]; + tensor var_3085 = const()[name = string("op_3085"), val = tensor([1, 8, 28, 139])]; + tensor matrix_bd_49_cast_fp16 = reshape(shape = var_3085, x = var_3084_cast_fp16)[name = string("matrix_bd_49_cast_fp16")]; + bool matrix_ac_25_transpose_x_0 = const()[name = string("matrix_ac_25_transpose_x_0"), val = bool(false)]; + bool matrix_ac_25_transpose_y_0 = const()[name = string("matrix_ac_25_transpose_y_0"), val = bool(false)]; + tensor transpose_120_perm_0 = const()[name = string("transpose_120_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_121_perm_0 = const()[name = string("transpose_121_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_121 = transpose(perm = transpose_121_perm_0, x = k_49_cast_fp16)[name = string("transpose_252")]; + tensor transpose_120 = transpose(perm = transpose_120_perm_0, x = var_3068_cast_fp16)[name = string("transpose_253")]; + tensor matrix_ac_25_cast_fp16 = matmul(transpose_x = matrix_ac_25_transpose_x_0, transpose_y = matrix_ac_25_transpose_y_0, x = transpose_120, y = transpose_121)[name = string("matrix_ac_25_cast_fp16")]; + tensor matrix_bd_51_begin_0 = const()[name = string("matrix_bd_51_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_51_end_0 = const()[name = string("matrix_bd_51_end_0"), val = tensor([1, 8, 28, 70])]; + tensor matrix_bd_51_end_mask_0 = const()[name = string("matrix_bd_51_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_51_cast_fp16 = slice_by_index(begin = matrix_bd_51_begin_0, end = matrix_bd_51_end_0, end_mask = matrix_bd_51_end_mask_0, x = matrix_bd_49_cast_fp16)[name = string("matrix_bd_51_cast_fp16")]; + tensor var_3094_cast_fp16 = add(x = matrix_ac_25_cast_fp16, y = matrix_bd_51_cast_fp16)[name = string("op_3094_cast_fp16")]; + fp16 _inversed_scores_49_y_0_to_fp16 = const()[name = string("_inversed_scores_49_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_49_cast_fp16 = mul(x = var_3094_cast_fp16, y = _inversed_scores_49_y_0_to_fp16)[name = string("_inversed_scores_49_cast_fp16")]; + tensor scores_51_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_49_cast_fp16, cond = mask_11)[name = string("scores_51_cast_fp16")]; + tensor var_3100_cast_fp16 = softmax(axis = var_59, x = scores_51_cast_fp16)[name = string("op_3100_cast_fp16")]; + tensor input_665_cast_fp16 = select(a = var_44_to_fp16, b = var_3100_cast_fp16, cond = mask_11)[name = string("input_665_cast_fp16")]; + bool x_325_transpose_x_0 = const()[name = string("x_325_transpose_x_0"), val = bool(false)]; + bool x_325_transpose_y_0 = const()[name = string("x_325_transpose_y_0"), val = bool(false)]; + tensor value_33_cast_fp16 = transpose(perm = value_33_perm_0, x = v_25_cast_fp16)[name = string("transpose_251")]; + tensor x_325_cast_fp16 = matmul(transpose_x = x_325_transpose_x_0, transpose_y = x_325_transpose_y_0, x = input_665_cast_fp16, y = value_33_cast_fp16)[name = string("x_325_cast_fp16")]; + tensor var_3104_perm_0 = const()[name = string("op_3104_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_3105 = const()[name = string("op_3105"), val = tensor([1, -1, 1024])]; + tensor var_3104_cast_fp16 = transpose(perm = var_3104_perm_0, x = x_325_cast_fp16)[name = string("transpose_250")]; + tensor input_667_cast_fp16 = reshape(shape = var_3105, x = var_3104_cast_fp16)[name = string("input_667_cast_fp16")]; + tensor encoder_layers_12_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256021120))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256807616))))[name = string("encoder_layers_12_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_12_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_12_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256807808)))]; + tensor linear_115_cast_fp16 = linear(bias = encoder_layers_12_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_12_self_attn_linear_out_weight_to_fp16_palettized, x = input_667_cast_fp16)[name = string("linear_115_cast_fp16")]; + tensor input_671_cast_fp16 = add(x = input_661_cast_fp16, y = linear_115_cast_fp16)[name = string("input_671_cast_fp16")]; + tensor x_329_axes_0 = const()[name = string("x_329_axes_0"), val = tensor([-1])]; + tensor encoder_layers_12_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_12_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256809920)))]; + tensor encoder_layers_12_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_12_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256812032)))]; + tensor x_329_cast_fp16 = layer_norm(axes = x_329_axes_0, beta = encoder_layers_12_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_12_norm_conv_weight_to_fp16, x = input_671_cast_fp16)[name = string("x_329_cast_fp16")]; + tensor input_673_perm_0 = const()[name = string("input_673_perm_0"), val = tensor([0, 2, 1])]; + string input_675_pad_type_0 = const()[name = string("input_675_pad_type_0"), val = string("valid")]; + tensor input_675_strides_0 = const()[name = string("input_675_strides_0"), val = tensor([1])]; + tensor input_675_pad_0 = const()[name = string("input_675_pad_0"), val = tensor([0, 0])]; + tensor input_675_dilations_0 = const()[name = string("input_675_dilations_0"), val = tensor([1])]; + int32 input_675_groups_0 = const()[name = string("input_675_groups_0"), val = int32(1)]; + tensor encoder_layers_12_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256814144))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(258911360))))[name = string("encoder_layers_12_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_673_cast_fp16 = transpose(perm = input_673_perm_0, x = x_329_cast_fp16)[name = string("transpose_249")]; + tensor input_675_cast_fp16 = conv(dilations = input_675_dilations_0, groups = input_675_groups_0, pad = input_675_pad_0, pad_type = input_675_pad_type_0, strides = input_675_strides_0, weight = encoder_layers_12_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_673_cast_fp16)[name = string("input_675_cast_fp16")]; + int32 x_331_split_num_splits_0 = const()[name = string("x_331_split_num_splits_0"), val = int32(2)]; + int32 x_331_split_axis_0 = const()[name = string("x_331_split_axis_0"), val = int32(1)]; + tensor x_331_split_cast_fp16_0, tensor x_331_split_cast_fp16_1 = split(axis = x_331_split_axis_0, num_splits = x_331_split_num_splits_0, x = input_675_cast_fp16)[name = string("x_331_split_cast_fp16")]; + tensor x_331_split_1_sigmoid_cast_fp16 = sigmoid(x = x_331_split_cast_fp16_1)[name = string("x_331_split_1_sigmoid_cast_fp16")]; + tensor x_331_cast_fp16 = mul(x = x_331_split_cast_fp16_0, y = x_331_split_1_sigmoid_cast_fp16)[name = string("x_331_cast_fp16")]; + tensor input_677_cast_fp16 = select(a = var_44_to_fp16, b = x_331_cast_fp16, cond = var_575)[name = string("input_677_cast_fp16")]; + bool new_x_51_interleave_0 = const()[name = string("new_x_51_interleave_0"), val = bool(false)]; + tensor new_x_51_cast_fp16 = concat(axis = var_59, interleave = new_x_51_interleave_0, values = (cache_51_cast_fp16, input_677_cast_fp16))[name = string("new_x_51_cast_fp16")]; + tensor var_3144_begin_0 = const()[name = string("op_3144_begin_0"), val = tensor([0, 0, 28])]; + tensor var_3144_end_0 = const()[name = string("op_3144_end_0"), val = tensor([1, 1024, 36])]; + tensor var_3144_end_mask_0 = const()[name = string("op_3144_end_mask_0"), val = tensor([true, true, true])]; + tensor var_3144_cast_fp16 = slice_by_index(begin = var_3144_begin_0, end = var_3144_end_0, end_mask = var_3144_end_mask_0, x = new_x_51_cast_fp16)[name = string("op_3144_cast_fp16")]; + string x_333_pad_type_0 = const()[name = string("x_333_pad_type_0"), val = string("valid")]; + int32 x_333_groups_0 = const()[name = string("x_333_groups_0"), val = int32(1024)]; + tensor x_333_strides_0 = const()[name = string("x_333_strides_0"), val = tensor([1])]; + tensor x_333_pad_0 = const()[name = string("x_333_pad_0"), val = tensor([0, 0])]; + tensor x_333_dilations_0 = const()[name = string("x_333_dilations_0"), val = tensor([1])]; + tensor encoder_layers_12_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(258915520))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(258924800))))[name = string("encoder_layers_12_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_333_cast_fp16 = conv(dilations = x_333_dilations_0, groups = x_333_groups_0, pad = x_333_pad_0, pad_type = x_333_pad_type_0, strides = x_333_strides_0, weight = encoder_layers_12_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_51_cast_fp16)[name = string("x_333_cast_fp16")]; + tensor input_679_perm_0 = const()[name = string("input_679_perm_0"), val = tensor([0, 2, 1])]; + tensor x_335_axes_0 = const()[name = string("x_335_axes_0"), val = tensor([-1])]; + tensor encoder_layers_12_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_12_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(258926912)))]; + tensor encoder_layers_12_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_12_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(258929024)))]; + tensor input_679_cast_fp16 = transpose(perm = input_679_perm_0, x = x_333_cast_fp16)[name = string("transpose_248")]; + tensor x_335_cast_fp16 = layer_norm(axes = x_335_axes_0, beta = encoder_layers_12_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_12_conv_batch_norm_weight_to_fp16, x = input_679_cast_fp16)[name = string("x_335_cast_fp16")]; + tensor input_681_perm_0 = const()[name = string("input_681_perm_0"), val = tensor([0, 2, 1])]; + tensor input_681_cast_fp16 = transpose(perm = input_681_perm_0, x = x_335_cast_fp16)[name = string("transpose_247")]; + tensor input_683_cast_fp16 = silu(x = input_681_cast_fp16)[name = string("input_683_cast_fp16")]; + string x_337_pad_type_0 = const()[name = string("x_337_pad_type_0"), val = string("valid")]; + tensor x_337_strides_0 = const()[name = string("x_337_strides_0"), val = tensor([1])]; + tensor x_337_pad_0 = const()[name = string("x_337_pad_0"), val = tensor([0, 0])]; + tensor x_337_dilations_0 = const()[name = string("x_337_dilations_0"), val = tensor([1])]; + int32 x_337_groups_0 = const()[name = string("x_337_groups_0"), val = int32(1)]; + tensor encoder_layers_12_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(258931136))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259979776))))[name = string("encoder_layers_12_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_337_cast_fp16 = conv(dilations = x_337_dilations_0, groups = x_337_groups_0, pad = x_337_pad_0, pad_type = x_337_pad_type_0, strides = x_337_strides_0, weight = encoder_layers_12_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_683_cast_fp16)[name = string("x_337_cast_fp16")]; + tensor input_685_perm_0 = const()[name = string("input_685_perm_0"), val = tensor([0, 2, 1])]; + tensor input_685_cast_fp16 = transpose(perm = input_685_perm_0, x = x_337_cast_fp16)[name = string("transpose_246")]; + tensor input_687_cast_fp16 = add(x = input_671_cast_fp16, y = input_685_cast_fp16)[name = string("input_687_cast_fp16")]; + tensor input_689_axes_0 = const()[name = string("input_689_axes_0"), val = tensor([-1])]; + tensor encoder_layers_12_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_12_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259981888)))]; + tensor encoder_layers_12_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_12_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259984000)))]; + tensor input_689_cast_fp16 = layer_norm(axes = input_689_axes_0, beta = encoder_layers_12_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_12_norm_feed_forward2_weight_to_fp16, x = input_687_cast_fp16)[name = string("input_689_cast_fp16")]; + tensor encoder_layers_12_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259986112))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(263131904))))[name = string("encoder_layers_12_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_12_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_12_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(263132096)))]; + tensor linear_116_cast_fp16 = linear(bias = encoder_layers_12_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_12_feed_forward2_linear1_weight_to_fp16_palettized, x = input_689_cast_fp16)[name = string("linear_116_cast_fp16")]; + tensor input_693_cast_fp16 = silu(x = linear_116_cast_fp16)[name = string("input_693_cast_fp16")]; + tensor encoder_layers_12_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(263140352))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(266286144))))[name = string("encoder_layers_12_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_12_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_12_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(266286336)))]; + tensor linear_117_cast_fp16 = linear(bias = encoder_layers_12_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_12_feed_forward2_linear2_weight_to_fp16_palettized, x = input_693_cast_fp16)[name = string("linear_117_cast_fp16")]; + fp16 var_3187_to_fp16 = const()[name = string("op_3187_to_fp16"), val = fp16(0x1p-1)]; + tensor var_3188_cast_fp16 = mul(x = linear_117_cast_fp16, y = var_3187_to_fp16)[name = string("op_3188_cast_fp16")]; + tensor input_699_cast_fp16 = add(x = input_687_cast_fp16, y = var_3188_cast_fp16)[name = string("input_699_cast_fp16")]; + tensor input_701_axes_0 = const()[name = string("input_701_axes_0"), val = tensor([-1])]; + tensor encoder_layers_12_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_12_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(266288448)))]; + tensor encoder_layers_12_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_12_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(266290560)))]; + tensor input_701_cast_fp16 = layer_norm(axes = input_701_axes_0, beta = encoder_layers_12_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_12_norm_out_weight_to_fp16, x = input_699_cast_fp16)[name = string("input_701_cast_fp16")]; + tensor cache_53_begin_0 = const()[name = string("cache_53_begin_0"), val = tensor([13, 0, 0, 0])]; + tensor cache_53_end_0 = const()[name = string("cache_53_end_0"), val = tensor([14, 1, 42, 1024])]; + tensor cache_53_end_mask_0 = const()[name = string("cache_53_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_53_squeeze_mask_0 = const()[name = string("cache_53_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_53_cast_fp16 = slice_by_index(begin = cache_53_begin_0, end = cache_53_end_0, end_mask = cache_53_end_mask_0, squeeze_mask = cache_53_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_53_cast_fp16")]; + tensor cache_55_begin_0 = const()[name = string("cache_55_begin_0"), val = tensor([13, 0, 0, 0])]; + tensor cache_55_end_0 = const()[name = string("cache_55_end_0"), val = tensor([14, 1, 1024, 8])]; + tensor cache_55_end_mask_0 = const()[name = string("cache_55_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_55_squeeze_mask_0 = const()[name = string("cache_55_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_55_cast_fp16 = slice_by_index(begin = cache_55_begin_0, end = cache_55_end_0, end_mask = cache_55_end_mask_0, squeeze_mask = cache_55_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_55_cast_fp16")]; + tensor input_703_axes_0 = const()[name = string("input_703_axes_0"), val = tensor([-1])]; + tensor encoder_layers_13_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_13_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(266292672)))]; + tensor encoder_layers_13_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_13_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(266294784)))]; + tensor input_703_cast_fp16 = layer_norm(axes = input_703_axes_0, beta = encoder_layers_13_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_13_norm_feed_forward1_weight_to_fp16, x = input_701_cast_fp16)[name = string("input_703_cast_fp16")]; + tensor encoder_layers_13_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(266296896))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(269442688))))[name = string("encoder_layers_13_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_13_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_13_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(269442880)))]; + tensor linear_118_cast_fp16 = linear(bias = encoder_layers_13_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_13_feed_forward1_linear1_weight_to_fp16_palettized, x = input_703_cast_fp16)[name = string("linear_118_cast_fp16")]; + tensor input_707_cast_fp16 = silu(x = linear_118_cast_fp16)[name = string("input_707_cast_fp16")]; + tensor encoder_layers_13_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(269451136))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272596928))))[name = string("encoder_layers_13_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_13_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_13_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272597120)))]; + tensor linear_119_cast_fp16 = linear(bias = encoder_layers_13_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_13_feed_forward1_linear2_weight_to_fp16_palettized, x = input_707_cast_fp16)[name = string("linear_119_cast_fp16")]; + fp16 var_3224_to_fp16 = const()[name = string("op_3224_to_fp16"), val = fp16(0x1p-1)]; + tensor var_3225_cast_fp16 = mul(x = linear_119_cast_fp16, y = var_3224_to_fp16)[name = string("op_3225_cast_fp16")]; + tensor input_713_cast_fp16 = add(x = input_701_cast_fp16, y = var_3225_cast_fp16)[name = string("input_713_cast_fp16")]; + tensor key_27_axes_0 = const()[name = string("key_27_axes_0"), val = tensor([-1])]; + tensor encoder_layers_13_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_13_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272599232)))]; + tensor encoder_layers_13_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_13_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272601344)))]; + tensor key_27_cast_fp16 = layer_norm(axes = key_27_axes_0, beta = encoder_layers_13_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_13_norm_self_att_weight_to_fp16, x = input_713_cast_fp16)[name = string("key_27_cast_fp16")]; + bool input_715_interleave_0 = const()[name = string("input_715_interleave_0"), val = bool(false)]; + tensor input_715_cast_fp16 = concat(axis = var_68, interleave = input_715_interleave_0, values = (cache_53_cast_fp16, key_27_cast_fp16))[name = string("input_715_cast_fp16")]; + tensor var_3247_begin_0 = const()[name = string("op_3247_begin_0"), val = tensor([0, 28, 0])]; + tensor var_3247_end_0 = const()[name = string("op_3247_end_0"), val = tensor([1, 42, 1024])]; + tensor var_3247_end_mask_0 = const()[name = string("op_3247_end_mask_0"), val = tensor([true, true, true])]; + tensor var_3247_cast_fp16 = slice_by_index(begin = var_3247_begin_0, end = var_3247_end_0, end_mask = var_3247_end_mask_0, x = cache_53_cast_fp16)[name = string("op_3247_cast_fp16")]; + bool var_3253_interleave_0 = const()[name = string("op_3253_interleave_0"), val = bool(false)]; + tensor var_3253_cast_fp16 = concat(axis = var_68, interleave = var_3253_interleave_0, values = (var_3247_cast_fp16, key_27_cast_fp16))[name = string("op_3253_cast_fp16")]; + tensor encoder_layers_13_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272603456))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(273389952))))[name = string("encoder_layers_13_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_13_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_13_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(273390144)))]; + tensor linear_120_cast_fp16 = linear(bias = encoder_layers_13_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_13_self_attn_linear_q_weight_to_fp16_palettized, x = key_27_cast_fp16)[name = string("linear_120_cast_fp16")]; + tensor var_3258 = const()[name = string("op_3258"), val = tensor([1, -1, 8, 128])]; + tensor q_79_cast_fp16 = reshape(shape = var_3258, x = linear_120_cast_fp16)[name = string("q_79_cast_fp16")]; + tensor encoder_layers_13_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(273392256))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274178752))))[name = string("encoder_layers_13_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_13_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_13_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274178944)))]; + tensor linear_121_cast_fp16 = linear(bias = encoder_layers_13_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_13_self_attn_linear_k_weight_to_fp16_palettized, x = input_715_cast_fp16)[name = string("linear_121_cast_fp16")]; + tensor var_3263 = const()[name = string("op_3263"), val = tensor([1, -1, 8, 128])]; + tensor k_53_cast_fp16 = reshape(shape = var_3263, x = linear_121_cast_fp16)[name = string("k_53_cast_fp16")]; + tensor encoder_layers_13_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274181056))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274967552))))[name = string("encoder_layers_13_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_13_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_13_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274967744)))]; + tensor linear_122_cast_fp16 = linear(bias = encoder_layers_13_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_13_self_attn_linear_v_weight_to_fp16_palettized, x = input_715_cast_fp16)[name = string("linear_122_cast_fp16")]; + tensor var_3268 = const()[name = string("op_3268"), val = tensor([1, -1, 8, 128])]; + tensor v_27_cast_fp16 = reshape(shape = var_3268, x = linear_122_cast_fp16)[name = string("v_27_cast_fp16")]; + tensor value_35_perm_0 = const()[name = string("value_35_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_13_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_13_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274969856)))]; + tensor var_3281_cast_fp16 = add(x = q_79_cast_fp16, y = encoder_layers_13_self_attn_pos_bias_u_to_fp16)[name = string("op_3281_cast_fp16")]; + tensor encoder_layers_13_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_13_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274971968)))]; + tensor var_3283_cast_fp16 = add(x = q_79_cast_fp16, y = encoder_layers_13_self_attn_pos_bias_v_to_fp16)[name = string("op_3283_cast_fp16")]; + tensor q_with_bias_v_27_perm_0 = const()[name = string("q_with_bias_v_27_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_345_transpose_x_0 = const()[name = string("x_345_transpose_x_0"), val = bool(false)]; + bool x_345_transpose_y_0 = const()[name = string("x_345_transpose_y_0"), val = bool(false)]; + tensor op_3285_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274974080))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275116480))))[name = string("op_3285_to_fp16_quantized")]; + tensor q_with_bias_v_27_cast_fp16 = transpose(perm = q_with_bias_v_27_perm_0, x = var_3283_cast_fp16)[name = string("transpose_245")]; + tensor x_345_cast_fp16 = matmul(transpose_x = x_345_transpose_x_0, transpose_y = x_345_transpose_y_0, x = q_with_bias_v_27_cast_fp16, y = op_3285_to_fp16_quantized)[name = string("x_345_cast_fp16")]; + tensor x_347_pad_0 = const()[name = string("x_347_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_347_mode_0 = const()[name = string("x_347_mode_0"), val = string("constant")]; + fp16 const_248_to_fp16 = const()[name = string("const_248_to_fp16"), val = fp16(0x0p+0)]; + tensor x_347_cast_fp16 = pad(constant_val = const_248_to_fp16, mode = x_347_mode_0, pad = x_347_pad_0, x = x_345_cast_fp16)[name = string("x_347_cast_fp16")]; + tensor var_3293 = const()[name = string("op_3293"), val = tensor([1, 8, -1, 28])]; + tensor x_349_cast_fp16 = reshape(shape = var_3293, x = x_347_cast_fp16)[name = string("x_349_cast_fp16")]; + tensor var_3297_begin_0 = const()[name = string("op_3297_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_3297_end_0 = const()[name = string("op_3297_end_0"), val = tensor([1, 8, 140, 28])]; + tensor var_3297_end_mask_0 = const()[name = string("op_3297_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_3297_cast_fp16 = slice_by_index(begin = var_3297_begin_0, end = var_3297_end_0, end_mask = var_3297_end_mask_0, x = x_349_cast_fp16)[name = string("op_3297_cast_fp16")]; + tensor var_3298 = const()[name = string("op_3298"), val = tensor([1, 8, 28, 139])]; + tensor matrix_bd_53_cast_fp16 = reshape(shape = var_3298, x = var_3297_cast_fp16)[name = string("matrix_bd_53_cast_fp16")]; + bool matrix_ac_27_transpose_x_0 = const()[name = string("matrix_ac_27_transpose_x_0"), val = bool(false)]; + bool matrix_ac_27_transpose_y_0 = const()[name = string("matrix_ac_27_transpose_y_0"), val = bool(false)]; + tensor transpose_122_perm_0 = const()[name = string("transpose_122_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_123_perm_0 = const()[name = string("transpose_123_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_123 = transpose(perm = transpose_123_perm_0, x = k_53_cast_fp16)[name = string("transpose_243")]; + tensor transpose_122 = transpose(perm = transpose_122_perm_0, x = var_3281_cast_fp16)[name = string("transpose_244")]; + tensor matrix_ac_27_cast_fp16 = matmul(transpose_x = matrix_ac_27_transpose_x_0, transpose_y = matrix_ac_27_transpose_y_0, x = transpose_122, y = transpose_123)[name = string("matrix_ac_27_cast_fp16")]; + tensor matrix_bd_55_begin_0 = const()[name = string("matrix_bd_55_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_55_end_0 = const()[name = string("matrix_bd_55_end_0"), val = tensor([1, 8, 28, 70])]; + tensor matrix_bd_55_end_mask_0 = const()[name = string("matrix_bd_55_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_55_cast_fp16 = slice_by_index(begin = matrix_bd_55_begin_0, end = matrix_bd_55_end_0, end_mask = matrix_bd_55_end_mask_0, x = matrix_bd_53_cast_fp16)[name = string("matrix_bd_55_cast_fp16")]; + tensor var_3307_cast_fp16 = add(x = matrix_ac_27_cast_fp16, y = matrix_bd_55_cast_fp16)[name = string("op_3307_cast_fp16")]; + fp16 _inversed_scores_53_y_0_to_fp16 = const()[name = string("_inversed_scores_53_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_53_cast_fp16 = mul(x = var_3307_cast_fp16, y = _inversed_scores_53_y_0_to_fp16)[name = string("_inversed_scores_53_cast_fp16")]; + tensor scores_55_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_53_cast_fp16, cond = mask_11)[name = string("scores_55_cast_fp16")]; + tensor var_3313_cast_fp16 = softmax(axis = var_59, x = scores_55_cast_fp16)[name = string("op_3313_cast_fp16")]; + tensor input_717_cast_fp16 = select(a = var_44_to_fp16, b = var_3313_cast_fp16, cond = mask_11)[name = string("input_717_cast_fp16")]; + bool x_351_transpose_x_0 = const()[name = string("x_351_transpose_x_0"), val = bool(false)]; + bool x_351_transpose_y_0 = const()[name = string("x_351_transpose_y_0"), val = bool(false)]; + tensor value_35_cast_fp16 = transpose(perm = value_35_perm_0, x = v_27_cast_fp16)[name = string("transpose_242")]; + tensor x_351_cast_fp16 = matmul(transpose_x = x_351_transpose_x_0, transpose_y = x_351_transpose_y_0, x = input_717_cast_fp16, y = value_35_cast_fp16)[name = string("x_351_cast_fp16")]; + tensor var_3317_perm_0 = const()[name = string("op_3317_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_3318 = const()[name = string("op_3318"), val = tensor([1, -1, 1024])]; + tensor var_3317_cast_fp16 = transpose(perm = var_3317_perm_0, x = x_351_cast_fp16)[name = string("transpose_241")]; + tensor input_719_cast_fp16 = reshape(shape = var_3318, x = var_3317_cast_fp16)[name = string("input_719_cast_fp16")]; + tensor encoder_layers_13_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275116864))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275903360))))[name = string("encoder_layers_13_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_13_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_13_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275903552)))]; + tensor linear_124_cast_fp16 = linear(bias = encoder_layers_13_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_13_self_attn_linear_out_weight_to_fp16_palettized, x = input_719_cast_fp16)[name = string("linear_124_cast_fp16")]; + tensor input_723_cast_fp16 = add(x = input_713_cast_fp16, y = linear_124_cast_fp16)[name = string("input_723_cast_fp16")]; + tensor x_355_axes_0 = const()[name = string("x_355_axes_0"), val = tensor([-1])]; + tensor encoder_layers_13_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_13_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275905664)))]; + tensor encoder_layers_13_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_13_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275907776)))]; + tensor x_355_cast_fp16 = layer_norm(axes = x_355_axes_0, beta = encoder_layers_13_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_13_norm_conv_weight_to_fp16, x = input_723_cast_fp16)[name = string("x_355_cast_fp16")]; + tensor input_725_perm_0 = const()[name = string("input_725_perm_0"), val = tensor([0, 2, 1])]; + string input_727_pad_type_0 = const()[name = string("input_727_pad_type_0"), val = string("valid")]; + tensor input_727_strides_0 = const()[name = string("input_727_strides_0"), val = tensor([1])]; + tensor input_727_pad_0 = const()[name = string("input_727_pad_0"), val = tensor([0, 0])]; + tensor input_727_dilations_0 = const()[name = string("input_727_dilations_0"), val = tensor([1])]; + int32 input_727_groups_0 = const()[name = string("input_727_groups_0"), val = int32(1)]; + tensor encoder_layers_13_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275909888))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278007104))))[name = string("encoder_layers_13_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_725_cast_fp16 = transpose(perm = input_725_perm_0, x = x_355_cast_fp16)[name = string("transpose_240")]; + tensor input_727_cast_fp16 = conv(dilations = input_727_dilations_0, groups = input_727_groups_0, pad = input_727_pad_0, pad_type = input_727_pad_type_0, strides = input_727_strides_0, weight = encoder_layers_13_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_725_cast_fp16)[name = string("input_727_cast_fp16")]; + int32 x_357_split_num_splits_0 = const()[name = string("x_357_split_num_splits_0"), val = int32(2)]; + int32 x_357_split_axis_0 = const()[name = string("x_357_split_axis_0"), val = int32(1)]; + tensor x_357_split_cast_fp16_0, tensor x_357_split_cast_fp16_1 = split(axis = x_357_split_axis_0, num_splits = x_357_split_num_splits_0, x = input_727_cast_fp16)[name = string("x_357_split_cast_fp16")]; + tensor x_357_split_1_sigmoid_cast_fp16 = sigmoid(x = x_357_split_cast_fp16_1)[name = string("x_357_split_1_sigmoid_cast_fp16")]; + tensor x_357_cast_fp16 = mul(x = x_357_split_cast_fp16_0, y = x_357_split_1_sigmoid_cast_fp16)[name = string("x_357_cast_fp16")]; + tensor input_729_cast_fp16 = select(a = var_44_to_fp16, b = x_357_cast_fp16, cond = var_575)[name = string("input_729_cast_fp16")]; + bool new_x_55_interleave_0 = const()[name = string("new_x_55_interleave_0"), val = bool(false)]; + tensor new_x_55_cast_fp16 = concat(axis = var_59, interleave = new_x_55_interleave_0, values = (cache_55_cast_fp16, input_729_cast_fp16))[name = string("new_x_55_cast_fp16")]; + tensor var_3357_begin_0 = const()[name = string("op_3357_begin_0"), val = tensor([0, 0, 28])]; + tensor var_3357_end_0 = const()[name = string("op_3357_end_0"), val = tensor([1, 1024, 36])]; + tensor var_3357_end_mask_0 = const()[name = string("op_3357_end_mask_0"), val = tensor([true, true, true])]; + tensor var_3357_cast_fp16 = slice_by_index(begin = var_3357_begin_0, end = var_3357_end_0, end_mask = var_3357_end_mask_0, x = new_x_55_cast_fp16)[name = string("op_3357_cast_fp16")]; + string x_359_pad_type_0 = const()[name = string("x_359_pad_type_0"), val = string("valid")]; + int32 x_359_groups_0 = const()[name = string("x_359_groups_0"), val = int32(1024)]; + tensor x_359_strides_0 = const()[name = string("x_359_strides_0"), val = tensor([1])]; + tensor x_359_pad_0 = const()[name = string("x_359_pad_0"), val = tensor([0, 0])]; + tensor x_359_dilations_0 = const()[name = string("x_359_dilations_0"), val = tensor([1])]; + tensor encoder_layers_13_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278011264))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278020544))))[name = string("encoder_layers_13_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_359_cast_fp16 = conv(dilations = x_359_dilations_0, groups = x_359_groups_0, pad = x_359_pad_0, pad_type = x_359_pad_type_0, strides = x_359_strides_0, weight = encoder_layers_13_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_55_cast_fp16)[name = string("x_359_cast_fp16")]; + tensor input_731_perm_0 = const()[name = string("input_731_perm_0"), val = tensor([0, 2, 1])]; + tensor x_361_axes_0 = const()[name = string("x_361_axes_0"), val = tensor([-1])]; + tensor encoder_layers_13_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_13_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278022656)))]; + tensor encoder_layers_13_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_13_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278024768)))]; + tensor input_731_cast_fp16 = transpose(perm = input_731_perm_0, x = x_359_cast_fp16)[name = string("transpose_239")]; + tensor x_361_cast_fp16 = layer_norm(axes = x_361_axes_0, beta = encoder_layers_13_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_13_conv_batch_norm_weight_to_fp16, x = input_731_cast_fp16)[name = string("x_361_cast_fp16")]; + tensor input_733_perm_0 = const()[name = string("input_733_perm_0"), val = tensor([0, 2, 1])]; + tensor input_733_cast_fp16 = transpose(perm = input_733_perm_0, x = x_361_cast_fp16)[name = string("transpose_238")]; + tensor input_735_cast_fp16 = silu(x = input_733_cast_fp16)[name = string("input_735_cast_fp16")]; + string x_363_pad_type_0 = const()[name = string("x_363_pad_type_0"), val = string("valid")]; + tensor x_363_strides_0 = const()[name = string("x_363_strides_0"), val = tensor([1])]; + tensor x_363_pad_0 = const()[name = string("x_363_pad_0"), val = tensor([0, 0])]; + tensor x_363_dilations_0 = const()[name = string("x_363_dilations_0"), val = tensor([1])]; + int32 x_363_groups_0 = const()[name = string("x_363_groups_0"), val = int32(1)]; + tensor encoder_layers_13_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278026880))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(279075520))))[name = string("encoder_layers_13_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_363_cast_fp16 = conv(dilations = x_363_dilations_0, groups = x_363_groups_0, pad = x_363_pad_0, pad_type = x_363_pad_type_0, strides = x_363_strides_0, weight = encoder_layers_13_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_735_cast_fp16)[name = string("x_363_cast_fp16")]; + tensor input_737_perm_0 = const()[name = string("input_737_perm_0"), val = tensor([0, 2, 1])]; + tensor input_737_cast_fp16 = transpose(perm = input_737_perm_0, x = x_363_cast_fp16)[name = string("transpose_237")]; + tensor input_739_cast_fp16 = add(x = input_723_cast_fp16, y = input_737_cast_fp16)[name = string("input_739_cast_fp16")]; + tensor input_741_axes_0 = const()[name = string("input_741_axes_0"), val = tensor([-1])]; + tensor encoder_layers_13_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_13_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(279077632)))]; + tensor encoder_layers_13_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_13_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(279079744)))]; + tensor input_741_cast_fp16 = layer_norm(axes = input_741_axes_0, beta = encoder_layers_13_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_13_norm_feed_forward2_weight_to_fp16, x = input_739_cast_fp16)[name = string("input_741_cast_fp16")]; + tensor encoder_layers_13_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(279081856))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(282227648))))[name = string("encoder_layers_13_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_13_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_13_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(282227840)))]; + tensor linear_125_cast_fp16 = linear(bias = encoder_layers_13_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_13_feed_forward2_linear1_weight_to_fp16_palettized, x = input_741_cast_fp16)[name = string("linear_125_cast_fp16")]; + tensor input_745_cast_fp16 = silu(x = linear_125_cast_fp16)[name = string("input_745_cast_fp16")]; + tensor encoder_layers_13_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(282236096))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(285381888))))[name = string("encoder_layers_13_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_13_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_13_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(285382080)))]; + tensor linear_126_cast_fp16 = linear(bias = encoder_layers_13_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_13_feed_forward2_linear2_weight_to_fp16_palettized, x = input_745_cast_fp16)[name = string("linear_126_cast_fp16")]; + fp16 var_3400_to_fp16 = const()[name = string("op_3400_to_fp16"), val = fp16(0x1p-1)]; + tensor var_3401_cast_fp16 = mul(x = linear_126_cast_fp16, y = var_3400_to_fp16)[name = string("op_3401_cast_fp16")]; + tensor input_751_cast_fp16 = add(x = input_739_cast_fp16, y = var_3401_cast_fp16)[name = string("input_751_cast_fp16")]; + tensor input_753_axes_0 = const()[name = string("input_753_axes_0"), val = tensor([-1])]; + tensor encoder_layers_13_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_13_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(285384192)))]; + tensor encoder_layers_13_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_13_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(285386304)))]; + tensor input_753_cast_fp16 = layer_norm(axes = input_753_axes_0, beta = encoder_layers_13_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_13_norm_out_weight_to_fp16, x = input_751_cast_fp16)[name = string("input_753_cast_fp16")]; + tensor cache_57_begin_0 = const()[name = string("cache_57_begin_0"), val = tensor([14, 0, 0, 0])]; + tensor cache_57_end_0 = const()[name = string("cache_57_end_0"), val = tensor([15, 1, 42, 1024])]; + tensor cache_57_end_mask_0 = const()[name = string("cache_57_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_57_squeeze_mask_0 = const()[name = string("cache_57_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_57_cast_fp16 = slice_by_index(begin = cache_57_begin_0, end = cache_57_end_0, end_mask = cache_57_end_mask_0, squeeze_mask = cache_57_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_57_cast_fp16")]; + tensor cache_59_begin_0 = const()[name = string("cache_59_begin_0"), val = tensor([14, 0, 0, 0])]; + tensor cache_59_end_0 = const()[name = string("cache_59_end_0"), val = tensor([15, 1, 1024, 8])]; + tensor cache_59_end_mask_0 = const()[name = string("cache_59_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_59_squeeze_mask_0 = const()[name = string("cache_59_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_59_cast_fp16 = slice_by_index(begin = cache_59_begin_0, end = cache_59_end_0, end_mask = cache_59_end_mask_0, squeeze_mask = cache_59_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_59_cast_fp16")]; + tensor input_755_axes_0 = const()[name = string("input_755_axes_0"), val = tensor([-1])]; + tensor encoder_layers_14_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_14_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(285388416)))]; + tensor encoder_layers_14_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_14_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(285390528)))]; + tensor input_755_cast_fp16 = layer_norm(axes = input_755_axes_0, beta = encoder_layers_14_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_14_norm_feed_forward1_weight_to_fp16, x = input_753_cast_fp16)[name = string("input_755_cast_fp16")]; + tensor encoder_layers_14_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(285392640))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(288538432))))[name = string("encoder_layers_14_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_14_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_14_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(288538624)))]; + tensor linear_127_cast_fp16 = linear(bias = encoder_layers_14_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_14_feed_forward1_linear1_weight_to_fp16_palettized, x = input_755_cast_fp16)[name = string("linear_127_cast_fp16")]; + tensor input_759_cast_fp16 = silu(x = linear_127_cast_fp16)[name = string("input_759_cast_fp16")]; + tensor encoder_layers_14_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(288546880))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291692672))))[name = string("encoder_layers_14_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_14_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_14_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291692864)))]; + tensor linear_128_cast_fp16 = linear(bias = encoder_layers_14_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_14_feed_forward1_linear2_weight_to_fp16_palettized, x = input_759_cast_fp16)[name = string("linear_128_cast_fp16")]; + fp16 var_3437_to_fp16 = const()[name = string("op_3437_to_fp16"), val = fp16(0x1p-1)]; + tensor var_3438_cast_fp16 = mul(x = linear_128_cast_fp16, y = var_3437_to_fp16)[name = string("op_3438_cast_fp16")]; + tensor input_765_cast_fp16 = add(x = input_753_cast_fp16, y = var_3438_cast_fp16)[name = string("input_765_cast_fp16")]; + tensor key_29_axes_0 = const()[name = string("key_29_axes_0"), val = tensor([-1])]; + tensor encoder_layers_14_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_14_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291694976)))]; + tensor encoder_layers_14_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_14_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291697088)))]; + tensor key_29_cast_fp16 = layer_norm(axes = key_29_axes_0, beta = encoder_layers_14_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_14_norm_self_att_weight_to_fp16, x = input_765_cast_fp16)[name = string("key_29_cast_fp16")]; + bool input_767_interleave_0 = const()[name = string("input_767_interleave_0"), val = bool(false)]; + tensor input_767_cast_fp16 = concat(axis = var_68, interleave = input_767_interleave_0, values = (cache_57_cast_fp16, key_29_cast_fp16))[name = string("input_767_cast_fp16")]; + tensor var_3460_begin_0 = const()[name = string("op_3460_begin_0"), val = tensor([0, 28, 0])]; + tensor var_3460_end_0 = const()[name = string("op_3460_end_0"), val = tensor([1, 42, 1024])]; + tensor var_3460_end_mask_0 = const()[name = string("op_3460_end_mask_0"), val = tensor([true, true, true])]; + tensor var_3460_cast_fp16 = slice_by_index(begin = var_3460_begin_0, end = var_3460_end_0, end_mask = var_3460_end_mask_0, x = cache_57_cast_fp16)[name = string("op_3460_cast_fp16")]; + bool var_3466_interleave_0 = const()[name = string("op_3466_interleave_0"), val = bool(false)]; + tensor var_3466_cast_fp16 = concat(axis = var_68, interleave = var_3466_interleave_0, values = (var_3460_cast_fp16, key_29_cast_fp16))[name = string("op_3466_cast_fp16")]; + tensor encoder_layers_14_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291699200))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(292485696))))[name = string("encoder_layers_14_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_14_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_14_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(292485888)))]; + tensor linear_129_cast_fp16 = linear(bias = encoder_layers_14_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_14_self_attn_linear_q_weight_to_fp16_palettized, x = key_29_cast_fp16)[name = string("linear_129_cast_fp16")]; + tensor var_3471 = const()[name = string("op_3471"), val = tensor([1, -1, 8, 128])]; + tensor q_85_cast_fp16 = reshape(shape = var_3471, x = linear_129_cast_fp16)[name = string("q_85_cast_fp16")]; + tensor encoder_layers_14_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(292488000))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293274496))))[name = string("encoder_layers_14_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_14_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_14_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293274688)))]; + tensor linear_130_cast_fp16 = linear(bias = encoder_layers_14_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_14_self_attn_linear_k_weight_to_fp16_palettized, x = input_767_cast_fp16)[name = string("linear_130_cast_fp16")]; + tensor var_3476 = const()[name = string("op_3476"), val = tensor([1, -1, 8, 128])]; + tensor k_57_cast_fp16 = reshape(shape = var_3476, x = linear_130_cast_fp16)[name = string("k_57_cast_fp16")]; + tensor encoder_layers_14_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293276800))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294063296))))[name = string("encoder_layers_14_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_14_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_14_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294063488)))]; + tensor linear_131_cast_fp16 = linear(bias = encoder_layers_14_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_14_self_attn_linear_v_weight_to_fp16_palettized, x = input_767_cast_fp16)[name = string("linear_131_cast_fp16")]; + tensor var_3481 = const()[name = string("op_3481"), val = tensor([1, -1, 8, 128])]; + tensor v_29_cast_fp16 = reshape(shape = var_3481, x = linear_131_cast_fp16)[name = string("v_29_cast_fp16")]; + tensor value_37_perm_0 = const()[name = string("value_37_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_14_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_14_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294065600)))]; + tensor var_3494_cast_fp16 = add(x = q_85_cast_fp16, y = encoder_layers_14_self_attn_pos_bias_u_to_fp16)[name = string("op_3494_cast_fp16")]; + tensor encoder_layers_14_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_14_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294067712)))]; + tensor var_3496_cast_fp16 = add(x = q_85_cast_fp16, y = encoder_layers_14_self_attn_pos_bias_v_to_fp16)[name = string("op_3496_cast_fp16")]; + tensor q_with_bias_v_29_perm_0 = const()[name = string("q_with_bias_v_29_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_371_transpose_x_0 = const()[name = string("x_371_transpose_x_0"), val = bool(false)]; + bool x_371_transpose_y_0 = const()[name = string("x_371_transpose_y_0"), val = bool(false)]; + tensor op_3498_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294069824))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294212224))))[name = string("op_3498_to_fp16_quantized")]; + tensor q_with_bias_v_29_cast_fp16 = transpose(perm = q_with_bias_v_29_perm_0, x = var_3496_cast_fp16)[name = string("transpose_236")]; + tensor x_371_cast_fp16 = matmul(transpose_x = x_371_transpose_x_0, transpose_y = x_371_transpose_y_0, x = q_with_bias_v_29_cast_fp16, y = op_3498_to_fp16_quantized)[name = string("x_371_cast_fp16")]; + tensor x_373_pad_0 = const()[name = string("x_373_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_373_mode_0 = const()[name = string("x_373_mode_0"), val = string("constant")]; + fp16 const_261_to_fp16 = const()[name = string("const_261_to_fp16"), val = fp16(0x0p+0)]; + tensor x_373_cast_fp16 = pad(constant_val = const_261_to_fp16, mode = x_373_mode_0, pad = x_373_pad_0, x = x_371_cast_fp16)[name = string("x_373_cast_fp16")]; + tensor var_3506 = const()[name = string("op_3506"), val = tensor([1, 8, -1, 28])]; + tensor x_375_cast_fp16 = reshape(shape = var_3506, x = x_373_cast_fp16)[name = string("x_375_cast_fp16")]; + tensor var_3510_begin_0 = const()[name = string("op_3510_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_3510_end_0 = const()[name = string("op_3510_end_0"), val = tensor([1, 8, 140, 28])]; + tensor var_3510_end_mask_0 = const()[name = string("op_3510_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_3510_cast_fp16 = slice_by_index(begin = var_3510_begin_0, end = var_3510_end_0, end_mask = var_3510_end_mask_0, x = x_375_cast_fp16)[name = string("op_3510_cast_fp16")]; + tensor var_3511 = const()[name = string("op_3511"), val = tensor([1, 8, 28, 139])]; + tensor matrix_bd_57_cast_fp16 = reshape(shape = var_3511, x = var_3510_cast_fp16)[name = string("matrix_bd_57_cast_fp16")]; + bool matrix_ac_29_transpose_x_0 = const()[name = string("matrix_ac_29_transpose_x_0"), val = bool(false)]; + bool matrix_ac_29_transpose_y_0 = const()[name = string("matrix_ac_29_transpose_y_0"), val = bool(false)]; + tensor transpose_124_perm_0 = const()[name = string("transpose_124_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_125_perm_0 = const()[name = string("transpose_125_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_125 = transpose(perm = transpose_125_perm_0, x = k_57_cast_fp16)[name = string("transpose_234")]; + tensor transpose_124 = transpose(perm = transpose_124_perm_0, x = var_3494_cast_fp16)[name = string("transpose_235")]; + tensor matrix_ac_29_cast_fp16 = matmul(transpose_x = matrix_ac_29_transpose_x_0, transpose_y = matrix_ac_29_transpose_y_0, x = transpose_124, y = transpose_125)[name = string("matrix_ac_29_cast_fp16")]; + tensor matrix_bd_59_begin_0 = const()[name = string("matrix_bd_59_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_59_end_0 = const()[name = string("matrix_bd_59_end_0"), val = tensor([1, 8, 28, 70])]; + tensor matrix_bd_59_end_mask_0 = const()[name = string("matrix_bd_59_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_59_cast_fp16 = slice_by_index(begin = matrix_bd_59_begin_0, end = matrix_bd_59_end_0, end_mask = matrix_bd_59_end_mask_0, x = matrix_bd_57_cast_fp16)[name = string("matrix_bd_59_cast_fp16")]; + tensor var_3520_cast_fp16 = add(x = matrix_ac_29_cast_fp16, y = matrix_bd_59_cast_fp16)[name = string("op_3520_cast_fp16")]; + fp16 _inversed_scores_57_y_0_to_fp16 = const()[name = string("_inversed_scores_57_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_57_cast_fp16 = mul(x = var_3520_cast_fp16, y = _inversed_scores_57_y_0_to_fp16)[name = string("_inversed_scores_57_cast_fp16")]; + tensor scores_59_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_57_cast_fp16, cond = mask_11)[name = string("scores_59_cast_fp16")]; + tensor var_3526_cast_fp16 = softmax(axis = var_59, x = scores_59_cast_fp16)[name = string("op_3526_cast_fp16")]; + tensor input_769_cast_fp16 = select(a = var_44_to_fp16, b = var_3526_cast_fp16, cond = mask_11)[name = string("input_769_cast_fp16")]; + bool x_377_transpose_x_0 = const()[name = string("x_377_transpose_x_0"), val = bool(false)]; + bool x_377_transpose_y_0 = const()[name = string("x_377_transpose_y_0"), val = bool(false)]; + tensor value_37_cast_fp16 = transpose(perm = value_37_perm_0, x = v_29_cast_fp16)[name = string("transpose_233")]; + tensor x_377_cast_fp16 = matmul(transpose_x = x_377_transpose_x_0, transpose_y = x_377_transpose_y_0, x = input_769_cast_fp16, y = value_37_cast_fp16)[name = string("x_377_cast_fp16")]; + tensor var_3530_perm_0 = const()[name = string("op_3530_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_3531 = const()[name = string("op_3531"), val = tensor([1, -1, 1024])]; + tensor var_3530_cast_fp16 = transpose(perm = var_3530_perm_0, x = x_377_cast_fp16)[name = string("transpose_232")]; + tensor input_771_cast_fp16 = reshape(shape = var_3531, x = var_3530_cast_fp16)[name = string("input_771_cast_fp16")]; + tensor encoder_layers_14_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294212608))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294999104))))[name = string("encoder_layers_14_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_14_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_14_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294999296)))]; + tensor linear_133_cast_fp16 = linear(bias = encoder_layers_14_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_14_self_attn_linear_out_weight_to_fp16_palettized, x = input_771_cast_fp16)[name = string("linear_133_cast_fp16")]; + tensor input_775_cast_fp16 = add(x = input_765_cast_fp16, y = linear_133_cast_fp16)[name = string("input_775_cast_fp16")]; + tensor x_381_axes_0 = const()[name = string("x_381_axes_0"), val = tensor([-1])]; + tensor encoder_layers_14_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_14_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(295001408)))]; + tensor encoder_layers_14_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_14_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(295003520)))]; + tensor x_381_cast_fp16 = layer_norm(axes = x_381_axes_0, beta = encoder_layers_14_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_14_norm_conv_weight_to_fp16, x = input_775_cast_fp16)[name = string("x_381_cast_fp16")]; + tensor input_777_perm_0 = const()[name = string("input_777_perm_0"), val = tensor([0, 2, 1])]; + string input_779_pad_type_0 = const()[name = string("input_779_pad_type_0"), val = string("valid")]; + tensor input_779_strides_0 = const()[name = string("input_779_strides_0"), val = tensor([1])]; + tensor input_779_pad_0 = const()[name = string("input_779_pad_0"), val = tensor([0, 0])]; + tensor input_779_dilations_0 = const()[name = string("input_779_dilations_0"), val = tensor([1])]; + int32 input_779_groups_0 = const()[name = string("input_779_groups_0"), val = int32(1)]; + tensor encoder_layers_14_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(295005632))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297102848))))[name = string("encoder_layers_14_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_777_cast_fp16 = transpose(perm = input_777_perm_0, x = x_381_cast_fp16)[name = string("transpose_231")]; + tensor input_779_cast_fp16 = conv(dilations = input_779_dilations_0, groups = input_779_groups_0, pad = input_779_pad_0, pad_type = input_779_pad_type_0, strides = input_779_strides_0, weight = encoder_layers_14_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_777_cast_fp16)[name = string("input_779_cast_fp16")]; + int32 x_383_split_num_splits_0 = const()[name = string("x_383_split_num_splits_0"), val = int32(2)]; + int32 x_383_split_axis_0 = const()[name = string("x_383_split_axis_0"), val = int32(1)]; + tensor x_383_split_cast_fp16_0, tensor x_383_split_cast_fp16_1 = split(axis = x_383_split_axis_0, num_splits = x_383_split_num_splits_0, x = input_779_cast_fp16)[name = string("x_383_split_cast_fp16")]; + tensor x_383_split_1_sigmoid_cast_fp16 = sigmoid(x = x_383_split_cast_fp16_1)[name = string("x_383_split_1_sigmoid_cast_fp16")]; + tensor x_383_cast_fp16 = mul(x = x_383_split_cast_fp16_0, y = x_383_split_1_sigmoid_cast_fp16)[name = string("x_383_cast_fp16")]; + tensor input_781_cast_fp16 = select(a = var_44_to_fp16, b = x_383_cast_fp16, cond = var_575)[name = string("input_781_cast_fp16")]; + bool new_x_59_interleave_0 = const()[name = string("new_x_59_interleave_0"), val = bool(false)]; + tensor new_x_59_cast_fp16 = concat(axis = var_59, interleave = new_x_59_interleave_0, values = (cache_59_cast_fp16, input_781_cast_fp16))[name = string("new_x_59_cast_fp16")]; + tensor var_3570_begin_0 = const()[name = string("op_3570_begin_0"), val = tensor([0, 0, 28])]; + tensor var_3570_end_0 = const()[name = string("op_3570_end_0"), val = tensor([1, 1024, 36])]; + tensor var_3570_end_mask_0 = const()[name = string("op_3570_end_mask_0"), val = tensor([true, true, true])]; + tensor var_3570_cast_fp16 = slice_by_index(begin = var_3570_begin_0, end = var_3570_end_0, end_mask = var_3570_end_mask_0, x = new_x_59_cast_fp16)[name = string("op_3570_cast_fp16")]; + string x_385_pad_type_0 = const()[name = string("x_385_pad_type_0"), val = string("valid")]; + int32 x_385_groups_0 = const()[name = string("x_385_groups_0"), val = int32(1024)]; + tensor x_385_strides_0 = const()[name = string("x_385_strides_0"), val = tensor([1])]; + tensor x_385_pad_0 = const()[name = string("x_385_pad_0"), val = tensor([0, 0])]; + tensor x_385_dilations_0 = const()[name = string("x_385_dilations_0"), val = tensor([1])]; + tensor encoder_layers_14_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297107008))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297116288))))[name = string("encoder_layers_14_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_385_cast_fp16 = conv(dilations = x_385_dilations_0, groups = x_385_groups_0, pad = x_385_pad_0, pad_type = x_385_pad_type_0, strides = x_385_strides_0, weight = encoder_layers_14_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_59_cast_fp16)[name = string("x_385_cast_fp16")]; + tensor input_783_perm_0 = const()[name = string("input_783_perm_0"), val = tensor([0, 2, 1])]; + tensor x_387_axes_0 = const()[name = string("x_387_axes_0"), val = tensor([-1])]; + tensor encoder_layers_14_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_14_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297118400)))]; + tensor encoder_layers_14_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_14_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297120512)))]; + tensor input_783_cast_fp16 = transpose(perm = input_783_perm_0, x = x_385_cast_fp16)[name = string("transpose_230")]; + tensor x_387_cast_fp16 = layer_norm(axes = x_387_axes_0, beta = encoder_layers_14_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_14_conv_batch_norm_weight_to_fp16, x = input_783_cast_fp16)[name = string("x_387_cast_fp16")]; + tensor input_785_perm_0 = const()[name = string("input_785_perm_0"), val = tensor([0, 2, 1])]; + tensor input_785_cast_fp16 = transpose(perm = input_785_perm_0, x = x_387_cast_fp16)[name = string("transpose_229")]; + tensor input_787_cast_fp16 = silu(x = input_785_cast_fp16)[name = string("input_787_cast_fp16")]; + string x_389_pad_type_0 = const()[name = string("x_389_pad_type_0"), val = string("valid")]; + tensor x_389_strides_0 = const()[name = string("x_389_strides_0"), val = tensor([1])]; + tensor x_389_pad_0 = const()[name = string("x_389_pad_0"), val = tensor([0, 0])]; + tensor x_389_dilations_0 = const()[name = string("x_389_dilations_0"), val = tensor([1])]; + int32 x_389_groups_0 = const()[name = string("x_389_groups_0"), val = int32(1)]; + tensor encoder_layers_14_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297122624))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(298171264))))[name = string("encoder_layers_14_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_389_cast_fp16 = conv(dilations = x_389_dilations_0, groups = x_389_groups_0, pad = x_389_pad_0, pad_type = x_389_pad_type_0, strides = x_389_strides_0, weight = encoder_layers_14_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_787_cast_fp16)[name = string("x_389_cast_fp16")]; + tensor input_789_perm_0 = const()[name = string("input_789_perm_0"), val = tensor([0, 2, 1])]; + tensor input_789_cast_fp16 = transpose(perm = input_789_perm_0, x = x_389_cast_fp16)[name = string("transpose_228")]; + tensor input_791_cast_fp16 = add(x = input_775_cast_fp16, y = input_789_cast_fp16)[name = string("input_791_cast_fp16")]; + tensor input_793_axes_0 = const()[name = string("input_793_axes_0"), val = tensor([-1])]; + tensor encoder_layers_14_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_14_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(298173376)))]; + tensor encoder_layers_14_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_14_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(298175488)))]; + tensor input_793_cast_fp16 = layer_norm(axes = input_793_axes_0, beta = encoder_layers_14_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_14_norm_feed_forward2_weight_to_fp16, x = input_791_cast_fp16)[name = string("input_793_cast_fp16")]; + tensor encoder_layers_14_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(298177600))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301323392))))[name = string("encoder_layers_14_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_14_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_14_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301323584)))]; + tensor linear_134_cast_fp16 = linear(bias = encoder_layers_14_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_14_feed_forward2_linear1_weight_to_fp16_palettized, x = input_793_cast_fp16)[name = string("linear_134_cast_fp16")]; + tensor input_797_cast_fp16 = silu(x = linear_134_cast_fp16)[name = string("input_797_cast_fp16")]; + tensor encoder_layers_14_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301331840))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(304477632))))[name = string("encoder_layers_14_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_14_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_14_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(304477824)))]; + tensor linear_135_cast_fp16 = linear(bias = encoder_layers_14_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_14_feed_forward2_linear2_weight_to_fp16_palettized, x = input_797_cast_fp16)[name = string("linear_135_cast_fp16")]; + fp16 var_3613_to_fp16 = const()[name = string("op_3613_to_fp16"), val = fp16(0x1p-1)]; + tensor var_3614_cast_fp16 = mul(x = linear_135_cast_fp16, y = var_3613_to_fp16)[name = string("op_3614_cast_fp16")]; + tensor input_803_cast_fp16 = add(x = input_791_cast_fp16, y = var_3614_cast_fp16)[name = string("input_803_cast_fp16")]; + tensor input_805_axes_0 = const()[name = string("input_805_axes_0"), val = tensor([-1])]; + tensor encoder_layers_14_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_14_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(304479936)))]; + tensor encoder_layers_14_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_14_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(304482048)))]; + tensor input_805_cast_fp16 = layer_norm(axes = input_805_axes_0, beta = encoder_layers_14_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_14_norm_out_weight_to_fp16, x = input_803_cast_fp16)[name = string("input_805_cast_fp16")]; + tensor cache_61_begin_0 = const()[name = string("cache_61_begin_0"), val = tensor([15, 0, 0, 0])]; + tensor cache_61_end_0 = const()[name = string("cache_61_end_0"), val = tensor([16, 1, 42, 1024])]; + tensor cache_61_end_mask_0 = const()[name = string("cache_61_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_61_squeeze_mask_0 = const()[name = string("cache_61_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_61_cast_fp16 = slice_by_index(begin = cache_61_begin_0, end = cache_61_end_0, end_mask = cache_61_end_mask_0, squeeze_mask = cache_61_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_61_cast_fp16")]; + tensor cache_63_begin_0 = const()[name = string("cache_63_begin_0"), val = tensor([15, 0, 0, 0])]; + tensor cache_63_end_0 = const()[name = string("cache_63_end_0"), val = tensor([16, 1, 1024, 8])]; + tensor cache_63_end_mask_0 = const()[name = string("cache_63_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_63_squeeze_mask_0 = const()[name = string("cache_63_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_63_cast_fp16 = slice_by_index(begin = cache_63_begin_0, end = cache_63_end_0, end_mask = cache_63_end_mask_0, squeeze_mask = cache_63_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_63_cast_fp16")]; + tensor input_807_axes_0 = const()[name = string("input_807_axes_0"), val = tensor([-1])]; + tensor encoder_layers_15_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_15_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(304484160)))]; + tensor encoder_layers_15_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_15_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(304486272)))]; + tensor input_807_cast_fp16 = layer_norm(axes = input_807_axes_0, beta = encoder_layers_15_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_15_norm_feed_forward1_weight_to_fp16, x = input_805_cast_fp16)[name = string("input_807_cast_fp16")]; + tensor encoder_layers_15_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(304488384))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307634176))))[name = string("encoder_layers_15_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_15_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_15_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307634368)))]; + tensor linear_136_cast_fp16 = linear(bias = encoder_layers_15_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_15_feed_forward1_linear1_weight_to_fp16_palettized, x = input_807_cast_fp16)[name = string("linear_136_cast_fp16")]; + tensor input_811_cast_fp16 = silu(x = linear_136_cast_fp16)[name = string("input_811_cast_fp16")]; + tensor encoder_layers_15_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307642624))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310788416))))[name = string("encoder_layers_15_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_15_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_15_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310788608)))]; + tensor linear_137_cast_fp16 = linear(bias = encoder_layers_15_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_15_feed_forward1_linear2_weight_to_fp16_palettized, x = input_811_cast_fp16)[name = string("linear_137_cast_fp16")]; + fp16 var_3650_to_fp16 = const()[name = string("op_3650_to_fp16"), val = fp16(0x1p-1)]; + tensor var_3651_cast_fp16 = mul(x = linear_137_cast_fp16, y = var_3650_to_fp16)[name = string("op_3651_cast_fp16")]; + tensor input_817_cast_fp16 = add(x = input_805_cast_fp16, y = var_3651_cast_fp16)[name = string("input_817_cast_fp16")]; + tensor key_31_axes_0 = const()[name = string("key_31_axes_0"), val = tensor([-1])]; + tensor encoder_layers_15_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_15_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310790720)))]; + tensor encoder_layers_15_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_15_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310792832)))]; + tensor key_31_cast_fp16 = layer_norm(axes = key_31_axes_0, beta = encoder_layers_15_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_15_norm_self_att_weight_to_fp16, x = input_817_cast_fp16)[name = string("key_31_cast_fp16")]; + bool input_819_interleave_0 = const()[name = string("input_819_interleave_0"), val = bool(false)]; + tensor input_819_cast_fp16 = concat(axis = var_68, interleave = input_819_interleave_0, values = (cache_61_cast_fp16, key_31_cast_fp16))[name = string("input_819_cast_fp16")]; + tensor var_3673_begin_0 = const()[name = string("op_3673_begin_0"), val = tensor([0, 28, 0])]; + tensor var_3673_end_0 = const()[name = string("op_3673_end_0"), val = tensor([1, 42, 1024])]; + tensor var_3673_end_mask_0 = const()[name = string("op_3673_end_mask_0"), val = tensor([true, true, true])]; + tensor var_3673_cast_fp16 = slice_by_index(begin = var_3673_begin_0, end = var_3673_end_0, end_mask = var_3673_end_mask_0, x = cache_61_cast_fp16)[name = string("op_3673_cast_fp16")]; + bool var_3679_interleave_0 = const()[name = string("op_3679_interleave_0"), val = bool(false)]; + tensor var_3679_cast_fp16 = concat(axis = var_68, interleave = var_3679_interleave_0, values = (var_3673_cast_fp16, key_31_cast_fp16))[name = string("op_3679_cast_fp16")]; + tensor encoder_layers_15_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310794944))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311581440))))[name = string("encoder_layers_15_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_15_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_15_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311581632)))]; + tensor linear_138_cast_fp16 = linear(bias = encoder_layers_15_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_15_self_attn_linear_q_weight_to_fp16_palettized, x = key_31_cast_fp16)[name = string("linear_138_cast_fp16")]; + tensor var_3684 = const()[name = string("op_3684"), val = tensor([1, -1, 8, 128])]; + tensor q_91_cast_fp16 = reshape(shape = var_3684, x = linear_138_cast_fp16)[name = string("q_91_cast_fp16")]; + tensor encoder_layers_15_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311583744))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(312370240))))[name = string("encoder_layers_15_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_15_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_15_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(312370432)))]; + tensor linear_139_cast_fp16 = linear(bias = encoder_layers_15_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_15_self_attn_linear_k_weight_to_fp16_palettized, x = input_819_cast_fp16)[name = string("linear_139_cast_fp16")]; + tensor var_3689 = const()[name = string("op_3689"), val = tensor([1, -1, 8, 128])]; + tensor k_61_cast_fp16 = reshape(shape = var_3689, x = linear_139_cast_fp16)[name = string("k_61_cast_fp16")]; + tensor encoder_layers_15_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(312372544))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(313159040))))[name = string("encoder_layers_15_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_15_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_15_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(313159232)))]; + tensor linear_140_cast_fp16 = linear(bias = encoder_layers_15_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_15_self_attn_linear_v_weight_to_fp16_palettized, x = input_819_cast_fp16)[name = string("linear_140_cast_fp16")]; + tensor var_3694 = const()[name = string("op_3694"), val = tensor([1, -1, 8, 128])]; + tensor v_31_cast_fp16 = reshape(shape = var_3694, x = linear_140_cast_fp16)[name = string("v_31_cast_fp16")]; + tensor value_39_perm_0 = const()[name = string("value_39_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_15_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_15_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(313161344)))]; + tensor var_3707_cast_fp16 = add(x = q_91_cast_fp16, y = encoder_layers_15_self_attn_pos_bias_u_to_fp16)[name = string("op_3707_cast_fp16")]; + tensor encoder_layers_15_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_15_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(313163456)))]; + tensor var_3709_cast_fp16 = add(x = q_91_cast_fp16, y = encoder_layers_15_self_attn_pos_bias_v_to_fp16)[name = string("op_3709_cast_fp16")]; + tensor q_with_bias_v_31_perm_0 = const()[name = string("q_with_bias_v_31_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_397_transpose_x_0 = const()[name = string("x_397_transpose_x_0"), val = bool(false)]; + bool x_397_transpose_y_0 = const()[name = string("x_397_transpose_y_0"), val = bool(false)]; + tensor op_3711_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(313165568))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(313307968))))[name = string("op_3711_to_fp16_quantized")]; + tensor q_with_bias_v_31_cast_fp16 = transpose(perm = q_with_bias_v_31_perm_0, x = var_3709_cast_fp16)[name = string("transpose_227")]; + tensor x_397_cast_fp16 = matmul(transpose_x = x_397_transpose_x_0, transpose_y = x_397_transpose_y_0, x = q_with_bias_v_31_cast_fp16, y = op_3711_to_fp16_quantized)[name = string("x_397_cast_fp16")]; + tensor x_399_pad_0 = const()[name = string("x_399_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_399_mode_0 = const()[name = string("x_399_mode_0"), val = string("constant")]; + fp16 const_274_to_fp16 = const()[name = string("const_274_to_fp16"), val = fp16(0x0p+0)]; + tensor x_399_cast_fp16 = pad(constant_val = const_274_to_fp16, mode = x_399_mode_0, pad = x_399_pad_0, x = x_397_cast_fp16)[name = string("x_399_cast_fp16")]; + tensor var_3719 = const()[name = string("op_3719"), val = tensor([1, 8, -1, 28])]; + tensor x_401_cast_fp16 = reshape(shape = var_3719, x = x_399_cast_fp16)[name = string("x_401_cast_fp16")]; + tensor var_3723_begin_0 = const()[name = string("op_3723_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_3723_end_0 = const()[name = string("op_3723_end_0"), val = tensor([1, 8, 140, 28])]; + tensor var_3723_end_mask_0 = const()[name = string("op_3723_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_3723_cast_fp16 = slice_by_index(begin = var_3723_begin_0, end = var_3723_end_0, end_mask = var_3723_end_mask_0, x = x_401_cast_fp16)[name = string("op_3723_cast_fp16")]; + tensor var_3724 = const()[name = string("op_3724"), val = tensor([1, 8, 28, 139])]; + tensor matrix_bd_61_cast_fp16 = reshape(shape = var_3724, x = var_3723_cast_fp16)[name = string("matrix_bd_61_cast_fp16")]; + bool matrix_ac_31_transpose_x_0 = const()[name = string("matrix_ac_31_transpose_x_0"), val = bool(false)]; + bool matrix_ac_31_transpose_y_0 = const()[name = string("matrix_ac_31_transpose_y_0"), val = bool(false)]; + tensor transpose_126_perm_0 = const()[name = string("transpose_126_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_127_perm_0 = const()[name = string("transpose_127_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_127 = transpose(perm = transpose_127_perm_0, x = k_61_cast_fp16)[name = string("transpose_225")]; + tensor transpose_126 = transpose(perm = transpose_126_perm_0, x = var_3707_cast_fp16)[name = string("transpose_226")]; + tensor matrix_ac_31_cast_fp16 = matmul(transpose_x = matrix_ac_31_transpose_x_0, transpose_y = matrix_ac_31_transpose_y_0, x = transpose_126, y = transpose_127)[name = string("matrix_ac_31_cast_fp16")]; + tensor matrix_bd_63_begin_0 = const()[name = string("matrix_bd_63_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_63_end_0 = const()[name = string("matrix_bd_63_end_0"), val = tensor([1, 8, 28, 70])]; + tensor matrix_bd_63_end_mask_0 = const()[name = string("matrix_bd_63_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_63_cast_fp16 = slice_by_index(begin = matrix_bd_63_begin_0, end = matrix_bd_63_end_0, end_mask = matrix_bd_63_end_mask_0, x = matrix_bd_61_cast_fp16)[name = string("matrix_bd_63_cast_fp16")]; + tensor var_3733_cast_fp16 = add(x = matrix_ac_31_cast_fp16, y = matrix_bd_63_cast_fp16)[name = string("op_3733_cast_fp16")]; + fp16 _inversed_scores_61_y_0_to_fp16 = const()[name = string("_inversed_scores_61_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_61_cast_fp16 = mul(x = var_3733_cast_fp16, y = _inversed_scores_61_y_0_to_fp16)[name = string("_inversed_scores_61_cast_fp16")]; + tensor scores_63_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_61_cast_fp16, cond = mask_11)[name = string("scores_63_cast_fp16")]; + tensor var_3739_cast_fp16 = softmax(axis = var_59, x = scores_63_cast_fp16)[name = string("op_3739_cast_fp16")]; + tensor input_821_cast_fp16 = select(a = var_44_to_fp16, b = var_3739_cast_fp16, cond = mask_11)[name = string("input_821_cast_fp16")]; + bool x_403_transpose_x_0 = const()[name = string("x_403_transpose_x_0"), val = bool(false)]; + bool x_403_transpose_y_0 = const()[name = string("x_403_transpose_y_0"), val = bool(false)]; + tensor value_39_cast_fp16 = transpose(perm = value_39_perm_0, x = v_31_cast_fp16)[name = string("transpose_224")]; + tensor x_403_cast_fp16 = matmul(transpose_x = x_403_transpose_x_0, transpose_y = x_403_transpose_y_0, x = input_821_cast_fp16, y = value_39_cast_fp16)[name = string("x_403_cast_fp16")]; + tensor var_3743_perm_0 = const()[name = string("op_3743_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_3744 = const()[name = string("op_3744"), val = tensor([1, -1, 1024])]; + tensor var_3743_cast_fp16 = transpose(perm = var_3743_perm_0, x = x_403_cast_fp16)[name = string("transpose_223")]; + tensor input_823_cast_fp16 = reshape(shape = var_3744, x = var_3743_cast_fp16)[name = string("input_823_cast_fp16")]; + tensor encoder_layers_15_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(313308352))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(314094848))))[name = string("encoder_layers_15_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_15_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_15_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(314095040)))]; + tensor linear_142_cast_fp16 = linear(bias = encoder_layers_15_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_15_self_attn_linear_out_weight_to_fp16_palettized, x = input_823_cast_fp16)[name = string("linear_142_cast_fp16")]; + tensor input_827_cast_fp16 = add(x = input_817_cast_fp16, y = linear_142_cast_fp16)[name = string("input_827_cast_fp16")]; + tensor x_407_axes_0 = const()[name = string("x_407_axes_0"), val = tensor([-1])]; + tensor encoder_layers_15_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_15_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(314097152)))]; + tensor encoder_layers_15_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_15_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(314099264)))]; + tensor x_407_cast_fp16 = layer_norm(axes = x_407_axes_0, beta = encoder_layers_15_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_15_norm_conv_weight_to_fp16, x = input_827_cast_fp16)[name = string("x_407_cast_fp16")]; + tensor input_829_perm_0 = const()[name = string("input_829_perm_0"), val = tensor([0, 2, 1])]; + string input_831_pad_type_0 = const()[name = string("input_831_pad_type_0"), val = string("valid")]; + tensor input_831_strides_0 = const()[name = string("input_831_strides_0"), val = tensor([1])]; + tensor input_831_pad_0 = const()[name = string("input_831_pad_0"), val = tensor([0, 0])]; + tensor input_831_dilations_0 = const()[name = string("input_831_dilations_0"), val = tensor([1])]; + int32 input_831_groups_0 = const()[name = string("input_831_groups_0"), val = int32(1)]; + tensor encoder_layers_15_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(314101376))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316198592))))[name = string("encoder_layers_15_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_829_cast_fp16 = transpose(perm = input_829_perm_0, x = x_407_cast_fp16)[name = string("transpose_222")]; + tensor input_831_cast_fp16 = conv(dilations = input_831_dilations_0, groups = input_831_groups_0, pad = input_831_pad_0, pad_type = input_831_pad_type_0, strides = input_831_strides_0, weight = encoder_layers_15_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_829_cast_fp16)[name = string("input_831_cast_fp16")]; + int32 x_409_split_num_splits_0 = const()[name = string("x_409_split_num_splits_0"), val = int32(2)]; + int32 x_409_split_axis_0 = const()[name = string("x_409_split_axis_0"), val = int32(1)]; + tensor x_409_split_cast_fp16_0, tensor x_409_split_cast_fp16_1 = split(axis = x_409_split_axis_0, num_splits = x_409_split_num_splits_0, x = input_831_cast_fp16)[name = string("x_409_split_cast_fp16")]; + tensor x_409_split_1_sigmoid_cast_fp16 = sigmoid(x = x_409_split_cast_fp16_1)[name = string("x_409_split_1_sigmoid_cast_fp16")]; + tensor x_409_cast_fp16 = mul(x = x_409_split_cast_fp16_0, y = x_409_split_1_sigmoid_cast_fp16)[name = string("x_409_cast_fp16")]; + tensor input_833_cast_fp16 = select(a = var_44_to_fp16, b = x_409_cast_fp16, cond = var_575)[name = string("input_833_cast_fp16")]; + bool new_x_63_interleave_0 = const()[name = string("new_x_63_interleave_0"), val = bool(false)]; + tensor new_x_63_cast_fp16 = concat(axis = var_59, interleave = new_x_63_interleave_0, values = (cache_63_cast_fp16, input_833_cast_fp16))[name = string("new_x_63_cast_fp16")]; + tensor var_3783_begin_0 = const()[name = string("op_3783_begin_0"), val = tensor([0, 0, 28])]; + tensor var_3783_end_0 = const()[name = string("op_3783_end_0"), val = tensor([1, 1024, 36])]; + tensor var_3783_end_mask_0 = const()[name = string("op_3783_end_mask_0"), val = tensor([true, true, true])]; + tensor var_3783_cast_fp16 = slice_by_index(begin = var_3783_begin_0, end = var_3783_end_0, end_mask = var_3783_end_mask_0, x = new_x_63_cast_fp16)[name = string("op_3783_cast_fp16")]; + string x_411_pad_type_0 = const()[name = string("x_411_pad_type_0"), val = string("valid")]; + int32 x_411_groups_0 = const()[name = string("x_411_groups_0"), val = int32(1024)]; + tensor x_411_strides_0 = const()[name = string("x_411_strides_0"), val = tensor([1])]; + tensor x_411_pad_0 = const()[name = string("x_411_pad_0"), val = tensor([0, 0])]; + tensor x_411_dilations_0 = const()[name = string("x_411_dilations_0"), val = tensor([1])]; + tensor encoder_layers_15_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316202752))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316212032))))[name = string("encoder_layers_15_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_411_cast_fp16 = conv(dilations = x_411_dilations_0, groups = x_411_groups_0, pad = x_411_pad_0, pad_type = x_411_pad_type_0, strides = x_411_strides_0, weight = encoder_layers_15_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_63_cast_fp16)[name = string("x_411_cast_fp16")]; + tensor input_835_perm_0 = const()[name = string("input_835_perm_0"), val = tensor([0, 2, 1])]; + tensor x_413_axes_0 = const()[name = string("x_413_axes_0"), val = tensor([-1])]; + tensor encoder_layers_15_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_15_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316214144)))]; + tensor encoder_layers_15_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_15_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316216256)))]; + tensor input_835_cast_fp16 = transpose(perm = input_835_perm_0, x = x_411_cast_fp16)[name = string("transpose_221")]; + tensor x_413_cast_fp16 = layer_norm(axes = x_413_axes_0, beta = encoder_layers_15_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_15_conv_batch_norm_weight_to_fp16, x = input_835_cast_fp16)[name = string("x_413_cast_fp16")]; + tensor input_837_perm_0 = const()[name = string("input_837_perm_0"), val = tensor([0, 2, 1])]; + tensor input_837_cast_fp16 = transpose(perm = input_837_perm_0, x = x_413_cast_fp16)[name = string("transpose_220")]; + tensor input_839_cast_fp16 = silu(x = input_837_cast_fp16)[name = string("input_839_cast_fp16")]; + string x_415_pad_type_0 = const()[name = string("x_415_pad_type_0"), val = string("valid")]; + tensor x_415_strides_0 = const()[name = string("x_415_strides_0"), val = tensor([1])]; + tensor x_415_pad_0 = const()[name = string("x_415_pad_0"), val = tensor([0, 0])]; + tensor x_415_dilations_0 = const()[name = string("x_415_dilations_0"), val = tensor([1])]; + int32 x_415_groups_0 = const()[name = string("x_415_groups_0"), val = int32(1)]; + tensor encoder_layers_15_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316218368))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(317267008))))[name = string("encoder_layers_15_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_415_cast_fp16 = conv(dilations = x_415_dilations_0, groups = x_415_groups_0, pad = x_415_pad_0, pad_type = x_415_pad_type_0, strides = x_415_strides_0, weight = encoder_layers_15_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_839_cast_fp16)[name = string("x_415_cast_fp16")]; + tensor input_841_perm_0 = const()[name = string("input_841_perm_0"), val = tensor([0, 2, 1])]; + tensor input_841_cast_fp16 = transpose(perm = input_841_perm_0, x = x_415_cast_fp16)[name = string("transpose_219")]; + tensor input_843_cast_fp16 = add(x = input_827_cast_fp16, y = input_841_cast_fp16)[name = string("input_843_cast_fp16")]; + tensor input_845_axes_0 = const()[name = string("input_845_axes_0"), val = tensor([-1])]; + tensor encoder_layers_15_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_15_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(317269120)))]; + tensor encoder_layers_15_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_15_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(317271232)))]; + tensor input_845_cast_fp16 = layer_norm(axes = input_845_axes_0, beta = encoder_layers_15_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_15_norm_feed_forward2_weight_to_fp16, x = input_843_cast_fp16)[name = string("input_845_cast_fp16")]; + tensor encoder_layers_15_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(317273344))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(320419136))))[name = string("encoder_layers_15_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_15_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_15_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(320419328)))]; + tensor linear_143_cast_fp16 = linear(bias = encoder_layers_15_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_15_feed_forward2_linear1_weight_to_fp16_palettized, x = input_845_cast_fp16)[name = string("linear_143_cast_fp16")]; + tensor input_849_cast_fp16 = silu(x = linear_143_cast_fp16)[name = string("input_849_cast_fp16")]; + tensor encoder_layers_15_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(320427584))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(323573376))))[name = string("encoder_layers_15_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_15_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_15_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(323573568)))]; + tensor linear_144_cast_fp16 = linear(bias = encoder_layers_15_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_15_feed_forward2_linear2_weight_to_fp16_palettized, x = input_849_cast_fp16)[name = string("linear_144_cast_fp16")]; + fp16 var_3826_to_fp16 = const()[name = string("op_3826_to_fp16"), val = fp16(0x1p-1)]; + tensor var_3827_cast_fp16 = mul(x = linear_144_cast_fp16, y = var_3826_to_fp16)[name = string("op_3827_cast_fp16")]; + tensor input_855_cast_fp16 = add(x = input_843_cast_fp16, y = var_3827_cast_fp16)[name = string("input_855_cast_fp16")]; + tensor input_857_axes_0 = const()[name = string("input_857_axes_0"), val = tensor([-1])]; + tensor encoder_layers_15_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_15_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(323575680)))]; + tensor encoder_layers_15_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_15_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(323577792)))]; + tensor input_857_cast_fp16 = layer_norm(axes = input_857_axes_0, beta = encoder_layers_15_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_15_norm_out_weight_to_fp16, x = input_855_cast_fp16)[name = string("input_857_cast_fp16")]; + tensor cache_65_begin_0 = const()[name = string("cache_65_begin_0"), val = tensor([16, 0, 0, 0])]; + tensor cache_65_end_0 = const()[name = string("cache_65_end_0"), val = tensor([17, 1, 42, 1024])]; + tensor cache_65_end_mask_0 = const()[name = string("cache_65_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_65_squeeze_mask_0 = const()[name = string("cache_65_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_65_cast_fp16 = slice_by_index(begin = cache_65_begin_0, end = cache_65_end_0, end_mask = cache_65_end_mask_0, squeeze_mask = cache_65_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_65_cast_fp16")]; + tensor cache_67_begin_0 = const()[name = string("cache_67_begin_0"), val = tensor([16, 0, 0, 0])]; + tensor cache_67_end_0 = const()[name = string("cache_67_end_0"), val = tensor([17, 1, 1024, 8])]; + tensor cache_67_end_mask_0 = const()[name = string("cache_67_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_67_squeeze_mask_0 = const()[name = string("cache_67_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_67_cast_fp16 = slice_by_index(begin = cache_67_begin_0, end = cache_67_end_0, end_mask = cache_67_end_mask_0, squeeze_mask = cache_67_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_67_cast_fp16")]; + tensor input_859_axes_0 = const()[name = string("input_859_axes_0"), val = tensor([-1])]; + tensor encoder_layers_16_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_16_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(323579904)))]; + tensor encoder_layers_16_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_16_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(323582016)))]; + tensor input_859_cast_fp16 = layer_norm(axes = input_859_axes_0, beta = encoder_layers_16_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_16_norm_feed_forward1_weight_to_fp16, x = input_857_cast_fp16)[name = string("input_859_cast_fp16")]; + tensor encoder_layers_16_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(323584128))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(326729920))))[name = string("encoder_layers_16_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_16_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_16_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(326730112)))]; + tensor linear_145_cast_fp16 = linear(bias = encoder_layers_16_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_16_feed_forward1_linear1_weight_to_fp16_palettized, x = input_859_cast_fp16)[name = string("linear_145_cast_fp16")]; + tensor input_863_cast_fp16 = silu(x = linear_145_cast_fp16)[name = string("input_863_cast_fp16")]; + tensor encoder_layers_16_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(326738368))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(329884160))))[name = string("encoder_layers_16_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_16_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_16_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(329884352)))]; + tensor linear_146_cast_fp16 = linear(bias = encoder_layers_16_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_16_feed_forward1_linear2_weight_to_fp16_palettized, x = input_863_cast_fp16)[name = string("linear_146_cast_fp16")]; + fp16 var_3863_to_fp16 = const()[name = string("op_3863_to_fp16"), val = fp16(0x1p-1)]; + tensor var_3864_cast_fp16 = mul(x = linear_146_cast_fp16, y = var_3863_to_fp16)[name = string("op_3864_cast_fp16")]; + tensor input_869_cast_fp16 = add(x = input_857_cast_fp16, y = var_3864_cast_fp16)[name = string("input_869_cast_fp16")]; + tensor key_33_axes_0 = const()[name = string("key_33_axes_0"), val = tensor([-1])]; + tensor encoder_layers_16_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_16_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(329886464)))]; + tensor encoder_layers_16_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_16_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(329888576)))]; + tensor key_33_cast_fp16 = layer_norm(axes = key_33_axes_0, beta = encoder_layers_16_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_16_norm_self_att_weight_to_fp16, x = input_869_cast_fp16)[name = string("key_33_cast_fp16")]; + bool input_871_interleave_0 = const()[name = string("input_871_interleave_0"), val = bool(false)]; + tensor input_871_cast_fp16 = concat(axis = var_68, interleave = input_871_interleave_0, values = (cache_65_cast_fp16, key_33_cast_fp16))[name = string("input_871_cast_fp16")]; + tensor var_3886_begin_0 = const()[name = string("op_3886_begin_0"), val = tensor([0, 28, 0])]; + tensor var_3886_end_0 = const()[name = string("op_3886_end_0"), val = tensor([1, 42, 1024])]; + tensor var_3886_end_mask_0 = const()[name = string("op_3886_end_mask_0"), val = tensor([true, true, true])]; + tensor var_3886_cast_fp16 = slice_by_index(begin = var_3886_begin_0, end = var_3886_end_0, end_mask = var_3886_end_mask_0, x = cache_65_cast_fp16)[name = string("op_3886_cast_fp16")]; + bool var_3892_interleave_0 = const()[name = string("op_3892_interleave_0"), val = bool(false)]; + tensor var_3892_cast_fp16 = concat(axis = var_68, interleave = var_3892_interleave_0, values = (var_3886_cast_fp16, key_33_cast_fp16))[name = string("op_3892_cast_fp16")]; + tensor encoder_layers_16_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(329890688))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(330677184))))[name = string("encoder_layers_16_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_16_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_16_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(330677376)))]; + tensor linear_147_cast_fp16 = linear(bias = encoder_layers_16_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_16_self_attn_linear_q_weight_to_fp16_palettized, x = key_33_cast_fp16)[name = string("linear_147_cast_fp16")]; + tensor var_3897 = const()[name = string("op_3897"), val = tensor([1, -1, 8, 128])]; + tensor q_97_cast_fp16 = reshape(shape = var_3897, x = linear_147_cast_fp16)[name = string("q_97_cast_fp16")]; + tensor encoder_layers_16_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(330679488))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331465984))))[name = string("encoder_layers_16_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_16_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_16_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331466176)))]; + tensor linear_148_cast_fp16 = linear(bias = encoder_layers_16_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_16_self_attn_linear_k_weight_to_fp16_palettized, x = input_871_cast_fp16)[name = string("linear_148_cast_fp16")]; + tensor var_3902 = const()[name = string("op_3902"), val = tensor([1, -1, 8, 128])]; + tensor k_65_cast_fp16 = reshape(shape = var_3902, x = linear_148_cast_fp16)[name = string("k_65_cast_fp16")]; + tensor encoder_layers_16_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331468288))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332254784))))[name = string("encoder_layers_16_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_16_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_16_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332254976)))]; + tensor linear_149_cast_fp16 = linear(bias = encoder_layers_16_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_16_self_attn_linear_v_weight_to_fp16_palettized, x = input_871_cast_fp16)[name = string("linear_149_cast_fp16")]; + tensor var_3907 = const()[name = string("op_3907"), val = tensor([1, -1, 8, 128])]; + tensor v_33_cast_fp16 = reshape(shape = var_3907, x = linear_149_cast_fp16)[name = string("v_33_cast_fp16")]; + tensor value_41_perm_0 = const()[name = string("value_41_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_16_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_16_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332257088)))]; + tensor var_3920_cast_fp16 = add(x = q_97_cast_fp16, y = encoder_layers_16_self_attn_pos_bias_u_to_fp16)[name = string("op_3920_cast_fp16")]; + tensor encoder_layers_16_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_16_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332259200)))]; + tensor var_3922_cast_fp16 = add(x = q_97_cast_fp16, y = encoder_layers_16_self_attn_pos_bias_v_to_fp16)[name = string("op_3922_cast_fp16")]; + tensor q_with_bias_v_33_perm_0 = const()[name = string("q_with_bias_v_33_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_423_transpose_x_0 = const()[name = string("x_423_transpose_x_0"), val = bool(false)]; + bool x_423_transpose_y_0 = const()[name = string("x_423_transpose_y_0"), val = bool(false)]; + tensor op_3924_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332261312))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332403712))))[name = string("op_3924_to_fp16_quantized")]; + tensor q_with_bias_v_33_cast_fp16 = transpose(perm = q_with_bias_v_33_perm_0, x = var_3922_cast_fp16)[name = string("transpose_218")]; + tensor x_423_cast_fp16 = matmul(transpose_x = x_423_transpose_x_0, transpose_y = x_423_transpose_y_0, x = q_with_bias_v_33_cast_fp16, y = op_3924_to_fp16_quantized)[name = string("x_423_cast_fp16")]; + tensor x_425_pad_0 = const()[name = string("x_425_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_425_mode_0 = const()[name = string("x_425_mode_0"), val = string("constant")]; + fp16 const_287_to_fp16 = const()[name = string("const_287_to_fp16"), val = fp16(0x0p+0)]; + tensor x_425_cast_fp16 = pad(constant_val = const_287_to_fp16, mode = x_425_mode_0, pad = x_425_pad_0, x = x_423_cast_fp16)[name = string("x_425_cast_fp16")]; + tensor var_3932 = const()[name = string("op_3932"), val = tensor([1, 8, -1, 28])]; + tensor x_427_cast_fp16 = reshape(shape = var_3932, x = x_425_cast_fp16)[name = string("x_427_cast_fp16")]; + tensor var_3936_begin_0 = const()[name = string("op_3936_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_3936_end_0 = const()[name = string("op_3936_end_0"), val = tensor([1, 8, 140, 28])]; + tensor var_3936_end_mask_0 = const()[name = string("op_3936_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_3936_cast_fp16 = slice_by_index(begin = var_3936_begin_0, end = var_3936_end_0, end_mask = var_3936_end_mask_0, x = x_427_cast_fp16)[name = string("op_3936_cast_fp16")]; + tensor var_3937 = const()[name = string("op_3937"), val = tensor([1, 8, 28, 139])]; + tensor matrix_bd_65_cast_fp16 = reshape(shape = var_3937, x = var_3936_cast_fp16)[name = string("matrix_bd_65_cast_fp16")]; + bool matrix_ac_33_transpose_x_0 = const()[name = string("matrix_ac_33_transpose_x_0"), val = bool(false)]; + bool matrix_ac_33_transpose_y_0 = const()[name = string("matrix_ac_33_transpose_y_0"), val = bool(false)]; + tensor transpose_128_perm_0 = const()[name = string("transpose_128_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_129_perm_0 = const()[name = string("transpose_129_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_129 = transpose(perm = transpose_129_perm_0, x = k_65_cast_fp16)[name = string("transpose_216")]; + tensor transpose_128 = transpose(perm = transpose_128_perm_0, x = var_3920_cast_fp16)[name = string("transpose_217")]; + tensor matrix_ac_33_cast_fp16 = matmul(transpose_x = matrix_ac_33_transpose_x_0, transpose_y = matrix_ac_33_transpose_y_0, x = transpose_128, y = transpose_129)[name = string("matrix_ac_33_cast_fp16")]; + tensor matrix_bd_67_begin_0 = const()[name = string("matrix_bd_67_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_67_end_0 = const()[name = string("matrix_bd_67_end_0"), val = tensor([1, 8, 28, 70])]; + tensor matrix_bd_67_end_mask_0 = const()[name = string("matrix_bd_67_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_67_cast_fp16 = slice_by_index(begin = matrix_bd_67_begin_0, end = matrix_bd_67_end_0, end_mask = matrix_bd_67_end_mask_0, x = matrix_bd_65_cast_fp16)[name = string("matrix_bd_67_cast_fp16")]; + tensor var_3946_cast_fp16 = add(x = matrix_ac_33_cast_fp16, y = matrix_bd_67_cast_fp16)[name = string("op_3946_cast_fp16")]; + fp16 _inversed_scores_65_y_0_to_fp16 = const()[name = string("_inversed_scores_65_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_65_cast_fp16 = mul(x = var_3946_cast_fp16, y = _inversed_scores_65_y_0_to_fp16)[name = string("_inversed_scores_65_cast_fp16")]; + tensor scores_67_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_65_cast_fp16, cond = mask_11)[name = string("scores_67_cast_fp16")]; + tensor var_3952_cast_fp16 = softmax(axis = var_59, x = scores_67_cast_fp16)[name = string("op_3952_cast_fp16")]; + tensor input_873_cast_fp16 = select(a = var_44_to_fp16, b = var_3952_cast_fp16, cond = mask_11)[name = string("input_873_cast_fp16")]; + bool x_429_transpose_x_0 = const()[name = string("x_429_transpose_x_0"), val = bool(false)]; + bool x_429_transpose_y_0 = const()[name = string("x_429_transpose_y_0"), val = bool(false)]; + tensor value_41_cast_fp16 = transpose(perm = value_41_perm_0, x = v_33_cast_fp16)[name = string("transpose_215")]; + tensor x_429_cast_fp16 = matmul(transpose_x = x_429_transpose_x_0, transpose_y = x_429_transpose_y_0, x = input_873_cast_fp16, y = value_41_cast_fp16)[name = string("x_429_cast_fp16")]; + tensor var_3956_perm_0 = const()[name = string("op_3956_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_3957 = const()[name = string("op_3957"), val = tensor([1, -1, 1024])]; + tensor var_3956_cast_fp16 = transpose(perm = var_3956_perm_0, x = x_429_cast_fp16)[name = string("transpose_214")]; + tensor input_875_cast_fp16 = reshape(shape = var_3957, x = var_3956_cast_fp16)[name = string("input_875_cast_fp16")]; + tensor encoder_layers_16_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332404096))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(333190592))))[name = string("encoder_layers_16_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_16_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_16_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(333190784)))]; + tensor linear_151_cast_fp16 = linear(bias = encoder_layers_16_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_16_self_attn_linear_out_weight_to_fp16_palettized, x = input_875_cast_fp16)[name = string("linear_151_cast_fp16")]; + tensor input_879_cast_fp16 = add(x = input_869_cast_fp16, y = linear_151_cast_fp16)[name = string("input_879_cast_fp16")]; + tensor x_433_axes_0 = const()[name = string("x_433_axes_0"), val = tensor([-1])]; + tensor encoder_layers_16_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_16_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(333192896)))]; + tensor encoder_layers_16_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_16_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(333195008)))]; + tensor x_433_cast_fp16 = layer_norm(axes = x_433_axes_0, beta = encoder_layers_16_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_16_norm_conv_weight_to_fp16, x = input_879_cast_fp16)[name = string("x_433_cast_fp16")]; + tensor input_881_perm_0 = const()[name = string("input_881_perm_0"), val = tensor([0, 2, 1])]; + string input_883_pad_type_0 = const()[name = string("input_883_pad_type_0"), val = string("valid")]; + tensor input_883_strides_0 = const()[name = string("input_883_strides_0"), val = tensor([1])]; + tensor input_883_pad_0 = const()[name = string("input_883_pad_0"), val = tensor([0, 0])]; + tensor input_883_dilations_0 = const()[name = string("input_883_dilations_0"), val = tensor([1])]; + int32 input_883_groups_0 = const()[name = string("input_883_groups_0"), val = int32(1)]; + tensor encoder_layers_16_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(333197120))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(335294336))))[name = string("encoder_layers_16_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_881_cast_fp16 = transpose(perm = input_881_perm_0, x = x_433_cast_fp16)[name = string("transpose_213")]; + tensor input_883_cast_fp16 = conv(dilations = input_883_dilations_0, groups = input_883_groups_0, pad = input_883_pad_0, pad_type = input_883_pad_type_0, strides = input_883_strides_0, weight = encoder_layers_16_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_881_cast_fp16)[name = string("input_883_cast_fp16")]; + int32 x_435_split_num_splits_0 = const()[name = string("x_435_split_num_splits_0"), val = int32(2)]; + int32 x_435_split_axis_0 = const()[name = string("x_435_split_axis_0"), val = int32(1)]; + tensor x_435_split_cast_fp16_0, tensor x_435_split_cast_fp16_1 = split(axis = x_435_split_axis_0, num_splits = x_435_split_num_splits_0, x = input_883_cast_fp16)[name = string("x_435_split_cast_fp16")]; + tensor x_435_split_1_sigmoid_cast_fp16 = sigmoid(x = x_435_split_cast_fp16_1)[name = string("x_435_split_1_sigmoid_cast_fp16")]; + tensor x_435_cast_fp16 = mul(x = x_435_split_cast_fp16_0, y = x_435_split_1_sigmoid_cast_fp16)[name = string("x_435_cast_fp16")]; + tensor input_885_cast_fp16 = select(a = var_44_to_fp16, b = x_435_cast_fp16, cond = var_575)[name = string("input_885_cast_fp16")]; + bool new_x_67_interleave_0 = const()[name = string("new_x_67_interleave_0"), val = bool(false)]; + tensor new_x_67_cast_fp16 = concat(axis = var_59, interleave = new_x_67_interleave_0, values = (cache_67_cast_fp16, input_885_cast_fp16))[name = string("new_x_67_cast_fp16")]; + tensor var_3996_begin_0 = const()[name = string("op_3996_begin_0"), val = tensor([0, 0, 28])]; + tensor var_3996_end_0 = const()[name = string("op_3996_end_0"), val = tensor([1, 1024, 36])]; + tensor var_3996_end_mask_0 = const()[name = string("op_3996_end_mask_0"), val = tensor([true, true, true])]; + tensor var_3996_cast_fp16 = slice_by_index(begin = var_3996_begin_0, end = var_3996_end_0, end_mask = var_3996_end_mask_0, x = new_x_67_cast_fp16)[name = string("op_3996_cast_fp16")]; + string x_437_pad_type_0 = const()[name = string("x_437_pad_type_0"), val = string("valid")]; + int32 x_437_groups_0 = const()[name = string("x_437_groups_0"), val = int32(1024)]; + tensor x_437_strides_0 = const()[name = string("x_437_strides_0"), val = tensor([1])]; + tensor x_437_pad_0 = const()[name = string("x_437_pad_0"), val = tensor([0, 0])]; + tensor x_437_dilations_0 = const()[name = string("x_437_dilations_0"), val = tensor([1])]; + tensor encoder_layers_16_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(335298496))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(335307776))))[name = string("encoder_layers_16_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_437_cast_fp16 = conv(dilations = x_437_dilations_0, groups = x_437_groups_0, pad = x_437_pad_0, pad_type = x_437_pad_type_0, strides = x_437_strides_0, weight = encoder_layers_16_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_67_cast_fp16)[name = string("x_437_cast_fp16")]; + tensor input_887_perm_0 = const()[name = string("input_887_perm_0"), val = tensor([0, 2, 1])]; + tensor x_439_axes_0 = const()[name = string("x_439_axes_0"), val = tensor([-1])]; + tensor encoder_layers_16_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_16_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(335309888)))]; + tensor encoder_layers_16_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_16_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(335312000)))]; + tensor input_887_cast_fp16 = transpose(perm = input_887_perm_0, x = x_437_cast_fp16)[name = string("transpose_212")]; + tensor x_439_cast_fp16 = layer_norm(axes = x_439_axes_0, beta = encoder_layers_16_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_16_conv_batch_norm_weight_to_fp16, x = input_887_cast_fp16)[name = string("x_439_cast_fp16")]; + tensor input_889_perm_0 = const()[name = string("input_889_perm_0"), val = tensor([0, 2, 1])]; + tensor input_889_cast_fp16 = transpose(perm = input_889_perm_0, x = x_439_cast_fp16)[name = string("transpose_211")]; + tensor input_891_cast_fp16 = silu(x = input_889_cast_fp16)[name = string("input_891_cast_fp16")]; + string x_441_pad_type_0 = const()[name = string("x_441_pad_type_0"), val = string("valid")]; + tensor x_441_strides_0 = const()[name = string("x_441_strides_0"), val = tensor([1])]; + tensor x_441_pad_0 = const()[name = string("x_441_pad_0"), val = tensor([0, 0])]; + tensor x_441_dilations_0 = const()[name = string("x_441_dilations_0"), val = tensor([1])]; + int32 x_441_groups_0 = const()[name = string("x_441_groups_0"), val = int32(1)]; + tensor encoder_layers_16_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(335314112))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(336362752))))[name = string("encoder_layers_16_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_441_cast_fp16 = conv(dilations = x_441_dilations_0, groups = x_441_groups_0, pad = x_441_pad_0, pad_type = x_441_pad_type_0, strides = x_441_strides_0, weight = encoder_layers_16_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_891_cast_fp16)[name = string("x_441_cast_fp16")]; + tensor input_893_perm_0 = const()[name = string("input_893_perm_0"), val = tensor([0, 2, 1])]; + tensor input_893_cast_fp16 = transpose(perm = input_893_perm_0, x = x_441_cast_fp16)[name = string("transpose_210")]; + tensor input_895_cast_fp16 = add(x = input_879_cast_fp16, y = input_893_cast_fp16)[name = string("input_895_cast_fp16")]; + tensor input_897_axes_0 = const()[name = string("input_897_axes_0"), val = tensor([-1])]; + tensor encoder_layers_16_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_16_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(336364864)))]; + tensor encoder_layers_16_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_16_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(336366976)))]; + tensor input_897_cast_fp16 = layer_norm(axes = input_897_axes_0, beta = encoder_layers_16_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_16_norm_feed_forward2_weight_to_fp16, x = input_895_cast_fp16)[name = string("input_897_cast_fp16")]; + tensor encoder_layers_16_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(336369088))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(339514880))))[name = string("encoder_layers_16_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_16_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_16_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(339515072)))]; + tensor linear_152_cast_fp16 = linear(bias = encoder_layers_16_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_16_feed_forward2_linear1_weight_to_fp16_palettized, x = input_897_cast_fp16)[name = string("linear_152_cast_fp16")]; + tensor input_901_cast_fp16 = silu(x = linear_152_cast_fp16)[name = string("input_901_cast_fp16")]; + tensor encoder_layers_16_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(339523328))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(342669120))))[name = string("encoder_layers_16_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_16_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_16_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(342669312)))]; + tensor linear_153_cast_fp16 = linear(bias = encoder_layers_16_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_16_feed_forward2_linear2_weight_to_fp16_palettized, x = input_901_cast_fp16)[name = string("linear_153_cast_fp16")]; + fp16 var_4039_to_fp16 = const()[name = string("op_4039_to_fp16"), val = fp16(0x1p-1)]; + tensor var_4040_cast_fp16 = mul(x = linear_153_cast_fp16, y = var_4039_to_fp16)[name = string("op_4040_cast_fp16")]; + tensor input_907_cast_fp16 = add(x = input_895_cast_fp16, y = var_4040_cast_fp16)[name = string("input_907_cast_fp16")]; + tensor input_909_axes_0 = const()[name = string("input_909_axes_0"), val = tensor([-1])]; + tensor encoder_layers_16_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_16_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(342671424)))]; + tensor encoder_layers_16_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_16_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(342673536)))]; + tensor input_909_cast_fp16 = layer_norm(axes = input_909_axes_0, beta = encoder_layers_16_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_16_norm_out_weight_to_fp16, x = input_907_cast_fp16)[name = string("input_909_cast_fp16")]; + tensor cache_69_begin_0 = const()[name = string("cache_69_begin_0"), val = tensor([17, 0, 0, 0])]; + tensor cache_69_end_0 = const()[name = string("cache_69_end_0"), val = tensor([18, 1, 42, 1024])]; + tensor cache_69_end_mask_0 = const()[name = string("cache_69_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_69_squeeze_mask_0 = const()[name = string("cache_69_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_69_cast_fp16 = slice_by_index(begin = cache_69_begin_0, end = cache_69_end_0, end_mask = cache_69_end_mask_0, squeeze_mask = cache_69_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_69_cast_fp16")]; + tensor cache_71_begin_0 = const()[name = string("cache_71_begin_0"), val = tensor([17, 0, 0, 0])]; + tensor cache_71_end_0 = const()[name = string("cache_71_end_0"), val = tensor([18, 1, 1024, 8])]; + tensor cache_71_end_mask_0 = const()[name = string("cache_71_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_71_squeeze_mask_0 = const()[name = string("cache_71_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_71_cast_fp16 = slice_by_index(begin = cache_71_begin_0, end = cache_71_end_0, end_mask = cache_71_end_mask_0, squeeze_mask = cache_71_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_71_cast_fp16")]; + tensor input_911_axes_0 = const()[name = string("input_911_axes_0"), val = tensor([-1])]; + tensor encoder_layers_17_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_17_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(342675648)))]; + tensor encoder_layers_17_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_17_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(342677760)))]; + tensor input_911_cast_fp16 = layer_norm(axes = input_911_axes_0, beta = encoder_layers_17_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_17_norm_feed_forward1_weight_to_fp16, x = input_909_cast_fp16)[name = string("input_911_cast_fp16")]; + tensor encoder_layers_17_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(342679872))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(345825664))))[name = string("encoder_layers_17_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_17_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_17_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(345825856)))]; + tensor linear_154_cast_fp16 = linear(bias = encoder_layers_17_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_17_feed_forward1_linear1_weight_to_fp16_palettized, x = input_911_cast_fp16)[name = string("linear_154_cast_fp16")]; + tensor input_915_cast_fp16 = silu(x = linear_154_cast_fp16)[name = string("input_915_cast_fp16")]; + tensor encoder_layers_17_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(345834112))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(348979904))))[name = string("encoder_layers_17_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_17_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_17_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(348980096)))]; + tensor linear_155_cast_fp16 = linear(bias = encoder_layers_17_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_17_feed_forward1_linear2_weight_to_fp16_palettized, x = input_915_cast_fp16)[name = string("linear_155_cast_fp16")]; + fp16 var_4076_to_fp16 = const()[name = string("op_4076_to_fp16"), val = fp16(0x1p-1)]; + tensor var_4077_cast_fp16 = mul(x = linear_155_cast_fp16, y = var_4076_to_fp16)[name = string("op_4077_cast_fp16")]; + tensor input_921_cast_fp16 = add(x = input_909_cast_fp16, y = var_4077_cast_fp16)[name = string("input_921_cast_fp16")]; + tensor key_35_axes_0 = const()[name = string("key_35_axes_0"), val = tensor([-1])]; + tensor encoder_layers_17_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_17_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(348982208)))]; + tensor encoder_layers_17_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_17_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(348984320)))]; + tensor key_35_cast_fp16 = layer_norm(axes = key_35_axes_0, beta = encoder_layers_17_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_17_norm_self_att_weight_to_fp16, x = input_921_cast_fp16)[name = string("key_35_cast_fp16")]; + bool input_923_interleave_0 = const()[name = string("input_923_interleave_0"), val = bool(false)]; + tensor input_923_cast_fp16 = concat(axis = var_68, interleave = input_923_interleave_0, values = (cache_69_cast_fp16, key_35_cast_fp16))[name = string("input_923_cast_fp16")]; + tensor var_4099_begin_0 = const()[name = string("op_4099_begin_0"), val = tensor([0, 28, 0])]; + tensor var_4099_end_0 = const()[name = string("op_4099_end_0"), val = tensor([1, 42, 1024])]; + tensor var_4099_end_mask_0 = const()[name = string("op_4099_end_mask_0"), val = tensor([true, true, true])]; + tensor var_4099_cast_fp16 = slice_by_index(begin = var_4099_begin_0, end = var_4099_end_0, end_mask = var_4099_end_mask_0, x = cache_69_cast_fp16)[name = string("op_4099_cast_fp16")]; + bool var_4105_interleave_0 = const()[name = string("op_4105_interleave_0"), val = bool(false)]; + tensor var_4105_cast_fp16 = concat(axis = var_68, interleave = var_4105_interleave_0, values = (var_4099_cast_fp16, key_35_cast_fp16))[name = string("op_4105_cast_fp16")]; + tensor encoder_layers_17_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(348986432))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(349772928))))[name = string("encoder_layers_17_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_17_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_17_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(349773120)))]; + tensor linear_156_cast_fp16 = linear(bias = encoder_layers_17_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_17_self_attn_linear_q_weight_to_fp16_palettized, x = key_35_cast_fp16)[name = string("linear_156_cast_fp16")]; + tensor var_4110 = const()[name = string("op_4110"), val = tensor([1, -1, 8, 128])]; + tensor q_103_cast_fp16 = reshape(shape = var_4110, x = linear_156_cast_fp16)[name = string("q_103_cast_fp16")]; + tensor encoder_layers_17_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(349775232))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350561728))))[name = string("encoder_layers_17_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_17_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_17_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350561920)))]; + tensor linear_157_cast_fp16 = linear(bias = encoder_layers_17_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_17_self_attn_linear_k_weight_to_fp16_palettized, x = input_923_cast_fp16)[name = string("linear_157_cast_fp16")]; + tensor var_4115 = const()[name = string("op_4115"), val = tensor([1, -1, 8, 128])]; + tensor k_69_cast_fp16 = reshape(shape = var_4115, x = linear_157_cast_fp16)[name = string("k_69_cast_fp16")]; + tensor encoder_layers_17_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350564032))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351350528))))[name = string("encoder_layers_17_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_17_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_17_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351350720)))]; + tensor linear_158_cast_fp16 = linear(bias = encoder_layers_17_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_17_self_attn_linear_v_weight_to_fp16_palettized, x = input_923_cast_fp16)[name = string("linear_158_cast_fp16")]; + tensor var_4120 = const()[name = string("op_4120"), val = tensor([1, -1, 8, 128])]; + tensor v_35_cast_fp16 = reshape(shape = var_4120, x = linear_158_cast_fp16)[name = string("v_35_cast_fp16")]; + tensor value_43_perm_0 = const()[name = string("value_43_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_17_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_17_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351352832)))]; + tensor var_4133_cast_fp16 = add(x = q_103_cast_fp16, y = encoder_layers_17_self_attn_pos_bias_u_to_fp16)[name = string("op_4133_cast_fp16")]; + tensor encoder_layers_17_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_17_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351354944)))]; + tensor var_4135_cast_fp16 = add(x = q_103_cast_fp16, y = encoder_layers_17_self_attn_pos_bias_v_to_fp16)[name = string("op_4135_cast_fp16")]; + tensor q_with_bias_v_35_perm_0 = const()[name = string("q_with_bias_v_35_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_449_transpose_x_0 = const()[name = string("x_449_transpose_x_0"), val = bool(false)]; + bool x_449_transpose_y_0 = const()[name = string("x_449_transpose_y_0"), val = bool(false)]; + tensor op_4137_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351357056))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351499456))))[name = string("op_4137_to_fp16_quantized")]; + tensor q_with_bias_v_35_cast_fp16 = transpose(perm = q_with_bias_v_35_perm_0, x = var_4135_cast_fp16)[name = string("transpose_209")]; + tensor x_449_cast_fp16 = matmul(transpose_x = x_449_transpose_x_0, transpose_y = x_449_transpose_y_0, x = q_with_bias_v_35_cast_fp16, y = op_4137_to_fp16_quantized)[name = string("x_449_cast_fp16")]; + tensor x_451_pad_0 = const()[name = string("x_451_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_451_mode_0 = const()[name = string("x_451_mode_0"), val = string("constant")]; + fp16 const_300_to_fp16 = const()[name = string("const_300_to_fp16"), val = fp16(0x0p+0)]; + tensor x_451_cast_fp16 = pad(constant_val = const_300_to_fp16, mode = x_451_mode_0, pad = x_451_pad_0, x = x_449_cast_fp16)[name = string("x_451_cast_fp16")]; + tensor var_4145 = const()[name = string("op_4145"), val = tensor([1, 8, -1, 28])]; + tensor x_453_cast_fp16 = reshape(shape = var_4145, x = x_451_cast_fp16)[name = string("x_453_cast_fp16")]; + tensor var_4149_begin_0 = const()[name = string("op_4149_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_4149_end_0 = const()[name = string("op_4149_end_0"), val = tensor([1, 8, 140, 28])]; + tensor var_4149_end_mask_0 = const()[name = string("op_4149_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_4149_cast_fp16 = slice_by_index(begin = var_4149_begin_0, end = var_4149_end_0, end_mask = var_4149_end_mask_0, x = x_453_cast_fp16)[name = string("op_4149_cast_fp16")]; + tensor var_4150 = const()[name = string("op_4150"), val = tensor([1, 8, 28, 139])]; + tensor matrix_bd_69_cast_fp16 = reshape(shape = var_4150, x = var_4149_cast_fp16)[name = string("matrix_bd_69_cast_fp16")]; + bool matrix_ac_35_transpose_x_0 = const()[name = string("matrix_ac_35_transpose_x_0"), val = bool(false)]; + bool matrix_ac_35_transpose_y_0 = const()[name = string("matrix_ac_35_transpose_y_0"), val = bool(false)]; + tensor transpose_130_perm_0 = const()[name = string("transpose_130_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_131_perm_0 = const()[name = string("transpose_131_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_131 = transpose(perm = transpose_131_perm_0, x = k_69_cast_fp16)[name = string("transpose_207")]; + tensor transpose_130 = transpose(perm = transpose_130_perm_0, x = var_4133_cast_fp16)[name = string("transpose_208")]; + tensor matrix_ac_35_cast_fp16 = matmul(transpose_x = matrix_ac_35_transpose_x_0, transpose_y = matrix_ac_35_transpose_y_0, x = transpose_130, y = transpose_131)[name = string("matrix_ac_35_cast_fp16")]; + tensor matrix_bd_71_begin_0 = const()[name = string("matrix_bd_71_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_71_end_0 = const()[name = string("matrix_bd_71_end_0"), val = tensor([1, 8, 28, 70])]; + tensor matrix_bd_71_end_mask_0 = const()[name = string("matrix_bd_71_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_71_cast_fp16 = slice_by_index(begin = matrix_bd_71_begin_0, end = matrix_bd_71_end_0, end_mask = matrix_bd_71_end_mask_0, x = matrix_bd_69_cast_fp16)[name = string("matrix_bd_71_cast_fp16")]; + tensor var_4159_cast_fp16 = add(x = matrix_ac_35_cast_fp16, y = matrix_bd_71_cast_fp16)[name = string("op_4159_cast_fp16")]; + fp16 _inversed_scores_69_y_0_to_fp16 = const()[name = string("_inversed_scores_69_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_69_cast_fp16 = mul(x = var_4159_cast_fp16, y = _inversed_scores_69_y_0_to_fp16)[name = string("_inversed_scores_69_cast_fp16")]; + tensor scores_71_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_69_cast_fp16, cond = mask_11)[name = string("scores_71_cast_fp16")]; + tensor var_4165_cast_fp16 = softmax(axis = var_59, x = scores_71_cast_fp16)[name = string("op_4165_cast_fp16")]; + tensor input_925_cast_fp16 = select(a = var_44_to_fp16, b = var_4165_cast_fp16, cond = mask_11)[name = string("input_925_cast_fp16")]; + bool x_455_transpose_x_0 = const()[name = string("x_455_transpose_x_0"), val = bool(false)]; + bool x_455_transpose_y_0 = const()[name = string("x_455_transpose_y_0"), val = bool(false)]; + tensor value_43_cast_fp16 = transpose(perm = value_43_perm_0, x = v_35_cast_fp16)[name = string("transpose_206")]; + tensor x_455_cast_fp16 = matmul(transpose_x = x_455_transpose_x_0, transpose_y = x_455_transpose_y_0, x = input_925_cast_fp16, y = value_43_cast_fp16)[name = string("x_455_cast_fp16")]; + tensor var_4169_perm_0 = const()[name = string("op_4169_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_4170 = const()[name = string("op_4170"), val = tensor([1, -1, 1024])]; + tensor var_4169_cast_fp16 = transpose(perm = var_4169_perm_0, x = x_455_cast_fp16)[name = string("transpose_205")]; + tensor input_927_cast_fp16 = reshape(shape = var_4170, x = var_4169_cast_fp16)[name = string("input_927_cast_fp16")]; + tensor encoder_layers_17_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351499840))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(352286336))))[name = string("encoder_layers_17_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_17_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_17_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(352286528)))]; + tensor linear_160_cast_fp16 = linear(bias = encoder_layers_17_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_17_self_attn_linear_out_weight_to_fp16_palettized, x = input_927_cast_fp16)[name = string("linear_160_cast_fp16")]; + tensor input_931_cast_fp16 = add(x = input_921_cast_fp16, y = linear_160_cast_fp16)[name = string("input_931_cast_fp16")]; + tensor x_459_axes_0 = const()[name = string("x_459_axes_0"), val = tensor([-1])]; + tensor encoder_layers_17_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_17_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(352288640)))]; + tensor encoder_layers_17_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_17_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(352290752)))]; + tensor x_459_cast_fp16 = layer_norm(axes = x_459_axes_0, beta = encoder_layers_17_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_17_norm_conv_weight_to_fp16, x = input_931_cast_fp16)[name = string("x_459_cast_fp16")]; + tensor input_933_perm_0 = const()[name = string("input_933_perm_0"), val = tensor([0, 2, 1])]; + string input_935_pad_type_0 = const()[name = string("input_935_pad_type_0"), val = string("valid")]; + tensor input_935_strides_0 = const()[name = string("input_935_strides_0"), val = tensor([1])]; + tensor input_935_pad_0 = const()[name = string("input_935_pad_0"), val = tensor([0, 0])]; + tensor input_935_dilations_0 = const()[name = string("input_935_dilations_0"), val = tensor([1])]; + int32 input_935_groups_0 = const()[name = string("input_935_groups_0"), val = int32(1)]; + tensor encoder_layers_17_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(352292864))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(354390080))))[name = string("encoder_layers_17_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_933_cast_fp16 = transpose(perm = input_933_perm_0, x = x_459_cast_fp16)[name = string("transpose_204")]; + tensor input_935_cast_fp16 = conv(dilations = input_935_dilations_0, groups = input_935_groups_0, pad = input_935_pad_0, pad_type = input_935_pad_type_0, strides = input_935_strides_0, weight = encoder_layers_17_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_933_cast_fp16)[name = string("input_935_cast_fp16")]; + int32 x_461_split_num_splits_0 = const()[name = string("x_461_split_num_splits_0"), val = int32(2)]; + int32 x_461_split_axis_0 = const()[name = string("x_461_split_axis_0"), val = int32(1)]; + tensor x_461_split_cast_fp16_0, tensor x_461_split_cast_fp16_1 = split(axis = x_461_split_axis_0, num_splits = x_461_split_num_splits_0, x = input_935_cast_fp16)[name = string("x_461_split_cast_fp16")]; + tensor x_461_split_1_sigmoid_cast_fp16 = sigmoid(x = x_461_split_cast_fp16_1)[name = string("x_461_split_1_sigmoid_cast_fp16")]; + tensor x_461_cast_fp16 = mul(x = x_461_split_cast_fp16_0, y = x_461_split_1_sigmoid_cast_fp16)[name = string("x_461_cast_fp16")]; + tensor input_937_cast_fp16 = select(a = var_44_to_fp16, b = x_461_cast_fp16, cond = var_575)[name = string("input_937_cast_fp16")]; + bool new_x_71_interleave_0 = const()[name = string("new_x_71_interleave_0"), val = bool(false)]; + tensor new_x_71_cast_fp16 = concat(axis = var_59, interleave = new_x_71_interleave_0, values = (cache_71_cast_fp16, input_937_cast_fp16))[name = string("new_x_71_cast_fp16")]; + tensor var_4209_begin_0 = const()[name = string("op_4209_begin_0"), val = tensor([0, 0, 28])]; + tensor var_4209_end_0 = const()[name = string("op_4209_end_0"), val = tensor([1, 1024, 36])]; + tensor var_4209_end_mask_0 = const()[name = string("op_4209_end_mask_0"), val = tensor([true, true, true])]; + tensor var_4209_cast_fp16 = slice_by_index(begin = var_4209_begin_0, end = var_4209_end_0, end_mask = var_4209_end_mask_0, x = new_x_71_cast_fp16)[name = string("op_4209_cast_fp16")]; + string x_463_pad_type_0 = const()[name = string("x_463_pad_type_0"), val = string("valid")]; + int32 x_463_groups_0 = const()[name = string("x_463_groups_0"), val = int32(1024)]; + tensor x_463_strides_0 = const()[name = string("x_463_strides_0"), val = tensor([1])]; + tensor x_463_pad_0 = const()[name = string("x_463_pad_0"), val = tensor([0, 0])]; + tensor x_463_dilations_0 = const()[name = string("x_463_dilations_0"), val = tensor([1])]; + tensor encoder_layers_17_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(354394240))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(354403520))))[name = string("encoder_layers_17_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_463_cast_fp16 = conv(dilations = x_463_dilations_0, groups = x_463_groups_0, pad = x_463_pad_0, pad_type = x_463_pad_type_0, strides = x_463_strides_0, weight = encoder_layers_17_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_71_cast_fp16)[name = string("x_463_cast_fp16")]; + tensor input_939_perm_0 = const()[name = string("input_939_perm_0"), val = tensor([0, 2, 1])]; + tensor x_465_axes_0 = const()[name = string("x_465_axes_0"), val = tensor([-1])]; + tensor encoder_layers_17_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_17_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(354405632)))]; + tensor encoder_layers_17_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_17_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(354407744)))]; + tensor input_939_cast_fp16 = transpose(perm = input_939_perm_0, x = x_463_cast_fp16)[name = string("transpose_203")]; + tensor x_465_cast_fp16 = layer_norm(axes = x_465_axes_0, beta = encoder_layers_17_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_17_conv_batch_norm_weight_to_fp16, x = input_939_cast_fp16)[name = string("x_465_cast_fp16")]; + tensor input_941_perm_0 = const()[name = string("input_941_perm_0"), val = tensor([0, 2, 1])]; + tensor input_941_cast_fp16 = transpose(perm = input_941_perm_0, x = x_465_cast_fp16)[name = string("transpose_202")]; + tensor input_943_cast_fp16 = silu(x = input_941_cast_fp16)[name = string("input_943_cast_fp16")]; + string x_467_pad_type_0 = const()[name = string("x_467_pad_type_0"), val = string("valid")]; + tensor x_467_strides_0 = const()[name = string("x_467_strides_0"), val = tensor([1])]; + tensor x_467_pad_0 = const()[name = string("x_467_pad_0"), val = tensor([0, 0])]; + tensor x_467_dilations_0 = const()[name = string("x_467_dilations_0"), val = tensor([1])]; + int32 x_467_groups_0 = const()[name = string("x_467_groups_0"), val = int32(1)]; + tensor encoder_layers_17_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(354409856))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(355458496))))[name = string("encoder_layers_17_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_467_cast_fp16 = conv(dilations = x_467_dilations_0, groups = x_467_groups_0, pad = x_467_pad_0, pad_type = x_467_pad_type_0, strides = x_467_strides_0, weight = encoder_layers_17_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_943_cast_fp16)[name = string("x_467_cast_fp16")]; + tensor input_945_perm_0 = const()[name = string("input_945_perm_0"), val = tensor([0, 2, 1])]; + tensor input_945_cast_fp16 = transpose(perm = input_945_perm_0, x = x_467_cast_fp16)[name = string("transpose_201")]; + tensor input_947_cast_fp16 = add(x = input_931_cast_fp16, y = input_945_cast_fp16)[name = string("input_947_cast_fp16")]; + tensor input_949_axes_0 = const()[name = string("input_949_axes_0"), val = tensor([-1])]; + tensor encoder_layers_17_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_17_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(355460608)))]; + tensor encoder_layers_17_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_17_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(355462720)))]; + tensor input_949_cast_fp16 = layer_norm(axes = input_949_axes_0, beta = encoder_layers_17_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_17_norm_feed_forward2_weight_to_fp16, x = input_947_cast_fp16)[name = string("input_949_cast_fp16")]; + tensor encoder_layers_17_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(355464832))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(358610624))))[name = string("encoder_layers_17_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_17_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_17_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(358610816)))]; + tensor linear_161_cast_fp16 = linear(bias = encoder_layers_17_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_17_feed_forward2_linear1_weight_to_fp16_palettized, x = input_949_cast_fp16)[name = string("linear_161_cast_fp16")]; + tensor input_953_cast_fp16 = silu(x = linear_161_cast_fp16)[name = string("input_953_cast_fp16")]; + tensor encoder_layers_17_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(358619072))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361764864))))[name = string("encoder_layers_17_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_17_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_17_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361765056)))]; + tensor linear_162_cast_fp16 = linear(bias = encoder_layers_17_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_17_feed_forward2_linear2_weight_to_fp16_palettized, x = input_953_cast_fp16)[name = string("linear_162_cast_fp16")]; + fp16 var_4252_to_fp16 = const()[name = string("op_4252_to_fp16"), val = fp16(0x1p-1)]; + tensor var_4253_cast_fp16 = mul(x = linear_162_cast_fp16, y = var_4252_to_fp16)[name = string("op_4253_cast_fp16")]; + tensor input_959_cast_fp16 = add(x = input_947_cast_fp16, y = var_4253_cast_fp16)[name = string("input_959_cast_fp16")]; + tensor input_961_axes_0 = const()[name = string("input_961_axes_0"), val = tensor([-1])]; + tensor encoder_layers_17_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_17_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361767168)))]; + tensor encoder_layers_17_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_17_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361769280)))]; + tensor input_961_cast_fp16 = layer_norm(axes = input_961_axes_0, beta = encoder_layers_17_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_17_norm_out_weight_to_fp16, x = input_959_cast_fp16)[name = string("input_961_cast_fp16")]; + tensor cache_73_begin_0 = const()[name = string("cache_73_begin_0"), val = tensor([18, 0, 0, 0])]; + tensor cache_73_end_0 = const()[name = string("cache_73_end_0"), val = tensor([19, 1, 42, 1024])]; + tensor cache_73_end_mask_0 = const()[name = string("cache_73_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_73_squeeze_mask_0 = const()[name = string("cache_73_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_73_cast_fp16 = slice_by_index(begin = cache_73_begin_0, end = cache_73_end_0, end_mask = cache_73_end_mask_0, squeeze_mask = cache_73_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_73_cast_fp16")]; + tensor cache_75_begin_0 = const()[name = string("cache_75_begin_0"), val = tensor([18, 0, 0, 0])]; + tensor cache_75_end_0 = const()[name = string("cache_75_end_0"), val = tensor([19, 1, 1024, 8])]; + tensor cache_75_end_mask_0 = const()[name = string("cache_75_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_75_squeeze_mask_0 = const()[name = string("cache_75_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_75_cast_fp16 = slice_by_index(begin = cache_75_begin_0, end = cache_75_end_0, end_mask = cache_75_end_mask_0, squeeze_mask = cache_75_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_75_cast_fp16")]; + tensor input_963_axes_0 = const()[name = string("input_963_axes_0"), val = tensor([-1])]; + tensor encoder_layers_18_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_18_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361771392)))]; + tensor encoder_layers_18_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_18_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361773504)))]; + tensor input_963_cast_fp16 = layer_norm(axes = input_963_axes_0, beta = encoder_layers_18_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_18_norm_feed_forward1_weight_to_fp16, x = input_961_cast_fp16)[name = string("input_963_cast_fp16")]; + tensor encoder_layers_18_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361775616))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(364921408))))[name = string("encoder_layers_18_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_18_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_18_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(364921600)))]; + tensor linear_163_cast_fp16 = linear(bias = encoder_layers_18_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_18_feed_forward1_linear1_weight_to_fp16_palettized, x = input_963_cast_fp16)[name = string("linear_163_cast_fp16")]; + tensor input_967_cast_fp16 = silu(x = linear_163_cast_fp16)[name = string("input_967_cast_fp16")]; + tensor encoder_layers_18_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(364929856))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368075648))))[name = string("encoder_layers_18_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_18_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_18_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368075840)))]; + tensor linear_164_cast_fp16 = linear(bias = encoder_layers_18_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_18_feed_forward1_linear2_weight_to_fp16_palettized, x = input_967_cast_fp16)[name = string("linear_164_cast_fp16")]; + fp16 var_4289_to_fp16 = const()[name = string("op_4289_to_fp16"), val = fp16(0x1p-1)]; + tensor var_4290_cast_fp16 = mul(x = linear_164_cast_fp16, y = var_4289_to_fp16)[name = string("op_4290_cast_fp16")]; + tensor input_973_cast_fp16 = add(x = input_961_cast_fp16, y = var_4290_cast_fp16)[name = string("input_973_cast_fp16")]; + tensor key_37_axes_0 = const()[name = string("key_37_axes_0"), val = tensor([-1])]; + tensor encoder_layers_18_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_18_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368077952)))]; + tensor encoder_layers_18_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_18_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368080064)))]; + tensor key_37_cast_fp16 = layer_norm(axes = key_37_axes_0, beta = encoder_layers_18_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_18_norm_self_att_weight_to_fp16, x = input_973_cast_fp16)[name = string("key_37_cast_fp16")]; + bool input_975_interleave_0 = const()[name = string("input_975_interleave_0"), val = bool(false)]; + tensor input_975_cast_fp16 = concat(axis = var_68, interleave = input_975_interleave_0, values = (cache_73_cast_fp16, key_37_cast_fp16))[name = string("input_975_cast_fp16")]; + tensor var_4312_begin_0 = const()[name = string("op_4312_begin_0"), val = tensor([0, 28, 0])]; + tensor var_4312_end_0 = const()[name = string("op_4312_end_0"), val = tensor([1, 42, 1024])]; + tensor var_4312_end_mask_0 = const()[name = string("op_4312_end_mask_0"), val = tensor([true, true, true])]; + tensor var_4312_cast_fp16 = slice_by_index(begin = var_4312_begin_0, end = var_4312_end_0, end_mask = var_4312_end_mask_0, x = cache_73_cast_fp16)[name = string("op_4312_cast_fp16")]; + bool var_4318_interleave_0 = const()[name = string("op_4318_interleave_0"), val = bool(false)]; + tensor var_4318_cast_fp16 = concat(axis = var_68, interleave = var_4318_interleave_0, values = (var_4312_cast_fp16, key_37_cast_fp16))[name = string("op_4318_cast_fp16")]; + tensor encoder_layers_18_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368082176))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368868672))))[name = string("encoder_layers_18_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_18_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_18_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368868864)))]; + tensor linear_165_cast_fp16 = linear(bias = encoder_layers_18_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_18_self_attn_linear_q_weight_to_fp16_palettized, x = key_37_cast_fp16)[name = string("linear_165_cast_fp16")]; + tensor var_4323 = const()[name = string("op_4323"), val = tensor([1, -1, 8, 128])]; + tensor q_109_cast_fp16 = reshape(shape = var_4323, x = linear_165_cast_fp16)[name = string("q_109_cast_fp16")]; + tensor encoder_layers_18_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368870976))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369657472))))[name = string("encoder_layers_18_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_18_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_18_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369657664)))]; + tensor linear_166_cast_fp16 = linear(bias = encoder_layers_18_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_18_self_attn_linear_k_weight_to_fp16_palettized, x = input_975_cast_fp16)[name = string("linear_166_cast_fp16")]; + tensor var_4328 = const()[name = string("op_4328"), val = tensor([1, -1, 8, 128])]; + tensor k_73_cast_fp16 = reshape(shape = var_4328, x = linear_166_cast_fp16)[name = string("k_73_cast_fp16")]; + tensor encoder_layers_18_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369659776))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(370446272))))[name = string("encoder_layers_18_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_18_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_18_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(370446464)))]; + tensor linear_167_cast_fp16 = linear(bias = encoder_layers_18_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_18_self_attn_linear_v_weight_to_fp16_palettized, x = input_975_cast_fp16)[name = string("linear_167_cast_fp16")]; + tensor var_4333 = const()[name = string("op_4333"), val = tensor([1, -1, 8, 128])]; + tensor v_37_cast_fp16 = reshape(shape = var_4333, x = linear_167_cast_fp16)[name = string("v_37_cast_fp16")]; + tensor value_45_perm_0 = const()[name = string("value_45_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_18_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_18_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(370448576)))]; + tensor var_4346_cast_fp16 = add(x = q_109_cast_fp16, y = encoder_layers_18_self_attn_pos_bias_u_to_fp16)[name = string("op_4346_cast_fp16")]; + tensor encoder_layers_18_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_18_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(370450688)))]; + tensor var_4348_cast_fp16 = add(x = q_109_cast_fp16, y = encoder_layers_18_self_attn_pos_bias_v_to_fp16)[name = string("op_4348_cast_fp16")]; + tensor q_with_bias_v_37_perm_0 = const()[name = string("q_with_bias_v_37_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_475_transpose_x_0 = const()[name = string("x_475_transpose_x_0"), val = bool(false)]; + bool x_475_transpose_y_0 = const()[name = string("x_475_transpose_y_0"), val = bool(false)]; + tensor op_4350_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(370452800))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(370595200))))[name = string("op_4350_to_fp16_quantized")]; + tensor q_with_bias_v_37_cast_fp16 = transpose(perm = q_with_bias_v_37_perm_0, x = var_4348_cast_fp16)[name = string("transpose_200")]; + tensor x_475_cast_fp16 = matmul(transpose_x = x_475_transpose_x_0, transpose_y = x_475_transpose_y_0, x = q_with_bias_v_37_cast_fp16, y = op_4350_to_fp16_quantized)[name = string("x_475_cast_fp16")]; + tensor x_477_pad_0 = const()[name = string("x_477_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_477_mode_0 = const()[name = string("x_477_mode_0"), val = string("constant")]; + fp16 const_313_to_fp16 = const()[name = string("const_313_to_fp16"), val = fp16(0x0p+0)]; + tensor x_477_cast_fp16 = pad(constant_val = const_313_to_fp16, mode = x_477_mode_0, pad = x_477_pad_0, x = x_475_cast_fp16)[name = string("x_477_cast_fp16")]; + tensor var_4358 = const()[name = string("op_4358"), val = tensor([1, 8, -1, 28])]; + tensor x_479_cast_fp16 = reshape(shape = var_4358, x = x_477_cast_fp16)[name = string("x_479_cast_fp16")]; + tensor var_4362_begin_0 = const()[name = string("op_4362_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_4362_end_0 = const()[name = string("op_4362_end_0"), val = tensor([1, 8, 140, 28])]; + tensor var_4362_end_mask_0 = const()[name = string("op_4362_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_4362_cast_fp16 = slice_by_index(begin = var_4362_begin_0, end = var_4362_end_0, end_mask = var_4362_end_mask_0, x = x_479_cast_fp16)[name = string("op_4362_cast_fp16")]; + tensor var_4363 = const()[name = string("op_4363"), val = tensor([1, 8, 28, 139])]; + tensor matrix_bd_73_cast_fp16 = reshape(shape = var_4363, x = var_4362_cast_fp16)[name = string("matrix_bd_73_cast_fp16")]; + bool matrix_ac_37_transpose_x_0 = const()[name = string("matrix_ac_37_transpose_x_0"), val = bool(false)]; + bool matrix_ac_37_transpose_y_0 = const()[name = string("matrix_ac_37_transpose_y_0"), val = bool(false)]; + tensor transpose_132_perm_0 = const()[name = string("transpose_132_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_133_perm_0 = const()[name = string("transpose_133_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_133 = transpose(perm = transpose_133_perm_0, x = k_73_cast_fp16)[name = string("transpose_198")]; + tensor transpose_132 = transpose(perm = transpose_132_perm_0, x = var_4346_cast_fp16)[name = string("transpose_199")]; + tensor matrix_ac_37_cast_fp16 = matmul(transpose_x = matrix_ac_37_transpose_x_0, transpose_y = matrix_ac_37_transpose_y_0, x = transpose_132, y = transpose_133)[name = string("matrix_ac_37_cast_fp16")]; + tensor matrix_bd_75_begin_0 = const()[name = string("matrix_bd_75_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_75_end_0 = const()[name = string("matrix_bd_75_end_0"), val = tensor([1, 8, 28, 70])]; + tensor matrix_bd_75_end_mask_0 = const()[name = string("matrix_bd_75_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_75_cast_fp16 = slice_by_index(begin = matrix_bd_75_begin_0, end = matrix_bd_75_end_0, end_mask = matrix_bd_75_end_mask_0, x = matrix_bd_73_cast_fp16)[name = string("matrix_bd_75_cast_fp16")]; + tensor var_4372_cast_fp16 = add(x = matrix_ac_37_cast_fp16, y = matrix_bd_75_cast_fp16)[name = string("op_4372_cast_fp16")]; + fp16 _inversed_scores_73_y_0_to_fp16 = const()[name = string("_inversed_scores_73_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_73_cast_fp16 = mul(x = var_4372_cast_fp16, y = _inversed_scores_73_y_0_to_fp16)[name = string("_inversed_scores_73_cast_fp16")]; + tensor scores_75_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_73_cast_fp16, cond = mask_11)[name = string("scores_75_cast_fp16")]; + tensor var_4378_cast_fp16 = softmax(axis = var_59, x = scores_75_cast_fp16)[name = string("op_4378_cast_fp16")]; + tensor input_977_cast_fp16 = select(a = var_44_to_fp16, b = var_4378_cast_fp16, cond = mask_11)[name = string("input_977_cast_fp16")]; + bool x_481_transpose_x_0 = const()[name = string("x_481_transpose_x_0"), val = bool(false)]; + bool x_481_transpose_y_0 = const()[name = string("x_481_transpose_y_0"), val = bool(false)]; + tensor value_45_cast_fp16 = transpose(perm = value_45_perm_0, x = v_37_cast_fp16)[name = string("transpose_197")]; + tensor x_481_cast_fp16 = matmul(transpose_x = x_481_transpose_x_0, transpose_y = x_481_transpose_y_0, x = input_977_cast_fp16, y = value_45_cast_fp16)[name = string("x_481_cast_fp16")]; + tensor var_4382_perm_0 = const()[name = string("op_4382_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_4383 = const()[name = string("op_4383"), val = tensor([1, -1, 1024])]; + tensor var_4382_cast_fp16 = transpose(perm = var_4382_perm_0, x = x_481_cast_fp16)[name = string("transpose_196")]; + tensor input_979_cast_fp16 = reshape(shape = var_4383, x = var_4382_cast_fp16)[name = string("input_979_cast_fp16")]; + tensor encoder_layers_18_self_attn_linear_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(370595584))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371644224))))[name = string("encoder_layers_18_self_attn_linear_out_weight_to_fp16_quantized")]; + tensor encoder_layers_18_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_18_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371646336)))]; + tensor linear_169_cast_fp16 = linear(bias = encoder_layers_18_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_18_self_attn_linear_out_weight_to_fp16_quantized, x = input_979_cast_fp16)[name = string("linear_169_cast_fp16")]; + tensor input_983_cast_fp16 = add(x = input_973_cast_fp16, y = linear_169_cast_fp16)[name = string("input_983_cast_fp16")]; + tensor x_485_axes_0 = const()[name = string("x_485_axes_0"), val = tensor([-1])]; + tensor encoder_layers_18_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_18_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371648448)))]; + tensor encoder_layers_18_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_18_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371650560)))]; + tensor x_485_cast_fp16 = layer_norm(axes = x_485_axes_0, beta = encoder_layers_18_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_18_norm_conv_weight_to_fp16, x = input_983_cast_fp16)[name = string("x_485_cast_fp16")]; + tensor input_985_perm_0 = const()[name = string("input_985_perm_0"), val = tensor([0, 2, 1])]; + string input_987_pad_type_0 = const()[name = string("input_987_pad_type_0"), val = string("valid")]; + tensor input_987_strides_0 = const()[name = string("input_987_strides_0"), val = tensor([1])]; + tensor input_987_pad_0 = const()[name = string("input_987_pad_0"), val = tensor([0, 0])]; + tensor input_987_dilations_0 = const()[name = string("input_987_dilations_0"), val = tensor([1])]; + int32 input_987_groups_0 = const()[name = string("input_987_groups_0"), val = int32(1)]; + tensor encoder_layers_18_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371652672))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(373749888))))[name = string("encoder_layers_18_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_985_cast_fp16 = transpose(perm = input_985_perm_0, x = x_485_cast_fp16)[name = string("transpose_195")]; + tensor input_987_cast_fp16 = conv(dilations = input_987_dilations_0, groups = input_987_groups_0, pad = input_987_pad_0, pad_type = input_987_pad_type_0, strides = input_987_strides_0, weight = encoder_layers_18_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_985_cast_fp16)[name = string("input_987_cast_fp16")]; + int32 x_487_split_num_splits_0 = const()[name = string("x_487_split_num_splits_0"), val = int32(2)]; + int32 x_487_split_axis_0 = const()[name = string("x_487_split_axis_0"), val = int32(1)]; + tensor x_487_split_cast_fp16_0, tensor x_487_split_cast_fp16_1 = split(axis = x_487_split_axis_0, num_splits = x_487_split_num_splits_0, x = input_987_cast_fp16)[name = string("x_487_split_cast_fp16")]; + tensor x_487_split_1_sigmoid_cast_fp16 = sigmoid(x = x_487_split_cast_fp16_1)[name = string("x_487_split_1_sigmoid_cast_fp16")]; + tensor x_487_cast_fp16 = mul(x = x_487_split_cast_fp16_0, y = x_487_split_1_sigmoid_cast_fp16)[name = string("x_487_cast_fp16")]; + tensor input_989_cast_fp16 = select(a = var_44_to_fp16, b = x_487_cast_fp16, cond = var_575)[name = string("input_989_cast_fp16")]; + bool new_x_75_interleave_0 = const()[name = string("new_x_75_interleave_0"), val = bool(false)]; + tensor new_x_75_cast_fp16 = concat(axis = var_59, interleave = new_x_75_interleave_0, values = (cache_75_cast_fp16, input_989_cast_fp16))[name = string("new_x_75_cast_fp16")]; + tensor var_4422_begin_0 = const()[name = string("op_4422_begin_0"), val = tensor([0, 0, 28])]; + tensor var_4422_end_0 = const()[name = string("op_4422_end_0"), val = tensor([1, 1024, 36])]; + tensor var_4422_end_mask_0 = const()[name = string("op_4422_end_mask_0"), val = tensor([true, true, true])]; + tensor var_4422_cast_fp16 = slice_by_index(begin = var_4422_begin_0, end = var_4422_end_0, end_mask = var_4422_end_mask_0, x = new_x_75_cast_fp16)[name = string("op_4422_cast_fp16")]; + string x_489_pad_type_0 = const()[name = string("x_489_pad_type_0"), val = string("valid")]; + int32 x_489_groups_0 = const()[name = string("x_489_groups_0"), val = int32(1024)]; + tensor x_489_strides_0 = const()[name = string("x_489_strides_0"), val = tensor([1])]; + tensor x_489_pad_0 = const()[name = string("x_489_pad_0"), val = tensor([0, 0])]; + tensor x_489_dilations_0 = const()[name = string("x_489_dilations_0"), val = tensor([1])]; + tensor encoder_layers_18_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(373754048))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(373763328))))[name = string("encoder_layers_18_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_489_cast_fp16 = conv(dilations = x_489_dilations_0, groups = x_489_groups_0, pad = x_489_pad_0, pad_type = x_489_pad_type_0, strides = x_489_strides_0, weight = encoder_layers_18_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_75_cast_fp16)[name = string("x_489_cast_fp16")]; + tensor input_991_perm_0 = const()[name = string("input_991_perm_0"), val = tensor([0, 2, 1])]; + tensor x_491_axes_0 = const()[name = string("x_491_axes_0"), val = tensor([-1])]; + tensor encoder_layers_18_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_18_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(373765440)))]; + tensor encoder_layers_18_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_18_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(373767552)))]; + tensor input_991_cast_fp16 = transpose(perm = input_991_perm_0, x = x_489_cast_fp16)[name = string("transpose_194")]; + tensor x_491_cast_fp16 = layer_norm(axes = x_491_axes_0, beta = encoder_layers_18_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_18_conv_batch_norm_weight_to_fp16, x = input_991_cast_fp16)[name = string("x_491_cast_fp16")]; + tensor input_993_perm_0 = const()[name = string("input_993_perm_0"), val = tensor([0, 2, 1])]; + tensor input_993_cast_fp16 = transpose(perm = input_993_perm_0, x = x_491_cast_fp16)[name = string("transpose_193")]; + tensor input_995_cast_fp16 = silu(x = input_993_cast_fp16)[name = string("input_995_cast_fp16")]; + string x_493_pad_type_0 = const()[name = string("x_493_pad_type_0"), val = string("valid")]; + tensor x_493_strides_0 = const()[name = string("x_493_strides_0"), val = tensor([1])]; + tensor x_493_pad_0 = const()[name = string("x_493_pad_0"), val = tensor([0, 0])]; + tensor x_493_dilations_0 = const()[name = string("x_493_dilations_0"), val = tensor([1])]; + int32 x_493_groups_0 = const()[name = string("x_493_groups_0"), val = int32(1)]; + tensor encoder_layers_18_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(373769664))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(374818304))))[name = string("encoder_layers_18_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_493_cast_fp16 = conv(dilations = x_493_dilations_0, groups = x_493_groups_0, pad = x_493_pad_0, pad_type = x_493_pad_type_0, strides = x_493_strides_0, weight = encoder_layers_18_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_995_cast_fp16)[name = string("x_493_cast_fp16")]; + tensor input_997_perm_0 = const()[name = string("input_997_perm_0"), val = tensor([0, 2, 1])]; + tensor input_997_cast_fp16 = transpose(perm = input_997_perm_0, x = x_493_cast_fp16)[name = string("transpose_192")]; + tensor input_999_cast_fp16 = add(x = input_983_cast_fp16, y = input_997_cast_fp16)[name = string("input_999_cast_fp16")]; + tensor input_1001_axes_0 = const()[name = string("input_1001_axes_0"), val = tensor([-1])]; + tensor encoder_layers_18_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_18_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(374820416)))]; + tensor encoder_layers_18_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_18_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(374822528)))]; + tensor input_1001_cast_fp16 = layer_norm(axes = input_1001_axes_0, beta = encoder_layers_18_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_18_norm_feed_forward2_weight_to_fp16, x = input_999_cast_fp16)[name = string("input_1001_cast_fp16")]; + tensor encoder_layers_18_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(374824640))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(379019008))))[name = string("encoder_layers_18_feed_forward2_linear1_weight_to_fp16_quantized")]; + tensor encoder_layers_18_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_18_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(379027264)))]; + tensor linear_170_cast_fp16 = linear(bias = encoder_layers_18_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_18_feed_forward2_linear1_weight_to_fp16_quantized, x = input_1001_cast_fp16)[name = string("linear_170_cast_fp16")]; + tensor input_1005_cast_fp16 = silu(x = linear_170_cast_fp16)[name = string("input_1005_cast_fp16")]; + tensor encoder_layers_18_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(379035520))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(383229888))))[name = string("encoder_layers_18_feed_forward2_linear2_weight_to_fp16_quantized")]; + tensor encoder_layers_18_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_18_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(383232000)))]; + tensor linear_171_cast_fp16 = linear(bias = encoder_layers_18_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_18_feed_forward2_linear2_weight_to_fp16_quantized, x = input_1005_cast_fp16)[name = string("linear_171_cast_fp16")]; + fp16 var_4465_to_fp16 = const()[name = string("op_4465_to_fp16"), val = fp16(0x1p-1)]; + tensor var_4466_cast_fp16 = mul(x = linear_171_cast_fp16, y = var_4465_to_fp16)[name = string("op_4466_cast_fp16")]; + tensor input_1011_cast_fp16 = add(x = input_999_cast_fp16, y = var_4466_cast_fp16)[name = string("input_1011_cast_fp16")]; + tensor input_1013_axes_0 = const()[name = string("input_1013_axes_0"), val = tensor([-1])]; + tensor encoder_layers_18_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_18_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(383234112)))]; + tensor encoder_layers_18_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_18_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(383236224)))]; + tensor input_1013_cast_fp16 = layer_norm(axes = input_1013_axes_0, beta = encoder_layers_18_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_18_norm_out_weight_to_fp16, x = input_1011_cast_fp16)[name = string("input_1013_cast_fp16")]; + tensor cache_77_begin_0 = const()[name = string("cache_77_begin_0"), val = tensor([19, 0, 0, 0])]; + tensor cache_77_end_0 = const()[name = string("cache_77_end_0"), val = tensor([20, 1, 42, 1024])]; + tensor cache_77_end_mask_0 = const()[name = string("cache_77_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_77_squeeze_mask_0 = const()[name = string("cache_77_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_77_cast_fp16 = slice_by_index(begin = cache_77_begin_0, end = cache_77_end_0, end_mask = cache_77_end_mask_0, squeeze_mask = cache_77_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_77_cast_fp16")]; + tensor cache_79_begin_0 = const()[name = string("cache_79_begin_0"), val = tensor([19, 0, 0, 0])]; + tensor cache_79_end_0 = const()[name = string("cache_79_end_0"), val = tensor([20, 1, 1024, 8])]; + tensor cache_79_end_mask_0 = const()[name = string("cache_79_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_79_squeeze_mask_0 = const()[name = string("cache_79_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_79_cast_fp16 = slice_by_index(begin = cache_79_begin_0, end = cache_79_end_0, end_mask = cache_79_end_mask_0, squeeze_mask = cache_79_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_79_cast_fp16")]; + tensor input_1015_axes_0 = const()[name = string("input_1015_axes_0"), val = tensor([-1])]; + tensor encoder_layers_19_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_19_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(383238336)))]; + tensor encoder_layers_19_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_19_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(383240448)))]; + tensor input_1015_cast_fp16 = layer_norm(axes = input_1015_axes_0, beta = encoder_layers_19_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_19_norm_feed_forward1_weight_to_fp16, x = input_1013_cast_fp16)[name = string("input_1015_cast_fp16")]; + tensor encoder_layers_19_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(383242560))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(387436928))))[name = string("encoder_layers_19_feed_forward1_linear1_weight_to_fp16_quantized")]; + tensor encoder_layers_19_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_19_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(387445184)))]; + tensor linear_172_cast_fp16 = linear(bias = encoder_layers_19_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_19_feed_forward1_linear1_weight_to_fp16_quantized, x = input_1015_cast_fp16)[name = string("linear_172_cast_fp16")]; + tensor input_1019_cast_fp16 = silu(x = linear_172_cast_fp16)[name = string("input_1019_cast_fp16")]; + tensor encoder_layers_19_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(387453440))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391647808))))[name = string("encoder_layers_19_feed_forward1_linear2_weight_to_fp16_quantized")]; + tensor encoder_layers_19_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_19_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391649920)))]; + tensor linear_173_cast_fp16 = linear(bias = encoder_layers_19_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_19_feed_forward1_linear2_weight_to_fp16_quantized, x = input_1019_cast_fp16)[name = string("linear_173_cast_fp16")]; + fp16 var_4502_to_fp16 = const()[name = string("op_4502_to_fp16"), val = fp16(0x1p-1)]; + tensor var_4503_cast_fp16 = mul(x = linear_173_cast_fp16, y = var_4502_to_fp16)[name = string("op_4503_cast_fp16")]; + tensor input_1025_cast_fp16 = add(x = input_1013_cast_fp16, y = var_4503_cast_fp16)[name = string("input_1025_cast_fp16")]; + tensor key_39_axes_0 = const()[name = string("key_39_axes_0"), val = tensor([-1])]; + tensor encoder_layers_19_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_19_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391652032)))]; + tensor encoder_layers_19_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_19_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391654144)))]; + tensor key_39_cast_fp16 = layer_norm(axes = key_39_axes_0, beta = encoder_layers_19_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_19_norm_self_att_weight_to_fp16, x = input_1025_cast_fp16)[name = string("key_39_cast_fp16")]; + bool input_1027_interleave_0 = const()[name = string("input_1027_interleave_0"), val = bool(false)]; + tensor input_1027_cast_fp16 = concat(axis = var_68, interleave = input_1027_interleave_0, values = (cache_77_cast_fp16, key_39_cast_fp16))[name = string("input_1027_cast_fp16")]; + tensor var_4525_begin_0 = const()[name = string("op_4525_begin_0"), val = tensor([0, 28, 0])]; + tensor var_4525_end_0 = const()[name = string("op_4525_end_0"), val = tensor([1, 42, 1024])]; + tensor var_4525_end_mask_0 = const()[name = string("op_4525_end_mask_0"), val = tensor([true, true, true])]; + tensor var_4525_cast_fp16 = slice_by_index(begin = var_4525_begin_0, end = var_4525_end_0, end_mask = var_4525_end_mask_0, x = cache_77_cast_fp16)[name = string("op_4525_cast_fp16")]; + bool var_4531_interleave_0 = const()[name = string("op_4531_interleave_0"), val = bool(false)]; + tensor var_4531_cast_fp16 = concat(axis = var_68, interleave = var_4531_interleave_0, values = (var_4525_cast_fp16, key_39_cast_fp16))[name = string("op_4531_cast_fp16")]; + tensor encoder_layers_19_self_attn_linear_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391656256))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(392704896))))[name = string("encoder_layers_19_self_attn_linear_q_weight_to_fp16_quantized")]; + tensor encoder_layers_19_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_19_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(392707008)))]; + tensor linear_174_cast_fp16 = linear(bias = encoder_layers_19_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_19_self_attn_linear_q_weight_to_fp16_quantized, x = key_39_cast_fp16)[name = string("linear_174_cast_fp16")]; + tensor var_4536 = const()[name = string("op_4536"), val = tensor([1, -1, 8, 128])]; + tensor q_115_cast_fp16 = reshape(shape = var_4536, x = linear_174_cast_fp16)[name = string("q_115_cast_fp16")]; + tensor encoder_layers_19_self_attn_linear_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(392709120))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(393757760))))[name = string("encoder_layers_19_self_attn_linear_k_weight_to_fp16_quantized")]; + tensor encoder_layers_19_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_19_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(393759872)))]; + tensor linear_175_cast_fp16 = linear(bias = encoder_layers_19_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_19_self_attn_linear_k_weight_to_fp16_quantized, x = input_1027_cast_fp16)[name = string("linear_175_cast_fp16")]; + tensor var_4541 = const()[name = string("op_4541"), val = tensor([1, -1, 8, 128])]; + tensor k_77_cast_fp16 = reshape(shape = var_4541, x = linear_175_cast_fp16)[name = string("k_77_cast_fp16")]; + tensor encoder_layers_19_self_attn_linear_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(393761984))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(394810624))))[name = string("encoder_layers_19_self_attn_linear_v_weight_to_fp16_quantized")]; + tensor encoder_layers_19_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_19_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(394812736)))]; + tensor linear_176_cast_fp16 = linear(bias = encoder_layers_19_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_19_self_attn_linear_v_weight_to_fp16_quantized, x = input_1027_cast_fp16)[name = string("linear_176_cast_fp16")]; + tensor var_4546 = const()[name = string("op_4546"), val = tensor([1, -1, 8, 128])]; + tensor v_39_cast_fp16 = reshape(shape = var_4546, x = linear_176_cast_fp16)[name = string("v_39_cast_fp16")]; + tensor value_47_perm_0 = const()[name = string("value_47_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_19_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_19_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(394814848)))]; + tensor var_4559_cast_fp16 = add(x = q_115_cast_fp16, y = encoder_layers_19_self_attn_pos_bias_u_to_fp16)[name = string("op_4559_cast_fp16")]; + tensor encoder_layers_19_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_19_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(394816960)))]; + tensor var_4561_cast_fp16 = add(x = q_115_cast_fp16, y = encoder_layers_19_self_attn_pos_bias_v_to_fp16)[name = string("op_4561_cast_fp16")]; + tensor q_with_bias_v_39_perm_0 = const()[name = string("q_with_bias_v_39_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_501_transpose_x_0 = const()[name = string("x_501_transpose_x_0"), val = bool(false)]; + bool x_501_transpose_y_0 = const()[name = string("x_501_transpose_y_0"), val = bool(false)]; + tensor op_4563_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(394819072))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(394961472))))[name = string("op_4563_to_fp16_quantized")]; + tensor q_with_bias_v_39_cast_fp16 = transpose(perm = q_with_bias_v_39_perm_0, x = var_4561_cast_fp16)[name = string("transpose_191")]; + tensor x_501_cast_fp16 = matmul(transpose_x = x_501_transpose_x_0, transpose_y = x_501_transpose_y_0, x = q_with_bias_v_39_cast_fp16, y = op_4563_to_fp16_quantized)[name = string("x_501_cast_fp16")]; + tensor x_503_pad_0 = const()[name = string("x_503_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_503_mode_0 = const()[name = string("x_503_mode_0"), val = string("constant")]; + fp16 const_326_to_fp16 = const()[name = string("const_326_to_fp16"), val = fp16(0x0p+0)]; + tensor x_503_cast_fp16 = pad(constant_val = const_326_to_fp16, mode = x_503_mode_0, pad = x_503_pad_0, x = x_501_cast_fp16)[name = string("x_503_cast_fp16")]; + tensor var_4571 = const()[name = string("op_4571"), val = tensor([1, 8, -1, 28])]; + tensor x_505_cast_fp16 = reshape(shape = var_4571, x = x_503_cast_fp16)[name = string("x_505_cast_fp16")]; + tensor var_4575_begin_0 = const()[name = string("op_4575_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_4575_end_0 = const()[name = string("op_4575_end_0"), val = tensor([1, 8, 140, 28])]; + tensor var_4575_end_mask_0 = const()[name = string("op_4575_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_4575_cast_fp16 = slice_by_index(begin = var_4575_begin_0, end = var_4575_end_0, end_mask = var_4575_end_mask_0, x = x_505_cast_fp16)[name = string("op_4575_cast_fp16")]; + tensor var_4576 = const()[name = string("op_4576"), val = tensor([1, 8, 28, 139])]; + tensor matrix_bd_77_cast_fp16 = reshape(shape = var_4576, x = var_4575_cast_fp16)[name = string("matrix_bd_77_cast_fp16")]; + bool matrix_ac_39_transpose_x_0 = const()[name = string("matrix_ac_39_transpose_x_0"), val = bool(false)]; + bool matrix_ac_39_transpose_y_0 = const()[name = string("matrix_ac_39_transpose_y_0"), val = bool(false)]; + tensor transpose_134_perm_0 = const()[name = string("transpose_134_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_135_perm_0 = const()[name = string("transpose_135_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_135 = transpose(perm = transpose_135_perm_0, x = k_77_cast_fp16)[name = string("transpose_189")]; + tensor transpose_134 = transpose(perm = transpose_134_perm_0, x = var_4559_cast_fp16)[name = string("transpose_190")]; + tensor matrix_ac_39_cast_fp16 = matmul(transpose_x = matrix_ac_39_transpose_x_0, transpose_y = matrix_ac_39_transpose_y_0, x = transpose_134, y = transpose_135)[name = string("matrix_ac_39_cast_fp16")]; + tensor matrix_bd_79_begin_0 = const()[name = string("matrix_bd_79_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_79_end_0 = const()[name = string("matrix_bd_79_end_0"), val = tensor([1, 8, 28, 70])]; + tensor matrix_bd_79_end_mask_0 = const()[name = string("matrix_bd_79_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_79_cast_fp16 = slice_by_index(begin = matrix_bd_79_begin_0, end = matrix_bd_79_end_0, end_mask = matrix_bd_79_end_mask_0, x = matrix_bd_77_cast_fp16)[name = string("matrix_bd_79_cast_fp16")]; + tensor var_4585_cast_fp16 = add(x = matrix_ac_39_cast_fp16, y = matrix_bd_79_cast_fp16)[name = string("op_4585_cast_fp16")]; + fp16 _inversed_scores_77_y_0_to_fp16 = const()[name = string("_inversed_scores_77_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_77_cast_fp16 = mul(x = var_4585_cast_fp16, y = _inversed_scores_77_y_0_to_fp16)[name = string("_inversed_scores_77_cast_fp16")]; + tensor scores_79_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_77_cast_fp16, cond = mask_11)[name = string("scores_79_cast_fp16")]; + tensor var_4591_cast_fp16 = softmax(axis = var_59, x = scores_79_cast_fp16)[name = string("op_4591_cast_fp16")]; + tensor input_1029_cast_fp16 = select(a = var_44_to_fp16, b = var_4591_cast_fp16, cond = mask_11)[name = string("input_1029_cast_fp16")]; + bool x_507_transpose_x_0 = const()[name = string("x_507_transpose_x_0"), val = bool(false)]; + bool x_507_transpose_y_0 = const()[name = string("x_507_transpose_y_0"), val = bool(false)]; + tensor value_47_cast_fp16 = transpose(perm = value_47_perm_0, x = v_39_cast_fp16)[name = string("transpose_188")]; + tensor x_507_cast_fp16 = matmul(transpose_x = x_507_transpose_x_0, transpose_y = x_507_transpose_y_0, x = input_1029_cast_fp16, y = value_47_cast_fp16)[name = string("x_507_cast_fp16")]; + tensor var_4595_perm_0 = const()[name = string("op_4595_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_4596 = const()[name = string("op_4596"), val = tensor([1, -1, 1024])]; + tensor var_4595_cast_fp16 = transpose(perm = var_4595_perm_0, x = x_507_cast_fp16)[name = string("transpose_187")]; + tensor input_1031_cast_fp16 = reshape(shape = var_4596, x = var_4595_cast_fp16)[name = string("input_1031_cast_fp16")]; + tensor encoder_layers_19_self_attn_linear_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(394961856))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(396010496))))[name = string("encoder_layers_19_self_attn_linear_out_weight_to_fp16_quantized")]; + tensor encoder_layers_19_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_19_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(396012608)))]; + tensor linear_178_cast_fp16 = linear(bias = encoder_layers_19_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_19_self_attn_linear_out_weight_to_fp16_quantized, x = input_1031_cast_fp16)[name = string("linear_178_cast_fp16")]; + tensor input_1035_cast_fp16 = add(x = input_1025_cast_fp16, y = linear_178_cast_fp16)[name = string("input_1035_cast_fp16")]; + tensor x_511_axes_0 = const()[name = string("x_511_axes_0"), val = tensor([-1])]; + tensor encoder_layers_19_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_19_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(396014720)))]; + tensor encoder_layers_19_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_19_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(396016832)))]; + tensor x_511_cast_fp16 = layer_norm(axes = x_511_axes_0, beta = encoder_layers_19_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_19_norm_conv_weight_to_fp16, x = input_1035_cast_fp16)[name = string("x_511_cast_fp16")]; + tensor input_1037_perm_0 = const()[name = string("input_1037_perm_0"), val = tensor([0, 2, 1])]; + string input_1039_pad_type_0 = const()[name = string("input_1039_pad_type_0"), val = string("valid")]; + tensor input_1039_strides_0 = const()[name = string("input_1039_strides_0"), val = tensor([1])]; + tensor input_1039_pad_0 = const()[name = string("input_1039_pad_0"), val = tensor([0, 0])]; + tensor input_1039_dilations_0 = const()[name = string("input_1039_dilations_0"), val = tensor([1])]; + int32 input_1039_groups_0 = const()[name = string("input_1039_groups_0"), val = int32(1)]; + tensor encoder_layers_19_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(396018944))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(398116160))))[name = string("encoder_layers_19_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_1037_cast_fp16 = transpose(perm = input_1037_perm_0, x = x_511_cast_fp16)[name = string("transpose_186")]; + tensor input_1039_cast_fp16 = conv(dilations = input_1039_dilations_0, groups = input_1039_groups_0, pad = input_1039_pad_0, pad_type = input_1039_pad_type_0, strides = input_1039_strides_0, weight = encoder_layers_19_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_1037_cast_fp16)[name = string("input_1039_cast_fp16")]; + int32 x_513_split_num_splits_0 = const()[name = string("x_513_split_num_splits_0"), val = int32(2)]; + int32 x_513_split_axis_0 = const()[name = string("x_513_split_axis_0"), val = int32(1)]; + tensor x_513_split_cast_fp16_0, tensor x_513_split_cast_fp16_1 = split(axis = x_513_split_axis_0, num_splits = x_513_split_num_splits_0, x = input_1039_cast_fp16)[name = string("x_513_split_cast_fp16")]; + tensor x_513_split_1_sigmoid_cast_fp16 = sigmoid(x = x_513_split_cast_fp16_1)[name = string("x_513_split_1_sigmoid_cast_fp16")]; + tensor x_513_cast_fp16 = mul(x = x_513_split_cast_fp16_0, y = x_513_split_1_sigmoid_cast_fp16)[name = string("x_513_cast_fp16")]; + tensor input_1041_cast_fp16 = select(a = var_44_to_fp16, b = x_513_cast_fp16, cond = var_575)[name = string("input_1041_cast_fp16")]; + bool new_x_79_interleave_0 = const()[name = string("new_x_79_interleave_0"), val = bool(false)]; + tensor new_x_79_cast_fp16 = concat(axis = var_59, interleave = new_x_79_interleave_0, values = (cache_79_cast_fp16, input_1041_cast_fp16))[name = string("new_x_79_cast_fp16")]; + tensor var_4635_begin_0 = const()[name = string("op_4635_begin_0"), val = tensor([0, 0, 28])]; + tensor var_4635_end_0 = const()[name = string("op_4635_end_0"), val = tensor([1, 1024, 36])]; + tensor var_4635_end_mask_0 = const()[name = string("op_4635_end_mask_0"), val = tensor([true, true, true])]; + tensor var_4635_cast_fp16 = slice_by_index(begin = var_4635_begin_0, end = var_4635_end_0, end_mask = var_4635_end_mask_0, x = new_x_79_cast_fp16)[name = string("op_4635_cast_fp16")]; + string x_515_pad_type_0 = const()[name = string("x_515_pad_type_0"), val = string("valid")]; + int32 x_515_groups_0 = const()[name = string("x_515_groups_0"), val = int32(1024)]; + tensor x_515_strides_0 = const()[name = string("x_515_strides_0"), val = tensor([1])]; + tensor x_515_pad_0 = const()[name = string("x_515_pad_0"), val = tensor([0, 0])]; + tensor x_515_dilations_0 = const()[name = string("x_515_dilations_0"), val = tensor([1])]; + tensor encoder_layers_19_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(398120320))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(398129600))))[name = string("encoder_layers_19_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_515_cast_fp16 = conv(dilations = x_515_dilations_0, groups = x_515_groups_0, pad = x_515_pad_0, pad_type = x_515_pad_type_0, strides = x_515_strides_0, weight = encoder_layers_19_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_79_cast_fp16)[name = string("x_515_cast_fp16")]; + tensor input_1043_perm_0 = const()[name = string("input_1043_perm_0"), val = tensor([0, 2, 1])]; + tensor x_517_axes_0 = const()[name = string("x_517_axes_0"), val = tensor([-1])]; + tensor encoder_layers_19_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_19_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(398131712)))]; + tensor encoder_layers_19_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_19_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(398133824)))]; + tensor input_1043_cast_fp16 = transpose(perm = input_1043_perm_0, x = x_515_cast_fp16)[name = string("transpose_185")]; + tensor x_517_cast_fp16 = layer_norm(axes = x_517_axes_0, beta = encoder_layers_19_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_19_conv_batch_norm_weight_to_fp16, x = input_1043_cast_fp16)[name = string("x_517_cast_fp16")]; + tensor input_1045_perm_0 = const()[name = string("input_1045_perm_0"), val = tensor([0, 2, 1])]; + tensor input_1045_cast_fp16 = transpose(perm = input_1045_perm_0, x = x_517_cast_fp16)[name = string("transpose_184")]; + tensor input_1047_cast_fp16 = silu(x = input_1045_cast_fp16)[name = string("input_1047_cast_fp16")]; + string x_519_pad_type_0 = const()[name = string("x_519_pad_type_0"), val = string("valid")]; + tensor x_519_strides_0 = const()[name = string("x_519_strides_0"), val = tensor([1])]; + tensor x_519_pad_0 = const()[name = string("x_519_pad_0"), val = tensor([0, 0])]; + tensor x_519_dilations_0 = const()[name = string("x_519_dilations_0"), val = tensor([1])]; + int32 x_519_groups_0 = const()[name = string("x_519_groups_0"), val = int32(1)]; + tensor encoder_layers_19_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(398135936))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(399184576))))[name = string("encoder_layers_19_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_519_cast_fp16 = conv(dilations = x_519_dilations_0, groups = x_519_groups_0, pad = x_519_pad_0, pad_type = x_519_pad_type_0, strides = x_519_strides_0, weight = encoder_layers_19_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_1047_cast_fp16)[name = string("x_519_cast_fp16")]; + tensor input_1049_perm_0 = const()[name = string("input_1049_perm_0"), val = tensor([0, 2, 1])]; + tensor input_1049_cast_fp16 = transpose(perm = input_1049_perm_0, x = x_519_cast_fp16)[name = string("transpose_183")]; + tensor input_1051_cast_fp16 = add(x = input_1035_cast_fp16, y = input_1049_cast_fp16)[name = string("input_1051_cast_fp16")]; + tensor input_1053_axes_0 = const()[name = string("input_1053_axes_0"), val = tensor([-1])]; + tensor encoder_layers_19_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_19_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(399186688)))]; + tensor encoder_layers_19_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_19_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(399188800)))]; + tensor input_1053_cast_fp16 = layer_norm(axes = input_1053_axes_0, beta = encoder_layers_19_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_19_norm_feed_forward2_weight_to_fp16, x = input_1051_cast_fp16)[name = string("input_1053_cast_fp16")]; + tensor encoder_layers_19_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(399190912))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(403385280))))[name = string("encoder_layers_19_feed_forward2_linear1_weight_to_fp16_quantized")]; + tensor encoder_layers_19_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_19_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(403393536)))]; + tensor linear_179_cast_fp16 = linear(bias = encoder_layers_19_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_19_feed_forward2_linear1_weight_to_fp16_quantized, x = input_1053_cast_fp16)[name = string("linear_179_cast_fp16")]; + tensor input_1057_cast_fp16 = silu(x = linear_179_cast_fp16)[name = string("input_1057_cast_fp16")]; + tensor encoder_layers_19_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(403401792))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407596160))))[name = string("encoder_layers_19_feed_forward2_linear2_weight_to_fp16_quantized")]; + tensor encoder_layers_19_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_19_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407598272)))]; + tensor linear_180_cast_fp16 = linear(bias = encoder_layers_19_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_19_feed_forward2_linear2_weight_to_fp16_quantized, x = input_1057_cast_fp16)[name = string("linear_180_cast_fp16")]; + fp16 var_4678_to_fp16 = const()[name = string("op_4678_to_fp16"), val = fp16(0x1p-1)]; + tensor var_4679_cast_fp16 = mul(x = linear_180_cast_fp16, y = var_4678_to_fp16)[name = string("op_4679_cast_fp16")]; + tensor input_1063_cast_fp16 = add(x = input_1051_cast_fp16, y = var_4679_cast_fp16)[name = string("input_1063_cast_fp16")]; + tensor input_1065_axes_0 = const()[name = string("input_1065_axes_0"), val = tensor([-1])]; + tensor encoder_layers_19_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_19_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407600384)))]; + tensor encoder_layers_19_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_19_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407602496)))]; + tensor input_1065_cast_fp16 = layer_norm(axes = input_1065_axes_0, beta = encoder_layers_19_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_19_norm_out_weight_to_fp16, x = input_1063_cast_fp16)[name = string("input_1065_cast_fp16")]; + tensor cache_81_begin_0 = const()[name = string("cache_81_begin_0"), val = tensor([20, 0, 0, 0])]; + tensor cache_81_end_0 = const()[name = string("cache_81_end_0"), val = tensor([21, 1, 42, 1024])]; + tensor cache_81_end_mask_0 = const()[name = string("cache_81_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_81_squeeze_mask_0 = const()[name = string("cache_81_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_81_cast_fp16 = slice_by_index(begin = cache_81_begin_0, end = cache_81_end_0, end_mask = cache_81_end_mask_0, squeeze_mask = cache_81_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_81_cast_fp16")]; + tensor cache_83_begin_0 = const()[name = string("cache_83_begin_0"), val = tensor([20, 0, 0, 0])]; + tensor cache_83_end_0 = const()[name = string("cache_83_end_0"), val = tensor([21, 1, 1024, 8])]; + tensor cache_83_end_mask_0 = const()[name = string("cache_83_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_83_squeeze_mask_0 = const()[name = string("cache_83_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_83_cast_fp16 = slice_by_index(begin = cache_83_begin_0, end = cache_83_end_0, end_mask = cache_83_end_mask_0, squeeze_mask = cache_83_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_83_cast_fp16")]; + tensor input_1067_axes_0 = const()[name = string("input_1067_axes_0"), val = tensor([-1])]; + tensor encoder_layers_20_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_20_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407604608)))]; + tensor encoder_layers_20_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_20_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407606720)))]; + tensor input_1067_cast_fp16 = layer_norm(axes = input_1067_axes_0, beta = encoder_layers_20_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_20_norm_feed_forward1_weight_to_fp16, x = input_1065_cast_fp16)[name = string("input_1067_cast_fp16")]; + tensor encoder_layers_20_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407608832))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(411803200))))[name = string("encoder_layers_20_feed_forward1_linear1_weight_to_fp16_quantized")]; + tensor encoder_layers_20_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_20_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(411811456)))]; + tensor linear_181_cast_fp16 = linear(bias = encoder_layers_20_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_20_feed_forward1_linear1_weight_to_fp16_quantized, x = input_1067_cast_fp16)[name = string("linear_181_cast_fp16")]; + tensor input_1071_cast_fp16 = silu(x = linear_181_cast_fp16)[name = string("input_1071_cast_fp16")]; + tensor encoder_layers_20_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(411819712))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(416014080))))[name = string("encoder_layers_20_feed_forward1_linear2_weight_to_fp16_quantized")]; + tensor encoder_layers_20_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_20_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(416016192)))]; + tensor linear_182_cast_fp16 = linear(bias = encoder_layers_20_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_20_feed_forward1_linear2_weight_to_fp16_quantized, x = input_1071_cast_fp16)[name = string("linear_182_cast_fp16")]; + fp16 var_4715_to_fp16 = const()[name = string("op_4715_to_fp16"), val = fp16(0x1p-1)]; + tensor var_4716_cast_fp16 = mul(x = linear_182_cast_fp16, y = var_4715_to_fp16)[name = string("op_4716_cast_fp16")]; + tensor input_1077_cast_fp16 = add(x = input_1065_cast_fp16, y = var_4716_cast_fp16)[name = string("input_1077_cast_fp16")]; + tensor key_41_axes_0 = const()[name = string("key_41_axes_0"), val = tensor([-1])]; + tensor encoder_layers_20_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_20_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(416018304)))]; + tensor encoder_layers_20_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_20_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(416020416)))]; + tensor key_41_cast_fp16 = layer_norm(axes = key_41_axes_0, beta = encoder_layers_20_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_20_norm_self_att_weight_to_fp16, x = input_1077_cast_fp16)[name = string("key_41_cast_fp16")]; + bool input_1079_interleave_0 = const()[name = string("input_1079_interleave_0"), val = bool(false)]; + tensor input_1079_cast_fp16 = concat(axis = var_68, interleave = input_1079_interleave_0, values = (cache_81_cast_fp16, key_41_cast_fp16))[name = string("input_1079_cast_fp16")]; + tensor var_4738_begin_0 = const()[name = string("op_4738_begin_0"), val = tensor([0, 28, 0])]; + tensor var_4738_end_0 = const()[name = string("op_4738_end_0"), val = tensor([1, 42, 1024])]; + tensor var_4738_end_mask_0 = const()[name = string("op_4738_end_mask_0"), val = tensor([true, true, true])]; + tensor var_4738_cast_fp16 = slice_by_index(begin = var_4738_begin_0, end = var_4738_end_0, end_mask = var_4738_end_mask_0, x = cache_81_cast_fp16)[name = string("op_4738_cast_fp16")]; + bool var_4744_interleave_0 = const()[name = string("op_4744_interleave_0"), val = bool(false)]; + tensor var_4744_cast_fp16 = concat(axis = var_68, interleave = var_4744_interleave_0, values = (var_4738_cast_fp16, key_41_cast_fp16))[name = string("op_4744_cast_fp16")]; + tensor encoder_layers_20_self_attn_linear_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(416022528))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(417071168))))[name = string("encoder_layers_20_self_attn_linear_q_weight_to_fp16_quantized")]; + tensor encoder_layers_20_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_20_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(417073280)))]; + tensor linear_183_cast_fp16 = linear(bias = encoder_layers_20_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_20_self_attn_linear_q_weight_to_fp16_quantized, x = key_41_cast_fp16)[name = string("linear_183_cast_fp16")]; + tensor var_4749 = const()[name = string("op_4749"), val = tensor([1, -1, 8, 128])]; + tensor q_121_cast_fp16 = reshape(shape = var_4749, x = linear_183_cast_fp16)[name = string("q_121_cast_fp16")]; + tensor encoder_layers_20_self_attn_linear_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(417075392))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(418124032))))[name = string("encoder_layers_20_self_attn_linear_k_weight_to_fp16_quantized")]; + tensor encoder_layers_20_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_20_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(418126144)))]; + tensor linear_184_cast_fp16 = linear(bias = encoder_layers_20_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_20_self_attn_linear_k_weight_to_fp16_quantized, x = input_1079_cast_fp16)[name = string("linear_184_cast_fp16")]; + tensor var_4754 = const()[name = string("op_4754"), val = tensor([1, -1, 8, 128])]; + tensor k_81_cast_fp16 = reshape(shape = var_4754, x = linear_184_cast_fp16)[name = string("k_81_cast_fp16")]; + tensor encoder_layers_20_self_attn_linear_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(418128256))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(419176896))))[name = string("encoder_layers_20_self_attn_linear_v_weight_to_fp16_quantized")]; + tensor encoder_layers_20_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_20_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(419179008)))]; + tensor linear_185_cast_fp16 = linear(bias = encoder_layers_20_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_20_self_attn_linear_v_weight_to_fp16_quantized, x = input_1079_cast_fp16)[name = string("linear_185_cast_fp16")]; + tensor var_4759 = const()[name = string("op_4759"), val = tensor([1, -1, 8, 128])]; + tensor v_41_cast_fp16 = reshape(shape = var_4759, x = linear_185_cast_fp16)[name = string("v_41_cast_fp16")]; + tensor value_49_perm_0 = const()[name = string("value_49_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_20_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_20_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(419181120)))]; + tensor var_4772_cast_fp16 = add(x = q_121_cast_fp16, y = encoder_layers_20_self_attn_pos_bias_u_to_fp16)[name = string("op_4772_cast_fp16")]; + tensor encoder_layers_20_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_20_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(419183232)))]; + tensor var_4774_cast_fp16 = add(x = q_121_cast_fp16, y = encoder_layers_20_self_attn_pos_bias_v_to_fp16)[name = string("op_4774_cast_fp16")]; + tensor q_with_bias_v_41_perm_0 = const()[name = string("q_with_bias_v_41_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_527_transpose_x_0 = const()[name = string("x_527_transpose_x_0"), val = bool(false)]; + bool x_527_transpose_y_0 = const()[name = string("x_527_transpose_y_0"), val = bool(false)]; + tensor op_4776_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(419185344))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(419327744))))[name = string("op_4776_to_fp16_quantized")]; + tensor q_with_bias_v_41_cast_fp16 = transpose(perm = q_with_bias_v_41_perm_0, x = var_4774_cast_fp16)[name = string("transpose_182")]; + tensor x_527_cast_fp16 = matmul(transpose_x = x_527_transpose_x_0, transpose_y = x_527_transpose_y_0, x = q_with_bias_v_41_cast_fp16, y = op_4776_to_fp16_quantized)[name = string("x_527_cast_fp16")]; + tensor x_529_pad_0 = const()[name = string("x_529_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_529_mode_0 = const()[name = string("x_529_mode_0"), val = string("constant")]; + fp16 const_339_to_fp16 = const()[name = string("const_339_to_fp16"), val = fp16(0x0p+0)]; + tensor x_529_cast_fp16 = pad(constant_val = const_339_to_fp16, mode = x_529_mode_0, pad = x_529_pad_0, x = x_527_cast_fp16)[name = string("x_529_cast_fp16")]; + tensor var_4784 = const()[name = string("op_4784"), val = tensor([1, 8, -1, 28])]; + tensor x_531_cast_fp16 = reshape(shape = var_4784, x = x_529_cast_fp16)[name = string("x_531_cast_fp16")]; + tensor var_4788_begin_0 = const()[name = string("op_4788_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_4788_end_0 = const()[name = string("op_4788_end_0"), val = tensor([1, 8, 140, 28])]; + tensor var_4788_end_mask_0 = const()[name = string("op_4788_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_4788_cast_fp16 = slice_by_index(begin = var_4788_begin_0, end = var_4788_end_0, end_mask = var_4788_end_mask_0, x = x_531_cast_fp16)[name = string("op_4788_cast_fp16")]; + tensor var_4789 = const()[name = string("op_4789"), val = tensor([1, 8, 28, 139])]; + tensor matrix_bd_81_cast_fp16 = reshape(shape = var_4789, x = var_4788_cast_fp16)[name = string("matrix_bd_81_cast_fp16")]; + bool matrix_ac_41_transpose_x_0 = const()[name = string("matrix_ac_41_transpose_x_0"), val = bool(false)]; + bool matrix_ac_41_transpose_y_0 = const()[name = string("matrix_ac_41_transpose_y_0"), val = bool(false)]; + tensor transpose_136_perm_0 = const()[name = string("transpose_136_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_137_perm_0 = const()[name = string("transpose_137_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_137 = transpose(perm = transpose_137_perm_0, x = k_81_cast_fp16)[name = string("transpose_180")]; + tensor transpose_136 = transpose(perm = transpose_136_perm_0, x = var_4772_cast_fp16)[name = string("transpose_181")]; + tensor matrix_ac_41_cast_fp16 = matmul(transpose_x = matrix_ac_41_transpose_x_0, transpose_y = matrix_ac_41_transpose_y_0, x = transpose_136, y = transpose_137)[name = string("matrix_ac_41_cast_fp16")]; + tensor matrix_bd_83_begin_0 = const()[name = string("matrix_bd_83_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_83_end_0 = const()[name = string("matrix_bd_83_end_0"), val = tensor([1, 8, 28, 70])]; + tensor matrix_bd_83_end_mask_0 = const()[name = string("matrix_bd_83_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_83_cast_fp16 = slice_by_index(begin = matrix_bd_83_begin_0, end = matrix_bd_83_end_0, end_mask = matrix_bd_83_end_mask_0, x = matrix_bd_81_cast_fp16)[name = string("matrix_bd_83_cast_fp16")]; + tensor var_4798_cast_fp16 = add(x = matrix_ac_41_cast_fp16, y = matrix_bd_83_cast_fp16)[name = string("op_4798_cast_fp16")]; + fp16 _inversed_scores_81_y_0_to_fp16 = const()[name = string("_inversed_scores_81_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_81_cast_fp16 = mul(x = var_4798_cast_fp16, y = _inversed_scores_81_y_0_to_fp16)[name = string("_inversed_scores_81_cast_fp16")]; + tensor scores_83_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_81_cast_fp16, cond = mask_11)[name = string("scores_83_cast_fp16")]; + tensor var_4804_cast_fp16 = softmax(axis = var_59, x = scores_83_cast_fp16)[name = string("op_4804_cast_fp16")]; + tensor input_1081_cast_fp16 = select(a = var_44_to_fp16, b = var_4804_cast_fp16, cond = mask_11)[name = string("input_1081_cast_fp16")]; + bool x_533_transpose_x_0 = const()[name = string("x_533_transpose_x_0"), val = bool(false)]; + bool x_533_transpose_y_0 = const()[name = string("x_533_transpose_y_0"), val = bool(false)]; + tensor value_49_cast_fp16 = transpose(perm = value_49_perm_0, x = v_41_cast_fp16)[name = string("transpose_179")]; + tensor x_533_cast_fp16 = matmul(transpose_x = x_533_transpose_x_0, transpose_y = x_533_transpose_y_0, x = input_1081_cast_fp16, y = value_49_cast_fp16)[name = string("x_533_cast_fp16")]; + tensor var_4808_perm_0 = const()[name = string("op_4808_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_4809 = const()[name = string("op_4809"), val = tensor([1, -1, 1024])]; + tensor var_4808_cast_fp16 = transpose(perm = var_4808_perm_0, x = x_533_cast_fp16)[name = string("transpose_178")]; + tensor input_1083_cast_fp16 = reshape(shape = var_4809, x = var_4808_cast_fp16)[name = string("input_1083_cast_fp16")]; + tensor encoder_layers_20_self_attn_linear_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(419328128))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(420376768))))[name = string("encoder_layers_20_self_attn_linear_out_weight_to_fp16_quantized")]; + tensor encoder_layers_20_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_20_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(420378880)))]; + tensor linear_187_cast_fp16 = linear(bias = encoder_layers_20_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_20_self_attn_linear_out_weight_to_fp16_quantized, x = input_1083_cast_fp16)[name = string("linear_187_cast_fp16")]; + tensor input_1087_cast_fp16 = add(x = input_1077_cast_fp16, y = linear_187_cast_fp16)[name = string("input_1087_cast_fp16")]; + tensor x_537_axes_0 = const()[name = string("x_537_axes_0"), val = tensor([-1])]; + tensor encoder_layers_20_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_20_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(420380992)))]; + tensor encoder_layers_20_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_20_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(420383104)))]; + tensor x_537_cast_fp16 = layer_norm(axes = x_537_axes_0, beta = encoder_layers_20_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_20_norm_conv_weight_to_fp16, x = input_1087_cast_fp16)[name = string("x_537_cast_fp16")]; + tensor input_1089_perm_0 = const()[name = string("input_1089_perm_0"), val = tensor([0, 2, 1])]; + string input_1091_pad_type_0 = const()[name = string("input_1091_pad_type_0"), val = string("valid")]; + tensor input_1091_strides_0 = const()[name = string("input_1091_strides_0"), val = tensor([1])]; + tensor input_1091_pad_0 = const()[name = string("input_1091_pad_0"), val = tensor([0, 0])]; + tensor input_1091_dilations_0 = const()[name = string("input_1091_dilations_0"), val = tensor([1])]; + int32 input_1091_groups_0 = const()[name = string("input_1091_groups_0"), val = int32(1)]; + tensor encoder_layers_20_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(420385216))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(422482432))))[name = string("encoder_layers_20_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_1089_cast_fp16 = transpose(perm = input_1089_perm_0, x = x_537_cast_fp16)[name = string("transpose_177")]; + tensor input_1091_cast_fp16 = conv(dilations = input_1091_dilations_0, groups = input_1091_groups_0, pad = input_1091_pad_0, pad_type = input_1091_pad_type_0, strides = input_1091_strides_0, weight = encoder_layers_20_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_1089_cast_fp16)[name = string("input_1091_cast_fp16")]; + int32 x_539_split_num_splits_0 = const()[name = string("x_539_split_num_splits_0"), val = int32(2)]; + int32 x_539_split_axis_0 = const()[name = string("x_539_split_axis_0"), val = int32(1)]; + tensor x_539_split_cast_fp16_0, tensor x_539_split_cast_fp16_1 = split(axis = x_539_split_axis_0, num_splits = x_539_split_num_splits_0, x = input_1091_cast_fp16)[name = string("x_539_split_cast_fp16")]; + tensor x_539_split_1_sigmoid_cast_fp16 = sigmoid(x = x_539_split_cast_fp16_1)[name = string("x_539_split_1_sigmoid_cast_fp16")]; + tensor x_539_cast_fp16 = mul(x = x_539_split_cast_fp16_0, y = x_539_split_1_sigmoid_cast_fp16)[name = string("x_539_cast_fp16")]; + tensor input_1093_cast_fp16 = select(a = var_44_to_fp16, b = x_539_cast_fp16, cond = var_575)[name = string("input_1093_cast_fp16")]; + bool new_x_83_interleave_0 = const()[name = string("new_x_83_interleave_0"), val = bool(false)]; + tensor new_x_83_cast_fp16 = concat(axis = var_59, interleave = new_x_83_interleave_0, values = (cache_83_cast_fp16, input_1093_cast_fp16))[name = string("new_x_83_cast_fp16")]; + tensor var_4848_begin_0 = const()[name = string("op_4848_begin_0"), val = tensor([0, 0, 28])]; + tensor var_4848_end_0 = const()[name = string("op_4848_end_0"), val = tensor([1, 1024, 36])]; + tensor var_4848_end_mask_0 = const()[name = string("op_4848_end_mask_0"), val = tensor([true, true, true])]; + tensor var_4848_cast_fp16 = slice_by_index(begin = var_4848_begin_0, end = var_4848_end_0, end_mask = var_4848_end_mask_0, x = new_x_83_cast_fp16)[name = string("op_4848_cast_fp16")]; + string x_541_pad_type_0 = const()[name = string("x_541_pad_type_0"), val = string("valid")]; + int32 x_541_groups_0 = const()[name = string("x_541_groups_0"), val = int32(1024)]; + tensor x_541_strides_0 = const()[name = string("x_541_strides_0"), val = tensor([1])]; + tensor x_541_pad_0 = const()[name = string("x_541_pad_0"), val = tensor([0, 0])]; + tensor x_541_dilations_0 = const()[name = string("x_541_dilations_0"), val = tensor([1])]; + tensor encoder_layers_20_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(422486592))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(422495872))))[name = string("encoder_layers_20_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_541_cast_fp16 = conv(dilations = x_541_dilations_0, groups = x_541_groups_0, pad = x_541_pad_0, pad_type = x_541_pad_type_0, strides = x_541_strides_0, weight = encoder_layers_20_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_83_cast_fp16)[name = string("x_541_cast_fp16")]; + tensor input_1095_perm_0 = const()[name = string("input_1095_perm_0"), val = tensor([0, 2, 1])]; + tensor x_543_axes_0 = const()[name = string("x_543_axes_0"), val = tensor([-1])]; + tensor encoder_layers_20_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_20_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(422497984)))]; + tensor encoder_layers_20_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_20_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(422500096)))]; + tensor input_1095_cast_fp16 = transpose(perm = input_1095_perm_0, x = x_541_cast_fp16)[name = string("transpose_176")]; + tensor x_543_cast_fp16 = layer_norm(axes = x_543_axes_0, beta = encoder_layers_20_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_20_conv_batch_norm_weight_to_fp16, x = input_1095_cast_fp16)[name = string("x_543_cast_fp16")]; + tensor input_1097_perm_0 = const()[name = string("input_1097_perm_0"), val = tensor([0, 2, 1])]; + tensor input_1097_cast_fp16 = transpose(perm = input_1097_perm_0, x = x_543_cast_fp16)[name = string("transpose_175")]; + tensor input_1099_cast_fp16 = silu(x = input_1097_cast_fp16)[name = string("input_1099_cast_fp16")]; + string x_545_pad_type_0 = const()[name = string("x_545_pad_type_0"), val = string("valid")]; + tensor x_545_strides_0 = const()[name = string("x_545_strides_0"), val = tensor([1])]; + tensor x_545_pad_0 = const()[name = string("x_545_pad_0"), val = tensor([0, 0])]; + tensor x_545_dilations_0 = const()[name = string("x_545_dilations_0"), val = tensor([1])]; + int32 x_545_groups_0 = const()[name = string("x_545_groups_0"), val = int32(1)]; + tensor encoder_layers_20_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(422502208))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(423550848))))[name = string("encoder_layers_20_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_545_cast_fp16 = conv(dilations = x_545_dilations_0, groups = x_545_groups_0, pad = x_545_pad_0, pad_type = x_545_pad_type_0, strides = x_545_strides_0, weight = encoder_layers_20_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_1099_cast_fp16)[name = string("x_545_cast_fp16")]; + tensor input_1101_perm_0 = const()[name = string("input_1101_perm_0"), val = tensor([0, 2, 1])]; + tensor input_1101_cast_fp16 = transpose(perm = input_1101_perm_0, x = x_545_cast_fp16)[name = string("transpose_174")]; + tensor input_1103_cast_fp16 = add(x = input_1087_cast_fp16, y = input_1101_cast_fp16)[name = string("input_1103_cast_fp16")]; + tensor input_1105_axes_0 = const()[name = string("input_1105_axes_0"), val = tensor([-1])]; + tensor encoder_layers_20_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_20_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(423552960)))]; + tensor encoder_layers_20_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_20_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(423555072)))]; + tensor input_1105_cast_fp16 = layer_norm(axes = input_1105_axes_0, beta = encoder_layers_20_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_20_norm_feed_forward2_weight_to_fp16, x = input_1103_cast_fp16)[name = string("input_1105_cast_fp16")]; + tensor encoder_layers_20_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(423557184))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427751552))))[name = string("encoder_layers_20_feed_forward2_linear1_weight_to_fp16_quantized")]; + tensor encoder_layers_20_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_20_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427759808)))]; + tensor linear_188_cast_fp16 = linear(bias = encoder_layers_20_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_20_feed_forward2_linear1_weight_to_fp16_quantized, x = input_1105_cast_fp16)[name = string("linear_188_cast_fp16")]; + tensor input_1109_cast_fp16 = silu(x = linear_188_cast_fp16)[name = string("input_1109_cast_fp16")]; + tensor encoder_layers_20_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427768064))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(431962432))))[name = string("encoder_layers_20_feed_forward2_linear2_weight_to_fp16_quantized")]; + tensor encoder_layers_20_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_20_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(431964544)))]; + tensor linear_189_cast_fp16 = linear(bias = encoder_layers_20_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_20_feed_forward2_linear2_weight_to_fp16_quantized, x = input_1109_cast_fp16)[name = string("linear_189_cast_fp16")]; + fp16 var_4891_to_fp16 = const()[name = string("op_4891_to_fp16"), val = fp16(0x1p-1)]; + tensor var_4892_cast_fp16 = mul(x = linear_189_cast_fp16, y = var_4891_to_fp16)[name = string("op_4892_cast_fp16")]; + tensor input_1115_cast_fp16 = add(x = input_1103_cast_fp16, y = var_4892_cast_fp16)[name = string("input_1115_cast_fp16")]; + tensor input_1117_axes_0 = const()[name = string("input_1117_axes_0"), val = tensor([-1])]; + tensor encoder_layers_20_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_20_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(431966656)))]; + tensor encoder_layers_20_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_20_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(431968768)))]; + tensor input_1117_cast_fp16 = layer_norm(axes = input_1117_axes_0, beta = encoder_layers_20_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_20_norm_out_weight_to_fp16, x = input_1115_cast_fp16)[name = string("input_1117_cast_fp16")]; + tensor cache_85_begin_0 = const()[name = string("cache_85_begin_0"), val = tensor([21, 0, 0, 0])]; + tensor cache_85_end_0 = const()[name = string("cache_85_end_0"), val = tensor([22, 1, 42, 1024])]; + tensor cache_85_end_mask_0 = const()[name = string("cache_85_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_85_squeeze_mask_0 = const()[name = string("cache_85_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_85_cast_fp16 = slice_by_index(begin = cache_85_begin_0, end = cache_85_end_0, end_mask = cache_85_end_mask_0, squeeze_mask = cache_85_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_85_cast_fp16")]; + tensor cache_87_begin_0 = const()[name = string("cache_87_begin_0"), val = tensor([21, 0, 0, 0])]; + tensor cache_87_end_0 = const()[name = string("cache_87_end_0"), val = tensor([22, 1, 1024, 8])]; + tensor cache_87_end_mask_0 = const()[name = string("cache_87_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_87_squeeze_mask_0 = const()[name = string("cache_87_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_87_cast_fp16 = slice_by_index(begin = cache_87_begin_0, end = cache_87_end_0, end_mask = cache_87_end_mask_0, squeeze_mask = cache_87_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_87_cast_fp16")]; + tensor input_1119_axes_0 = const()[name = string("input_1119_axes_0"), val = tensor([-1])]; + tensor encoder_layers_21_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_21_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(431970880)))]; + tensor encoder_layers_21_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_21_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(431972992)))]; + tensor input_1119_cast_fp16 = layer_norm(axes = input_1119_axes_0, beta = encoder_layers_21_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_21_norm_feed_forward1_weight_to_fp16, x = input_1117_cast_fp16)[name = string("input_1119_cast_fp16")]; + tensor encoder_layers_21_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(431975104))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436169472))))[name = string("encoder_layers_21_feed_forward1_linear1_weight_to_fp16_quantized")]; + tensor encoder_layers_21_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_21_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436177728)))]; + tensor linear_190_cast_fp16 = linear(bias = encoder_layers_21_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_21_feed_forward1_linear1_weight_to_fp16_quantized, x = input_1119_cast_fp16)[name = string("linear_190_cast_fp16")]; + tensor input_1123_cast_fp16 = silu(x = linear_190_cast_fp16)[name = string("input_1123_cast_fp16")]; + tensor encoder_layers_21_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436185984))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(440380352))))[name = string("encoder_layers_21_feed_forward1_linear2_weight_to_fp16_quantized")]; + tensor encoder_layers_21_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_21_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(440382464)))]; + tensor linear_191_cast_fp16 = linear(bias = encoder_layers_21_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_21_feed_forward1_linear2_weight_to_fp16_quantized, x = input_1123_cast_fp16)[name = string("linear_191_cast_fp16")]; + fp16 var_4928_to_fp16 = const()[name = string("op_4928_to_fp16"), val = fp16(0x1p-1)]; + tensor var_4929_cast_fp16 = mul(x = linear_191_cast_fp16, y = var_4928_to_fp16)[name = string("op_4929_cast_fp16")]; + tensor input_1129_cast_fp16 = add(x = input_1117_cast_fp16, y = var_4929_cast_fp16)[name = string("input_1129_cast_fp16")]; + tensor key_43_axes_0 = const()[name = string("key_43_axes_0"), val = tensor([-1])]; + tensor encoder_layers_21_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_21_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(440384576)))]; + tensor encoder_layers_21_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_21_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(440386688)))]; + tensor key_43_cast_fp16 = layer_norm(axes = key_43_axes_0, beta = encoder_layers_21_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_21_norm_self_att_weight_to_fp16, x = input_1129_cast_fp16)[name = string("key_43_cast_fp16")]; + bool input_1131_interleave_0 = const()[name = string("input_1131_interleave_0"), val = bool(false)]; + tensor input_1131_cast_fp16 = concat(axis = var_68, interleave = input_1131_interleave_0, values = (cache_85_cast_fp16, key_43_cast_fp16))[name = string("input_1131_cast_fp16")]; + tensor var_4951_begin_0 = const()[name = string("op_4951_begin_0"), val = tensor([0, 28, 0])]; + tensor var_4951_end_0 = const()[name = string("op_4951_end_0"), val = tensor([1, 42, 1024])]; + tensor var_4951_end_mask_0 = const()[name = string("op_4951_end_mask_0"), val = tensor([true, true, true])]; + tensor var_4951_cast_fp16 = slice_by_index(begin = var_4951_begin_0, end = var_4951_end_0, end_mask = var_4951_end_mask_0, x = cache_85_cast_fp16)[name = string("op_4951_cast_fp16")]; + bool var_4957_interleave_0 = const()[name = string("op_4957_interleave_0"), val = bool(false)]; + tensor var_4957_cast_fp16 = concat(axis = var_68, interleave = var_4957_interleave_0, values = (var_4951_cast_fp16, key_43_cast_fp16))[name = string("op_4957_cast_fp16")]; + tensor encoder_layers_21_self_attn_linear_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(440388800))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(441437440))))[name = string("encoder_layers_21_self_attn_linear_q_weight_to_fp16_quantized")]; + tensor encoder_layers_21_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_21_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(441439552)))]; + tensor linear_192_cast_fp16 = linear(bias = encoder_layers_21_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_21_self_attn_linear_q_weight_to_fp16_quantized, x = key_43_cast_fp16)[name = string("linear_192_cast_fp16")]; + tensor var_4962 = const()[name = string("op_4962"), val = tensor([1, -1, 8, 128])]; + tensor q_127_cast_fp16 = reshape(shape = var_4962, x = linear_192_cast_fp16)[name = string("q_127_cast_fp16")]; + tensor encoder_layers_21_self_attn_linear_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(441441664))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(442490304))))[name = string("encoder_layers_21_self_attn_linear_k_weight_to_fp16_quantized")]; + tensor encoder_layers_21_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_21_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(442492416)))]; + tensor linear_193_cast_fp16 = linear(bias = encoder_layers_21_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_21_self_attn_linear_k_weight_to_fp16_quantized, x = input_1131_cast_fp16)[name = string("linear_193_cast_fp16")]; + tensor var_4967 = const()[name = string("op_4967"), val = tensor([1, -1, 8, 128])]; + tensor k_85_cast_fp16 = reshape(shape = var_4967, x = linear_193_cast_fp16)[name = string("k_85_cast_fp16")]; + tensor encoder_layers_21_self_attn_linear_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(442494528))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(443543168))))[name = string("encoder_layers_21_self_attn_linear_v_weight_to_fp16_quantized")]; + tensor encoder_layers_21_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_21_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(443545280)))]; + tensor linear_194_cast_fp16 = linear(bias = encoder_layers_21_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_21_self_attn_linear_v_weight_to_fp16_quantized, x = input_1131_cast_fp16)[name = string("linear_194_cast_fp16")]; + tensor var_4972 = const()[name = string("op_4972"), val = tensor([1, -1, 8, 128])]; + tensor v_43_cast_fp16 = reshape(shape = var_4972, x = linear_194_cast_fp16)[name = string("v_43_cast_fp16")]; + tensor value_51_perm_0 = const()[name = string("value_51_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_21_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_21_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(443547392)))]; + tensor var_4985_cast_fp16 = add(x = q_127_cast_fp16, y = encoder_layers_21_self_attn_pos_bias_u_to_fp16)[name = string("op_4985_cast_fp16")]; + tensor encoder_layers_21_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_21_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(443549504)))]; + tensor var_4987_cast_fp16 = add(x = q_127_cast_fp16, y = encoder_layers_21_self_attn_pos_bias_v_to_fp16)[name = string("op_4987_cast_fp16")]; + tensor q_with_bias_v_43_perm_0 = const()[name = string("q_with_bias_v_43_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_553_transpose_x_0 = const()[name = string("x_553_transpose_x_0"), val = bool(false)]; + bool x_553_transpose_y_0 = const()[name = string("x_553_transpose_y_0"), val = bool(false)]; + tensor op_4989_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(443551616))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(443694016))))[name = string("op_4989_to_fp16_quantized")]; + tensor q_with_bias_v_43_cast_fp16 = transpose(perm = q_with_bias_v_43_perm_0, x = var_4987_cast_fp16)[name = string("transpose_173")]; + tensor x_553_cast_fp16 = matmul(transpose_x = x_553_transpose_x_0, transpose_y = x_553_transpose_y_0, x = q_with_bias_v_43_cast_fp16, y = op_4989_to_fp16_quantized)[name = string("x_553_cast_fp16")]; + tensor x_555_pad_0 = const()[name = string("x_555_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_555_mode_0 = const()[name = string("x_555_mode_0"), val = string("constant")]; + fp16 const_352_to_fp16 = const()[name = string("const_352_to_fp16"), val = fp16(0x0p+0)]; + tensor x_555_cast_fp16 = pad(constant_val = const_352_to_fp16, mode = x_555_mode_0, pad = x_555_pad_0, x = x_553_cast_fp16)[name = string("x_555_cast_fp16")]; + tensor var_4997 = const()[name = string("op_4997"), val = tensor([1, 8, -1, 28])]; + tensor x_557_cast_fp16 = reshape(shape = var_4997, x = x_555_cast_fp16)[name = string("x_557_cast_fp16")]; + tensor var_5001_begin_0 = const()[name = string("op_5001_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_5001_end_0 = const()[name = string("op_5001_end_0"), val = tensor([1, 8, 140, 28])]; + tensor var_5001_end_mask_0 = const()[name = string("op_5001_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_5001_cast_fp16 = slice_by_index(begin = var_5001_begin_0, end = var_5001_end_0, end_mask = var_5001_end_mask_0, x = x_557_cast_fp16)[name = string("op_5001_cast_fp16")]; + tensor var_5002 = const()[name = string("op_5002"), val = tensor([1, 8, 28, 139])]; + tensor matrix_bd_85_cast_fp16 = reshape(shape = var_5002, x = var_5001_cast_fp16)[name = string("matrix_bd_85_cast_fp16")]; + bool matrix_ac_43_transpose_x_0 = const()[name = string("matrix_ac_43_transpose_x_0"), val = bool(false)]; + bool matrix_ac_43_transpose_y_0 = const()[name = string("matrix_ac_43_transpose_y_0"), val = bool(false)]; + tensor transpose_138_perm_0 = const()[name = string("transpose_138_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_139_perm_0 = const()[name = string("transpose_139_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_139 = transpose(perm = transpose_139_perm_0, x = k_85_cast_fp16)[name = string("transpose_171")]; + tensor transpose_138 = transpose(perm = transpose_138_perm_0, x = var_4985_cast_fp16)[name = string("transpose_172")]; + tensor matrix_ac_43_cast_fp16 = matmul(transpose_x = matrix_ac_43_transpose_x_0, transpose_y = matrix_ac_43_transpose_y_0, x = transpose_138, y = transpose_139)[name = string("matrix_ac_43_cast_fp16")]; + tensor matrix_bd_87_begin_0 = const()[name = string("matrix_bd_87_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_87_end_0 = const()[name = string("matrix_bd_87_end_0"), val = tensor([1, 8, 28, 70])]; + tensor matrix_bd_87_end_mask_0 = const()[name = string("matrix_bd_87_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_87_cast_fp16 = slice_by_index(begin = matrix_bd_87_begin_0, end = matrix_bd_87_end_0, end_mask = matrix_bd_87_end_mask_0, x = matrix_bd_85_cast_fp16)[name = string("matrix_bd_87_cast_fp16")]; + tensor var_5011_cast_fp16 = add(x = matrix_ac_43_cast_fp16, y = matrix_bd_87_cast_fp16)[name = string("op_5011_cast_fp16")]; + fp16 _inversed_scores_85_y_0_to_fp16 = const()[name = string("_inversed_scores_85_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_85_cast_fp16 = mul(x = var_5011_cast_fp16, y = _inversed_scores_85_y_0_to_fp16)[name = string("_inversed_scores_85_cast_fp16")]; + tensor scores_87_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_85_cast_fp16, cond = mask_11)[name = string("scores_87_cast_fp16")]; + tensor var_5017_cast_fp16 = softmax(axis = var_59, x = scores_87_cast_fp16)[name = string("op_5017_cast_fp16")]; + tensor input_1133_cast_fp16 = select(a = var_44_to_fp16, b = var_5017_cast_fp16, cond = mask_11)[name = string("input_1133_cast_fp16")]; + bool x_559_transpose_x_0 = const()[name = string("x_559_transpose_x_0"), val = bool(false)]; + bool x_559_transpose_y_0 = const()[name = string("x_559_transpose_y_0"), val = bool(false)]; + tensor value_51_cast_fp16 = transpose(perm = value_51_perm_0, x = v_43_cast_fp16)[name = string("transpose_170")]; + tensor x_559_cast_fp16 = matmul(transpose_x = x_559_transpose_x_0, transpose_y = x_559_transpose_y_0, x = input_1133_cast_fp16, y = value_51_cast_fp16)[name = string("x_559_cast_fp16")]; + tensor var_5021_perm_0 = const()[name = string("op_5021_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_5022 = const()[name = string("op_5022"), val = tensor([1, -1, 1024])]; + tensor var_5021_cast_fp16 = transpose(perm = var_5021_perm_0, x = x_559_cast_fp16)[name = string("transpose_169")]; + tensor input_1135_cast_fp16 = reshape(shape = var_5022, x = var_5021_cast_fp16)[name = string("input_1135_cast_fp16")]; + tensor encoder_layers_21_self_attn_linear_out_weight_to_fp16 = const()[name = string("encoder_layers_21_self_attn_linear_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(443694400)))]; + tensor encoder_layers_21_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_21_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(445791616)))]; + tensor linear_196_cast_fp16 = linear(bias = encoder_layers_21_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_21_self_attn_linear_out_weight_to_fp16, x = input_1135_cast_fp16)[name = string("linear_196_cast_fp16")]; + tensor input_1139_cast_fp16 = add(x = input_1129_cast_fp16, y = linear_196_cast_fp16)[name = string("input_1139_cast_fp16")]; + tensor x_563_axes_0 = const()[name = string("x_563_axes_0"), val = tensor([-1])]; + tensor encoder_layers_21_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_21_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(445793728)))]; + tensor encoder_layers_21_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_21_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(445795840)))]; + tensor x_563_cast_fp16 = layer_norm(axes = x_563_axes_0, beta = encoder_layers_21_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_21_norm_conv_weight_to_fp16, x = input_1139_cast_fp16)[name = string("x_563_cast_fp16")]; + tensor input_1141_perm_0 = const()[name = string("input_1141_perm_0"), val = tensor([0, 2, 1])]; + string input_1143_pad_type_0 = const()[name = string("input_1143_pad_type_0"), val = string("valid")]; + tensor input_1143_strides_0 = const()[name = string("input_1143_strides_0"), val = tensor([1])]; + tensor input_1143_pad_0 = const()[name = string("input_1143_pad_0"), val = tensor([0, 0])]; + tensor input_1143_dilations_0 = const()[name = string("input_1143_dilations_0"), val = tensor([1])]; + int32 input_1143_groups_0 = const()[name = string("input_1143_groups_0"), val = int32(1)]; + tensor encoder_layers_21_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(445797952))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(447895168))))[name = string("encoder_layers_21_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_1141_cast_fp16 = transpose(perm = input_1141_perm_0, x = x_563_cast_fp16)[name = string("transpose_168")]; + tensor input_1143_cast_fp16 = conv(dilations = input_1143_dilations_0, groups = input_1143_groups_0, pad = input_1143_pad_0, pad_type = input_1143_pad_type_0, strides = input_1143_strides_0, weight = encoder_layers_21_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_1141_cast_fp16)[name = string("input_1143_cast_fp16")]; + int32 x_565_split_num_splits_0 = const()[name = string("x_565_split_num_splits_0"), val = int32(2)]; + int32 x_565_split_axis_0 = const()[name = string("x_565_split_axis_0"), val = int32(1)]; + tensor x_565_split_cast_fp16_0, tensor x_565_split_cast_fp16_1 = split(axis = x_565_split_axis_0, num_splits = x_565_split_num_splits_0, x = input_1143_cast_fp16)[name = string("x_565_split_cast_fp16")]; + tensor x_565_split_1_sigmoid_cast_fp16 = sigmoid(x = x_565_split_cast_fp16_1)[name = string("x_565_split_1_sigmoid_cast_fp16")]; + tensor x_565_cast_fp16 = mul(x = x_565_split_cast_fp16_0, y = x_565_split_1_sigmoid_cast_fp16)[name = string("x_565_cast_fp16")]; + tensor input_1145_cast_fp16 = select(a = var_44_to_fp16, b = x_565_cast_fp16, cond = var_575)[name = string("input_1145_cast_fp16")]; + bool new_x_87_interleave_0 = const()[name = string("new_x_87_interleave_0"), val = bool(false)]; + tensor new_x_87_cast_fp16 = concat(axis = var_59, interleave = new_x_87_interleave_0, values = (cache_87_cast_fp16, input_1145_cast_fp16))[name = string("new_x_87_cast_fp16")]; + tensor var_5061_begin_0 = const()[name = string("op_5061_begin_0"), val = tensor([0, 0, 28])]; + tensor var_5061_end_0 = const()[name = string("op_5061_end_0"), val = tensor([1, 1024, 36])]; + tensor var_5061_end_mask_0 = const()[name = string("op_5061_end_mask_0"), val = tensor([true, true, true])]; + tensor var_5061_cast_fp16 = slice_by_index(begin = var_5061_begin_0, end = var_5061_end_0, end_mask = var_5061_end_mask_0, x = new_x_87_cast_fp16)[name = string("op_5061_cast_fp16")]; + string x_567_pad_type_0 = const()[name = string("x_567_pad_type_0"), val = string("valid")]; + int32 x_567_groups_0 = const()[name = string("x_567_groups_0"), val = int32(1024)]; + tensor x_567_strides_0 = const()[name = string("x_567_strides_0"), val = tensor([1])]; + tensor x_567_pad_0 = const()[name = string("x_567_pad_0"), val = tensor([0, 0])]; + tensor x_567_dilations_0 = const()[name = string("x_567_dilations_0"), val = tensor([1])]; + tensor encoder_layers_21_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(447899328))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(447908608))))[name = string("encoder_layers_21_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_567_cast_fp16 = conv(dilations = x_567_dilations_0, groups = x_567_groups_0, pad = x_567_pad_0, pad_type = x_567_pad_type_0, strides = x_567_strides_0, weight = encoder_layers_21_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_87_cast_fp16)[name = string("x_567_cast_fp16")]; + tensor input_1147_perm_0 = const()[name = string("input_1147_perm_0"), val = tensor([0, 2, 1])]; + tensor x_569_axes_0 = const()[name = string("x_569_axes_0"), val = tensor([-1])]; + tensor encoder_layers_21_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_21_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(447910720)))]; + tensor encoder_layers_21_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_21_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(447912832)))]; + tensor input_1147_cast_fp16 = transpose(perm = input_1147_perm_0, x = x_567_cast_fp16)[name = string("transpose_167")]; + tensor x_569_cast_fp16 = layer_norm(axes = x_569_axes_0, beta = encoder_layers_21_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_21_conv_batch_norm_weight_to_fp16, x = input_1147_cast_fp16)[name = string("x_569_cast_fp16")]; + tensor input_1149_perm_0 = const()[name = string("input_1149_perm_0"), val = tensor([0, 2, 1])]; + tensor input_1149_cast_fp16 = transpose(perm = input_1149_perm_0, x = x_569_cast_fp16)[name = string("transpose_166")]; + tensor input_1151_cast_fp16 = silu(x = input_1149_cast_fp16)[name = string("input_1151_cast_fp16")]; + string x_571_pad_type_0 = const()[name = string("x_571_pad_type_0"), val = string("valid")]; + tensor x_571_strides_0 = const()[name = string("x_571_strides_0"), val = tensor([1])]; + tensor x_571_pad_0 = const()[name = string("x_571_pad_0"), val = tensor([0, 0])]; + tensor x_571_dilations_0 = const()[name = string("x_571_dilations_0"), val = tensor([1])]; + int32 x_571_groups_0 = const()[name = string("x_571_groups_0"), val = int32(1)]; + tensor encoder_layers_21_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(447914944))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(448963584))))[name = string("encoder_layers_21_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_571_cast_fp16 = conv(dilations = x_571_dilations_0, groups = x_571_groups_0, pad = x_571_pad_0, pad_type = x_571_pad_type_0, strides = x_571_strides_0, weight = encoder_layers_21_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_1151_cast_fp16)[name = string("x_571_cast_fp16")]; + tensor input_1153_perm_0 = const()[name = string("input_1153_perm_0"), val = tensor([0, 2, 1])]; + tensor input_1153_cast_fp16 = transpose(perm = input_1153_perm_0, x = x_571_cast_fp16)[name = string("transpose_165")]; + tensor input_1155_cast_fp16 = add(x = input_1139_cast_fp16, y = input_1153_cast_fp16)[name = string("input_1155_cast_fp16")]; + tensor input_1157_axes_0 = const()[name = string("input_1157_axes_0"), val = tensor([-1])]; + tensor encoder_layers_21_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_21_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(448965696)))]; + tensor encoder_layers_21_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_21_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(448967808)))]; + tensor input_1157_cast_fp16 = layer_norm(axes = input_1157_axes_0, beta = encoder_layers_21_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_21_norm_feed_forward2_weight_to_fp16, x = input_1155_cast_fp16)[name = string("input_1157_cast_fp16")]; + tensor encoder_layers_21_feed_forward2_linear1_weight_to_fp16 = const()[name = string("encoder_layers_21_feed_forward2_linear1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(448969920)))]; + tensor encoder_layers_21_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_21_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(457358592)))]; + tensor linear_197_cast_fp16 = linear(bias = encoder_layers_21_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_21_feed_forward2_linear1_weight_to_fp16, x = input_1157_cast_fp16)[name = string("linear_197_cast_fp16")]; + tensor input_1161_cast_fp16 = silu(x = linear_197_cast_fp16)[name = string("input_1161_cast_fp16")]; + tensor encoder_layers_21_feed_forward2_linear2_weight_to_fp16 = const()[name = string("encoder_layers_21_feed_forward2_linear2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(457366848)))]; + tensor encoder_layers_21_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_21_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465755520)))]; + tensor linear_198_cast_fp16 = linear(bias = encoder_layers_21_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_21_feed_forward2_linear2_weight_to_fp16, x = input_1161_cast_fp16)[name = string("linear_198_cast_fp16")]; + fp16 var_5104_to_fp16 = const()[name = string("op_5104_to_fp16"), val = fp16(0x1p-1)]; + tensor var_5105_cast_fp16 = mul(x = linear_198_cast_fp16, y = var_5104_to_fp16)[name = string("op_5105_cast_fp16")]; + tensor input_1167_cast_fp16 = add(x = input_1155_cast_fp16, y = var_5105_cast_fp16)[name = string("input_1167_cast_fp16")]; + tensor input_1169_axes_0 = const()[name = string("input_1169_axes_0"), val = tensor([-1])]; + tensor encoder_layers_21_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_21_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465757632)))]; + tensor encoder_layers_21_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_21_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465759744)))]; + tensor input_1169_cast_fp16 = layer_norm(axes = input_1169_axes_0, beta = encoder_layers_21_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_21_norm_out_weight_to_fp16, x = input_1167_cast_fp16)[name = string("input_1169_cast_fp16")]; + tensor cache_89_begin_0 = const()[name = string("cache_89_begin_0"), val = tensor([22, 0, 0, 0])]; + tensor cache_89_end_0 = const()[name = string("cache_89_end_0"), val = tensor([23, 1, 42, 1024])]; + tensor cache_89_end_mask_0 = const()[name = string("cache_89_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_89_squeeze_mask_0 = const()[name = string("cache_89_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_89_cast_fp16 = slice_by_index(begin = cache_89_begin_0, end = cache_89_end_0, end_mask = cache_89_end_mask_0, squeeze_mask = cache_89_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_89_cast_fp16")]; + tensor cache_91_begin_0 = const()[name = string("cache_91_begin_0"), val = tensor([22, 0, 0, 0])]; + tensor cache_91_end_0 = const()[name = string("cache_91_end_0"), val = tensor([23, 1, 1024, 8])]; + tensor cache_91_end_mask_0 = const()[name = string("cache_91_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_91_squeeze_mask_0 = const()[name = string("cache_91_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_91_cast_fp16 = slice_by_index(begin = cache_91_begin_0, end = cache_91_end_0, end_mask = cache_91_end_mask_0, squeeze_mask = cache_91_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_91_cast_fp16")]; + tensor input_1171_axes_0 = const()[name = string("input_1171_axes_0"), val = tensor([-1])]; + tensor encoder_layers_22_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_22_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465761856)))]; + tensor encoder_layers_22_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_22_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465763968)))]; + tensor input_1171_cast_fp16 = layer_norm(axes = input_1171_axes_0, beta = encoder_layers_22_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_22_norm_feed_forward1_weight_to_fp16, x = input_1169_cast_fp16)[name = string("input_1171_cast_fp16")]; + tensor encoder_layers_22_feed_forward1_linear1_weight_to_fp16 = const()[name = string("encoder_layers_22_feed_forward1_linear1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465766080)))]; + tensor encoder_layers_22_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_22_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(474154752)))]; + tensor linear_199_cast_fp16 = linear(bias = encoder_layers_22_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_22_feed_forward1_linear1_weight_to_fp16, x = input_1171_cast_fp16)[name = string("linear_199_cast_fp16")]; + tensor input_1175_cast_fp16 = silu(x = linear_199_cast_fp16)[name = string("input_1175_cast_fp16")]; + tensor encoder_layers_22_feed_forward1_linear2_weight_to_fp16 = const()[name = string("encoder_layers_22_feed_forward1_linear2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(474163008)))]; + tensor encoder_layers_22_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_22_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(482551680)))]; + tensor linear_200_cast_fp16 = linear(bias = encoder_layers_22_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_22_feed_forward1_linear2_weight_to_fp16, x = input_1175_cast_fp16)[name = string("linear_200_cast_fp16")]; + fp16 var_5141_to_fp16 = const()[name = string("op_5141_to_fp16"), val = fp16(0x1p-1)]; + tensor var_5142_cast_fp16 = mul(x = linear_200_cast_fp16, y = var_5141_to_fp16)[name = string("op_5142_cast_fp16")]; + tensor input_1181_cast_fp16 = add(x = input_1169_cast_fp16, y = var_5142_cast_fp16)[name = string("input_1181_cast_fp16")]; + tensor key_45_axes_0 = const()[name = string("key_45_axes_0"), val = tensor([-1])]; + tensor encoder_layers_22_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_22_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(482553792)))]; + tensor encoder_layers_22_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_22_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(482555904)))]; + tensor key_45_cast_fp16 = layer_norm(axes = key_45_axes_0, beta = encoder_layers_22_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_22_norm_self_att_weight_to_fp16, x = input_1181_cast_fp16)[name = string("key_45_cast_fp16")]; + bool input_1183_interleave_0 = const()[name = string("input_1183_interleave_0"), val = bool(false)]; + tensor input_1183_cast_fp16 = concat(axis = var_68, interleave = input_1183_interleave_0, values = (cache_89_cast_fp16, key_45_cast_fp16))[name = string("input_1183_cast_fp16")]; + tensor var_5164_begin_0 = const()[name = string("op_5164_begin_0"), val = tensor([0, 28, 0])]; + tensor var_5164_end_0 = const()[name = string("op_5164_end_0"), val = tensor([1, 42, 1024])]; + tensor var_5164_end_mask_0 = const()[name = string("op_5164_end_mask_0"), val = tensor([true, true, true])]; + tensor var_5164_cast_fp16 = slice_by_index(begin = var_5164_begin_0, end = var_5164_end_0, end_mask = var_5164_end_mask_0, x = cache_89_cast_fp16)[name = string("op_5164_cast_fp16")]; + bool var_5170_interleave_0 = const()[name = string("op_5170_interleave_0"), val = bool(false)]; + tensor var_5170_cast_fp16 = concat(axis = var_68, interleave = var_5170_interleave_0, values = (var_5164_cast_fp16, key_45_cast_fp16))[name = string("op_5170_cast_fp16")]; + tensor encoder_layers_22_self_attn_linear_q_weight_to_fp16 = const()[name = string("encoder_layers_22_self_attn_linear_q_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(482558016)))]; + tensor encoder_layers_22_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_22_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(484655232)))]; + tensor linear_201_cast_fp16 = linear(bias = encoder_layers_22_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_22_self_attn_linear_q_weight_to_fp16, x = key_45_cast_fp16)[name = string("linear_201_cast_fp16")]; + tensor var_5175 = const()[name = string("op_5175"), val = tensor([1, -1, 8, 128])]; + tensor q_133_cast_fp16 = reshape(shape = var_5175, x = linear_201_cast_fp16)[name = string("q_133_cast_fp16")]; + tensor encoder_layers_22_self_attn_linear_k_weight_to_fp16 = const()[name = string("encoder_layers_22_self_attn_linear_k_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(484657344)))]; + tensor encoder_layers_22_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_22_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(486754560)))]; + tensor linear_202_cast_fp16 = linear(bias = encoder_layers_22_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_22_self_attn_linear_k_weight_to_fp16, x = input_1183_cast_fp16)[name = string("linear_202_cast_fp16")]; + tensor var_5180 = const()[name = string("op_5180"), val = tensor([1, -1, 8, 128])]; + tensor k_89_cast_fp16 = reshape(shape = var_5180, x = linear_202_cast_fp16)[name = string("k_89_cast_fp16")]; + tensor encoder_layers_22_self_attn_linear_v_weight_to_fp16 = const()[name = string("encoder_layers_22_self_attn_linear_v_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(486756672)))]; + tensor encoder_layers_22_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_22_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(488853888)))]; + tensor linear_203_cast_fp16 = linear(bias = encoder_layers_22_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_22_self_attn_linear_v_weight_to_fp16, x = input_1183_cast_fp16)[name = string("linear_203_cast_fp16")]; + tensor var_5185 = const()[name = string("op_5185"), val = tensor([1, -1, 8, 128])]; + tensor v_45_cast_fp16 = reshape(shape = var_5185, x = linear_203_cast_fp16)[name = string("v_45_cast_fp16")]; + tensor value_53_perm_0 = const()[name = string("value_53_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_22_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_22_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(488856000)))]; + tensor var_5198_cast_fp16 = add(x = q_133_cast_fp16, y = encoder_layers_22_self_attn_pos_bias_u_to_fp16)[name = string("op_5198_cast_fp16")]; + tensor encoder_layers_22_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_22_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(488858112)))]; + tensor var_5200_cast_fp16 = add(x = q_133_cast_fp16, y = encoder_layers_22_self_attn_pos_bias_v_to_fp16)[name = string("op_5200_cast_fp16")]; + tensor q_with_bias_v_45_perm_0 = const()[name = string("q_with_bias_v_45_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_579_transpose_x_0 = const()[name = string("x_579_transpose_x_0"), val = bool(false)]; + bool x_579_transpose_y_0 = const()[name = string("x_579_transpose_y_0"), val = bool(false)]; + tensor op_5202_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(488860224))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(489002624))))[name = string("op_5202_to_fp16_quantized")]; + tensor q_with_bias_v_45_cast_fp16 = transpose(perm = q_with_bias_v_45_perm_0, x = var_5200_cast_fp16)[name = string("transpose_164")]; + tensor x_579_cast_fp16 = matmul(transpose_x = x_579_transpose_x_0, transpose_y = x_579_transpose_y_0, x = q_with_bias_v_45_cast_fp16, y = op_5202_to_fp16_quantized)[name = string("x_579_cast_fp16")]; + tensor x_581_pad_0 = const()[name = string("x_581_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_581_mode_0 = const()[name = string("x_581_mode_0"), val = string("constant")]; + fp16 const_365_to_fp16 = const()[name = string("const_365_to_fp16"), val = fp16(0x0p+0)]; + tensor x_581_cast_fp16 = pad(constant_val = const_365_to_fp16, mode = x_581_mode_0, pad = x_581_pad_0, x = x_579_cast_fp16)[name = string("x_581_cast_fp16")]; + tensor var_5210 = const()[name = string("op_5210"), val = tensor([1, 8, -1, 28])]; + tensor x_583_cast_fp16 = reshape(shape = var_5210, x = x_581_cast_fp16)[name = string("x_583_cast_fp16")]; + tensor var_5214_begin_0 = const()[name = string("op_5214_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_5214_end_0 = const()[name = string("op_5214_end_0"), val = tensor([1, 8, 140, 28])]; + tensor var_5214_end_mask_0 = const()[name = string("op_5214_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_5214_cast_fp16 = slice_by_index(begin = var_5214_begin_0, end = var_5214_end_0, end_mask = var_5214_end_mask_0, x = x_583_cast_fp16)[name = string("op_5214_cast_fp16")]; + tensor var_5215 = const()[name = string("op_5215"), val = tensor([1, 8, 28, 139])]; + tensor matrix_bd_89_cast_fp16 = reshape(shape = var_5215, x = var_5214_cast_fp16)[name = string("matrix_bd_89_cast_fp16")]; + bool matrix_ac_45_transpose_x_0 = const()[name = string("matrix_ac_45_transpose_x_0"), val = bool(false)]; + bool matrix_ac_45_transpose_y_0 = const()[name = string("matrix_ac_45_transpose_y_0"), val = bool(false)]; + tensor transpose_140_perm_0 = const()[name = string("transpose_140_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_141_perm_0 = const()[name = string("transpose_141_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_141 = transpose(perm = transpose_141_perm_0, x = k_89_cast_fp16)[name = string("transpose_162")]; + tensor transpose_140 = transpose(perm = transpose_140_perm_0, x = var_5198_cast_fp16)[name = string("transpose_163")]; + tensor matrix_ac_45_cast_fp16 = matmul(transpose_x = matrix_ac_45_transpose_x_0, transpose_y = matrix_ac_45_transpose_y_0, x = transpose_140, y = transpose_141)[name = string("matrix_ac_45_cast_fp16")]; + tensor matrix_bd_91_begin_0 = const()[name = string("matrix_bd_91_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_91_end_0 = const()[name = string("matrix_bd_91_end_0"), val = tensor([1, 8, 28, 70])]; + tensor matrix_bd_91_end_mask_0 = const()[name = string("matrix_bd_91_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_91_cast_fp16 = slice_by_index(begin = matrix_bd_91_begin_0, end = matrix_bd_91_end_0, end_mask = matrix_bd_91_end_mask_0, x = matrix_bd_89_cast_fp16)[name = string("matrix_bd_91_cast_fp16")]; + tensor var_5224_cast_fp16 = add(x = matrix_ac_45_cast_fp16, y = matrix_bd_91_cast_fp16)[name = string("op_5224_cast_fp16")]; + fp16 _inversed_scores_89_y_0_to_fp16 = const()[name = string("_inversed_scores_89_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_89_cast_fp16 = mul(x = var_5224_cast_fp16, y = _inversed_scores_89_y_0_to_fp16)[name = string("_inversed_scores_89_cast_fp16")]; + tensor scores_91_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_89_cast_fp16, cond = mask_11)[name = string("scores_91_cast_fp16")]; + tensor var_5230_cast_fp16 = softmax(axis = var_59, x = scores_91_cast_fp16)[name = string("op_5230_cast_fp16")]; + tensor input_1185_cast_fp16 = select(a = var_44_to_fp16, b = var_5230_cast_fp16, cond = mask_11)[name = string("input_1185_cast_fp16")]; + bool x_585_transpose_x_0 = const()[name = string("x_585_transpose_x_0"), val = bool(false)]; + bool x_585_transpose_y_0 = const()[name = string("x_585_transpose_y_0"), val = bool(false)]; + tensor value_53_cast_fp16 = transpose(perm = value_53_perm_0, x = v_45_cast_fp16)[name = string("transpose_161")]; + tensor x_585_cast_fp16 = matmul(transpose_x = x_585_transpose_x_0, transpose_y = x_585_transpose_y_0, x = input_1185_cast_fp16, y = value_53_cast_fp16)[name = string("x_585_cast_fp16")]; + tensor var_5234_perm_0 = const()[name = string("op_5234_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_5235 = const()[name = string("op_5235"), val = tensor([1, -1, 1024])]; + tensor var_5234_cast_fp16 = transpose(perm = var_5234_perm_0, x = x_585_cast_fp16)[name = string("transpose_160")]; + tensor input_1187_cast_fp16 = reshape(shape = var_5235, x = var_5234_cast_fp16)[name = string("input_1187_cast_fp16")]; + tensor encoder_layers_22_self_attn_linear_out_weight_to_fp16 = const()[name = string("encoder_layers_22_self_attn_linear_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(489003008)))]; + tensor encoder_layers_22_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_22_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(491100224)))]; + tensor linear_205_cast_fp16 = linear(bias = encoder_layers_22_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_22_self_attn_linear_out_weight_to_fp16, x = input_1187_cast_fp16)[name = string("linear_205_cast_fp16")]; + tensor input_1191_cast_fp16 = add(x = input_1181_cast_fp16, y = linear_205_cast_fp16)[name = string("input_1191_cast_fp16")]; + tensor x_589_axes_0 = const()[name = string("x_589_axes_0"), val = tensor([-1])]; + tensor encoder_layers_22_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_22_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(491102336)))]; + tensor encoder_layers_22_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_22_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(491104448)))]; + tensor x_589_cast_fp16 = layer_norm(axes = x_589_axes_0, beta = encoder_layers_22_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_22_norm_conv_weight_to_fp16, x = input_1191_cast_fp16)[name = string("x_589_cast_fp16")]; + tensor input_1193_perm_0 = const()[name = string("input_1193_perm_0"), val = tensor([0, 2, 1])]; + string input_1195_pad_type_0 = const()[name = string("input_1195_pad_type_0"), val = string("valid")]; + tensor input_1195_strides_0 = const()[name = string("input_1195_strides_0"), val = tensor([1])]; + tensor input_1195_pad_0 = const()[name = string("input_1195_pad_0"), val = tensor([0, 0])]; + tensor input_1195_dilations_0 = const()[name = string("input_1195_dilations_0"), val = tensor([1])]; + int32 input_1195_groups_0 = const()[name = string("input_1195_groups_0"), val = int32(1)]; + tensor encoder_layers_22_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(491106560))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(493203776))))[name = string("encoder_layers_22_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_1193_cast_fp16 = transpose(perm = input_1193_perm_0, x = x_589_cast_fp16)[name = string("transpose_159")]; + tensor input_1195_cast_fp16 = conv(dilations = input_1195_dilations_0, groups = input_1195_groups_0, pad = input_1195_pad_0, pad_type = input_1195_pad_type_0, strides = input_1195_strides_0, weight = encoder_layers_22_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_1193_cast_fp16)[name = string("input_1195_cast_fp16")]; + int32 x_591_split_num_splits_0 = const()[name = string("x_591_split_num_splits_0"), val = int32(2)]; + int32 x_591_split_axis_0 = const()[name = string("x_591_split_axis_0"), val = int32(1)]; + tensor x_591_split_cast_fp16_0, tensor x_591_split_cast_fp16_1 = split(axis = x_591_split_axis_0, num_splits = x_591_split_num_splits_0, x = input_1195_cast_fp16)[name = string("x_591_split_cast_fp16")]; + tensor x_591_split_1_sigmoid_cast_fp16 = sigmoid(x = x_591_split_cast_fp16_1)[name = string("x_591_split_1_sigmoid_cast_fp16")]; + tensor x_591_cast_fp16 = mul(x = x_591_split_cast_fp16_0, y = x_591_split_1_sigmoid_cast_fp16)[name = string("x_591_cast_fp16")]; + tensor input_1197_cast_fp16 = select(a = var_44_to_fp16, b = x_591_cast_fp16, cond = var_575)[name = string("input_1197_cast_fp16")]; + bool new_x_91_interleave_0 = const()[name = string("new_x_91_interleave_0"), val = bool(false)]; + tensor new_x_91_cast_fp16 = concat(axis = var_59, interleave = new_x_91_interleave_0, values = (cache_91_cast_fp16, input_1197_cast_fp16))[name = string("new_x_91_cast_fp16")]; + tensor var_5274_begin_0 = const()[name = string("op_5274_begin_0"), val = tensor([0, 0, 28])]; + tensor var_5274_end_0 = const()[name = string("op_5274_end_0"), val = tensor([1, 1024, 36])]; + tensor var_5274_end_mask_0 = const()[name = string("op_5274_end_mask_0"), val = tensor([true, true, true])]; + tensor var_5274_cast_fp16 = slice_by_index(begin = var_5274_begin_0, end = var_5274_end_0, end_mask = var_5274_end_mask_0, x = new_x_91_cast_fp16)[name = string("op_5274_cast_fp16")]; + string x_593_pad_type_0 = const()[name = string("x_593_pad_type_0"), val = string("valid")]; + int32 x_593_groups_0 = const()[name = string("x_593_groups_0"), val = int32(1024)]; + tensor x_593_strides_0 = const()[name = string("x_593_strides_0"), val = tensor([1])]; + tensor x_593_pad_0 = const()[name = string("x_593_pad_0"), val = tensor([0, 0])]; + tensor x_593_dilations_0 = const()[name = string("x_593_dilations_0"), val = tensor([1])]; + tensor encoder_layers_22_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(493207936))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(493217216))))[name = string("encoder_layers_22_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_593_cast_fp16 = conv(dilations = x_593_dilations_0, groups = x_593_groups_0, pad = x_593_pad_0, pad_type = x_593_pad_type_0, strides = x_593_strides_0, weight = encoder_layers_22_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_91_cast_fp16)[name = string("x_593_cast_fp16")]; + tensor input_1199_perm_0 = const()[name = string("input_1199_perm_0"), val = tensor([0, 2, 1])]; + tensor x_595_axes_0 = const()[name = string("x_595_axes_0"), val = tensor([-1])]; + tensor encoder_layers_22_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_22_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(493219328)))]; + tensor encoder_layers_22_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_22_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(493221440)))]; + tensor input_1199_cast_fp16 = transpose(perm = input_1199_perm_0, x = x_593_cast_fp16)[name = string("transpose_158")]; + tensor x_595_cast_fp16 = layer_norm(axes = x_595_axes_0, beta = encoder_layers_22_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_22_conv_batch_norm_weight_to_fp16, x = input_1199_cast_fp16)[name = string("x_595_cast_fp16")]; + tensor input_1201_perm_0 = const()[name = string("input_1201_perm_0"), val = tensor([0, 2, 1])]; + tensor input_1201_cast_fp16 = transpose(perm = input_1201_perm_0, x = x_595_cast_fp16)[name = string("transpose_157")]; + tensor input_1203_cast_fp16 = silu(x = input_1201_cast_fp16)[name = string("input_1203_cast_fp16")]; + string x_597_pad_type_0 = const()[name = string("x_597_pad_type_0"), val = string("valid")]; + tensor x_597_strides_0 = const()[name = string("x_597_strides_0"), val = tensor([1])]; + tensor x_597_pad_0 = const()[name = string("x_597_pad_0"), val = tensor([0, 0])]; + tensor x_597_dilations_0 = const()[name = string("x_597_dilations_0"), val = tensor([1])]; + int32 x_597_groups_0 = const()[name = string("x_597_groups_0"), val = int32(1)]; + tensor encoder_layers_22_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(493223552))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(494272192))))[name = string("encoder_layers_22_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_597_cast_fp16 = conv(dilations = x_597_dilations_0, groups = x_597_groups_0, pad = x_597_pad_0, pad_type = x_597_pad_type_0, strides = x_597_strides_0, weight = encoder_layers_22_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_1203_cast_fp16)[name = string("x_597_cast_fp16")]; + tensor input_1205_perm_0 = const()[name = string("input_1205_perm_0"), val = tensor([0, 2, 1])]; + tensor input_1205_cast_fp16 = transpose(perm = input_1205_perm_0, x = x_597_cast_fp16)[name = string("transpose_156")]; + tensor input_1207_cast_fp16 = add(x = input_1191_cast_fp16, y = input_1205_cast_fp16)[name = string("input_1207_cast_fp16")]; + tensor input_1209_axes_0 = const()[name = string("input_1209_axes_0"), val = tensor([-1])]; + tensor encoder_layers_22_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_22_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(494274304)))]; + tensor encoder_layers_22_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_22_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(494276416)))]; + tensor input_1209_cast_fp16 = layer_norm(axes = input_1209_axes_0, beta = encoder_layers_22_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_22_norm_feed_forward2_weight_to_fp16, x = input_1207_cast_fp16)[name = string("input_1209_cast_fp16")]; + tensor encoder_layers_22_feed_forward2_linear1_weight_to_fp16 = const()[name = string("encoder_layers_22_feed_forward2_linear1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(494278528)))]; + tensor encoder_layers_22_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_22_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(502667200)))]; + tensor linear_206_cast_fp16 = linear(bias = encoder_layers_22_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_22_feed_forward2_linear1_weight_to_fp16, x = input_1209_cast_fp16)[name = string("linear_206_cast_fp16")]; + tensor input_1213_cast_fp16 = silu(x = linear_206_cast_fp16)[name = string("input_1213_cast_fp16")]; + tensor encoder_layers_22_feed_forward2_linear2_weight_to_fp16 = const()[name = string("encoder_layers_22_feed_forward2_linear2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(502675456)))]; + tensor encoder_layers_22_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_22_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(511064128)))]; + tensor linear_207_cast_fp16 = linear(bias = encoder_layers_22_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_22_feed_forward2_linear2_weight_to_fp16, x = input_1213_cast_fp16)[name = string("linear_207_cast_fp16")]; + fp16 var_5317_to_fp16 = const()[name = string("op_5317_to_fp16"), val = fp16(0x1p-1)]; + tensor var_5318_cast_fp16 = mul(x = linear_207_cast_fp16, y = var_5317_to_fp16)[name = string("op_5318_cast_fp16")]; + tensor input_1219_cast_fp16 = add(x = input_1207_cast_fp16, y = var_5318_cast_fp16)[name = string("input_1219_cast_fp16")]; + tensor input_1221_axes_0 = const()[name = string("input_1221_axes_0"), val = tensor([-1])]; + tensor encoder_layers_22_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_22_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(511066240)))]; + tensor encoder_layers_22_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_22_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(511068352)))]; + tensor input_1221_cast_fp16 = layer_norm(axes = input_1221_axes_0, beta = encoder_layers_22_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_22_norm_out_weight_to_fp16, x = input_1219_cast_fp16)[name = string("input_1221_cast_fp16")]; + tensor cache_93_begin_0 = const()[name = string("cache_93_begin_0"), val = tensor([23, 0, 0, 0])]; + tensor cache_93_end_0 = const()[name = string("cache_93_end_0"), val = tensor([24, 1, 42, 1024])]; + tensor cache_93_end_mask_0 = const()[name = string("cache_93_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_93_squeeze_mask_0 = const()[name = string("cache_93_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_93_cast_fp16 = slice_by_index(begin = cache_93_begin_0, end = cache_93_end_0, end_mask = cache_93_end_mask_0, squeeze_mask = cache_93_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_93_cast_fp16")]; + tensor cache_begin_0 = const()[name = string("cache_begin_0"), val = tensor([23, 0, 0, 0])]; + tensor cache_end_0 = const()[name = string("cache_end_0"), val = tensor([24, 1, 1024, 8])]; + tensor cache_end_mask_0 = const()[name = string("cache_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_squeeze_mask_0 = const()[name = string("cache_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_cast_fp16 = slice_by_index(begin = cache_begin_0, end = cache_end_0, end_mask = cache_end_mask_0, squeeze_mask = cache_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_cast_fp16")]; + tensor input_1223_axes_0 = const()[name = string("input_1223_axes_0"), val = tensor([-1])]; + tensor encoder_layers_23_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_23_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(511070464)))]; + tensor encoder_layers_23_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_23_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(511072576)))]; + tensor input_1223_cast_fp16 = layer_norm(axes = input_1223_axes_0, beta = encoder_layers_23_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_23_norm_feed_forward1_weight_to_fp16, x = input_1221_cast_fp16)[name = string("input_1223_cast_fp16")]; + tensor encoder_layers_23_feed_forward1_linear1_weight_to_fp16 = const()[name = string("encoder_layers_23_feed_forward1_linear1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(511074688)))]; + tensor encoder_layers_23_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_23_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(519463360)))]; + tensor linear_208_cast_fp16 = linear(bias = encoder_layers_23_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_23_feed_forward1_linear1_weight_to_fp16, x = input_1223_cast_fp16)[name = string("linear_208_cast_fp16")]; + tensor input_1227_cast_fp16 = silu(x = linear_208_cast_fp16)[name = string("input_1227_cast_fp16")]; + tensor encoder_layers_23_feed_forward1_linear2_weight_to_fp16 = const()[name = string("encoder_layers_23_feed_forward1_linear2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(519471616)))]; + tensor encoder_layers_23_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_23_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(527860288)))]; + tensor linear_209_cast_fp16 = linear(bias = encoder_layers_23_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_23_feed_forward1_linear2_weight_to_fp16, x = input_1227_cast_fp16)[name = string("linear_209_cast_fp16")]; + fp16 var_5354_to_fp16 = const()[name = string("op_5354_to_fp16"), val = fp16(0x1p-1)]; + tensor var_5355_cast_fp16 = mul(x = linear_209_cast_fp16, y = var_5354_to_fp16)[name = string("op_5355_cast_fp16")]; + tensor input_1233_cast_fp16 = add(x = input_1221_cast_fp16, y = var_5355_cast_fp16)[name = string("input_1233_cast_fp16")]; + tensor key_axes_0 = const()[name = string("key_axes_0"), val = tensor([-1])]; + tensor encoder_layers_23_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_23_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(527862400)))]; + tensor encoder_layers_23_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_23_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(527864512)))]; + tensor key_cast_fp16 = layer_norm(axes = key_axes_0, beta = encoder_layers_23_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_23_norm_self_att_weight_to_fp16, x = input_1233_cast_fp16)[name = string("key_cast_fp16")]; + bool input_1235_interleave_0 = const()[name = string("input_1235_interleave_0"), val = bool(false)]; + tensor input_1235_cast_fp16 = concat(axis = var_68, interleave = input_1235_interleave_0, values = (cache_93_cast_fp16, key_cast_fp16))[name = string("input_1235_cast_fp16")]; + tensor var_5377_begin_0 = const()[name = string("op_5377_begin_0"), val = tensor([0, 28, 0])]; + tensor var_5377_end_0 = const()[name = string("op_5377_end_0"), val = tensor([1, 42, 1024])]; + tensor var_5377_end_mask_0 = const()[name = string("op_5377_end_mask_0"), val = tensor([true, true, true])]; + tensor var_5377_cast_fp16 = slice_by_index(begin = var_5377_begin_0, end = var_5377_end_0, end_mask = var_5377_end_mask_0, x = cache_93_cast_fp16)[name = string("op_5377_cast_fp16")]; + bool cache_last_channel_cur_interleave_0 = const()[name = string("cache_last_channel_cur_interleave_0"), val = bool(false)]; + tensor cache_last_channel_cur_cast_fp16 = concat(axis = var_68, interleave = cache_last_channel_cur_interleave_0, values = (var_5377_cast_fp16, key_cast_fp16))[name = string("cache_last_channel_cur_cast_fp16")]; + tensor encoder_layers_23_self_attn_linear_q_weight_to_fp16 = const()[name = string("encoder_layers_23_self_attn_linear_q_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(527866624)))]; + tensor encoder_layers_23_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_23_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(529963840)))]; + tensor linear_210_cast_fp16 = linear(bias = encoder_layers_23_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_23_self_attn_linear_q_weight_to_fp16, x = key_cast_fp16)[name = string("linear_210_cast_fp16")]; + tensor var_5388 = const()[name = string("op_5388"), val = tensor([1, -1, 8, 128])]; + tensor q_139_cast_fp16 = reshape(shape = var_5388, x = linear_210_cast_fp16)[name = string("q_139_cast_fp16")]; + tensor encoder_layers_23_self_attn_linear_k_weight_to_fp16 = const()[name = string("encoder_layers_23_self_attn_linear_k_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(529965952)))]; + tensor encoder_layers_23_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_23_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(532063168)))]; + tensor linear_211_cast_fp16 = linear(bias = encoder_layers_23_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_23_self_attn_linear_k_weight_to_fp16, x = input_1235_cast_fp16)[name = string("linear_211_cast_fp16")]; + tensor var_5393 = const()[name = string("op_5393"), val = tensor([1, -1, 8, 128])]; + tensor k_93_cast_fp16 = reshape(shape = var_5393, x = linear_211_cast_fp16)[name = string("k_93_cast_fp16")]; + tensor encoder_layers_23_self_attn_linear_v_weight_to_fp16 = const()[name = string("encoder_layers_23_self_attn_linear_v_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(532065280)))]; + tensor encoder_layers_23_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_23_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(534162496)))]; + tensor linear_212_cast_fp16 = linear(bias = encoder_layers_23_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_23_self_attn_linear_v_weight_to_fp16, x = input_1235_cast_fp16)[name = string("linear_212_cast_fp16")]; + tensor var_5398 = const()[name = string("op_5398"), val = tensor([1, -1, 8, 128])]; + tensor v_cast_fp16 = reshape(shape = var_5398, x = linear_212_cast_fp16)[name = string("v_cast_fp16")]; + tensor value_perm_0 = const()[name = string("value_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_23_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_23_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(534164608)))]; + tensor var_5411_cast_fp16 = add(x = q_139_cast_fp16, y = encoder_layers_23_self_attn_pos_bias_u_to_fp16)[name = string("op_5411_cast_fp16")]; + tensor encoder_layers_23_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_23_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(534166720)))]; + tensor var_5413_cast_fp16 = add(x = q_139_cast_fp16, y = encoder_layers_23_self_attn_pos_bias_v_to_fp16)[name = string("op_5413_cast_fp16")]; + tensor q_with_bias_v_perm_0 = const()[name = string("q_with_bias_v_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_605_transpose_x_0 = const()[name = string("x_605_transpose_x_0"), val = bool(false)]; + bool x_605_transpose_y_0 = const()[name = string("x_605_transpose_y_0"), val = bool(false)]; + tensor op_5415_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(534168832))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(534311232))))[name = string("op_5415_to_fp16_quantized")]; + tensor q_with_bias_v_cast_fp16 = transpose(perm = q_with_bias_v_perm_0, x = var_5413_cast_fp16)[name = string("transpose_155")]; + tensor x_605_cast_fp16 = matmul(transpose_x = x_605_transpose_x_0, transpose_y = x_605_transpose_y_0, x = q_with_bias_v_cast_fp16, y = op_5415_to_fp16_quantized)[name = string("x_605_cast_fp16")]; + tensor x_607_pad_0 = const()[name = string("x_607_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_607_mode_0 = const()[name = string("x_607_mode_0"), val = string("constant")]; + fp16 const_378_to_fp16 = const()[name = string("const_378_to_fp16"), val = fp16(0x0p+0)]; + tensor x_607_cast_fp16 = pad(constant_val = const_378_to_fp16, mode = x_607_mode_0, pad = x_607_pad_0, x = x_605_cast_fp16)[name = string("x_607_cast_fp16")]; + tensor var_5423 = const()[name = string("op_5423"), val = tensor([1, 8, -1, 28])]; + tensor x_609_cast_fp16 = reshape(shape = var_5423, x = x_607_cast_fp16)[name = string("x_609_cast_fp16")]; + tensor var_5427_begin_0 = const()[name = string("op_5427_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_5427_end_0 = const()[name = string("op_5427_end_0"), val = tensor([1, 8, 140, 28])]; + tensor var_5427_end_mask_0 = const()[name = string("op_5427_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_5427_cast_fp16 = slice_by_index(begin = var_5427_begin_0, end = var_5427_end_0, end_mask = var_5427_end_mask_0, x = x_609_cast_fp16)[name = string("op_5427_cast_fp16")]; + tensor var_5428 = const()[name = string("op_5428"), val = tensor([1, 8, 28, 139])]; + tensor matrix_bd_93_cast_fp16 = reshape(shape = var_5428, x = var_5427_cast_fp16)[name = string("matrix_bd_93_cast_fp16")]; + bool matrix_ac_transpose_x_0 = const()[name = string("matrix_ac_transpose_x_0"), val = bool(false)]; + bool matrix_ac_transpose_y_0 = const()[name = string("matrix_ac_transpose_y_0"), val = bool(false)]; + tensor transpose_142_perm_0 = const()[name = string("transpose_142_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_143_perm_0 = const()[name = string("transpose_143_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_143 = transpose(perm = transpose_143_perm_0, x = k_93_cast_fp16)[name = string("transpose_153")]; + tensor transpose_142 = transpose(perm = transpose_142_perm_0, x = var_5411_cast_fp16)[name = string("transpose_154")]; + tensor matrix_ac_cast_fp16 = matmul(transpose_x = matrix_ac_transpose_x_0, transpose_y = matrix_ac_transpose_y_0, x = transpose_142, y = transpose_143)[name = string("matrix_ac_cast_fp16")]; + tensor matrix_bd_begin_0 = const()[name = string("matrix_bd_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_end_0 = const()[name = string("matrix_bd_end_0"), val = tensor([1, 8, 28, 70])]; + tensor matrix_bd_end_mask_0 = const()[name = string("matrix_bd_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_cast_fp16 = slice_by_index(begin = matrix_bd_begin_0, end = matrix_bd_end_0, end_mask = matrix_bd_end_mask_0, x = matrix_bd_93_cast_fp16)[name = string("matrix_bd_cast_fp16")]; + tensor var_5437_cast_fp16 = add(x = matrix_ac_cast_fp16, y = matrix_bd_cast_fp16)[name = string("op_5437_cast_fp16")]; + fp16 _inversed_scores_93_y_0_to_fp16 = const()[name = string("_inversed_scores_93_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_93_cast_fp16 = mul(x = var_5437_cast_fp16, y = _inversed_scores_93_y_0_to_fp16)[name = string("_inversed_scores_93_cast_fp16")]; + tensor scores_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_93_cast_fp16, cond = mask_11)[name = string("scores_cast_fp16")]; + tensor var_5443_cast_fp16 = softmax(axis = var_59, x = scores_cast_fp16)[name = string("op_5443_cast_fp16")]; + tensor input_1237_cast_fp16 = select(a = var_44_to_fp16, b = var_5443_cast_fp16, cond = mask_11)[name = string("input_1237_cast_fp16")]; + bool x_611_transpose_x_0 = const()[name = string("x_611_transpose_x_0"), val = bool(false)]; + bool x_611_transpose_y_0 = const()[name = string("x_611_transpose_y_0"), val = bool(false)]; + tensor value_cast_fp16 = transpose(perm = value_perm_0, x = v_cast_fp16)[name = string("transpose_152")]; + tensor x_611_cast_fp16 = matmul(transpose_x = x_611_transpose_x_0, transpose_y = x_611_transpose_y_0, x = input_1237_cast_fp16, y = value_cast_fp16)[name = string("x_611_cast_fp16")]; + tensor var_5447_perm_0 = const()[name = string("op_5447_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_5448 = const()[name = string("op_5448"), val = tensor([1, -1, 1024])]; + tensor var_5447_cast_fp16 = transpose(perm = var_5447_perm_0, x = x_611_cast_fp16)[name = string("transpose_151")]; + tensor input_1239_cast_fp16 = reshape(shape = var_5448, x = var_5447_cast_fp16)[name = string("input_1239_cast_fp16")]; + tensor encoder_layers_23_self_attn_linear_out_weight_to_fp16 = const()[name = string("encoder_layers_23_self_attn_linear_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(534311616)))]; + tensor encoder_layers_23_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_23_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(536408832)))]; + tensor linear_214_cast_fp16 = linear(bias = encoder_layers_23_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_23_self_attn_linear_out_weight_to_fp16, x = input_1239_cast_fp16)[name = string("linear_214_cast_fp16")]; + tensor input_1243_cast_fp16 = add(x = input_1233_cast_fp16, y = linear_214_cast_fp16)[name = string("input_1243_cast_fp16")]; + tensor x_615_axes_0 = const()[name = string("x_615_axes_0"), val = tensor([-1])]; + tensor encoder_layers_23_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_23_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(536410944)))]; + tensor encoder_layers_23_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_23_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(536413056)))]; + tensor x_615_cast_fp16 = layer_norm(axes = x_615_axes_0, beta = encoder_layers_23_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_23_norm_conv_weight_to_fp16, x = input_1243_cast_fp16)[name = string("x_615_cast_fp16")]; + tensor input_1245_perm_0 = const()[name = string("input_1245_perm_0"), val = tensor([0, 2, 1])]; + string input_1247_pad_type_0 = const()[name = string("input_1247_pad_type_0"), val = string("valid")]; + tensor input_1247_strides_0 = const()[name = string("input_1247_strides_0"), val = tensor([1])]; + tensor input_1247_pad_0 = const()[name = string("input_1247_pad_0"), val = tensor([0, 0])]; + tensor input_1247_dilations_0 = const()[name = string("input_1247_dilations_0"), val = tensor([1])]; + int32 input_1247_groups_0 = const()[name = string("input_1247_groups_0"), val = int32(1)]; + tensor encoder_layers_23_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(536415168))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(538512384))))[name = string("encoder_layers_23_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_1245_cast_fp16 = transpose(perm = input_1245_perm_0, x = x_615_cast_fp16)[name = string("transpose_150")]; + tensor input_1247_cast_fp16 = conv(dilations = input_1247_dilations_0, groups = input_1247_groups_0, pad = input_1247_pad_0, pad_type = input_1247_pad_type_0, strides = input_1247_strides_0, weight = encoder_layers_23_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_1245_cast_fp16)[name = string("input_1247_cast_fp16")]; + int32 x_617_split_num_splits_0 = const()[name = string("x_617_split_num_splits_0"), val = int32(2)]; + int32 x_617_split_axis_0 = const()[name = string("x_617_split_axis_0"), val = int32(1)]; + tensor x_617_split_cast_fp16_0, tensor x_617_split_cast_fp16_1 = split(axis = x_617_split_axis_0, num_splits = x_617_split_num_splits_0, x = input_1247_cast_fp16)[name = string("x_617_split_cast_fp16")]; + tensor x_617_split_1_sigmoid_cast_fp16 = sigmoid(x = x_617_split_cast_fp16_1)[name = string("x_617_split_1_sigmoid_cast_fp16")]; + tensor x_617_cast_fp16 = mul(x = x_617_split_cast_fp16_0, y = x_617_split_1_sigmoid_cast_fp16)[name = string("x_617_cast_fp16")]; + tensor input_1249_cast_fp16 = select(a = var_44_to_fp16, b = x_617_cast_fp16, cond = var_575)[name = string("input_1249_cast_fp16")]; + bool new_x_interleave_0 = const()[name = string("new_x_interleave_0"), val = bool(false)]; + tensor new_x_cast_fp16 = concat(axis = var_59, interleave = new_x_interleave_0, values = (cache_cast_fp16, input_1249_cast_fp16))[name = string("new_x_cast_fp16")]; + tensor cache_last_time_cur_begin_0 = const()[name = string("cache_last_time_cur_begin_0"), val = tensor([0, 0, 28])]; + tensor cache_last_time_cur_end_0 = const()[name = string("cache_last_time_cur_end_0"), val = tensor([1, 1024, 36])]; + tensor cache_last_time_cur_end_mask_0 = const()[name = string("cache_last_time_cur_end_mask_0"), val = tensor([true, true, true])]; + tensor cache_last_time_cur_cast_fp16 = slice_by_index(begin = cache_last_time_cur_begin_0, end = cache_last_time_cur_end_0, end_mask = cache_last_time_cur_end_mask_0, x = new_x_cast_fp16)[name = string("cache_last_time_cur_cast_fp16")]; + string x_619_pad_type_0 = const()[name = string("x_619_pad_type_0"), val = string("valid")]; + int32 x_619_groups_0 = const()[name = string("x_619_groups_0"), val = int32(1024)]; + tensor x_619_strides_0 = const()[name = string("x_619_strides_0"), val = tensor([1])]; + tensor x_619_pad_0 = const()[name = string("x_619_pad_0"), val = tensor([0, 0])]; + tensor x_619_dilations_0 = const()[name = string("x_619_dilations_0"), val = tensor([1])]; + tensor encoder_layers_23_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(538516544))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(538525824))))[name = string("encoder_layers_23_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_619_cast_fp16 = conv(dilations = x_619_dilations_0, groups = x_619_groups_0, pad = x_619_pad_0, pad_type = x_619_pad_type_0, strides = x_619_strides_0, weight = encoder_layers_23_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_cast_fp16)[name = string("x_619_cast_fp16")]; + tensor input_1251_perm_0 = const()[name = string("input_1251_perm_0"), val = tensor([0, 2, 1])]; + tensor x_621_axes_0 = const()[name = string("x_621_axes_0"), val = tensor([-1])]; + tensor encoder_layers_23_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_23_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(538527936)))]; + tensor encoder_layers_23_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_23_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(538530048)))]; + tensor input_1251_cast_fp16 = transpose(perm = input_1251_perm_0, x = x_619_cast_fp16)[name = string("transpose_149")]; + tensor x_621_cast_fp16 = layer_norm(axes = x_621_axes_0, beta = encoder_layers_23_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_23_conv_batch_norm_weight_to_fp16, x = input_1251_cast_fp16)[name = string("x_621_cast_fp16")]; + tensor input_1253_perm_0 = const()[name = string("input_1253_perm_0"), val = tensor([0, 2, 1])]; + tensor input_1253_cast_fp16 = transpose(perm = input_1253_perm_0, x = x_621_cast_fp16)[name = string("transpose_148")]; + tensor input_1255_cast_fp16 = silu(x = input_1253_cast_fp16)[name = string("input_1255_cast_fp16")]; + string x_623_pad_type_0 = const()[name = string("x_623_pad_type_0"), val = string("valid")]; + tensor x_623_strides_0 = const()[name = string("x_623_strides_0"), val = tensor([1])]; + tensor x_623_pad_0 = const()[name = string("x_623_pad_0"), val = tensor([0, 0])]; + tensor x_623_dilations_0 = const()[name = string("x_623_dilations_0"), val = tensor([1])]; + int32 x_623_groups_0 = const()[name = string("x_623_groups_0"), val = int32(1)]; + tensor encoder_layers_23_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(538532160))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(539580800))))[name = string("encoder_layers_23_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_623_cast_fp16 = conv(dilations = x_623_dilations_0, groups = x_623_groups_0, pad = x_623_pad_0, pad_type = x_623_pad_type_0, strides = x_623_strides_0, weight = encoder_layers_23_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_1255_cast_fp16)[name = string("x_623_cast_fp16")]; + tensor input_1257_perm_0 = const()[name = string("input_1257_perm_0"), val = tensor([0, 2, 1])]; + tensor input_1257_cast_fp16 = transpose(perm = input_1257_perm_0, x = x_623_cast_fp16)[name = string("transpose_147")]; + tensor input_1259_cast_fp16 = add(x = input_1243_cast_fp16, y = input_1257_cast_fp16)[name = string("input_1259_cast_fp16")]; + tensor input_1261_axes_0 = const()[name = string("input_1261_axes_0"), val = tensor([-1])]; + tensor encoder_layers_23_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_23_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(539582912)))]; + tensor encoder_layers_23_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_23_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(539585024)))]; + tensor input_1261_cast_fp16 = layer_norm(axes = input_1261_axes_0, beta = encoder_layers_23_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_23_norm_feed_forward2_weight_to_fp16, x = input_1259_cast_fp16)[name = string("input_1261_cast_fp16")]; + tensor encoder_layers_23_feed_forward2_linear1_weight_to_fp16 = const()[name = string("encoder_layers_23_feed_forward2_linear1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(539587136)))]; + tensor encoder_layers_23_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_23_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(547975808)))]; + tensor linear_215_cast_fp16 = linear(bias = encoder_layers_23_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_23_feed_forward2_linear1_weight_to_fp16, x = input_1261_cast_fp16)[name = string("linear_215_cast_fp16")]; + tensor input_1265_cast_fp16 = silu(x = linear_215_cast_fp16)[name = string("input_1265_cast_fp16")]; + tensor encoder_layers_23_feed_forward2_linear2_weight_to_fp16 = const()[name = string("encoder_layers_23_feed_forward2_linear2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(547984064)))]; + tensor encoder_layers_23_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_23_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(556372736)))]; + tensor linear_216_cast_fp16 = linear(bias = encoder_layers_23_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_23_feed_forward2_linear2_weight_to_fp16, x = input_1265_cast_fp16)[name = string("linear_216_cast_fp16")]; + fp16 var_5530_to_fp16 = const()[name = string("op_5530_to_fp16"), val = fp16(0x1p-1)]; + tensor var_5531_cast_fp16 = mul(x = linear_216_cast_fp16, y = var_5530_to_fp16)[name = string("op_5531_cast_fp16")]; + tensor input_1271_cast_fp16 = add(x = input_1259_cast_fp16, y = var_5531_cast_fp16)[name = string("input_1271_cast_fp16")]; + tensor audio_signal_axes_0 = const()[name = string("audio_signal_axes_0"), val = tensor([-1])]; + tensor encoder_layers_23_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_23_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(556374848)))]; + tensor encoder_layers_23_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_23_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(556376960)))]; + tensor audio_signal_cast_fp16 = layer_norm(axes = audio_signal_axes_0, beta = encoder_layers_23_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_23_norm_out_weight_to_fp16, x = input_1271_cast_fp16)[name = string("audio_signal_cast_fp16")]; + int32 obj_5_axis_0 = const()[name = string("obj_5_axis_0"), val = int32(0)]; + tensor obj_5_cast_fp16 = stack(axis = obj_5_axis_0, values = (var_484_cast_fp16, var_697_cast_fp16, var_910_cast_fp16, var_1123_cast_fp16, var_1336_cast_fp16, var_1549_cast_fp16, var_1762_cast_fp16, var_1975_cast_fp16, var_2188_cast_fp16, var_2401_cast_fp16, var_2614_cast_fp16, var_2827_cast_fp16, var_3040_cast_fp16, var_3253_cast_fp16, var_3466_cast_fp16, var_3679_cast_fp16, var_3892_cast_fp16, var_4105_cast_fp16, var_4318_cast_fp16, var_4531_cast_fp16, var_4744_cast_fp16, var_4957_cast_fp16, var_5170_cast_fp16, cache_last_channel_cur_cast_fp16))[name = string("obj_5_cast_fp16")]; + int32 obj_7_axis_0 = const()[name = string("obj_7_axis_0"), val = int32(0)]; + tensor obj_7_cast_fp16 = stack(axis = obj_7_axis_0, values = (var_588_cast_fp16, var_801_cast_fp16, var_1014_cast_fp16, var_1227_cast_fp16, var_1440_cast_fp16, var_1653_cast_fp16, var_1866_cast_fp16, var_2079_cast_fp16, var_2292_cast_fp16, var_2505_cast_fp16, var_2718_cast_fp16, var_2931_cast_fp16, var_3144_cast_fp16, var_3357_cast_fp16, var_3570_cast_fp16, var_3783_cast_fp16, var_3996_cast_fp16, var_4209_cast_fp16, var_4422_cast_fp16, var_4635_cast_fp16, var_4848_cast_fp16, var_5061_cast_fp16, var_5274_cast_fp16, cache_last_time_cur_cast_fp16))[name = string("obj_7_cast_fp16")]; + tensor var_5547 = add(x = cache_len, y = max_audio_length_1)[name = string("op_5547")]; + string var_5547_promoted_to_fp16_dtype_0 = const()[name = string("op_5547_promoted_to_fp16_dtype_0"), val = string("fp16")]; + fp16 const_384_to_fp16 = const()[name = string("const_384_to_fp16"), val = fp16(-inf)]; + fp16 var_49_promoted_to_fp16 = const()[name = string("op_49_promoted_to_fp16"), val = fp16(0x1.5p+5)]; + tensor var_5547_to_fp16 = cast(dtype = var_5547_promoted_to_fp16_dtype_0, x = var_5547)[name = string("cast_9")]; + tensor clip_1_cast_fp16 = clip(alpha = const_384_to_fp16, beta = var_49_promoted_to_fp16, x = var_5547_to_fp16)[name = string("clip_1_cast_fp16")]; + int32 one_hot_1_batch_dims_0 = const()[name = string("one_hot_1_batch_dims_0"), val = int32(0)]; + bool one_hot_1_validate_indices_0 = const()[name = string("one_hot_1_validate_indices_0"), val = bool(false)]; + tensor to_onehot_identity_table_to_fp16 = const()[name = string("to_onehot_identity_table_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(556379072)))]; + string prompt_id_to_int16_dtype_0 = const()[name = string("prompt_id_to_int16_dtype_0"), val = string("int16")]; + string cast_230_dtype_0 = const()[name = string("cast_230_dtype_0"), val = string("int32")]; + int32 greater_equal_0_y_0 = const()[name = string("greater_equal_0_y_0"), val = int32(0)]; + tensor prompt_id_to_int16 = cast(dtype = prompt_id_to_int16_dtype_0, x = prompt_id)[name = string("cast_8")]; + tensor cast_230 = cast(dtype = cast_230_dtype_0, x = prompt_id_to_int16)[name = string("cast_7")]; + tensor greater_equal_0 = greater_equal(x = cast_230, y = greater_equal_0_y_0)[name = string("greater_equal_0")]; + int32 slice_by_index_2 = const()[name = string("slice_by_index_2"), val = int32(128)]; + tensor add_0 = add(x = cast_230, y = slice_by_index_2)[name = string("add_0")]; + tensor select_0 = select(a = cast_230, b = add_0, cond = greater_equal_0)[name = string("select_0")]; + string select_0_to_int16_dtype_0 = const()[name = string("select_0_to_int16_dtype_0"), val = string("int16")]; + string cast_0_dtype_0 = const()[name = string("cast_0_dtype_0"), val = string("int32")]; + int32 greater_equal_0_y_0_1 = const()[name = string("greater_equal_0_y_0_1"), val = int32(0)]; + tensor select_0_to_int16 = cast(dtype = select_0_to_int16_dtype_0, x = select_0)[name = string("cast_6")]; + tensor cast_0 = cast(dtype = cast_0_dtype_0, x = select_0_to_int16)[name = string("cast_5")]; + tensor greater_equal_0_1 = greater_equal(x = cast_0, y = greater_equal_0_y_0_1)[name = string("greater_equal_0_1")]; + int32 slice_by_index_0 = const()[name = string("slice_by_index_0"), val = int32(128)]; + tensor add_0_1 = add(x = cast_0, y = slice_by_index_0)[name = string("add_0_1")]; + tensor select_0_1 = select(a = cast_0, b = add_0_1, cond = greater_equal_0_1)[name = string("select_0_1")]; + int32 greater_equal_0_y_0_2 = const()[name = string("greater_equal_0_y_0_2"), val = int32(0)]; + tensor greater_equal_0_2 = greater_equal(x = select_0_1, y = greater_equal_0_y_0_2)[name = string("greater_equal_0_2")]; + int32 slice_by_index_0_1 = const()[name = string("slice_by_index_0_1"), val = int32(128)]; + tensor add_0_2 = add(x = select_0_1, y = slice_by_index_0_1)[name = string("add_0_2")]; + tensor select_0_2 = select(a = select_0_1, b = add_0_2, cond = greater_equal_0_2)[name = string("select_0_2")]; + int32 one_hot_1_cast_fp16_cast_uint16_cast_uint16_axis_0 = const()[name = string("one_hot_1_cast_fp16_cast_uint16_cast_uint16_axis_0"), val = int32(0)]; + tensor one_hot_1_cast_fp16_cast_uint16_cast_uint16 = gather(axis = one_hot_1_cast_fp16_cast_uint16_cast_uint16_axis_0, batch_dims = one_hot_1_batch_dims_0, indices = select_0_2, validate_indices = one_hot_1_validate_indices_0, x = to_onehot_identity_table_to_fp16)[name = string("one_hot_1_cast_fp16_cast_uint16_cast_uint16")]; + tensor var_5593_axes_0 = const()[name = string("op_5593_axes_0"), val = tensor([1])]; + tensor var_5593_cast_fp16 = expand_dims(axes = var_5593_axes_0, x = one_hot_1_cast_fp16_cast_uint16_cast_uint16)[name = string("op_5593_cast_fp16")]; + tensor one_hot_reps_0 = const()[name = string("one_hot_reps_0"), val = tensor([1, 28, 1])]; + tensor one_hot_cast_fp16 = tile(reps = one_hot_reps_0, x = var_5593_cast_fp16)[name = string("one_hot_cast_fp16")]; + int32 var_5602 = const()[name = string("op_5602"), val = int32(-1)]; + bool input_1273_interleave_0 = const()[name = string("input_1273_interleave_0"), val = bool(false)]; + tensor input_1273_cast_fp16 = concat(axis = var_5602, interleave = input_1273_interleave_0, values = (audio_signal_cast_fp16, one_hot_cast_fp16))[name = string("input_1273_cast_fp16")]; + tensor prompt_kernel_0_weight_to_fp16 = const()[name = string("prompt_kernel_0_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(556411904)))]; + tensor prompt_kernel_0_bias_to_fp16 = const()[name = string("prompt_kernel_0_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(561130560)))]; + tensor linear_217_cast_fp16 = linear(bias = prompt_kernel_0_bias_to_fp16, weight = prompt_kernel_0_weight_to_fp16, x = input_1273_cast_fp16)[name = string("linear_217_cast_fp16")]; + tensor input_cast_fp16 = relu(x = linear_217_cast_fp16)[name = string("input_cast_fp16")]; + tensor prompt_kernel_2_weight_to_fp16 = const()[name = string("prompt_kernel_2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(561134720)))]; + tensor prompt_kernel_2_bias_to_fp16 = const()[name = string("prompt_kernel_2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(565329088)))]; + tensor linear_218_cast_fp16 = linear(bias = prompt_kernel_2_bias_to_fp16, weight = prompt_kernel_2_weight_to_fp16, x = input_cast_fp16)[name = string("linear_218_cast_fp16")]; + tensor var_5615_perm_0 = const()[name = string("op_5615_perm_0"), val = tensor([0, 2, 1])]; + string var_5615_cast_fp16_to_fp32_dtype_0 = const()[name = string("op_5615_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + string var_5620_dtype_0 = const()[name = string("op_5620_dtype_0"), val = string("int32")]; + tensor var_5623_perm_0 = const()[name = string("op_5623_perm_0"), val = tensor([1, 0, 2, 3])]; + string var_5623_cast_fp16_to_fp32_dtype_0 = const()[name = string("op_5623_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + tensor var_5626_perm_0 = const()[name = string("op_5626_perm_0"), val = tensor([1, 0, 2, 3])]; + string var_5626_cast_fp16_to_fp32_dtype_0 = const()[name = string("op_5626_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + string var_5631_dtype_0 = const()[name = string("op_5631_dtype_0"), val = string("int32")]; + tensor cache_len_out = cast(dtype = var_5631_dtype_0, x = clip_1_cast_fp16)[name = string("cast_0")]; + tensor var_5626_cast_fp16 = transpose(perm = var_5626_perm_0, x = obj_7_cast_fp16)[name = string("transpose_144")]; + tensor cache_time_out = cast(dtype = var_5626_cast_fp16_to_fp32_dtype_0, x = var_5626_cast_fp16)[name = string("cast_1")]; + tensor var_5623_cast_fp16 = transpose(perm = var_5623_perm_0, x = obj_5_cast_fp16)[name = string("transpose_145")]; + tensor cache_channel_out = cast(dtype = var_5623_cast_fp16_to_fp32_dtype_0, x = var_5623_cast_fp16)[name = string("cast_2")]; + tensor encoded_length = cast(dtype = var_5620_dtype_0, x = clip_0_cast_fp16)[name = string("cast_3")]; + tensor var_5615_cast_fp16 = transpose(perm = var_5615_perm_0, x = linear_218_cast_fp16)[name = string("transpose_146")]; + tensor encoded = cast(dtype = var_5615_cast_fp16_to_fp32_dtype_0, x = var_5615_cast_fp16)[name = string("cast_4")]; + } -> (encoded, encoded_length, cache_channel_out, cache_time_out, cache_len_out); +} \ No newline at end of file diff --git a/it/2240ms/encoder.mlmodelc/weights/weight.bin b/it/2240ms/encoder.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..7a12bc272fe95f2080e5cebe63d7189329748a69 --- /dev/null +++ b/it/2240ms/encoder.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:728cf78b85d85c573a2a33a8287ea44bc7a846fc8462d9038d1ed1b24a2c9ac8 +size 565331200 diff --git a/it/2240ms/encoder.mlpackage/Data/com.apple.CoreML/model.mlmodel b/it/2240ms/encoder.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..808c1ca295fe918d2a2372f1e2fadb3bc18930d1 --- /dev/null +++ b/it/2240ms/encoder.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a0653e7c223390a9fe23c5e6f6266b97488e294bbc759705aaa66c68bb15ff8f +size 805589 diff --git a/it/2240ms/encoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/it/2240ms/encoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..7a12bc272fe95f2080e5cebe63d7189329748a69 --- /dev/null +++ b/it/2240ms/encoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:728cf78b85d85c573a2a33a8287ea44bc7a846fc8462d9038d1ed1b24a2c9ac8 +size 565331200 diff --git a/it/2240ms/encoder.mlpackage/Manifest.json b/it/2240ms/encoder.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..955535c4db53bb00e2d26147a70036cd887e0451 --- /dev/null +++ b/it/2240ms/encoder.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "00BBBFAC-20B6-427A-B978-654939D77446": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + }, + "5707DA36-A222-47FA-8727-5AED43CBE002": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + } + }, + "rootModelIdentifier": "5707DA36-A222-47FA-8727-5AED43CBE002" +} diff --git a/it/2240ms/joint.mlmodelc/analytics/coremldata.bin b/it/2240ms/joint.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..ee72f6da06176879d781b1a46f441fa01048d662 --- /dev/null +++ b/it/2240ms/joint.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cae47ef3b8cb372625305f4cdd31159b5ce56be1470c94613ff946be342f6d7a +size 243 diff --git a/it/2240ms/joint.mlmodelc/coremldata.bin b/it/2240ms/joint.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..07f143747ee2f43103809647f4058203bf60dc56 --- /dev/null +++ b/it/2240ms/joint.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:215d8dd2e33da0c37f08e6b0c0a1e997a3e056f2cb9113fdcf17f8027a61216d +size 341 diff --git a/it/2240ms/joint.mlmodelc/model.mil b/it/2240ms/joint.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..bfde40ec94bf61746424d2d3e196a4fba198de2d --- /dev/null +++ b/it/2240ms/joint.mlmodelc/model.mil @@ -0,0 +1,31 @@ +program(1.3) +[buildInfo = dict({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.5.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.3.0"}})] +{ + func main(tensor decoder, tensor encoder) { + tensor input_1_perm_0 = const()[name = string("input_1_perm_0"), val = tensor([0, 2, 1])]; + string encoder_to_fp16_dtype_0 = const()[name = string("encoder_to_fp16_dtype_0"), val = string("fp16")]; + tensor input_3_perm_0 = const()[name = string("input_3_perm_0"), val = tensor([0, 2, 1])]; + string decoder_to_fp16_dtype_0 = const()[name = string("decoder_to_fp16_dtype_0"), val = string("fp16")]; + tensor module_enc_weight_to_fp16 = const()[name = string("module_enc_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))]; + tensor module_enc_bias_to_fp16 = const()[name = string("module_enc_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1310848)))]; + tensor encoder_to_fp16 = cast(dtype = encoder_to_fp16_dtype_0, x = encoder)[name = string("cast_2")]; + tensor input_1_cast_fp16 = transpose(perm = input_1_perm_0, x = encoder_to_fp16)[name = string("transpose_1")]; + tensor linear_0_cast_fp16 = linear(bias = module_enc_bias_to_fp16, weight = module_enc_weight_to_fp16, x = input_1_cast_fp16)[name = string("linear_0_cast_fp16")]; + tensor module_pred_weight_to_fp16 = const()[name = string("module_pred_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1312192)))]; + tensor module_pred_bias_to_fp16 = const()[name = string("module_pred_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2131456)))]; + tensor decoder_to_fp16 = cast(dtype = decoder_to_fp16_dtype_0, x = decoder)[name = string("cast_1")]; + tensor input_3_cast_fp16 = transpose(perm = input_3_perm_0, x = decoder_to_fp16)[name = string("transpose_0")]; + tensor linear_1_cast_fp16 = linear(bias = module_pred_bias_to_fp16, weight = module_pred_weight_to_fp16, x = input_3_cast_fp16)[name = string("linear_1_cast_fp16")]; + tensor var_23_axes_0 = const()[name = string("op_23_axes_0"), val = tensor([2])]; + tensor var_23_cast_fp16 = expand_dims(axes = var_23_axes_0, x = linear_0_cast_fp16)[name = string("op_23_cast_fp16")]; + tensor var_25_axes_0 = const()[name = string("op_25_axes_0"), val = tensor([1])]; + tensor var_25_cast_fp16 = expand_dims(axes = var_25_axes_0, x = linear_1_cast_fp16)[name = string("op_25_cast_fp16")]; + tensor input_5_cast_fp16 = add(x = var_23_cast_fp16, y = var_25_cast_fp16)[name = string("input_5_cast_fp16")]; + tensor input_7_cast_fp16 = relu(x = input_5_cast_fp16)[name = string("input_7_cast_fp16")]; + tensor module_joint_net_2_weight_to_fp16 = const()[name = string("module_joint_net_2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2132800)))]; + tensor module_joint_net_2_bias_to_fp16 = const()[name = string("module_joint_net_2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3164544)))]; + tensor linear_2_cast_fp16 = linear(bias = module_joint_net_2_bias_to_fp16, weight = module_joint_net_2_weight_to_fp16, x = input_7_cast_fp16)[name = string("linear_2_cast_fp16")]; + string linear_2_cast_fp16_to_fp32_dtype_0 = const()[name = string("linear_2_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + tensor logits = cast(dtype = linear_2_cast_fp16_to_fp32_dtype_0, x = linear_2_cast_fp16)[name = string("cast_0")]; + } -> (logits); +} \ No newline at end of file diff --git a/it/2240ms/joint.mlmodelc/weights/weight.bin b/it/2240ms/joint.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..3673a1d2a1060ecda8626c76ffacdfd89c5f00c1 --- /dev/null +++ b/it/2240ms/joint.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1740c99cbe34ebeaa0163c8421135b4586df09960ef07fe02abb2a94b5693411 +size 3166220 diff --git a/it/2240ms/joint.mlpackage/Data/com.apple.CoreML/model.mlmodel b/it/2240ms/joint.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..f7f468ddd814131e36b8af9ed7a3358576bffcf0 --- /dev/null +++ b/it/2240ms/joint.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d902ae83b8a93f0f4240a4a6939466dbd1a6b2291f1615d81d7ac26d9115bc23 +size 4486 diff --git a/it/2240ms/joint.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/it/2240ms/joint.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..3673a1d2a1060ecda8626c76ffacdfd89c5f00c1 --- /dev/null +++ b/it/2240ms/joint.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1740c99cbe34ebeaa0163c8421135b4586df09960ef07fe02abb2a94b5693411 +size 3166220 diff --git a/it/2240ms/joint.mlpackage/Manifest.json b/it/2240ms/joint.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..fea9f1ee7eee62ace96d28134fe38a74b32b40c9 --- /dev/null +++ b/it/2240ms/joint.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "6D4EAD3B-A17D-4807-8572-3E0A513C0C7E": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + }, + "8242AF76-3606-4A98-8B3D-E0BB730201A4": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + } + }, + "rootModelIdentifier": "6D4EAD3B-A17D-4807-8572-3E0A513C0C7E" +} diff --git a/it/2240ms/joint_noencproj_batched.mlmodelc/analytics/coremldata.bin b/it/2240ms/joint_noencproj_batched.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..d2ffa5c9579dd51ce490b9c8fe47e3c27a05c723 --- /dev/null +++ b/it/2240ms/joint_noencproj_batched.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:24510a0ca90eea355a0e446ba54d63004a3f2c51d30eb3e415e4af938ac1acba +size 243 diff --git a/it/2240ms/joint_noencproj_batched.mlmodelc/coremldata.bin b/it/2240ms/joint_noencproj_batched.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..03999c10ebcacf3633f026f01254ed98ae34fa25 --- /dev/null +++ b/it/2240ms/joint_noencproj_batched.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dc93bb8f241754f1359837f597d600b029d687ce0c57a4280fa586f8c8386337 +size 406 diff --git a/it/2240ms/joint_noencproj_batched.mlmodelc/model.mil b/it/2240ms/joint_noencproj_batched.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..58942374262618031d39a52f7a009b81c7f24c24 --- /dev/null +++ b/it/2240ms/joint_noencproj_batched.mlmodelc/model.mil @@ -0,0 +1,26 @@ +program(1.3) +[buildInfo = dict({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.10.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor decoder, tensor encoder_proj) { + tensor input_1_perm_0 = const()[name = string("input_1_perm_0"), val = tensor([0, 2, 1])]; + string decoder_to_fp16_dtype_0 = const()[name = string("decoder_to_fp16_dtype_0"), val = string("fp16")]; + tensor joint_module_pred_weight_to_fp16 = const()[name = string("joint_module_pred_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))]; + tensor joint_module_pred_bias_to_fp16 = const()[name = string("joint_module_pred_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(819328)))]; + tensor decoder_to_fp16 = cast(dtype = decoder_to_fp16_dtype_0, x = decoder)[name = string("cast_2")]; + tensor input_1_cast_fp16 = transpose(perm = input_1_perm_0, x = decoder_to_fp16)[name = string("transpose_0")]; + tensor linear_0_cast_fp16 = linear(bias = joint_module_pred_bias_to_fp16, weight = joint_module_pred_weight_to_fp16, x = input_1_cast_fp16)[name = string("linear_0_cast_fp16")]; + tensor var_15_axes_0 = const()[name = string("op_15_axes_0"), val = tensor([2])]; + string encoder_proj_to_fp16_dtype_0 = const()[name = string("encoder_proj_to_fp16_dtype_0"), val = string("fp16")]; + tensor encoder_proj_to_fp16 = cast(dtype = encoder_proj_to_fp16_dtype_0, x = encoder_proj)[name = string("cast_1")]; + tensor var_15_cast_fp16 = expand_dims(axes = var_15_axes_0, x = encoder_proj_to_fp16)[name = string("op_15_cast_fp16")]; + tensor var_17_axes_0 = const()[name = string("op_17_axes_0"), val = tensor([1])]; + tensor var_17_cast_fp16 = expand_dims(axes = var_17_axes_0, x = linear_0_cast_fp16)[name = string("op_17_cast_fp16")]; + tensor input_3_cast_fp16 = add(x = var_15_cast_fp16, y = var_17_cast_fp16)[name = string("input_3_cast_fp16")]; + tensor input_5_cast_fp16 = relu(x = input_3_cast_fp16)[name = string("input_5_cast_fp16")]; + tensor joint_module_joint_net_2_weight_to_fp16 = const()[name = string("joint_module_joint_net_2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(820672)))]; + tensor joint_module_joint_net_2_bias_to_fp16 = const()[name = string("joint_module_joint_net_2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1852416)))]; + tensor linear_1_cast_fp16 = linear(bias = joint_module_joint_net_2_bias_to_fp16, weight = joint_module_joint_net_2_weight_to_fp16, x = input_5_cast_fp16)[name = string("linear_1_cast_fp16")]; + string linear_1_cast_fp16_to_fp32_dtype_0 = const()[name = string("linear_1_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + tensor logits = cast(dtype = linear_1_cast_fp16_to_fp32_dtype_0, x = linear_1_cast_fp16)[name = string("cast_0")]; + } -> (logits); +} \ No newline at end of file diff --git a/it/2240ms/joint_noencproj_batched.mlmodelc/weights/weight.bin b/it/2240ms/joint_noencproj_batched.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..40ceadd152241059aa378e2ddb6cc9f649e0b59c --- /dev/null +++ b/it/2240ms/joint_noencproj_batched.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cd83e82dcfec315f28c8a8872b0d7f22e668a2c485821de86a0379ae2b3864ad +size 1854092 diff --git a/it/2240ms/joint_noencproj_batched.mlpackage/Data/com.apple.CoreML/model.mlmodel b/it/2240ms/joint_noencproj_batched.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..49f1fd816a2725ab32d473dfe600f911656f675b --- /dev/null +++ b/it/2240ms/joint_noencproj_batched.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:26a19f2e269811070e57eda978e8226ca9dcca6177b46b95f916538b5f815716 +size 3839 diff --git a/it/2240ms/joint_noencproj_batched.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/it/2240ms/joint_noencproj_batched.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..40ceadd152241059aa378e2ddb6cc9f649e0b59c --- /dev/null +++ b/it/2240ms/joint_noencproj_batched.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cd83e82dcfec315f28c8a8872b0d7f22e668a2c485821de86a0379ae2b3864ad +size 1854092 diff --git a/it/2240ms/joint_noencproj_batched.mlpackage/Manifest.json b/it/2240ms/joint_noencproj_batched.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..e3857a03022f3cec16823ed1d8a741fee56cfce4 --- /dev/null +++ b/it/2240ms/joint_noencproj_batched.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "4233CE8E-FB95-4FF9-BCD8-2A834D55C580": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + }, + "96E0F26C-90DC-49EE-B510-D0FB3FC812CC": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + } + }, + "rootModelIdentifier": "96E0F26C-90DC-49EE-B510-D0FB3FC812CC" +} diff --git a/it/2240ms/metadata.json b/it/2240ms/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..e7c46d827d13aabb635263d1632a6cdfb10f9107 --- /dev/null +++ b/it/2240ms/metadata.json @@ -0,0 +1,199 @@ +{ + "model": "nvidia/nemotron-asr-streaming-multilingual-0.6b", + "model_class": "nemo.collections.asr.models.rnnt_bpe_models_prompt.EncDecRNNTBPEModelWithPrompt", + "sample_rate": 16000, + "mel_features": 128, + "chunk_mel_frames": 224, + "pre_encode_cache": 9, + "total_mel_frames": 233, + "att_context_size": [ + 42, + 13 + ], + "vocab_size": 805, + "blank_idx": 805, + "vocab_pruned": true, + "vocab_pruned_original_size": 13087, + "cache_channel_shape": [ + 1, + 24, + 42, + 1024 + ], + "cache_time_shape": [ + 1, + 24, + 1024, + 8 + ], + "decoder_hidden": 640, + "decoder_layers": 2, + "encoder_dim": 1024, + "num_prompts": 128, + "prompt_dictionary": { + "af-ZA": 54, + "am-ET": 49, + "ar": 7, + "ar-AR": 7, + "auto": 101, + "ay-BO": 81, + "az-AZ": 66, + "bg": 30, + "bg-BG": 30, + "bn-IN": 36, + "cs": 22, + "cs-CZ": 22, + "da": 25, + "da-DK": 25, + "de": 9, + "de-DE": 9, + "el": 21, + "el-GR": 21, + "en": 0, + "en-GB": 1, + "en-US": 0, + "enGB": 1, + "es": 3, + "es-ES": 2, + "es-US": 3, + "esES": 2, + "et": 60, + "et-EE": 60, + "fa-IR": 38, + "fi": 26, + "fi-FI": 26, + "fr": 8, + "fr-CA": 100, + "fr-FR": 8, + "gn-PY": 82, + "gu-IN": 42, + "ha-NG": 50, + "haw-US": 97, + "he-IL": 64, + "hi": 6, + "hi-HI": 6, + "hi-IN": 6, + "hr": 29, + "hr-HR": 29, + "hu": 23, + "hu-HU": 23, + "hy-AM": 68, + "id-ID": 34, + "ig-NG": 53, + "it": 15, + "it-IT": 15, + "ja-JA": 10, + "ja-JP": 10, + "ka-GE": 67, + "km-KH": 47, + "kn-IN": 43, + "ko": 14, + "ko-KO": 14, + "ko-KR": 14, + "ku-TR": 65, + "ky-KG": 71, + "ln-CD": 58, + "lt": 31, + "lt-LT": 31, + "lv": 61, + "lv-LV": 61, + "mi-NZ": 96, + "ml-IN": 44, + "mr-IN": 41, + "ms-MY": 35, + "mt-MT": 102, + "nah-MX": 83, + "nb": 103, + "nb-NO": 103, + "ne-NP": 46, + "nl": 16, + "nl-NL": 16, + "nn": 104, + "nn-NO": 104, + "no": 27, + "no-NO": 27, + "ny-MW": 57, + "or-KE": 59, + "pl": 17, + "pl-PL": 17, + "pt": 13, + "pt-BR": 12, + "pt-PT": 13, + "qu-PE": 80, + "ro": 20, + "ro-RO": 20, + "ru": 11, + "ru-RU": 11, + "rw-RW": 55, + "si-LK": 45, + "sk": 28, + "sk-SK": 28, + "sl": 62, + "sl-SI": 62, + "sm-WS": 98, + "so-SO": 56, + "sv": 24, + "sv-SE": 24, + "sw-KE": 48, + "ta-IN": 39, + "te-IN": 40, + "tg-TJ": 70, + "th-TH": 32, + "to-TO": 99, + "tr": 18, + "tr-TR": 18, + "uk": 19, + "uk-UA": 19, + "ur-PK": 37, + "uz-UZ": 69, + "vi-VN": 33, + "yo-NG": 52, + "zh-CN": 4, + "zh-TW": 5, + "zh-ZH": 4, + "zu-ZA": 51 + }, + "default_prompt_id": 101, + "lang_tag_token_ids": [ + 1, + 63, + 115, + 167, + 226, + 227, + 259, + 276, + 328, + 353, + 368, + 462, + 481, + 499, + 518, + 542, + 571, + 602, + 603, + 612, + 624, + 646, + 647, + 667, + 689, + 699, + 720, + 727, + 747, + 748, + 750, + 751, + 752, + 756, + 774, + 787, + 788, + 801, + 802 + ], + "chunk_ms": 2240 +} \ No newline at end of file diff --git a/it/2240ms/preprocessor.mlmodelc/analytics/coremldata.bin b/it/2240ms/preprocessor.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..d4b1561ab413a9d87db506bc842f077779dcbded --- /dev/null +++ b/it/2240ms/preprocessor.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3b8033bec5ee01649f325b8f4c5aeef1b31c99b469ce56d46039c1b73f09585d +size 243 diff --git a/it/2240ms/preprocessor.mlmodelc/coremldata.bin b/it/2240ms/preprocessor.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..7ca85ca2509c42944d27f8afdfa784313a8a5791 --- /dev/null +++ b/it/2240ms/preprocessor.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6364156582bef7b1bfda73ee29fc10fdfa794f89d016a968d6f92aca74ccfb0f +size 431 diff --git a/it/2240ms/preprocessor.mlmodelc/model.mil b/it/2240ms/preprocessor.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..b1a0b2b9193c992de42e51245fc1ef433d345afc --- /dev/null +++ b/it/2240ms/preprocessor.mlmodelc/model.mil @@ -0,0 +1,122 @@ +program(1.3) +[buildInfo = dict({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.10.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor audio, tensor audio_length) [FlexibleShapeInformation = tuple>>, tuple, ?>>>>((("DefaultShapes", {{"audio", [1, 1]}}), ("RangeDims", {{"audio", [[1, 1], [1, 1280000]]}})))] { + int32 var_9 = const()[name = string("op_9"), val = int32(1)]; + int32 var_10 = const()[name = string("op_10"), val = int32(160)]; + int32 var_12 = const()[name = string("op_12"), val = int32(0)]; + int32 var_33 = const()[name = string("op_33"), val = int32(512)]; + tensor var_34 = add(x = audio_length, y = var_33)[name = string("op_34")]; + int32 var_35 = const()[name = string("op_35"), val = int32(512)]; + tensor var_36 = sub(x = var_34, y = var_35)[name = string("op_36")]; + tensor floor_div_0 = floor_div(x = var_36, y = var_10)[name = string("floor_div_0")]; + tensor var_39 = equal(x = audio_length, y = var_12)[name = string("op_39")]; + tensor var_40 = const()[name = string("op_40"), val = tensor([0])]; + tensor mel_length = select(a = var_40, b = floor_div_0, cond = var_39)[name = string("seq_len")]; + string audio_to_fp16_dtype_0 = const()[name = string("audio_to_fp16_dtype_0"), val = string("fp16")]; + tensor audio_to_fp16 = cast(dtype = audio_to_fp16_dtype_0, x = audio)[name = string("cast_10")]; + tensor var_42_shape_cast_fp16 = shape(x = audio_to_fp16)[name = string("op_42_shape_cast_fp16")]; + int32 gather_0_axis_0 = const()[name = string("gather_0_axis_0"), val = int32(0)]; + int32 gather_0_batch_dims_0 = const()[name = string("gather_0_batch_dims_0"), val = int32(0)]; + bool gather_0_validate_indices_0 = const()[name = string("gather_0_validate_indices_0"), val = bool(false)]; + string var_42_shape_cast_fp16_to_int16_dtype_0 = const()[name = string("op_42_shape_cast_fp16_to_int16_dtype_0"), val = string("int16")]; + uint16 gather_0_indices_0_to_uint16 = const()[name = string("gather_0_indices_0_to_uint16"), val = uint16(1)]; + tensor var_42_shape_cast_fp16_to_int16 = cast(dtype = var_42_shape_cast_fp16_to_int16_dtype_0, x = var_42_shape_cast_fp16)[name = string("cast_9")]; + int16 gather_0_cast_uint16 = gather(axis = gather_0_axis_0, batch_dims = gather_0_batch_dims_0, indices = gather_0_indices_0_to_uint16, validate_indices = gather_0_validate_indices_0, x = var_42_shape_cast_fp16_to_int16)[name = string("gather_0_cast_uint16")]; + string gather_0_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_0_cast_uint16_to_int32_dtype_0"), val = string("int32")]; + int32 const_0 = const()[name = string("const_0"), val = int32(0)]; + int32 const_1 = const()[name = string("const_1"), val = int32(1)]; + int32 gather_0_cast_uint16_to_int32 = cast(dtype = gather_0_cast_uint16_to_int32_dtype_0, x = gather_0_cast_uint16)[name = string("cast_8")]; + tensor var_43 = range_1d(end = gather_0_cast_uint16_to_int32, start = const_0, step = const_1)[name = string("op_43")]; + tensor var_44_axes_0 = const()[name = string("op_44_axes_0"), val = tensor([0])]; + tensor var_44 = expand_dims(axes = var_44_axes_0, x = var_43)[name = string("op_44")]; + tensor var_45_axes_0 = const()[name = string("op_45_axes_0"), val = tensor([1])]; + tensor var_45 = expand_dims(axes = var_45_axes_0, x = audio_length)[name = string("op_45")]; + tensor timemask = less(x = var_44, y = var_45)[name = string("timemask")]; + tensor var_48_begin_0 = const()[name = string("op_48_begin_0"), val = tensor([0, 0])]; + tensor var_48_end_0 = const()[name = string("op_48_end_0"), val = tensor([1, 1])]; + tensor var_48_end_mask_0 = const()[name = string("op_48_end_mask_0"), val = tensor([true, false])]; + tensor var_48_squeeze_mask_0 = const()[name = string("op_48_squeeze_mask_0"), val = tensor([false, true])]; + tensor var_48_cast_fp16 = slice_by_index(begin = var_48_begin_0, end = var_48_end_0, end_mask = var_48_end_mask_0, squeeze_mask = var_48_squeeze_mask_0, x = audio_to_fp16)[name = string("op_48_cast_fp16")]; + tensor var_49_axes_0 = const()[name = string("op_49_axes_0"), val = tensor([1])]; + tensor var_49_cast_fp16 = expand_dims(axes = var_49_axes_0, x = var_48_cast_fp16)[name = string("op_49_cast_fp16")]; + tensor var_51_begin_0 = const()[name = string("op_51_begin_0"), val = tensor([0, 1])]; + tensor var_51_end_0 = const()[name = string("op_51_end_0"), val = tensor([1, 0])]; + tensor var_51_end_mask_0 = const()[name = string("op_51_end_mask_0"), val = tensor([true, true])]; + tensor var_51_cast_fp16 = slice_by_index(begin = var_51_begin_0, end = var_51_end_0, end_mask = var_51_end_mask_0, x = audio_to_fp16)[name = string("op_51_cast_fp16")]; + tensor var_53_begin_0 = const()[name = string("op_53_begin_0"), val = tensor([0, 0])]; + tensor var_53_end_0 = const()[name = string("op_53_end_0"), val = tensor([1, -1])]; + tensor var_53_end_mask_0 = const()[name = string("op_53_end_mask_0"), val = tensor([true, false])]; + tensor var_53_cast_fp16 = slice_by_index(begin = var_53_begin_0, end = var_53_end_0, end_mask = var_53_end_mask_0, x = audio_to_fp16)[name = string("op_53_cast_fp16")]; + fp16 var_54_to_fp16 = const()[name = string("op_54_to_fp16"), val = fp16(0x1.f0cp-1)]; + tensor var_55_cast_fp16 = mul(x = var_53_cast_fp16, y = var_54_to_fp16)[name = string("op_55_cast_fp16")]; + tensor var_56_cast_fp16 = sub(x = var_51_cast_fp16, y = var_55_cast_fp16)[name = string("op_56_cast_fp16")]; + bool x_3_interleave_0 = const()[name = string("x_3_interleave_0"), val = bool(false)]; + tensor x_3_cast_fp16 = concat(axis = var_9, interleave = x_3_interleave_0, values = (var_49_cast_fp16, var_56_cast_fp16))[name = string("x_3_cast_fp16")]; + tensor var_59 = logical_not(x = timemask)[name = string("op_59")]; + fp16 var_16_to_fp16 = const()[name = string("op_16_to_fp16"), val = fp16(0x0p+0)]; + tensor input_1_cast_fp16 = select(a = var_16_to_fp16, b = x_3_cast_fp16, cond = var_59)[name = string("input_1_cast_fp16")]; + tensor concat_1x = const()[name = string("concat_1x"), val = tensor([1, 1, -1])]; + tensor input_3_cast_fp16 = reshape(shape = concat_1x, x = input_1_cast_fp16)[name = string("input_3_cast_fp16")]; + tensor input_5_pad_0 = const()[name = string("input_5_pad_0"), val = tensor([0, 0, 0, 0, 256, 256])]; + string input_5_mode_0 = const()[name = string("input_5_mode_0"), val = string("constant")]; + fp16 const_3_to_fp16 = const()[name = string("const_3_to_fp16"), val = fp16(0x0p+0)]; + tensor input_5_cast_fp16 = pad(constant_val = const_3_to_fp16, mode = input_5_mode_0, pad = input_5_pad_0, x = input_3_cast_fp16)[name = string("input_5_cast_fp16")]; + tensor concat_2x = const()[name = string("concat_2x"), val = tensor([1, -1])]; + tensor input_cast_fp16 = reshape(shape = concat_2x, x = input_5_cast_fp16)[name = string("input_cast_fp16")]; + tensor expand_dims_3 = const()[name = string("expand_dims_3"), val = tensor([160])]; + tensor expand_dims_4_axes_0 = const()[name = string("expand_dims_4_axes_0"), val = tensor([1])]; + tensor expand_dims_4_cast_fp16 = expand_dims(axes = expand_dims_4_axes_0, x = input_cast_fp16)[name = string("expand_dims_4_cast_fp16")]; + string conv_0_pad_type_0 = const()[name = string("conv_0_pad_type_0"), val = string("valid")]; + tensor conv_0_pad_0 = const()[name = string("conv_0_pad_0"), val = tensor([0, 0])]; + tensor conv_0_dilations_0 = const()[name = string("conv_0_dilations_0"), val = tensor([1])]; + int32 conv_0_groups_0 = const()[name = string("conv_0_groups_0"), val = int32(1)]; + tensor expand_dims_1_to_fp16 = const()[name = string("expand_dims_1_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))]; + tensor conv_0_cast_fp16 = conv(dilations = conv_0_dilations_0, groups = conv_0_groups_0, pad = conv_0_pad_0, pad_type = conv_0_pad_type_0, strides = expand_dims_3, weight = expand_dims_1_to_fp16, x = expand_dims_4_cast_fp16)[name = string("conv_0_cast_fp16")]; + string conv_1_pad_type_0 = const()[name = string("conv_1_pad_type_0"), val = string("valid")]; + tensor conv_1_pad_0 = const()[name = string("conv_1_pad_0"), val = tensor([0, 0])]; + tensor conv_1_dilations_0 = const()[name = string("conv_1_dilations_0"), val = tensor([1])]; + int32 conv_1_groups_0 = const()[name = string("conv_1_groups_0"), val = int32(1)]; + tensor expand_dims_2_to_fp16 = const()[name = string("expand_dims_2_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(263296)))]; + tensor conv_1_cast_fp16 = conv(dilations = conv_1_dilations_0, groups = conv_1_groups_0, pad = conv_1_pad_0, pad_type = conv_1_pad_type_0, strides = expand_dims_3, weight = expand_dims_2_to_fp16, x = expand_dims_4_cast_fp16)[name = string("conv_1_cast_fp16")]; + int32 stack_0_axis_0 = const()[name = string("stack_0_axis_0"), val = int32(-1)]; + tensor stack_0_cast_fp16 = stack(axis = stack_0_axis_0, values = (conv_0_cast_fp16, conv_1_cast_fp16))[name = string("stack_0_cast_fp16")]; + fp16 var_19_promoted_to_fp16 = const()[name = string("op_19_promoted_to_fp16"), val = fp16(0x1p+1)]; + tensor var_74_cast_fp16 = pow(x = stack_0_cast_fp16, y = var_19_promoted_to_fp16)[name = string("op_74_cast_fp16")]; + tensor var_76_axes_0 = const()[name = string("op_76_axes_0"), val = tensor([-1])]; + bool var_76_keep_dims_0 = const()[name = string("op_76_keep_dims_0"), val = bool(false)]; + tensor var_76_cast_fp16 = reduce_sum(axes = var_76_axes_0, keep_dims = var_76_keep_dims_0, x = var_74_cast_fp16)[name = string("op_76_cast_fp16")]; + tensor x_11_cast_fp16 = identity(x = var_76_cast_fp16)[name = string("x_11_cast_fp16")]; + bool x_13_transpose_x_0 = const()[name = string("x_13_transpose_x_0"), val = bool(false)]; + bool x_13_transpose_y_0 = const()[name = string("x_13_transpose_y_0"), val = bool(false)]; + tensor const_4_to_fp16 = const()[name = string("const_4_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(526528)))]; + tensor x_13_cast_fp16 = matmul(transpose_x = x_13_transpose_x_0, transpose_y = x_13_transpose_y_0, x = const_4_to_fp16, y = x_11_cast_fp16)[name = string("x_13_cast_fp16")]; + fp16 var_83_to_fp16 = const()[name = string("op_83_to_fp16"), val = fp16(0x1p-24)]; + tensor var_84_cast_fp16 = add(x = x_13_cast_fp16, y = var_83_to_fp16)[name = string("op_84_cast_fp16")]; + fp32 x_epsilon_0 = const()[name = string("x_epsilon_0"), val = fp32(0x1p-149)]; + tensor x_cast_fp16 = log(epsilon = x_epsilon_0, x = var_84_cast_fp16)[name = string("x_cast_fp16")]; + tensor var_86_shape_cast_fp16 = shape(x = x_cast_fp16)[name = string("op_86_shape_cast_fp16")]; + int32 gather_5_batch_dims_0 = const()[name = string("gather_5_batch_dims_0"), val = int32(0)]; + bool gather_5_validate_indices_0 = const()[name = string("gather_5_validate_indices_0"), val = bool(false)]; + string var_86_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_86_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")]; + int32 gather_5_cast_uint16_axis_0 = const()[name = string("gather_5_cast_uint16_axis_0"), val = int32(0)]; + uint16 select_0_to_uint16 = const()[name = string("select_0_to_uint16"), val = uint16(2)]; + tensor var_86_shape_cast_fp16_to_uint16 = cast(dtype = var_86_shape_cast_fp16_to_uint16_dtype_0, x = var_86_shape_cast_fp16)[name = string("cast_7")]; + uint16 gather_5_cast_uint16_cast_uint16 = gather(axis = gather_5_cast_uint16_axis_0, batch_dims = gather_5_batch_dims_0, indices = select_0_to_uint16, validate_indices = gather_5_validate_indices_0, x = var_86_shape_cast_fp16_to_uint16)[name = string("gather_5_cast_uint16_cast_uint16")]; + string gather_5_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_5_cast_uint16_to_int32_dtype_0"), val = string("int32")]; + int32 const_5 = const()[name = string("const_5"), val = int32(0)]; + int32 const_6 = const()[name = string("const_6"), val = int32(1)]; + int32 gather_5_cast_uint16_to_int32 = cast(dtype = gather_5_cast_uint16_to_int32_dtype_0, x = gather_5_cast_uint16_cast_uint16)[name = string("cast_6")]; + tensor mask_1 = range_1d(end = gather_5_cast_uint16_to_int32, start = const_5, step = const_6)[name = string("mask_1")]; + tensor expand_dims_0_axes_0 = const()[name = string("expand_dims_0_axes_0"), val = tensor([0])]; + tensor expand_dims_0 = expand_dims(axes = expand_dims_0_axes_0, x = mask_1)[name = string("expand_dims_0")]; + tensor var_91_axes_0 = const()[name = string("op_91_axes_0"), val = tensor([1])]; + tensor var_91 = expand_dims(axes = var_91_axes_0, x = mel_length)[name = string("op_91")]; + tensor mask = greater_equal(x = expand_dims_0, y = var_91)[name = string("mask")]; + tensor var_93_axes_0 = const()[name = string("op_93_axes_0"), val = tensor([1])]; + tensor var_93 = expand_dims(axes = var_93_axes_0, x = mask)[name = string("op_93")]; + tensor processed_signal_cast_fp16 = select(a = var_16_to_fp16, b = x_cast_fp16, cond = var_93)[name = string("processed_signal_cast_fp16")]; + string processed_signal_cast_fp16_to_fp32_dtype_0 = const()[name = string("processed_signal_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + tensor mel = cast(dtype = processed_signal_cast_fp16_to_fp32_dtype_0, x = processed_signal_cast_fp16)[name = string("cast_5")]; + } -> (mel, mel_length); +} \ No newline at end of file diff --git a/it/2240ms/preprocessor.mlmodelc/weights/weight.bin b/it/2240ms/preprocessor.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..86dd375f6649d262d58c9cf8b89006ceab216411 --- /dev/null +++ b/it/2240ms/preprocessor.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:297514e2b211d14b0e53cb97193d679bb89ead98d28e578f3f1d049ddbcc36b3 +size 592384 diff --git a/it/2240ms/preprocessor.mlpackage/Data/com.apple.CoreML/model.mlmodel b/it/2240ms/preprocessor.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..050fa97ca7a2aa4b7c4fa318f4fa2a51914287c4 --- /dev/null +++ b/it/2240ms/preprocessor.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0feb90f2cf9a5bc749fa763b6eb78a23755b6184b41a262fe08faebe4e709b3e +size 16035 diff --git a/it/2240ms/preprocessor.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/it/2240ms/preprocessor.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..86dd375f6649d262d58c9cf8b89006ceab216411 --- /dev/null +++ b/it/2240ms/preprocessor.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:297514e2b211d14b0e53cb97193d679bb89ead98d28e578f3f1d049ddbcc36b3 +size 592384 diff --git a/it/2240ms/preprocessor.mlpackage/Manifest.json b/it/2240ms/preprocessor.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..667b90978564849cb1c6d3c8b10f535ad524d4be --- /dev/null +++ b/it/2240ms/preprocessor.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "02ED6D10-6658-4E52-83B8-E0D3DEA2B8AC": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + }, + "8D115BDF-817A-4C02-9F24-4AF6137B0210": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + } + }, + "rootModelIdentifier": "8D115BDF-817A-4C02-9F24-4AF6137B0210" +} diff --git a/it/2240ms/tokenizer.json b/it/2240ms/tokenizer.json new file mode 100644 index 0000000000000000000000000000000000000000..033007249c80c55ca310e0c90339004b867968f6 --- /dev/null +++ b/it/2240ms/tokenizer.json @@ -0,0 +1,808 @@ +{ + "0": "", + "1": "", + "2": "▁", + "3": ".", + "4": ",", + "5": "e", + "6": "t", + "7": "a", + "8": "s", + "9": "o", + "10": "i", + "11": "r", + "12": "l", + "13": "u", + "14": "d", + "15": "c", + "16": "h", + "17": "m", + "18": "p", + "19": "n", + "20": "g", + "21": "f", + "22": "en", + "23": "in", + "24": "on", + "25": "y", + "26": "er", + "27": "an", + "28": "w", + "29": "0", + "30": "b", + "31": "v", + "32": "2", + "33": "1", + "34": "k", + "35": "3", + "36": "5", + "37": "q", + "38": "I", + "39": "é", + "40": "4", + "41": "z", + "42": "A", + "43": "6", + "44": "j", + "45": "E", + "46": "7", + "47": "8", + "48": "9", + "49": "S", + "50": "x", + "51": "C", + "52": "P", + "53": "U", + "54": "N", + "55": "O", + "56": "à", + "57": "L", + "58": "è", + "59": "V", + "60": "R", + "61": "ó", + "62": "J", + "63": "", + "64": "st", + "65": "ch", + "66": "le", + "67": "li", + "68": "▁po", + "69": "no", + "70": "to", + "71": "me", + "72": "te", + "73": "ho", + "74": "▁pro", + "75": "ro", + "76": "▁na", + "77": "ce", + "78": "la", + "79": "ni", + "80": "ra", + "81": "ti", + "82": "lo", + "83": "ko", + "84": "po", + "85": "je", + "86": "de", + "87": "na", + "88": "mi", + "89": "ci", + "90": "▁by", + "91": "ve", + "92": "▁za", + "93": "▁A", + "94": "re", + "95": "ou", + "96": "vo", + "97": "né", + "98": "va", + "99": "mo", + "100": "ze", + "101": "ne", + "102": "ka", + "103": "ky", + "104": "ovat", + "105": "▁jak", + "106": "ny", + "107": "vi", + "108": "ent", + "109": "prav", + "110": "Z", + "111": "▁Je", + "112": "oval", + "113": "ă", + "114": "X", + "115": "", + "116": "▁for", + "117": "▁det", + "118": "▁at", + "119": "et", + "120": "▁og", + "121": "▁vi", + "122": "al", + "123": "▁de", + "124": "▁der", + "125": "or", + "126": "om", + "127": "and", + "128": "▁har", + "129": "at", + "130": "▁af", + "131": "ge", + "132": "ar", + "133": "is", + "134": "▁med", + "135": "▁be", + "136": "un", + "137": "lig", + "138": "▁man", + "139": "ig", + "140": "▁som", + "141": "▁Og", + "142": "el", + "143": "ag", + "144": "erne", + "145": "▁den", + "146": "ste", + "147": "id", + "148": "▁kan", + "149": "ske", + "150": "iv", + "151": "ion", + "152": "am", + "153": "ur", + "154": "for", + "155": "else", + "156": "▁sig", + "157": "▁ind", + "158": "ende", + "159": "▁Vi", + "160": "ation", + "161": "mme", + "162": "▁op", + "163": "▁fra", + "164": "▁alle", + "165": "▁Men", + "166": "▁var", + "167": "", + "168": "▁die", + "169": "▁und", + "170": "sch", + "171": "it", + "172": "gen", + "173": "▁W", + "174": "▁B", + "175": "▁E", + "176": "▁F", + "177": "ll", + "178": "▁es", + "179": "▁K", + "180": "ie", + "181": "au", + "182": "ich", + "183": "ck", + "184": "ten", + "185": "mal", + "186": "ein", + "187": "▁T", + "188": "▁dann", + "189": "▁mit", + "190": "ter", + "191": "tz", + "192": "▁G", + "193": "ben", + "194": "um", + "195": "us", + "196": "il", + "197": "ut", + "198": "▁ver", + "199": "ri", + "200": "ach", + "201": "ol", + "202": "▁Da", + "203": "sp", + "204": "ell", + "205": "▁was", + "206": "▁ja", + "207": "wi", + "208": "rei", + "209": "▁Ge", + "210": "und", + "211": "▁St", + "212": "▁sie", + "213": "▁Ja", + "214": "▁du", + "215": "▁Sch", + "216": "▁Ma", + "217": "▁De", + "218": "▁Sie", + "219": "▁vor", + "220": "▁Le", + "221": "▁In", + "222": "▁Ver", + "223": "▁Re", + "224": "▁Mi", + "225": "▁Ha", + "226": "", + "227": "", + "228": "ma", + "229": "ta", + "230": "se", + "231": "da", + "232": "si", + "233": "ks", + "234": "ga", + "235": "he", + "236": "mu", + "237": "tu", + "238": "ha", + "239": "ja", + "240": "gi", + "241": "gu", + "242": "ju", + "243": "est", + "244": "▁pa", + "245": "tud", + "246": "nda", + "247": "ke", + "248": "sta", + "249": "sed", + "250": "di", + "251": "▁su", + "252": "ide", + "253": "pool", + "254": "val", + "255": "▁Me", + "256": "ment", + "257": "ndus", + "258": "Q", + "259": "", + "260": "ssa", + "261": "lla", + "262": "ki", + "263": "pa", + "264": "lle", + "265": "lu", + "266": "tta", + "267": "isi", + "268": "tte", + "269": "ista", + "270": "llis", + "271": "vu", + "272": "▁voi", + "273": "utta", + "274": "iden", + "275": "parlament", + "276": "", + "277": "▁est", + "278": "▁c", + "279": "▁que", + "280": "es", + "281": "▁un", + "282": "▁pas", + "283": "▁qui", + "284": "▁il", + "285": "▁des", + "286": "▁qu", + "287": "▁par", + "288": "ant", + "289": "▁C", + "290": "tre", + "291": "ir", + "292": "elle", + "293": "eur", + "294": "▁sur", + "295": "▁con", + "296": "ement", + "297": "tion", + "298": "mp", + "299": "▁comme", + "300": "ac", + "301": "▁là", + "302": "che", + "303": "que", + "304": "ul", + "305": "▁bien", + "306": "age", + "307": "▁mon", + "308": "end", + "309": "ver", + "310": "tra", + "311": "ille", + "312": "ff", + "313": "▁ex", + "314": "▁Il", + "315": "im", + "316": "▁dire", + "317": "ance", + "318": "aire", + "319": "▁app", + "320": "onne", + "321": "mb", + "322": "man", + "323": "▁quand", + "324": "port", + "325": "form", + "326": "ture", + "327": "ù", + "328": "", + "329": "ok", + "330": "gy", + "331": "ek", + "332": "em", + "333": "▁is", + "334": "os", + "335": "ak", + "336": "ban", + "337": "ik", + "338": "▁nem", + "339": "oz", + "340": "cs", + "341": "nek", + "342": "bb", + "343": "▁ha", + "344": "ott", + "345": "▁van", + "346": "▁fel", + "347": "leg", + "348": "▁ami", + "349": "tart", + "350": "rend", + "351": "▁fog", + "352": "▁volt", + "353": "", + "354": "▁bi", + "355": "▁sa", + "356": "ru", + "357": "go", + "358": "sti", + "359": "▁pri", + "360": "ima", + "361": "nu", + "362": "▁pre", + "363": "zi", + "364": "vr", + "365": "ca", + "366": "ba", + "367": "▁raz", + "368": "", + "369": "▁di", + "370": "▁che", + "371": "▁è", + "372": "co", + "373": "▁per", + "374": "▁non", + "375": "do", + "376": "gli", + "377": "so", + "378": "amo", + "379": "sa", + "380": "ndo", + "381": "▁una", + "382": "fi", + "383": "pi", + "384": "nti", + "385": "tto", + "386": "tro", + "387": "▁fa", + "388": "chi", + "389": "▁qua", + "390": "zione", + "391": "bi", + "392": "▁del", + "393": "mente", + "394": "pe", + "395": "ssi", + "396": "▁sono", + "397": "▁questo", + "398": "nte", + "399": "tti", + "400": "tà", + "401": "▁nel", + "402": "▁anche", + "403": "sso", + "404": "▁perché", + "405": "▁più", + "406": "nta", + "407": "▁come", + "408": "cu", + "409": "▁quindi", + "410": "ggi", + "411": "nza", + "412": "sto", + "413": "ò", + "414": "▁della", + "415": "gra", + "416": "▁fare", + "417": "spe", + "418": "cco", + "419": "nde", + "420": "mento", + "421": "fe", + "422": "gio", + "423": "pu", + "424": "▁questa", + "425": "zza", + "426": "sci", + "427": "▁dei", + "428": "▁poi", + "429": "sco", + "430": "stra", + "431": "▁quel", + "432": "qui", + "433": "▁delle", + "434": "▁cosa", + "435": "▁molto", + "436": "sse", + "437": "zioni", + "438": "▁inter", + "439": "sce", + "440": "▁fatto", + "441": "▁com", + "442": "▁quello", + "443": "▁essere", + "444": "▁due", + "445": "▁abbiamo", + "446": "▁comp", + "447": "▁tutti", + "448": "ì", + "449": "▁prima", + "450": "▁parte", + "451": "▁così", + "452": "▁sempre", + "453": "▁tutto", + "454": "▁video", + "455": "▁imp", + "456": "▁cui", + "457": "▁dove", + "458": "▁Quindi", + "459": "sione", + "460": "rebbe", + "461": "scri", + "462": "", + "463": "ai", + "464": "▁ir", + "465": "as", + "466": "▁tai", + "467": "uo", + "468": "tin", + "469": "▁vis", + "470": "ly", + "471": "gal", + "472": "tar", + "473": "▁Ir", + "474": "▁turi", + "475": "▁Tai", + "476": "▁nu", + "477": "▁mes", + "478": "išk", + "479": "imas", + "480": "▁pra", + "481": "", + "482": "iem", + "483": "▁pie", + "484": "ies", + "485": "ot", + "486": "▁vien", + "487": "▁Un", + "488": "iet", + "489": "▁Ta", + "490": "dar", + "491": "sim", + "492": "gan", + "493": "▁ap", + "494": "▁nav", + "495": "▁Nu", + "496": "▁cit", + "497": "▁Ne", + "498": "▁20", + "499": "", + "500": "▁dat", + "501": "▁we", + "502": "▁En", + "503": "▁dan", + "504": "▁zo", + "505": "▁met", + "506": "▁wat", + "507": "der", + "508": "ui", + "509": "den", + "510": "op", + "511": "oor", + "512": "▁of", + "513": "ven", + "514": "acht", + "515": "▁even", + "516": "▁wil", + "517": "vol", + "518": "", + "519": "nie", + "520": "wa", + "521": "cie", + "522": "nia", + "523": "wo", + "524": "rze", + "525": "by", + "526": "za", + "527": "dy", + "528": "ry", + "529": "ego", + "530": "mie", + "531": "rz", + "532": "pie", + "533": "▁bo", + "534": "bie", + "535": "▁Po", + "536": "ski", + "537": "nego", + "538": "▁No", + "539": "▁Na", + "540": "▁Was", + "541": "▁musi", + "542": "", + "543": "-", + "544": "▁para", + "545": "▁uma", + "546": "▁pe", + "547": "▁tem", + "548": "▁gente", + "549": "▁O", + "550": "▁ele", + "551": "pre", + "552": "ria", + "553": "▁fo", + "554": "mos", + "555": "bo", + "556": "▁por", + "557": "nto", + "558": "be", + "559": "▁esse", + "560": "ente", + "561": "▁essa", + "562": "▁mas", + "563": "qua", + "564": "fica", + "565": "▁Se", + "566": "▁Por", + "567": "▁Co", + "568": "iza", + "569": "▁sua", + "570": "▁quando", + "571": "", + "572": "▁mai", + "573": "sc", + "574": "are", + "575": "▁din", + "576": "▁este", + "577": "rea", + "578": "ele", + "579": "du", + "580": "▁M", + "581": "▁fac", + "582": "lor", + "583": "▁mult", + "584": "per", + "585": "cur", + "586": "tor", + "587": "inte", + "588": "▁sau", + "589": "tat", + "590": "ori", + "591": "▁prim", + "592": "▁spun", + "593": "▁lui", + "594": "▁sub", + "595": "itate", + "596": "▁prin", + "597": "▁alt", + "598": "stru", + "599": "▁vede", + "600": "fer", + "601": "▁chiar", + "602": "", + "603": "", + "604": "ov", + "605": "ob", + "606": "▁bol", + "607": "ali", + "608": "rov", + "609": "rob", + "610": "▁spo", + "611": "osti", + "612": "", + "613": "sl", + "614": "udi", + "615": "del", + "616": "▁sem", + "617": "▁samo", + "618": "▁pred", + "619": "nost", + "620": "▁Pre", + "621": "▁prot", + "622": "▁internet", + "623": "▁film", + "624": "", + "625": "▁att", + "626": "▁inte", + "627": "▁av", + "628": "all", + "629": "era", + "630": "pp", + "631": "▁upp", + "632": "isk", + "633": "het", + "634": "▁vill", + "635": "erna", + "636": "ande", + "637": "ade", + "638": "bil", + "639": "▁min", + "640": "▁alla", + "641": "lev", + "642": "▁oss", + "643": "land", + "644": "▁Vad", + "645": "person", + "646": "", + "647": "", + "648": "vy", + "649": "ft", + "650": "lige", + "651": "ved", + "652": "'", + "653": "▁H", + "654": "▁D", + "655": "aus", + "656": "▁N", + "657": "▁Be", + "658": "mm", + "659": "ab", + "660": "▁Er", + "661": "ssen", + "662": "rie", + "663": "lei", + "664": "▁An", + "665": "rau", + "666": "▁So", + "667": "", + "668": "▁and", + "669": "▁can", + "670": "ed", + "671": "ay", + "672": "th", + "673": "ic", + "674": "hi", + "675": "▁Oh", + "676": "▁not", + "677": "ight", + "678": "ex", + "679": "▁great", + "680": "ill", + "681": "▁don", + "682": "▁problem", + "683": "▁fine", + "684": "▁month", + "685": "▁check", + "686": "▁zero", + "687": "▁first", + "688": "▁question", + "689": "", + "690": "ive", + "691": "ate", + "692": "ad", + "693": "ng", + "694": "ity", + "695": "ther", + "696": "act", + "697": "side", + "698": "\"", + "699": "", + "700": "ción", + "701": "▁Es", + "702": "res", + "703": "▁La", + "704": "dos", + "705": "▁El", + "706": "▁las", + "707": "men", + "708": "par", + "709": "rio", + "710": "enta", + "711": "▁Ca", + "712": "▁Su", + "713": "▁son", + "714": "ncia", + "715": "▁Con", + "716": "ones", + "717": "▁San", + "718": "▁persona", + "719": "▁Com", + "720": "", + "721": "cia", + "722": "▁Y", + "723": "ron", + "724": "les", + "725": "cio", + "726": "bu", + "727": "", + "728": "ré", + "729": "▁Les", + "730": "our", + "731": "▁Ce", + "732": "com", + "733": "ale", + "734": "if", + "735": "iste", + "736": "▁parti", + "737": "avec", + "738": "app", + "739": "gue", + "740": "▁grand", + "741": "Une", + "742": "È", + "743": "av", + "744": "pri", + "745": "sion", + "746": "ard", + "747": "", + "748": "", + "749": "!", + "750": "", + "751": "", + "752": "", + "753": "ene", + "754": "opp", + "755": "▁han", + "756": "", + "757": "eg", + "758": "kk", + "759": "▁god", + "760": "dde", + "761": "inn", + "762": "dig", + "763": "ord", + "764": "▁tru", + "765": "▁sei", + "766": "ller", + "767": "car", + "768": "ito", + "769": "ram", + "770": "fa", + "771": "▁mil", + "772": "▁passa", + "773": "▁casa", + "774": "", + "775": "▁Pa", + "776": "tura", + "777": "forma", + "778": "tua", + "779": "mar", + "780": "este", + "781": "fun", + "782": "gua", + "783": "▁grande", + "784": "▁nome", + "785": "▁Sua", + "786": "var", + "787": "", + "788": "", + "789": "ş", + "790": "ğ", + "791": "ya", + "792": "▁ve", + "793": "lar", + "794": "ler", + "795": "leri", + "796": "▁bu", + "797": "lan", + "798": "ara", + "799": "▁Bu", + "800": "yo", + "801": "", + "802": "", + "803": "▁t", + "804": "nh", + "805": "" +} \ No newline at end of file diff --git a/it/4480ms/decoder.mlmodelc/analytics/coremldata.bin b/it/4480ms/decoder.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..85afd8d84c262c9e1ba71c6b460a5beb4d6b94c3 --- /dev/null +++ b/it/4480ms/decoder.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4cdca6bf678463f31354072f526088e5bdf5115ae94c04e387bb35b2c7a607d6 +size 243 diff --git a/it/4480ms/decoder.mlmodelc/coremldata.bin b/it/4480ms/decoder.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..15599a75fcde036d80411d67dd27c79119d63f7b --- /dev/null +++ b/it/4480ms/decoder.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:041e1367c92762e4788386bbababd5cce1bc8389eea17ceabc1f91655da101f0 +size 433 diff --git a/it/4480ms/decoder.mlmodelc/model.mil b/it/4480ms/decoder.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..e034376bbf9a1dff11539e03ae80e7a65ea4f393 --- /dev/null +++ b/it/4480ms/decoder.mlmodelc/model.mil @@ -0,0 +1,64 @@ +program(1.3) +[buildInfo = dict({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.5.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.3.0"}})] +{ + func main(tensor c_in, tensor h_in, tensor token, tensor token_length) { + int32 y_axis_0 = const()[name = string("y_axis_0"), val = int32(0)]; + int32 y_batch_dims_0 = const()[name = string("y_batch_dims_0"), val = int32(0)]; + bool y_validate_indices_0 = const()[name = string("y_validate_indices_0"), val = bool(false)]; + tensor module_prediction_embed_weight_to_fp16 = const()[name = string("module_prediction_embed_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))]; + string token_to_int16_dtype_0 = const()[name = string("token_to_int16_dtype_0"), val = string("int16")]; + tensor token_to_int16 = cast(dtype = token_to_int16_dtype_0, x = token)[name = string("cast_8")]; + tensor y_cast_fp16_cast_uint16 = gather(axis = y_axis_0, batch_dims = y_batch_dims_0, indices = token_to_int16, validate_indices = y_validate_indices_0, x = module_prediction_embed_weight_to_fp16)[name = string("y_cast_fp16_cast_uint16")]; + tensor input_3_perm_0 = const()[name = string("input_3_perm_0"), val = tensor([1, 0, 2])]; + int32 split_0_num_splits_0 = const()[name = string("split_0_num_splits_0"), val = int32(2)]; + int32 split_0_axis_0 = const()[name = string("split_0_axis_0"), val = int32(0)]; + string h_in_to_fp16_dtype_0 = const()[name = string("h_in_to_fp16_dtype_0"), val = string("fp16")]; + tensor h_in_to_fp16 = cast(dtype = h_in_to_fp16_dtype_0, x = h_in)[name = string("cast_7")]; + tensor split_0_cast_fp16_0, tensor split_0_cast_fp16_1 = split(axis = split_0_axis_0, num_splits = split_0_num_splits_0, x = h_in_to_fp16)[name = string("split_0_cast_fp16")]; + int32 split_1_num_splits_0 = const()[name = string("split_1_num_splits_0"), val = int32(2)]; + int32 split_1_axis_0 = const()[name = string("split_1_axis_0"), val = int32(0)]; + string c_in_to_fp16_dtype_0 = const()[name = string("c_in_to_fp16_dtype_0"), val = string("fp16")]; + tensor c_in_to_fp16 = cast(dtype = c_in_to_fp16_dtype_0, x = c_in)[name = string("cast_6")]; + tensor split_1_cast_fp16_0, tensor split_1_cast_fp16_1 = split(axis = split_1_axis_0, num_splits = split_1_num_splits_0, x = c_in_to_fp16)[name = string("split_1_cast_fp16")]; + tensor input_lstm_layer_0_lstm_h0_squeeze_axes_0 = const()[name = string("input_lstm_layer_0_lstm_h0_squeeze_axes_0"), val = tensor([0])]; + tensor input_lstm_layer_0_lstm_h0_squeeze_cast_fp16 = squeeze(axes = input_lstm_layer_0_lstm_h0_squeeze_axes_0, x = split_0_cast_fp16_0)[name = string("input_lstm_layer_0_lstm_h0_squeeze_cast_fp16")]; + tensor input_lstm_layer_0_lstm_c0_squeeze_axes_0 = const()[name = string("input_lstm_layer_0_lstm_c0_squeeze_axes_0"), val = tensor([0])]; + tensor input_lstm_layer_0_lstm_c0_squeeze_cast_fp16 = squeeze(axes = input_lstm_layer_0_lstm_c0_squeeze_axes_0, x = split_1_cast_fp16_0)[name = string("input_lstm_layer_0_lstm_c0_squeeze_cast_fp16")]; + string input_lstm_layer_0_direction_0 = const()[name = string("input_lstm_layer_0_direction_0"), val = string("forward")]; + bool input_lstm_layer_0_output_sequence_0 = const()[name = string("input_lstm_layer_0_output_sequence_0"), val = bool(true)]; + string input_lstm_layer_0_recurrent_activation_0 = const()[name = string("input_lstm_layer_0_recurrent_activation_0"), val = string("sigmoid")]; + string input_lstm_layer_0_cell_activation_0 = const()[name = string("input_lstm_layer_0_cell_activation_0"), val = string("tanh")]; + string input_lstm_layer_0_activation_0 = const()[name = string("input_lstm_layer_0_activation_0"), val = string("tanh")]; + tensor concat_1_to_fp16 = const()[name = string("concat_1_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1031808)))]; + tensor concat_2_to_fp16 = const()[name = string("concat_2_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4308672)))]; + tensor concat_0_to_fp16 = const()[name = string("concat_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7585536)))]; + tensor input_3_cast_fp16 = transpose(perm = input_3_perm_0, x = y_cast_fp16_cast_uint16)[name = string("transpose_2")]; + tensor input_lstm_layer_0_cast_fp16_0, tensor input_lstm_layer_0_cast_fp16_1, tensor input_lstm_layer_0_cast_fp16_2 = lstm(activation = input_lstm_layer_0_activation_0, bias = concat_0_to_fp16, cell_activation = input_lstm_layer_0_cell_activation_0, direction = input_lstm_layer_0_direction_0, initial_c = input_lstm_layer_0_lstm_c0_squeeze_cast_fp16, initial_h = input_lstm_layer_0_lstm_h0_squeeze_cast_fp16, output_sequence = input_lstm_layer_0_output_sequence_0, recurrent_activation = input_lstm_layer_0_recurrent_activation_0, weight_hh = concat_2_to_fp16, weight_ih = concat_1_to_fp16, x = input_3_cast_fp16)[name = string("input_lstm_layer_0_cast_fp16")]; + tensor input_lstm_h0_squeeze_axes_0 = const()[name = string("input_lstm_h0_squeeze_axes_0"), val = tensor([0])]; + tensor input_lstm_h0_squeeze_cast_fp16 = squeeze(axes = input_lstm_h0_squeeze_axes_0, x = split_0_cast_fp16_1)[name = string("input_lstm_h0_squeeze_cast_fp16")]; + tensor input_lstm_c0_squeeze_axes_0 = const()[name = string("input_lstm_c0_squeeze_axes_0"), val = tensor([0])]; + tensor input_lstm_c0_squeeze_cast_fp16 = squeeze(axes = input_lstm_c0_squeeze_axes_0, x = split_1_cast_fp16_1)[name = string("input_lstm_c0_squeeze_cast_fp16")]; + string input_direction_0 = const()[name = string("input_direction_0"), val = string("forward")]; + bool input_output_sequence_0 = const()[name = string("input_output_sequence_0"), val = bool(true)]; + string input_recurrent_activation_0 = const()[name = string("input_recurrent_activation_0"), val = string("sigmoid")]; + string input_cell_activation_0 = const()[name = string("input_cell_activation_0"), val = string("tanh")]; + string input_activation_0 = const()[name = string("input_activation_0"), val = string("tanh")]; + tensor concat_4_to_fp16 = const()[name = string("concat_4_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7590720)))]; + tensor concat_5_to_fp16 = const()[name = string("concat_5_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10867584)))]; + tensor concat_3_to_fp16 = const()[name = string("concat_3_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14144448)))]; + tensor input_cast_fp16_0, tensor input_cast_fp16_1, tensor input_cast_fp16_2 = lstm(activation = input_activation_0, bias = concat_3_to_fp16, cell_activation = input_cell_activation_0, direction = input_direction_0, initial_c = input_lstm_c0_squeeze_cast_fp16, initial_h = input_lstm_h0_squeeze_cast_fp16, output_sequence = input_output_sequence_0, recurrent_activation = input_recurrent_activation_0, weight_hh = concat_5_to_fp16, weight_ih = concat_4_to_fp16, x = input_lstm_layer_0_cast_fp16_0)[name = string("input_cast_fp16")]; + int32 obj_3_axis_0 = const()[name = string("obj_3_axis_0"), val = int32(0)]; + tensor obj_3_cast_fp16 = stack(axis = obj_3_axis_0, values = (input_lstm_layer_0_cast_fp16_1, input_cast_fp16_1))[name = string("obj_3_cast_fp16")]; + string obj_3_cast_fp16_to_fp32_dtype_0 = const()[name = string("obj_3_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + int32 obj_axis_0 = const()[name = string("obj_axis_0"), val = int32(0)]; + tensor obj_cast_fp16 = stack(axis = obj_axis_0, values = (input_lstm_layer_0_cast_fp16_2, input_cast_fp16_2))[name = string("obj_cast_fp16")]; + string obj_cast_fp16_to_fp32_dtype_0 = const()[name = string("obj_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + tensor transpose_0_perm_0 = const()[name = string("transpose_0_perm_0"), val = tensor([1, 2, 0])]; + string transpose_0_cast_fp16_to_fp32_dtype_0 = const()[name = string("transpose_0_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + tensor transpose_0_cast_fp16 = transpose(perm = transpose_0_perm_0, x = input_cast_fp16_0)[name = string("transpose_1")]; + tensor decoder_out = cast(dtype = transpose_0_cast_fp16_to_fp32_dtype_0, x = transpose_0_cast_fp16)[name = string("cast_3")]; + tensor c_out = cast(dtype = obj_cast_fp16_to_fp32_dtype_0, x = obj_cast_fp16)[name = string("cast_4")]; + tensor h_out = cast(dtype = obj_3_cast_fp16_to_fp32_dtype_0, x = obj_3_cast_fp16)[name = string("cast_5")]; + tensor token_length_tmp = identity(x = token_length)[name = string("token_length_tmp")]; + } -> (decoder_out, h_out, c_out); +} \ No newline at end of file diff --git a/it/4480ms/decoder.mlmodelc/weights/weight.bin b/it/4480ms/decoder.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..2fac0f3a92f8a80de4d92a62819f6dba98aa4983 --- /dev/null +++ b/it/4480ms/decoder.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1360d93c68c3e9c54bda4adaec860753949f3b0dc93bc98f4edc9d6f8dd5595c +size 14149632 diff --git a/it/4480ms/decoder.mlpackage/Data/com.apple.CoreML/model.mlmodel b/it/4480ms/decoder.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..567c038e1e42f382639a9ececec8bb38c22cbde0 --- /dev/null +++ b/it/4480ms/decoder.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:73bb3afa62698bc822b6d32b3731d0bc40521e03737e3139e10a768542fca1fe +size 10359 diff --git a/it/4480ms/decoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/it/4480ms/decoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..2fac0f3a92f8a80de4d92a62819f6dba98aa4983 --- /dev/null +++ b/it/4480ms/decoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1360d93c68c3e9c54bda4adaec860753949f3b0dc93bc98f4edc9d6f8dd5595c +size 14149632 diff --git a/it/4480ms/decoder.mlpackage/Manifest.json b/it/4480ms/decoder.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..8fc74b00a3e9885d54546160ecae1f6da7736d01 --- /dev/null +++ b/it/4480ms/decoder.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "7CBCED8D-FA6A-45B0-BF60-30DB0A653074": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + }, + "AFD197FC-BECC-451A-961C-C0CA05D58065": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + } + }, + "rootModelIdentifier": "AFD197FC-BECC-451A-961C-C0CA05D58065" +} diff --git a/it/4480ms/decoder_joint.mlmodelc/analytics/coremldata.bin b/it/4480ms/decoder_joint.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..6063f90a9756de97c8450a89ef53ef04317ef653 --- /dev/null +++ b/it/4480ms/decoder_joint.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:12f4bcf5114baa2b3a37b8ebeab6c519109bd857e50ec345c458b7a6c4deb20e +size 243 diff --git a/it/4480ms/decoder_joint.mlmodelc/coremldata.bin b/it/4480ms/decoder_joint.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..30521702055e705320aa5a67fef85c830c77b3fd --- /dev/null +++ b/it/4480ms/decoder_joint.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:526ac08a31eeba77b6cf21e2c1b28c5378edbe37535a3c57b4c7c0d0256c3ec4 +size 454 diff --git a/it/4480ms/decoder_joint.mlmodelc/model.mil b/it/4480ms/decoder_joint.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..9e96b62349b7d1c4bd97fe8db2d7755704041510 --- /dev/null +++ b/it/4480ms/decoder_joint.mlmodelc/model.mil @@ -0,0 +1,83 @@ +program(1.3) +[buildInfo = dict({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.5.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.3.0"}})] +{ + func main(tensor c_in, tensor encoder, tensor h_in, tensor token, tensor token_length) { + int32 y_axis_0 = const()[name = string("y_axis_0"), val = int32(0)]; + int32 y_batch_dims_0 = const()[name = string("y_batch_dims_0"), val = int32(0)]; + bool y_validate_indices_0 = const()[name = string("y_validate_indices_0"), val = bool(false)]; + tensor decoder_module_prediction_embed_weight_to_fp16 = const()[name = string("decoder_module_prediction_embed_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))]; + string token_to_int16_dtype_0 = const()[name = string("token_to_int16_dtype_0"), val = string("int16")]; + tensor token_to_int16 = cast(dtype = token_to_int16_dtype_0, x = token)[name = string("cast_9")]; + tensor y_cast_fp16_cast_uint16 = gather(axis = y_axis_0, batch_dims = y_batch_dims_0, indices = token_to_int16, validate_indices = y_validate_indices_0, x = decoder_module_prediction_embed_weight_to_fp16)[name = string("y_cast_fp16_cast_uint16")]; + tensor input_3_perm_0 = const()[name = string("input_3_perm_0"), val = tensor([1, 0, 2])]; + int32 split_0_num_splits_0 = const()[name = string("split_0_num_splits_0"), val = int32(2)]; + int32 split_0_axis_0 = const()[name = string("split_0_axis_0"), val = int32(0)]; + string h_in_to_fp16_dtype_0 = const()[name = string("h_in_to_fp16_dtype_0"), val = string("fp16")]; + tensor h_in_to_fp16 = cast(dtype = h_in_to_fp16_dtype_0, x = h_in)[name = string("cast_8")]; + tensor split_0_cast_fp16_0, tensor split_0_cast_fp16_1 = split(axis = split_0_axis_0, num_splits = split_0_num_splits_0, x = h_in_to_fp16)[name = string("split_0_cast_fp16")]; + int32 split_1_num_splits_0 = const()[name = string("split_1_num_splits_0"), val = int32(2)]; + int32 split_1_axis_0 = const()[name = string("split_1_axis_0"), val = int32(0)]; + string c_in_to_fp16_dtype_0 = const()[name = string("c_in_to_fp16_dtype_0"), val = string("fp16")]; + tensor c_in_to_fp16 = cast(dtype = c_in_to_fp16_dtype_0, x = c_in)[name = string("cast_7")]; + tensor split_1_cast_fp16_0, tensor split_1_cast_fp16_1 = split(axis = split_1_axis_0, num_splits = split_1_num_splits_0, x = c_in_to_fp16)[name = string("split_1_cast_fp16")]; + tensor input_5_lstm_layer_0_lstm_h0_squeeze_axes_0 = const()[name = string("input_5_lstm_layer_0_lstm_h0_squeeze_axes_0"), val = tensor([0])]; + tensor input_5_lstm_layer_0_lstm_h0_squeeze_cast_fp16 = squeeze(axes = input_5_lstm_layer_0_lstm_h0_squeeze_axes_0, x = split_0_cast_fp16_0)[name = string("input_5_lstm_layer_0_lstm_h0_squeeze_cast_fp16")]; + tensor input_5_lstm_layer_0_lstm_c0_squeeze_axes_0 = const()[name = string("input_5_lstm_layer_0_lstm_c0_squeeze_axes_0"), val = tensor([0])]; + tensor input_5_lstm_layer_0_lstm_c0_squeeze_cast_fp16 = squeeze(axes = input_5_lstm_layer_0_lstm_c0_squeeze_axes_0, x = split_1_cast_fp16_0)[name = string("input_5_lstm_layer_0_lstm_c0_squeeze_cast_fp16")]; + string input_5_lstm_layer_0_direction_0 = const()[name = string("input_5_lstm_layer_0_direction_0"), val = string("forward")]; + bool input_5_lstm_layer_0_output_sequence_0 = const()[name = string("input_5_lstm_layer_0_output_sequence_0"), val = bool(true)]; + string input_5_lstm_layer_0_recurrent_activation_0 = const()[name = string("input_5_lstm_layer_0_recurrent_activation_0"), val = string("sigmoid")]; + string input_5_lstm_layer_0_cell_activation_0 = const()[name = string("input_5_lstm_layer_0_cell_activation_0"), val = string("tanh")]; + string input_5_lstm_layer_0_activation_0 = const()[name = string("input_5_lstm_layer_0_activation_0"), val = string("tanh")]; + tensor concat_1_to_fp16 = const()[name = string("concat_1_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1031808)))]; + tensor concat_2_to_fp16 = const()[name = string("concat_2_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4308672)))]; + tensor concat_0_to_fp16 = const()[name = string("concat_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7585536)))]; + tensor input_3_cast_fp16 = transpose(perm = input_3_perm_0, x = y_cast_fp16_cast_uint16)[name = string("transpose_4")]; + tensor input_5_lstm_layer_0_cast_fp16_0, tensor input_5_lstm_layer_0_cast_fp16_1, tensor input_5_lstm_layer_0_cast_fp16_2 = lstm(activation = input_5_lstm_layer_0_activation_0, bias = concat_0_to_fp16, cell_activation = input_5_lstm_layer_0_cell_activation_0, direction = input_5_lstm_layer_0_direction_0, initial_c = input_5_lstm_layer_0_lstm_c0_squeeze_cast_fp16, initial_h = input_5_lstm_layer_0_lstm_h0_squeeze_cast_fp16, output_sequence = input_5_lstm_layer_0_output_sequence_0, recurrent_activation = input_5_lstm_layer_0_recurrent_activation_0, weight_hh = concat_2_to_fp16, weight_ih = concat_1_to_fp16, x = input_3_cast_fp16)[name = string("input_5_lstm_layer_0_cast_fp16")]; + tensor input_5_lstm_h0_squeeze_axes_0 = const()[name = string("input_5_lstm_h0_squeeze_axes_0"), val = tensor([0])]; + tensor input_5_lstm_h0_squeeze_cast_fp16 = squeeze(axes = input_5_lstm_h0_squeeze_axes_0, x = split_0_cast_fp16_1)[name = string("input_5_lstm_h0_squeeze_cast_fp16")]; + tensor input_5_lstm_c0_squeeze_axes_0 = const()[name = string("input_5_lstm_c0_squeeze_axes_0"), val = tensor([0])]; + tensor input_5_lstm_c0_squeeze_cast_fp16 = squeeze(axes = input_5_lstm_c0_squeeze_axes_0, x = split_1_cast_fp16_1)[name = string("input_5_lstm_c0_squeeze_cast_fp16")]; + string input_5_direction_0 = const()[name = string("input_5_direction_0"), val = string("forward")]; + bool input_5_output_sequence_0 = const()[name = string("input_5_output_sequence_0"), val = bool(true)]; + string input_5_recurrent_activation_0 = const()[name = string("input_5_recurrent_activation_0"), val = string("sigmoid")]; + string input_5_cell_activation_0 = const()[name = string("input_5_cell_activation_0"), val = string("tanh")]; + string input_5_activation_0 = const()[name = string("input_5_activation_0"), val = string("tanh")]; + tensor concat_4_to_fp16 = const()[name = string("concat_4_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7590720)))]; + tensor concat_5_to_fp16 = const()[name = string("concat_5_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10867584)))]; + tensor concat_3_to_fp16 = const()[name = string("concat_3_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14144448)))]; + tensor input_5_cast_fp16_0, tensor input_5_cast_fp16_1, tensor input_5_cast_fp16_2 = lstm(activation = input_5_activation_0, bias = concat_3_to_fp16, cell_activation = input_5_cell_activation_0, direction = input_5_direction_0, initial_c = input_5_lstm_c0_squeeze_cast_fp16, initial_h = input_5_lstm_h0_squeeze_cast_fp16, output_sequence = input_5_output_sequence_0, recurrent_activation = input_5_recurrent_activation_0, weight_hh = concat_5_to_fp16, weight_ih = concat_4_to_fp16, x = input_5_lstm_layer_0_cast_fp16_0)[name = string("input_5_cast_fp16")]; + int32 obj_3_axis_0 = const()[name = string("obj_3_axis_0"), val = int32(0)]; + tensor obj_3_cast_fp16 = stack(axis = obj_3_axis_0, values = (input_5_lstm_layer_0_cast_fp16_1, input_5_cast_fp16_1))[name = string("obj_3_cast_fp16")]; + string obj_3_cast_fp16_to_fp32_dtype_0 = const()[name = string("obj_3_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + int32 obj_axis_0 = const()[name = string("obj_axis_0"), val = int32(0)]; + tensor obj_cast_fp16 = stack(axis = obj_axis_0, values = (input_5_lstm_layer_0_cast_fp16_2, input_5_cast_fp16_2))[name = string("obj_cast_fp16")]; + string obj_cast_fp16_to_fp32_dtype_0 = const()[name = string("obj_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + tensor transpose_1_perm_0 = const()[name = string("transpose_1_perm_0"), val = tensor([1, 0, 2])]; + tensor input_7_perm_0 = const()[name = string("input_7_perm_0"), val = tensor([0, 2, 1])]; + string encoder_to_fp16_dtype_0 = const()[name = string("encoder_to_fp16_dtype_0"), val = string("fp16")]; + tensor joint_module_enc_weight_to_fp16 = const()[name = string("joint_module_enc_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14149632)))]; + tensor joint_module_enc_bias_to_fp16 = const()[name = string("joint_module_enc_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15460416)))]; + tensor encoder_to_fp16 = cast(dtype = encoder_to_fp16_dtype_0, x = encoder)[name = string("cast_4")]; + tensor input_7_cast_fp16 = transpose(perm = input_7_perm_0, x = encoder_to_fp16)[name = string("transpose_2")]; + tensor linear_0_cast_fp16 = linear(bias = joint_module_enc_bias_to_fp16, weight = joint_module_enc_weight_to_fp16, x = input_7_cast_fp16)[name = string("linear_0_cast_fp16")]; + tensor joint_module_pred_weight_to_fp16 = const()[name = string("joint_module_pred_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15461760)))]; + tensor joint_module_pred_bias_to_fp16 = const()[name = string("joint_module_pred_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16281024)))]; + tensor transpose_1_cast_fp16 = transpose(perm = transpose_1_perm_0, x = input_5_cast_fp16_0)[name = string("transpose_3")]; + tensor linear_1_cast_fp16 = linear(bias = joint_module_pred_bias_to_fp16, weight = joint_module_pred_weight_to_fp16, x = transpose_1_cast_fp16)[name = string("linear_1_cast_fp16")]; + tensor var_79_axes_0 = const()[name = string("op_79_axes_0"), val = tensor([2])]; + tensor var_79_cast_fp16 = expand_dims(axes = var_79_axes_0, x = linear_0_cast_fp16)[name = string("op_79_cast_fp16")]; + tensor var_80_axes_0 = const()[name = string("op_80_axes_0"), val = tensor([1])]; + tensor var_80_cast_fp16 = expand_dims(axes = var_80_axes_0, x = linear_1_cast_fp16)[name = string("op_80_cast_fp16")]; + tensor input_11_cast_fp16 = add(x = var_79_cast_fp16, y = var_80_cast_fp16)[name = string("input_11_cast_fp16")]; + tensor input_13_cast_fp16 = relu(x = input_11_cast_fp16)[name = string("input_13_cast_fp16")]; + tensor joint_module_joint_net_2_weight_to_fp16 = const()[name = string("joint_module_joint_net_2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16282368)))]; + tensor joint_module_joint_net_2_bias_to_fp16 = const()[name = string("joint_module_joint_net_2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17314112)))]; + tensor linear_2_cast_fp16 = linear(bias = joint_module_joint_net_2_bias_to_fp16, weight = joint_module_joint_net_2_weight_to_fp16, x = input_13_cast_fp16)[name = string("linear_2_cast_fp16")]; + string linear_2_cast_fp16_to_fp32_dtype_0 = const()[name = string("linear_2_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + tensor logits = cast(dtype = linear_2_cast_fp16_to_fp32_dtype_0, x = linear_2_cast_fp16)[name = string("cast_3")]; + tensor c_out = cast(dtype = obj_cast_fp16_to_fp32_dtype_0, x = obj_cast_fp16)[name = string("cast_5")]; + tensor h_out = cast(dtype = obj_3_cast_fp16_to_fp32_dtype_0, x = obj_3_cast_fp16)[name = string("cast_6")]; + tensor token_length_tmp = identity(x = token_length)[name = string("token_length_tmp")]; + } -> (logits, h_out, c_out); +} \ No newline at end of file diff --git a/it/4480ms/decoder_joint.mlmodelc/weights/weight.bin b/it/4480ms/decoder_joint.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..33ea2ec2f210db3873bae9b152a8fa5b13171f2e --- /dev/null +++ b/it/4480ms/decoder_joint.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e507a69196a04e30adfafc302b6a5f5f527e45c1965c65dd81d63a621cae2064 +size 17315788 diff --git a/it/4480ms/decoder_joint.mlpackage/Data/com.apple.CoreML/model.mlmodel b/it/4480ms/decoder_joint.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..cc3525ebd701acd827b72a5d6a05caf5ddff80e9 --- /dev/null +++ b/it/4480ms/decoder_joint.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5fa6a6d89c1c07ae16f162f5a3b6809b12aafe57a663f5cdb270be3dec7b1427 +size 13745 diff --git a/it/4480ms/decoder_joint.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/it/4480ms/decoder_joint.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..33ea2ec2f210db3873bae9b152a8fa5b13171f2e --- /dev/null +++ b/it/4480ms/decoder_joint.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e507a69196a04e30adfafc302b6a5f5f527e45c1965c65dd81d63a621cae2064 +size 17315788 diff --git a/it/4480ms/decoder_joint.mlpackage/Manifest.json b/it/4480ms/decoder_joint.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..48ab93415f34542ace6e74a66d563506a10f114a --- /dev/null +++ b/it/4480ms/decoder_joint.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "2B19A50C-1D16-4D97-BE3C-D9BCF35884CF": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + }, + "9CA734BC-CFD2-4F39-B068-BE69ABCAAD1F": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + } + }, + "rootModelIdentifier": "2B19A50C-1D16-4D97-BE3C-D9BCF35884CF" +} diff --git a/it/4480ms/decoder_joint_noencproj.mlmodelc/analytics/coremldata.bin b/it/4480ms/decoder_joint_noencproj.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..3573ed8dea8350501693449f8d9e59b9543d1e3b --- /dev/null +++ b/it/4480ms/decoder_joint_noencproj.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:40ec657603479e7dbf8cdb3d6368349eb8b766a52439a26a735d1fadf1b4281d +size 243 diff --git a/it/4480ms/decoder_joint_noencproj.mlmodelc/coremldata.bin b/it/4480ms/decoder_joint_noencproj.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..08e0a4b1a1f6fd4a609dc06e71726616f5e428ec --- /dev/null +++ b/it/4480ms/decoder_joint_noencproj.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:028789e96584213d83d1d5a4fe4d8fb4fa0f51341b8c46b51606781c4f4d4438 +size 519 diff --git a/it/4480ms/decoder_joint_noencproj.mlmodelc/model.mil b/it/4480ms/decoder_joint_noencproj.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..5d5cbd528590956dead59657945f5dab997a7da9 --- /dev/null +++ b/it/4480ms/decoder_joint_noencproj.mlmodelc/model.mil @@ -0,0 +1,91 @@ +program(1.3) +[buildInfo = dict({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.10.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor c_in, tensor encoder_proj, tensor h_in, tensor token, tensor token_length) { + int32 y_batch_dims_0 = const()[name = string("y_batch_dims_0"), val = int32(0)]; + bool y_validate_indices_0 = const()[name = string("y_validate_indices_0"), val = bool(false)]; + tensor decoder_module_prediction_embed_weight_to_fp16 = const()[name = string("decoder_module_prediction_embed_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))]; + string token_to_int16_dtype_0 = const()[name = string("token_to_int16_dtype_0"), val = string("int16")]; + string cast_1_dtype_0 = const()[name = string("cast_1_dtype_0"), val = string("int32")]; + int32 greater_equal_0_y_0 = const()[name = string("greater_equal_0_y_0"), val = int32(0)]; + tensor token_to_int16 = cast(dtype = token_to_int16_dtype_0, x = token)[name = string("cast_10")]; + tensor cast_1 = cast(dtype = cast_1_dtype_0, x = token_to_int16)[name = string("cast_9")]; + tensor greater_equal_0 = greater_equal(x = cast_1, y = greater_equal_0_y_0)[name = string("greater_equal_0")]; + int32 slice_by_index_0 = const()[name = string("slice_by_index_0"), val = int32(806)]; + tensor add_2 = add(x = cast_1, y = slice_by_index_0)[name = string("add_2")]; + tensor select_0 = select(a = cast_1, b = add_2, cond = greater_equal_0)[name = string("select_0")]; + int32 y_cast_fp16_cast_uint16_axis_0 = const()[name = string("y_cast_fp16_cast_uint16_axis_0"), val = int32(0)]; + string select_0_to_int16_dtype_0 = const()[name = string("select_0_to_int16_dtype_0"), val = string("int16")]; + tensor select_0_to_int16 = cast(dtype = select_0_to_int16_dtype_0, x = select_0)[name = string("cast_8")]; + tensor y_cast_fp16_cast_uint16_cast_uint16 = gather(axis = y_cast_fp16_cast_uint16_axis_0, batch_dims = y_batch_dims_0, indices = select_0_to_int16, validate_indices = y_validate_indices_0, x = decoder_module_prediction_embed_weight_to_fp16)[name = string("y_cast_fp16_cast_uint16_cast_uint16")]; + tensor input_3_perm_0 = const()[name = string("input_3_perm_0"), val = tensor([1, 0, 2])]; + int32 split_0_num_splits_0 = const()[name = string("split_0_num_splits_0"), val = int32(2)]; + int32 split_0_axis_0 = const()[name = string("split_0_axis_0"), val = int32(0)]; + string h_in_to_fp16_dtype_0 = const()[name = string("h_in_to_fp16_dtype_0"), val = string("fp16")]; + tensor h_in_to_fp16 = cast(dtype = h_in_to_fp16_dtype_0, x = h_in)[name = string("cast_7")]; + tensor split_0_cast_fp16_0, tensor split_0_cast_fp16_1 = split(axis = split_0_axis_0, num_splits = split_0_num_splits_0, x = h_in_to_fp16)[name = string("split_0_cast_fp16")]; + int32 split_1_num_splits_0 = const()[name = string("split_1_num_splits_0"), val = int32(2)]; + int32 split_1_axis_0 = const()[name = string("split_1_axis_0"), val = int32(0)]; + string c_in_to_fp16_dtype_0 = const()[name = string("c_in_to_fp16_dtype_0"), val = string("fp16")]; + tensor c_in_to_fp16 = cast(dtype = c_in_to_fp16_dtype_0, x = c_in)[name = string("cast_6")]; + tensor split_1_cast_fp16_0, tensor split_1_cast_fp16_1 = split(axis = split_1_axis_0, num_splits = split_1_num_splits_0, x = c_in_to_fp16)[name = string("split_1_cast_fp16")]; + tensor input_5_lstm_layer_0_lstm_h0_squeeze_axes_0 = const()[name = string("input_5_lstm_layer_0_lstm_h0_squeeze_axes_0"), val = tensor([0])]; + tensor input_5_lstm_layer_0_lstm_h0_squeeze_cast_fp16 = squeeze(axes = input_5_lstm_layer_0_lstm_h0_squeeze_axes_0, x = split_0_cast_fp16_0)[name = string("input_5_lstm_layer_0_lstm_h0_squeeze_cast_fp16")]; + tensor input_5_lstm_layer_0_lstm_c0_squeeze_axes_0 = const()[name = string("input_5_lstm_layer_0_lstm_c0_squeeze_axes_0"), val = tensor([0])]; + tensor input_5_lstm_layer_0_lstm_c0_squeeze_cast_fp16 = squeeze(axes = input_5_lstm_layer_0_lstm_c0_squeeze_axes_0, x = split_1_cast_fp16_0)[name = string("input_5_lstm_layer_0_lstm_c0_squeeze_cast_fp16")]; + string input_5_lstm_layer_0_direction_0 = const()[name = string("input_5_lstm_layer_0_direction_0"), val = string("forward")]; + bool input_5_lstm_layer_0_output_sequence_0 = const()[name = string("input_5_lstm_layer_0_output_sequence_0"), val = bool(true)]; + string input_5_lstm_layer_0_recurrent_activation_0 = const()[name = string("input_5_lstm_layer_0_recurrent_activation_0"), val = string("sigmoid")]; + string input_5_lstm_layer_0_cell_activation_0 = const()[name = string("input_5_lstm_layer_0_cell_activation_0"), val = string("tanh")]; + string input_5_lstm_layer_0_activation_0 = const()[name = string("input_5_lstm_layer_0_activation_0"), val = string("tanh")]; + tensor concat_1_to_fp16 = const()[name = string("concat_1_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1031808)))]; + tensor concat_2_to_fp16 = const()[name = string("concat_2_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4308672)))]; + tensor concat_0_to_fp16 = const()[name = string("concat_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7585536)))]; + tensor input_3_cast_fp16 = transpose(perm = input_3_perm_0, x = y_cast_fp16_cast_uint16_cast_uint16)[name = string("transpose_3")]; + tensor input_5_lstm_layer_0_cast_fp16_0, tensor input_5_lstm_layer_0_cast_fp16_1, tensor input_5_lstm_layer_0_cast_fp16_2 = lstm(activation = input_5_lstm_layer_0_activation_0, bias = concat_0_to_fp16, cell_activation = input_5_lstm_layer_0_cell_activation_0, direction = input_5_lstm_layer_0_direction_0, initial_c = input_5_lstm_layer_0_lstm_c0_squeeze_cast_fp16, initial_h = input_5_lstm_layer_0_lstm_h0_squeeze_cast_fp16, output_sequence = input_5_lstm_layer_0_output_sequence_0, recurrent_activation = input_5_lstm_layer_0_recurrent_activation_0, weight_hh = concat_2_to_fp16, weight_ih = concat_1_to_fp16, x = input_3_cast_fp16)[name = string("input_5_lstm_layer_0_cast_fp16")]; + tensor input_5_lstm_h0_squeeze_axes_0 = const()[name = string("input_5_lstm_h0_squeeze_axes_0"), val = tensor([0])]; + tensor input_5_lstm_h0_squeeze_cast_fp16 = squeeze(axes = input_5_lstm_h0_squeeze_axes_0, x = split_0_cast_fp16_1)[name = string("input_5_lstm_h0_squeeze_cast_fp16")]; + tensor input_5_lstm_c0_squeeze_axes_0 = const()[name = string("input_5_lstm_c0_squeeze_axes_0"), val = tensor([0])]; + tensor input_5_lstm_c0_squeeze_cast_fp16 = squeeze(axes = input_5_lstm_c0_squeeze_axes_0, x = split_1_cast_fp16_1)[name = string("input_5_lstm_c0_squeeze_cast_fp16")]; + string input_5_direction_0 = const()[name = string("input_5_direction_0"), val = string("forward")]; + bool input_5_output_sequence_0 = const()[name = string("input_5_output_sequence_0"), val = bool(true)]; + string input_5_recurrent_activation_0 = const()[name = string("input_5_recurrent_activation_0"), val = string("sigmoid")]; + string input_5_cell_activation_0 = const()[name = string("input_5_cell_activation_0"), val = string("tanh")]; + string input_5_activation_0 = const()[name = string("input_5_activation_0"), val = string("tanh")]; + tensor concat_4_to_fp16 = const()[name = string("concat_4_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7590720)))]; + tensor concat_5_to_fp16 = const()[name = string("concat_5_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10867584)))]; + tensor concat_3_to_fp16 = const()[name = string("concat_3_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14144448)))]; + tensor input_5_cast_fp16_0, tensor input_5_cast_fp16_1, tensor input_5_cast_fp16_2 = lstm(activation = input_5_activation_0, bias = concat_3_to_fp16, cell_activation = input_5_cell_activation_0, direction = input_5_direction_0, initial_c = input_5_lstm_c0_squeeze_cast_fp16, initial_h = input_5_lstm_h0_squeeze_cast_fp16, output_sequence = input_5_output_sequence_0, recurrent_activation = input_5_recurrent_activation_0, weight_hh = concat_5_to_fp16, weight_ih = concat_4_to_fp16, x = input_5_lstm_layer_0_cast_fp16_0)[name = string("input_5_cast_fp16")]; + int32 obj_3_axis_0 = const()[name = string("obj_3_axis_0"), val = int32(0)]; + tensor obj_3_cast_fp16 = stack(axis = obj_3_axis_0, values = (input_5_lstm_layer_0_cast_fp16_1, input_5_cast_fp16_1))[name = string("obj_3_cast_fp16")]; + string obj_3_cast_fp16_to_fp32_dtype_0 = const()[name = string("obj_3_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + int32 obj_axis_0 = const()[name = string("obj_axis_0"), val = int32(0)]; + tensor obj_cast_fp16 = stack(axis = obj_axis_0, values = (input_5_lstm_layer_0_cast_fp16_2, input_5_cast_fp16_2))[name = string("obj_cast_fp16")]; + string obj_cast_fp16_to_fp32_dtype_0 = const()[name = string("obj_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + tensor transpose_1_perm_0 = const()[name = string("transpose_1_perm_0"), val = tensor([1, 0, 2])]; + tensor joint_module_pred_weight_to_fp16 = const()[name = string("joint_module_pred_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14149632)))]; + tensor joint_module_pred_bias_to_fp16 = const()[name = string("joint_module_pred_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14968896)))]; + tensor transpose_1_cast_fp16 = transpose(perm = transpose_1_perm_0, x = input_5_cast_fp16_0)[name = string("transpose_2")]; + tensor linear_0_cast_fp16 = linear(bias = joint_module_pred_bias_to_fp16, weight = joint_module_pred_weight_to_fp16, x = transpose_1_cast_fp16)[name = string("linear_0_cast_fp16")]; + tensor f_axes_0 = const()[name = string("f_axes_0"), val = tensor([2])]; + string encoder_proj_to_fp16_dtype_0 = const()[name = string("encoder_proj_to_fp16_dtype_0"), val = string("fp16")]; + tensor encoder_proj_to_fp16 = cast(dtype = encoder_proj_to_fp16_dtype_0, x = encoder_proj)[name = string("cast_3")]; + tensor f_cast_fp16 = expand_dims(axes = f_axes_0, x = encoder_proj_to_fp16)[name = string("f_cast_fp16")]; + tensor g_axes_0 = const()[name = string("g_axes_0"), val = tensor([1])]; + tensor g_cast_fp16 = expand_dims(axes = g_axes_0, x = linear_0_cast_fp16)[name = string("g_cast_fp16")]; + tensor input_9_cast_fp16 = add(x = f_cast_fp16, y = g_cast_fp16)[name = string("input_9_cast_fp16")]; + tensor input_11_cast_fp16 = relu(x = input_9_cast_fp16)[name = string("input_11_cast_fp16")]; + tensor joint_module_joint_net_2_weight_to_fp16 = const()[name = string("joint_module_joint_net_2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14970240)))]; + tensor joint_module_joint_net_2_bias_to_fp16 = const()[name = string("joint_module_joint_net_2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16001984)))]; + tensor linear_1_cast_fp16 = linear(bias = joint_module_joint_net_2_bias_to_fp16, weight = joint_module_joint_net_2_weight_to_fp16, x = input_11_cast_fp16)[name = string("linear_1_cast_fp16")]; + int32 var_83 = const()[name = string("op_83"), val = int32(-1)]; + tensor var_85_softmax_cast_fp16 = softmax(axis = var_83, x = linear_1_cast_fp16)[name = string("op_85_softmax_cast_fp16")]; + fp32 var_85_epsilon_0 = const()[name = string("op_85_epsilon_0"), val = fp32(0x1p-149)]; + tensor var_85_cast_fp16 = log(epsilon = var_85_epsilon_0, x = var_85_softmax_cast_fp16)[name = string("op_85_cast_fp16")]; + string var_85_cast_fp16_to_fp32_dtype_0 = const()[name = string("op_85_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + tensor logits = cast(dtype = var_85_cast_fp16_to_fp32_dtype_0, x = var_85_cast_fp16)[name = string("cast_2")]; + tensor c_out = cast(dtype = obj_cast_fp16_to_fp32_dtype_0, x = obj_cast_fp16)[name = string("cast_4")]; + tensor h_out = cast(dtype = obj_3_cast_fp16_to_fp32_dtype_0, x = obj_3_cast_fp16)[name = string("cast_5")]; + tensor token_length_tmp = identity(x = token_length)[name = string("token_length_tmp")]; + } -> (logits, h_out, c_out); +} \ No newline at end of file diff --git a/it/4480ms/decoder_joint_noencproj.mlmodelc/weights/weight.bin b/it/4480ms/decoder_joint_noencproj.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..80af5cec724e7e6b117fcd6a7bc8046c27c26e75 --- /dev/null +++ b/it/4480ms/decoder_joint_noencproj.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5ca6a467ebf44612a8032c6c4ddf323e35a9ffaa15c01822528fb97144ec9439 +size 16003660 diff --git a/it/4480ms/decoder_joint_noencproj.mlpackage/Data/com.apple.CoreML/model.mlmodel b/it/4480ms/decoder_joint_noencproj.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..492f62656a116fa907f67e5963520bcaa37729f8 --- /dev/null +++ b/it/4480ms/decoder_joint_noencproj.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1a46ef6436e8c1e5af6f379658d871f2a40dff3a13c946d4f8df736b19cdc45e +size 14630 diff --git a/it/4480ms/decoder_joint_noencproj.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/it/4480ms/decoder_joint_noencproj.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..80af5cec724e7e6b117fcd6a7bc8046c27c26e75 --- /dev/null +++ b/it/4480ms/decoder_joint_noencproj.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5ca6a467ebf44612a8032c6c4ddf323e35a9ffaa15c01822528fb97144ec9439 +size 16003660 diff --git a/it/4480ms/decoder_joint_noencproj.mlpackage/Manifest.json b/it/4480ms/decoder_joint_noencproj.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..2135f9be1a1ec8555bef5683841b910c941cbd7c --- /dev/null +++ b/it/4480ms/decoder_joint_noencproj.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "E92AE661-EE59-447C-B9BE-52D50BC4FC37": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + }, + "F7138F3F-8633-4C54-A3F4-D0F6E8220E39": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + } + }, + "rootModelIdentifier": "F7138F3F-8633-4C54-A3F4-D0F6E8220E39" +} diff --git a/it/4480ms/encoder.mlmodelc/analytics/coremldata.bin b/it/4480ms/encoder.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..2a980d9b38cd6f02b06f8b4550b3668926af78ed --- /dev/null +++ b/it/4480ms/encoder.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:27e74941fe1531f34f2184e41169b6efc4eab0373a327474cd15d5485b97e2a0 +size 243 diff --git a/it/4480ms/encoder.mlmodelc/coremldata.bin b/it/4480ms/encoder.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..bcf5a96ae3bc5f26a4f6d74378d4848c921087b8 --- /dev/null +++ b/it/4480ms/encoder.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e4ab38512a03f40881980a1aea9e633b491b45656c79f6886aafe37e5c4deb30 +size 573 diff --git a/it/4480ms/encoder.mlmodelc/model.mil b/it/4480ms/encoder.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..ac7ff9f2bb92afa45e5a2fabeefa184a92678a46 --- /dev/null +++ b/it/4480ms/encoder.mlmodelc/model.mil @@ -0,0 +1,4321 @@ +program(1.3) +[buildInfo = dict({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}})] +{ + func main(tensor cache_channel, tensor cache_len, tensor cache_time, tensor mel, tensor mel_length, tensor prompt_id) { + tensor value_3_perm_0 = const()[name = string("value_3_perm_0"), val = tensor([1, 0, 2, 3])]; + string cache_channel_to_fp16_dtype_0 = const()[name = string("cache_channel_to_fp16_dtype_0"), val = string("fp16")]; + tensor value_5_perm_0 = const()[name = string("value_5_perm_0"), val = tensor([1, 0, 2, 3])]; + string cache_time_to_fp16_dtype_0 = const()[name = string("cache_time_to_fp16_dtype_0"), val = string("fp16")]; + int32 var_58 = const()[name = string("op_58"), val = int32(-1)]; + int32 var_67 = const()[name = string("op_67"), val = int32(1)]; + tensor x_1_perm_0 = const()[name = string("x_1_perm_0"), val = tensor([0, 2, 1])]; + string mel_to_fp16_dtype_0 = const()[name = string("mel_to_fp16_dtype_0"), val = string("fp16")]; + tensor tensor_1_axes_0 = const()[name = string("tensor_1_axes_0"), val = tensor([1])]; + tensor mel_to_fp16 = cast(dtype = mel_to_fp16_dtype_0, x = mel)[name = string("cast_18")]; + tensor x_1_cast_fp16 = transpose(perm = x_1_perm_0, x = mel_to_fp16)[name = string("transpose_367")]; + tensor tensor_1_cast_fp16 = expand_dims(axes = tensor_1_axes_0, x = x_1_cast_fp16)[name = string("tensor_1_cast_fp16")]; + tensor expand_dims_0 = const()[name = string("expand_dims_0"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))]; + tensor var_136_axes_0 = const()[name = string("op_136_axes_0"), val = tensor([1])]; + tensor var_136 = expand_dims(axes = var_136_axes_0, x = mel_length)[name = string("op_136")]; + tensor time_mask_1 = less(x = expand_dims_0, y = var_136)[name = string("time_mask_1")]; + tensor var_138_axes_0 = const()[name = string("op_138_axes_0"), val = tensor([-1])]; + tensor var_138 = expand_dims(axes = var_138_axes_0, x = time_mask_1)[name = string("op_138")]; + tensor var_140_reps_0 = const()[name = string("op_140_reps_0"), val = tensor([1, 1, 128])]; + tensor var_140 = tile(reps = var_140_reps_0, x = var_138)[name = string("op_140")]; + tensor var_146_axes_0 = const()[name = string("op_146_axes_0"), val = tensor([1])]; + string cast_4_to_fp16_dtype_0 = const()[name = string("cast_4_to_fp16_dtype_0"), val = string("fp16")]; + tensor var_140_to_fp16 = cast(dtype = cast_4_to_fp16_dtype_0, x = var_140)[name = string("cast_17")]; + tensor var_146_cast_fp16 = expand_dims(axes = var_146_axes_0, x = var_140_to_fp16)[name = string("op_146_cast_fp16")]; + tensor input_1_cast_fp16 = mul(x = tensor_1_cast_fp16, y = var_146_cast_fp16)[name = string("input_1_cast_fp16")]; + tensor input_3_pad_0 = const()[name = string("input_3_pad_0"), val = tensor([0, 0, 0, 0, 2, 1, 2, 1])]; + string input_3_mode_0 = const()[name = string("input_3_mode_0"), val = string("constant")]; + fp16 const_9_to_fp16 = const()[name = string("const_9_to_fp16"), val = fp16(0x0p+0)]; + tensor input_3_cast_fp16 = pad(constant_val = const_9_to_fp16, mode = input_3_mode_0, pad = input_3_pad_0, x = input_1_cast_fp16)[name = string("input_3_cast_fp16")]; + string tensor_3_pad_type_0 = const()[name = string("tensor_3_pad_type_0"), val = string("valid")]; + tensor tensor_3_strides_0 = const()[name = string("tensor_3_strides_0"), val = tensor([2, 2])]; + tensor tensor_3_pad_0 = const()[name = string("tensor_3_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor tensor_3_dilations_0 = const()[name = string("tensor_3_dilations_0"), val = tensor([1, 1])]; + int32 tensor_3_groups_0 = const()[name = string("tensor_3_groups_0"), val = int32(1)]; + tensor encoder_pre_encode_conv_0_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1984))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4352))))[name = string("encoder_pre_encode_conv_0_weight_to_fp16_quantized")]; + tensor encoder_pre_encode_conv_0_bias_to_fp16 = const()[name = string("encoder_pre_encode_conv_0_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4928)))]; + tensor tensor_3_cast_fp16 = conv(bias = encoder_pre_encode_conv_0_bias_to_fp16, dilations = tensor_3_dilations_0, groups = tensor_3_groups_0, pad = tensor_3_pad_0, pad_type = tensor_3_pad_type_0, strides = tensor_3_strides_0, weight = encoder_pre_encode_conv_0_weight_to_fp16_quantized, x = input_3_cast_fp16)[name = string("tensor_3_cast_fp16")]; + string cast_2_to_fp16_dtype_0 = const()[name = string("cast_2_to_fp16_dtype_0"), val = string("fp16")]; + fp16 var_159_promoted_to_fp16 = const()[name = string("op_159_promoted_to_fp16"), val = fp16(0x1p+1)]; + tensor mel_length_to_fp16 = cast(dtype = cast_2_to_fp16_dtype_0, x = mel_length)[name = string("cast_16")]; + tensor var_160_cast_fp16 = add(x = mel_length_to_fp16, y = var_159_promoted_to_fp16)[name = string("op_160_cast_fp16")]; + fp16 var_161_promoted_to_fp16 = const()[name = string("op_161_promoted_to_fp16"), val = fp16(0x1p+0)]; + tensor var_162_cast_fp16 = add(x = var_160_cast_fp16, y = var_161_promoted_to_fp16)[name = string("op_162_cast_fp16")]; + fp16 var_163_promoted_to_fp16 = const()[name = string("op_163_promoted_to_fp16"), val = fp16(0x1.8p+1)]; + tensor var_164_cast_fp16 = sub(x = var_162_cast_fp16, y = var_163_promoted_to_fp16)[name = string("op_164_cast_fp16")]; + fp16 var_55_promoted_to_fp16 = const()[name = string("op_55_promoted_to_fp16"), val = fp16(0x1p+1)]; + tensor floor_div_0_cast_fp16 = floor_div(x = var_164_cast_fp16, y = var_55_promoted_to_fp16)[name = string("floor_div_0_cast_fp16")]; + fp16 var_166_promoted_to_fp16 = const()[name = string("op_166_promoted_to_fp16"), val = fp16(0x1p+0)]; + tensor current_lengths_3_cast_fp16 = add(x = floor_div_0_cast_fp16, y = var_166_promoted_to_fp16)[name = string("current_lengths_3_cast_fp16")]; + string cast_5_dtype_0 = const()[name = string("cast_5_dtype_0"), val = string("int32")]; + tensor expand_dims_1 = const()[name = string("expand_dims_1"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5504)))]; + tensor var_175_axes_0 = const()[name = string("op_175_axes_0"), val = tensor([1])]; + tensor current_lengths_3_cast_fp16_to_int32 = cast(dtype = cast_5_dtype_0, x = current_lengths_3_cast_fp16)[name = string("cast_15")]; + tensor var_175 = expand_dims(axes = var_175_axes_0, x = current_lengths_3_cast_fp16_to_int32)[name = string("op_175")]; + tensor time_mask_3 = less(x = expand_dims_1, y = var_175)[name = string("time_mask_3")]; + tensor var_177_axes_0 = const()[name = string("op_177_axes_0"), val = tensor([-1])]; + tensor var_177 = expand_dims(axes = var_177_axes_0, x = time_mask_3)[name = string("op_177")]; + tensor var_179_reps_0 = const()[name = string("op_179_reps_0"), val = tensor([1, 1, 65])]; + tensor var_179 = tile(reps = var_179_reps_0, x = var_177)[name = string("op_179")]; + tensor var_185_axes_0 = const()[name = string("op_185_axes_0"), val = tensor([1])]; + string cast_6_to_fp16_dtype_0 = const()[name = string("cast_6_to_fp16_dtype_0"), val = string("fp16")]; + tensor var_179_to_fp16 = cast(dtype = cast_6_to_fp16_dtype_0, x = var_179)[name = string("cast_14")]; + tensor var_185_cast_fp16 = expand_dims(axes = var_185_axes_0, x = var_179_to_fp16)[name = string("op_185_cast_fp16")]; + tensor expanded_mask_3_reps_0 = const()[name = string("expanded_mask_3_reps_0"), val = tensor([1, 256, 1, 1])]; + tensor expanded_mask_3_cast_fp16 = tile(reps = expanded_mask_3_reps_0, x = var_185_cast_fp16)[name = string("expanded_mask_3_cast_fp16")]; + tensor input_5_cast_fp16 = mul(x = tensor_3_cast_fp16, y = expanded_mask_3_cast_fp16)[name = string("input_5_cast_fp16")]; + tensor tensor_5_cast_fp16 = relu(x = input_5_cast_fp16)[name = string("tensor_5_cast_fp16")]; + tensor input_7_cast_fp16 = mul(x = tensor_5_cast_fp16, y = expanded_mask_3_cast_fp16)[name = string("input_7_cast_fp16")]; + tensor input_9_pad_0 = const()[name = string("input_9_pad_0"), val = tensor([0, 0, 0, 0, 2, 1, 2, 1])]; + string input_9_mode_0 = const()[name = string("input_9_mode_0"), val = string("constant")]; + fp16 const_23_to_fp16 = const()[name = string("const_23_to_fp16"), val = fp16(0x0p+0)]; + tensor input_9_cast_fp16 = pad(constant_val = const_23_to_fp16, mode = input_9_mode_0, pad = input_9_pad_0, x = input_7_cast_fp16)[name = string("input_9_cast_fp16")]; + string tensor_7_pad_type_0 = const()[name = string("tensor_7_pad_type_0"), val = string("valid")]; + tensor tensor_7_strides_0 = const()[name = string("tensor_7_strides_0"), val = tensor([2, 2])]; + int32 tensor_7_groups_0 = const()[name = string("tensor_7_groups_0"), val = int32(256)]; + tensor tensor_7_pad_0 = const()[name = string("tensor_7_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor tensor_7_dilations_0 = const()[name = string("tensor_7_dilations_0"), val = tensor([1, 1])]; + tensor encoder_pre_encode_conv_2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6528))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8896))))[name = string("encoder_pre_encode_conv_2_weight_to_fp16_quantized")]; + tensor encoder_pre_encode_conv_2_bias_to_fp16 = const()[name = string("encoder_pre_encode_conv_2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(9472)))]; + tensor tensor_7_cast_fp16 = conv(bias = encoder_pre_encode_conv_2_bias_to_fp16, dilations = tensor_7_dilations_0, groups = tensor_7_groups_0, pad = tensor_7_pad_0, pad_type = tensor_7_pad_type_0, strides = tensor_7_strides_0, weight = encoder_pre_encode_conv_2_weight_to_fp16_quantized, x = input_9_cast_fp16)[name = string("tensor_7_cast_fp16")]; + fp16 var_207_promoted_to_fp16 = const()[name = string("op_207_promoted_to_fp16"), val = fp16(0x1p+1)]; + tensor var_208_cast_fp16 = add(x = current_lengths_3_cast_fp16, y = var_207_promoted_to_fp16)[name = string("op_208_cast_fp16")]; + fp16 var_209_promoted_to_fp16 = const()[name = string("op_209_promoted_to_fp16"), val = fp16(0x1p+0)]; + tensor var_210_cast_fp16 = add(x = var_208_cast_fp16, y = var_209_promoted_to_fp16)[name = string("op_210_cast_fp16")]; + fp16 var_211_promoted_to_fp16 = const()[name = string("op_211_promoted_to_fp16"), val = fp16(0x1.8p+1)]; + tensor var_212_cast_fp16 = sub(x = var_210_cast_fp16, y = var_211_promoted_to_fp16)[name = string("op_212_cast_fp16")]; + fp16 var_55_promoted_1_to_fp16 = const()[name = string("op_55_promoted_1_to_fp16"), val = fp16(0x1p+1)]; + tensor floor_div_1_cast_fp16 = floor_div(x = var_212_cast_fp16, y = var_55_promoted_1_to_fp16)[name = string("floor_div_1_cast_fp16")]; + fp16 var_214_promoted_to_fp16 = const()[name = string("op_214_promoted_to_fp16"), val = fp16(0x1p+0)]; + tensor current_lengths_5_cast_fp16 = add(x = floor_div_1_cast_fp16, y = var_214_promoted_to_fp16)[name = string("current_lengths_5_cast_fp16")]; + string cast_7_dtype_0 = const()[name = string("cast_7_dtype_0"), val = string("int32")]; + tensor expand_dims_2 = const()[name = string("expand_dims_2"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10048)))]; + tensor var_223_axes_0 = const()[name = string("op_223_axes_0"), val = tensor([1])]; + tensor current_lengths_5_cast_fp16_to_int32 = cast(dtype = cast_7_dtype_0, x = current_lengths_5_cast_fp16)[name = string("cast_13")]; + tensor var_223 = expand_dims(axes = var_223_axes_0, x = current_lengths_5_cast_fp16_to_int32)[name = string("op_223")]; + tensor time_mask_5 = less(x = expand_dims_2, y = var_223)[name = string("time_mask_5")]; + tensor var_225_axes_0 = const()[name = string("op_225_axes_0"), val = tensor([-1])]; + tensor var_225 = expand_dims(axes = var_225_axes_0, x = time_mask_5)[name = string("op_225")]; + tensor var_227_reps_0 = const()[name = string("op_227_reps_0"), val = tensor([1, 1, 33])]; + tensor var_227 = tile(reps = var_227_reps_0, x = var_225)[name = string("op_227")]; + tensor var_233_axes_0 = const()[name = string("op_233_axes_0"), val = tensor([1])]; + string cast_8_to_fp16_dtype_0 = const()[name = string("cast_8_to_fp16_dtype_0"), val = string("fp16")]; + tensor var_227_to_fp16 = cast(dtype = cast_8_to_fp16_dtype_0, x = var_227)[name = string("cast_12")]; + tensor var_233_cast_fp16 = expand_dims(axes = var_233_axes_0, x = var_227_to_fp16)[name = string("op_233_cast_fp16")]; + tensor expanded_mask_7_reps_0 = const()[name = string("expanded_mask_7_reps_0"), val = tensor([1, 256, 1, 1])]; + tensor expanded_mask_7_cast_fp16 = tile(reps = expanded_mask_7_reps_0, x = var_233_cast_fp16)[name = string("expanded_mask_7_cast_fp16")]; + tensor input_11_cast_fp16 = mul(x = tensor_7_cast_fp16, y = expanded_mask_7_cast_fp16)[name = string("input_11_cast_fp16")]; + string tensor_9_pad_type_0 = const()[name = string("tensor_9_pad_type_0"), val = string("valid")]; + tensor tensor_9_strides_0 = const()[name = string("tensor_9_strides_0"), val = tensor([1, 1])]; + tensor tensor_9_pad_0 = const()[name = string("tensor_9_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor tensor_9_dilations_0 = const()[name = string("tensor_9_dilations_0"), val = tensor([1, 1])]; + int32 tensor_9_groups_0 = const()[name = string("tensor_9_groups_0"), val = int32(1)]; + tensor encoder_pre_encode_conv_3_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10624))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76224))))[name = string("encoder_pre_encode_conv_3_weight_to_fp16_quantized")]; + tensor encoder_pre_encode_conv_3_bias_to_fp16 = const()[name = string("encoder_pre_encode_conv_3_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76800)))]; + tensor tensor_9_cast_fp16 = conv(bias = encoder_pre_encode_conv_3_bias_to_fp16, dilations = tensor_9_dilations_0, groups = tensor_9_groups_0, pad = tensor_9_pad_0, pad_type = tensor_9_pad_type_0, strides = tensor_9_strides_0, weight = encoder_pre_encode_conv_3_weight_to_fp16_quantized, x = input_11_cast_fp16)[name = string("tensor_9_cast_fp16")]; + tensor input_13_cast_fp16 = mul(x = tensor_9_cast_fp16, y = expanded_mask_7_cast_fp16)[name = string("input_13_cast_fp16")]; + tensor tensor_11_cast_fp16 = relu(x = input_13_cast_fp16)[name = string("tensor_11_cast_fp16")]; + tensor input_15_cast_fp16 = mul(x = tensor_11_cast_fp16, y = expanded_mask_7_cast_fp16)[name = string("input_15_cast_fp16")]; + tensor input_17_pad_0 = const()[name = string("input_17_pad_0"), val = tensor([0, 0, 0, 0, 2, 1, 2, 1])]; + string input_17_mode_0 = const()[name = string("input_17_mode_0"), val = string("constant")]; + fp16 const_41_to_fp16 = const()[name = string("const_41_to_fp16"), val = fp16(0x0p+0)]; + tensor input_17_cast_fp16 = pad(constant_val = const_41_to_fp16, mode = input_17_mode_0, pad = input_17_pad_0, x = input_15_cast_fp16)[name = string("input_17_cast_fp16")]; + string tensor_13_pad_type_0 = const()[name = string("tensor_13_pad_type_0"), val = string("valid")]; + tensor tensor_13_strides_0 = const()[name = string("tensor_13_strides_0"), val = tensor([2, 2])]; + int32 tensor_13_groups_0 = const()[name = string("tensor_13_groups_0"), val = int32(256)]; + tensor tensor_13_pad_0 = const()[name = string("tensor_13_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor tensor_13_dilations_0 = const()[name = string("tensor_13_dilations_0"), val = tensor([1, 1])]; + tensor encoder_pre_encode_conv_5_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77376))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(79744))))[name = string("encoder_pre_encode_conv_5_weight_to_fp16_quantized")]; + tensor encoder_pre_encode_conv_5_bias_to_fp16 = const()[name = string("encoder_pre_encode_conv_5_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(80320)))]; + tensor tensor_13_cast_fp16 = conv(bias = encoder_pre_encode_conv_5_bias_to_fp16, dilations = tensor_13_dilations_0, groups = tensor_13_groups_0, pad = tensor_13_pad_0, pad_type = tensor_13_pad_type_0, strides = tensor_13_strides_0, weight = encoder_pre_encode_conv_5_weight_to_fp16_quantized, x = input_17_cast_fp16)[name = string("tensor_13_cast_fp16")]; + fp16 var_270_promoted_to_fp16 = const()[name = string("op_270_promoted_to_fp16"), val = fp16(0x1p+1)]; + tensor var_271_cast_fp16 = add(x = current_lengths_5_cast_fp16, y = var_270_promoted_to_fp16)[name = string("op_271_cast_fp16")]; + fp16 var_272_promoted_to_fp16 = const()[name = string("op_272_promoted_to_fp16"), val = fp16(0x1p+0)]; + tensor var_273_cast_fp16 = add(x = var_271_cast_fp16, y = var_272_promoted_to_fp16)[name = string("op_273_cast_fp16")]; + fp16 var_274_promoted_to_fp16 = const()[name = string("op_274_promoted_to_fp16"), val = fp16(0x1.8p+1)]; + tensor var_275_cast_fp16 = sub(x = var_273_cast_fp16, y = var_274_promoted_to_fp16)[name = string("op_275_cast_fp16")]; + fp16 var_55_promoted_2_to_fp16 = const()[name = string("op_55_promoted_2_to_fp16"), val = fp16(0x1p+1)]; + tensor floor_div_2_cast_fp16 = floor_div(x = var_275_cast_fp16, y = var_55_promoted_2_to_fp16)[name = string("floor_div_2_cast_fp16")]; + fp16 var_277_promoted_to_fp16 = const()[name = string("op_277_promoted_to_fp16"), val = fp16(0x1p+0)]; + tensor current_lengths_cast_fp16 = add(x = floor_div_2_cast_fp16, y = var_277_promoted_to_fp16)[name = string("current_lengths_cast_fp16")]; + string cast_9_dtype_0 = const()[name = string("cast_9_dtype_0"), val = string("int32")]; + tensor expand_dims_3 = const()[name = string("expand_dims_3"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(80896)))]; + tensor var_286_axes_0 = const()[name = string("op_286_axes_0"), val = tensor([1])]; + tensor current_lengths_cast_fp16_to_int32 = cast(dtype = cast_9_dtype_0, x = current_lengths_cast_fp16)[name = string("cast_11")]; + tensor var_286 = expand_dims(axes = var_286_axes_0, x = current_lengths_cast_fp16_to_int32)[name = string("op_286")]; + tensor time_mask = less(x = expand_dims_3, y = var_286)[name = string("time_mask")]; + tensor var_288_axes_0 = const()[name = string("op_288_axes_0"), val = tensor([-1])]; + tensor var_288 = expand_dims(axes = var_288_axes_0, x = time_mask)[name = string("op_288")]; + tensor var_290_reps_0 = const()[name = string("op_290_reps_0"), val = tensor([1, 1, 17])]; + tensor var_290 = tile(reps = var_290_reps_0, x = var_288)[name = string("op_290")]; + tensor var_296_axes_0 = const()[name = string("op_296_axes_0"), val = tensor([1])]; + string cast_10_to_fp16_dtype_0 = const()[name = string("cast_10_to_fp16_dtype_0"), val = string("fp16")]; + tensor var_290_to_fp16 = cast(dtype = cast_10_to_fp16_dtype_0, x = var_290)[name = string("cast_10")]; + tensor var_296_cast_fp16 = expand_dims(axes = var_296_axes_0, x = var_290_to_fp16)[name = string("op_296_cast_fp16")]; + tensor expanded_mask_13_reps_0 = const()[name = string("expanded_mask_13_reps_0"), val = tensor([1, 256, 1, 1])]; + tensor expanded_mask_13_cast_fp16 = tile(reps = expanded_mask_13_reps_0, x = var_296_cast_fp16)[name = string("expanded_mask_13_cast_fp16")]; + tensor input_19_cast_fp16 = mul(x = tensor_13_cast_fp16, y = expanded_mask_13_cast_fp16)[name = string("input_19_cast_fp16")]; + string tensor_15_pad_type_0 = const()[name = string("tensor_15_pad_type_0"), val = string("valid")]; + tensor tensor_15_strides_0 = const()[name = string("tensor_15_strides_0"), val = tensor([1, 1])]; + tensor tensor_15_pad_0 = const()[name = string("tensor_15_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor tensor_15_dilations_0 = const()[name = string("tensor_15_dilations_0"), val = tensor([1, 1])]; + int32 tensor_15_groups_0 = const()[name = string("tensor_15_groups_0"), val = int32(1)]; + tensor encoder_pre_encode_conv_6_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81216))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(146816))))[name = string("encoder_pre_encode_conv_6_weight_to_fp16_quantized")]; + tensor encoder_pre_encode_conv_6_bias_to_fp16 = const()[name = string("encoder_pre_encode_conv_6_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(147392)))]; + tensor tensor_15_cast_fp16 = conv(bias = encoder_pre_encode_conv_6_bias_to_fp16, dilations = tensor_15_dilations_0, groups = tensor_15_groups_0, pad = tensor_15_pad_0, pad_type = tensor_15_pad_type_0, strides = tensor_15_strides_0, weight = encoder_pre_encode_conv_6_weight_to_fp16_quantized, x = input_19_cast_fp16)[name = string("tensor_15_cast_fp16")]; + tensor input_21_cast_fp16 = mul(x = tensor_15_cast_fp16, y = expanded_mask_13_cast_fp16)[name = string("input_21_cast_fp16")]; + tensor tensor_cast_fp16 = relu(x = input_21_cast_fp16)[name = string("tensor_cast_fp16")]; + tensor x_3_cast_fp16 = mul(x = tensor_cast_fp16, y = expanded_mask_13_cast_fp16)[name = string("x_3_cast_fp16")]; + tensor var_330_perm_0 = const()[name = string("op_330_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_331 = const()[name = string("op_331"), val = tensor([1, 58, -1])]; + tensor var_330_cast_fp16 = transpose(perm = var_330_perm_0, x = x_3_cast_fp16)[name = string("transpose_366")]; + tensor input_23_cast_fp16 = reshape(shape = var_331, x = var_330_cast_fp16)[name = string("input_23_cast_fp16")]; + tensor encoder_pre_encode_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(147968))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4604480))))[name = string("encoder_pre_encode_out_weight_to_fp16_quantized")]; + tensor encoder_pre_encode_out_bias_to_fp16 = const()[name = string("encoder_pre_encode_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4606592)))]; + tensor linear_0_cast_fp16 = linear(bias = encoder_pre_encode_out_bias_to_fp16, weight = encoder_pre_encode_out_weight_to_fp16_quantized, x = input_23_cast_fp16)[name = string("linear_0_cast_fp16")]; + tensor var_341_begin_0 = const()[name = string("op_341_begin_0"), val = tensor([0, 2, 0])]; + tensor var_341_end_0 = const()[name = string("op_341_end_0"), val = tensor([1, 58, 1024])]; + tensor var_341_end_mask_0 = const()[name = string("op_341_end_mask_0"), val = tensor([true, true, true])]; + tensor var_341_cast_fp16 = slice_by_index(begin = var_341_begin_0, end = var_341_end_0, end_mask = var_341_end_mask_0, x = linear_0_cast_fp16)[name = string("op_341_cast_fp16")]; + int32 var_343 = const()[name = string("op_343"), val = int32(2)]; + tensor var_344 = sub(x = current_lengths_cast_fp16_to_int32, y = var_343)[name = string("op_344")]; + string var_344_promoted_to_fp16_dtype_0 = const()[name = string("op_344_promoted_to_fp16_dtype_0"), val = string("fp16")]; + fp16 var_61_promoted_to_fp16 = const()[name = string("op_61_promoted_to_fp16"), val = fp16(0x0p+0)]; + fp16 const_61_to_fp16 = const()[name = string("const_61_to_fp16"), val = fp16(inf)]; + tensor var_344_to_fp16 = cast(dtype = var_344_promoted_to_fp16_dtype_0, x = var_344)[name = string("cast_9")]; + tensor clip_0_cast_fp16 = clip(alpha = var_61_promoted_to_fp16, beta = const_61_to_fp16, x = var_344_to_fp16)[name = string("clip_0_cast_fp16")]; + tensor max_audio_length_1 = const()[name = string("max_audio_length_1"), val = tensor([56])]; + fp16 var_360_promoted_to_fp16 = const()[name = string("op_360_promoted_to_fp16"), val = fp16(0x1.5p+5)]; + tensor padding_length_cast_fp16 = add(x = clip_0_cast_fp16, y = var_360_promoted_to_fp16)[name = string("padding_length_cast_fp16")]; + int32 const_63 = const()[name = string("const_63"), val = int32(-1)]; + tensor var_362 = mul(x = cache_len, y = const_63)[name = string("op_362")]; + int32 var_363 = const()[name = string("op_363"), val = int32(42)]; + tensor offset = add(x = var_362, y = var_363)[name = string("offset")]; + tensor var_403_axes_0 = const()[name = string("op_403_axes_0"), val = tensor([-1])]; + tensor var_403_cast_fp16 = expand_dims(axes = var_403_axes_0, x = padding_length_cast_fp16)[name = string("op_403_cast_fp16")]; + tensor var_402_promoted_to_fp16 = const()[name = string("op_402_promoted_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4608704)))]; + tensor pad_mask_1_cast_fp16 = less(x = var_402_promoted_to_fp16, y = var_403_cast_fp16)[name = string("pad_mask_1_cast_fp16")]; + tensor expand_dims_5 = const()[name = string("expand_dims_5"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4609024)))]; + tensor var_409_axes_0 = const()[name = string("op_409_axes_0"), val = tensor([-1])]; + tensor var_409 = expand_dims(axes = var_409_axes_0, x = offset)[name = string("op_409")]; + tensor pad_mask_off = greater_equal(x = expand_dims_5, y = var_409)[name = string("pad_mask_off")]; + tensor pad_mask_3 = logical_and(x = pad_mask_off, y = pad_mask_1_cast_fp16)[name = string("pad_mask_3")]; + tensor var_412_axes_0 = const()[name = string("op_412_axes_0"), val = tensor([1])]; + tensor var_412 = expand_dims(axes = var_412_axes_0, x = pad_mask_3)[name = string("op_412")]; + tensor var_413 = const()[name = string("op_413"), val = tensor([1, 98, 1])]; + tensor pad_mask_for_att_mask_1 = tile(reps = var_413, x = var_412)[name = string("pad_mask_for_att_mask_1")]; + tensor var_415_perm_0 = const()[name = string("op_415_perm_0"), val = tensor([0, 2, 1])]; + tensor var_415 = transpose(perm = var_415_perm_0, x = pad_mask_for_att_mask_1)[name = string("transpose_365")]; + tensor pad_mask_for_att_mask = logical_and(x = pad_mask_for_att_mask_1, y = var_415)[name = string("pad_mask_for_att_mask")]; + tensor const_71 = const()[name = string("const_71"), val = tensor([[[true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true]]])]; + tensor att_mask_9 = logical_and(x = pad_mask_for_att_mask, y = const_71)[name = string("att_mask_9")]; + tensor att_mask = logical_not(x = att_mask_9)[name = string("att_mask")]; + tensor pad_mask_5 = logical_not(x = pad_mask_3)[name = string("pad_mask_5")]; + tensor pad_mask_begin_0 = const()[name = string("pad_mask_begin_0"), val = tensor([0, 42])]; + tensor pad_mask_end_0 = const()[name = string("pad_mask_end_0"), val = tensor([1, 98])]; + tensor pad_mask_end_mask_0 = const()[name = string("pad_mask_end_mask_0"), val = tensor([true, true])]; + tensor pad_mask = slice_by_index(begin = pad_mask_begin_0, end = pad_mask_end_0, end_mask = pad_mask_end_mask_0, x = pad_mask_5)[name = string("pad_mask")]; + tensor mask_9_begin_0 = const()[name = string("mask_9_begin_0"), val = tensor([0, 42, 0])]; + tensor mask_9_end_0 = const()[name = string("mask_9_end_0"), val = tensor([1, 98, 98])]; + tensor mask_9_end_mask_0 = const()[name = string("mask_9_end_mask_0"), val = tensor([true, true, true])]; + tensor mask_9 = slice_by_index(begin = mask_9_begin_0, end = mask_9_end_0, end_mask = mask_9_end_mask_0, x = att_mask)[name = string("mask_9")]; + tensor cache_1_begin_0 = const()[name = string("cache_1_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor cache_1_end_0 = const()[name = string("cache_1_end_0"), val = tensor([1, 1, 42, 1024])]; + tensor cache_1_end_mask_0 = const()[name = string("cache_1_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_1_squeeze_mask_0 = const()[name = string("cache_1_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_channel_to_fp16 = cast(dtype = cache_channel_to_fp16_dtype_0, x = cache_channel)[name = string("cast_8")]; + tensor value_3_cast_fp16 = transpose(perm = value_3_perm_0, x = cache_channel_to_fp16)[name = string("transpose_364")]; + tensor cache_1_cast_fp16 = slice_by_index(begin = cache_1_begin_0, end = cache_1_end_0, end_mask = cache_1_end_mask_0, squeeze_mask = cache_1_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_1_cast_fp16")]; + tensor cache_3_begin_0 = const()[name = string("cache_3_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor cache_3_end_0 = const()[name = string("cache_3_end_0"), val = tensor([1, 1, 1024, 8])]; + tensor cache_3_end_mask_0 = const()[name = string("cache_3_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_3_squeeze_mask_0 = const()[name = string("cache_3_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_time_to_fp16 = cast(dtype = cache_time_to_fp16_dtype_0, x = cache_time)[name = string("cast_7")]; + tensor value_5_cast_fp16 = transpose(perm = value_5_perm_0, x = cache_time_to_fp16)[name = string("transpose_363")]; + tensor cache_3_cast_fp16 = slice_by_index(begin = cache_3_begin_0, end = cache_3_end_0, end_mask = cache_3_end_mask_0, squeeze_mask = cache_3_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_3_cast_fp16")]; + tensor input_27_axes_0 = const()[name = string("input_27_axes_0"), val = tensor([-1])]; + tensor encoder_layers_0_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_0_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4609536)))]; + tensor encoder_layers_0_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_0_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4611648)))]; + fp16 var_41_to_fp16 = const()[name = string("op_41_to_fp16"), val = fp16(0x1.5p-17)]; + tensor input_27_cast_fp16 = layer_norm(axes = input_27_axes_0, beta = encoder_layers_0_norm_feed_forward1_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_0_norm_feed_forward1_weight_to_fp16, x = var_341_cast_fp16)[name = string("input_27_cast_fp16")]; + tensor encoder_layers_0_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4613760))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8808128))))[name = string("encoder_layers_0_feed_forward1_linear1_weight_to_fp16_quantized")]; + tensor encoder_layers_0_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_0_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8816384)))]; + tensor linear_1_cast_fp16 = linear(bias = encoder_layers_0_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_0_feed_forward1_linear1_weight_to_fp16_quantized, x = input_27_cast_fp16)[name = string("linear_1_cast_fp16")]; + tensor input_31_cast_fp16 = silu(x = linear_1_cast_fp16)[name = string("input_31_cast_fp16")]; + tensor encoder_layers_0_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8824640))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13019008))))[name = string("encoder_layers_0_feed_forward1_linear2_weight_to_fp16_quantized")]; + tensor encoder_layers_0_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_0_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13021120)))]; + tensor linear_2_cast_fp16 = linear(bias = encoder_layers_0_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_0_feed_forward1_linear2_weight_to_fp16_quantized, x = input_31_cast_fp16)[name = string("linear_2_cast_fp16")]; + fp16 var_454_to_fp16 = const()[name = string("op_454_to_fp16"), val = fp16(0x1p-1)]; + tensor var_455_cast_fp16 = mul(x = linear_2_cast_fp16, y = var_454_to_fp16)[name = string("op_455_cast_fp16")]; + tensor input_37_cast_fp16 = add(x = var_341_cast_fp16, y = var_455_cast_fp16)[name = string("input_37_cast_fp16")]; + tensor key_1_axes_0 = const()[name = string("key_1_axes_0"), val = tensor([-1])]; + tensor encoder_layers_0_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_0_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13023232)))]; + tensor encoder_layers_0_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_0_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13025344)))]; + tensor key_1_cast_fp16 = layer_norm(axes = key_1_axes_0, beta = encoder_layers_0_norm_self_att_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_0_norm_self_att_weight_to_fp16, x = input_37_cast_fp16)[name = string("key_1_cast_fp16")]; + bool input_39_interleave_0 = const()[name = string("input_39_interleave_0"), val = bool(false)]; + tensor input_39_cast_fp16 = concat(axis = var_67, interleave = input_39_interleave_0, values = (cache_1_cast_fp16, key_1_cast_fp16))[name = string("input_39_cast_fp16")]; + bool var_483_interleave_0 = const()[name = string("op_483_interleave_0"), val = bool(false)]; + tensor var_483_cast_fp16 = concat(axis = var_67, interleave = var_483_interleave_0, values = key_1_cast_fp16)[name = string("op_483_cast_fp16")]; + tensor encoder_layers_0_self_attn_linear_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13027456))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14076096))))[name = string("encoder_layers_0_self_attn_linear_q_weight_to_fp16_quantized")]; + tensor encoder_layers_0_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_0_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14078208)))]; + tensor linear_3_cast_fp16 = linear(bias = encoder_layers_0_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_0_self_attn_linear_q_weight_to_fp16_quantized, x = key_1_cast_fp16)[name = string("linear_3_cast_fp16")]; + tensor var_488 = const()[name = string("op_488"), val = tensor([1, -1, 8, 128])]; + tensor q_1_cast_fp16 = reshape(shape = var_488, x = linear_3_cast_fp16)[name = string("q_1_cast_fp16")]; + tensor encoder_layers_0_self_attn_linear_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14080320))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15128960))))[name = string("encoder_layers_0_self_attn_linear_k_weight_to_fp16_quantized")]; + tensor encoder_layers_0_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_0_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15131072)))]; + tensor linear_4_cast_fp16 = linear(bias = encoder_layers_0_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_0_self_attn_linear_k_weight_to_fp16_quantized, x = input_39_cast_fp16)[name = string("linear_4_cast_fp16")]; + tensor var_493 = const()[name = string("op_493"), val = tensor([1, -1, 8, 128])]; + tensor k_1_cast_fp16 = reshape(shape = var_493, x = linear_4_cast_fp16)[name = string("k_1_cast_fp16")]; + tensor encoder_layers_0_self_attn_linear_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15133184))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16181824))))[name = string("encoder_layers_0_self_attn_linear_v_weight_to_fp16_quantized")]; + tensor encoder_layers_0_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_0_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16183936)))]; + tensor linear_5_cast_fp16 = linear(bias = encoder_layers_0_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_0_self_attn_linear_v_weight_to_fp16_quantized, x = input_39_cast_fp16)[name = string("linear_5_cast_fp16")]; + tensor var_498 = const()[name = string("op_498"), val = tensor([1, -1, 8, 128])]; + tensor v_1_cast_fp16 = reshape(shape = var_498, x = linear_5_cast_fp16)[name = string("v_1_cast_fp16")]; + tensor value_9_perm_0 = const()[name = string("value_9_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_0_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_0_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16186048)))]; + tensor var_511_cast_fp16 = add(x = q_1_cast_fp16, y = encoder_layers_0_self_attn_pos_bias_u_to_fp16)[name = string("op_511_cast_fp16")]; + tensor encoder_layers_0_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_0_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16188160)))]; + tensor var_513_cast_fp16 = add(x = q_1_cast_fp16, y = encoder_layers_0_self_attn_pos_bias_v_to_fp16)[name = string("op_513_cast_fp16")]; + tensor q_with_bias_v_1_perm_0 = const()[name = string("q_with_bias_v_1_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_7_transpose_x_0 = const()[name = string("x_7_transpose_x_0"), val = bool(false)]; + bool x_7_transpose_y_0 = const()[name = string("x_7_transpose_y_0"), val = bool(false)]; + tensor op_515_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16190272))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16390016))))[name = string("op_515_to_fp16_quantized")]; + tensor q_with_bias_v_1_cast_fp16 = transpose(perm = q_with_bias_v_1_perm_0, x = var_513_cast_fp16)[name = string("transpose_362")]; + tensor x_7_cast_fp16 = matmul(transpose_x = x_7_transpose_x_0, transpose_y = x_7_transpose_y_0, x = q_with_bias_v_1_cast_fp16, y = op_515_to_fp16_quantized)[name = string("x_7_cast_fp16")]; + tensor x_9_pad_0 = const()[name = string("x_9_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_9_mode_0 = const()[name = string("x_9_mode_0"), val = string("constant")]; + fp16 const_79_to_fp16 = const()[name = string("const_79_to_fp16"), val = fp16(0x0p+0)]; + tensor x_9_cast_fp16 = pad(constant_val = const_79_to_fp16, mode = x_9_mode_0, pad = x_9_pad_0, x = x_7_cast_fp16)[name = string("x_9_cast_fp16")]; + tensor var_523 = const()[name = string("op_523"), val = tensor([1, 8, -1, 56])]; + tensor x_11_cast_fp16 = reshape(shape = var_523, x = x_9_cast_fp16)[name = string("x_11_cast_fp16")]; + tensor var_527_begin_0 = const()[name = string("op_527_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_527_end_0 = const()[name = string("op_527_end_0"), val = tensor([1, 8, 196, 56])]; + tensor var_527_end_mask_0 = const()[name = string("op_527_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_527_cast_fp16 = slice_by_index(begin = var_527_begin_0, end = var_527_end_0, end_mask = var_527_end_mask_0, x = x_11_cast_fp16)[name = string("op_527_cast_fp16")]; + tensor var_528 = const()[name = string("op_528"), val = tensor([1, 8, 56, 195])]; + tensor matrix_bd_1_cast_fp16 = reshape(shape = var_528, x = var_527_cast_fp16)[name = string("matrix_bd_1_cast_fp16")]; + bool matrix_ac_1_transpose_x_0 = const()[name = string("matrix_ac_1_transpose_x_0"), val = bool(false)]; + bool matrix_ac_1_transpose_y_0 = const()[name = string("matrix_ac_1_transpose_y_0"), val = bool(false)]; + tensor transpose_96_perm_0 = const()[name = string("transpose_96_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_97_perm_0 = const()[name = string("transpose_97_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_97 = transpose(perm = transpose_97_perm_0, x = k_1_cast_fp16)[name = string("transpose_360")]; + tensor transpose_96 = transpose(perm = transpose_96_perm_0, x = var_511_cast_fp16)[name = string("transpose_361")]; + tensor matrix_ac_1_cast_fp16 = matmul(transpose_x = matrix_ac_1_transpose_x_0, transpose_y = matrix_ac_1_transpose_y_0, x = transpose_96, y = transpose_97)[name = string("matrix_ac_1_cast_fp16")]; + tensor matrix_bd_3_begin_0 = const()[name = string("matrix_bd_3_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_3_end_0 = const()[name = string("matrix_bd_3_end_0"), val = tensor([1, 8, 56, 98])]; + tensor matrix_bd_3_end_mask_0 = const()[name = string("matrix_bd_3_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_3_cast_fp16 = slice_by_index(begin = matrix_bd_3_begin_0, end = matrix_bd_3_end_0, end_mask = matrix_bd_3_end_mask_0, x = matrix_bd_1_cast_fp16)[name = string("matrix_bd_3_cast_fp16")]; + tensor var_537_cast_fp16 = add(x = matrix_ac_1_cast_fp16, y = matrix_bd_3_cast_fp16)[name = string("op_537_cast_fp16")]; + fp16 _inversed_scores_1_y_0_to_fp16 = const()[name = string("_inversed_scores_1_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_1_cast_fp16 = mul(x = var_537_cast_fp16, y = _inversed_scores_1_y_0_to_fp16)[name = string("_inversed_scores_1_cast_fp16")]; + tensor mask_11_axes_0 = const()[name = string("mask_11_axes_0"), val = tensor([1])]; + tensor mask_11 = expand_dims(axes = mask_11_axes_0, x = mask_9)[name = string("mask_11")]; + fp16 var_44_to_fp16 = const()[name = string("op_44_to_fp16"), val = fp16(-0x1.388p+13)]; + tensor scores_3_cast_fp16 = select(a = var_44_to_fp16, b = _inversed_scores_1_cast_fp16, cond = mask_11)[name = string("scores_3_cast_fp16")]; + tensor var_543_cast_fp16 = softmax(axis = var_58, x = scores_3_cast_fp16)[name = string("op_543_cast_fp16")]; + fp16 var_43_to_fp16 = const()[name = string("op_43_to_fp16"), val = fp16(0x0p+0)]; + tensor input_41_cast_fp16 = select(a = var_43_to_fp16, b = var_543_cast_fp16, cond = mask_11)[name = string("input_41_cast_fp16")]; + bool x_13_transpose_x_0 = const()[name = string("x_13_transpose_x_0"), val = bool(false)]; + bool x_13_transpose_y_0 = const()[name = string("x_13_transpose_y_0"), val = bool(false)]; + tensor value_9_cast_fp16 = transpose(perm = value_9_perm_0, x = v_1_cast_fp16)[name = string("transpose_359")]; + tensor x_13_cast_fp16 = matmul(transpose_x = x_13_transpose_x_0, transpose_y = x_13_transpose_y_0, x = input_41_cast_fp16, y = value_9_cast_fp16)[name = string("x_13_cast_fp16")]; + tensor var_547_perm_0 = const()[name = string("op_547_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_548 = const()[name = string("op_548"), val = tensor([1, -1, 1024])]; + tensor var_547_cast_fp16 = transpose(perm = var_547_perm_0, x = x_13_cast_fp16)[name = string("transpose_358")]; + tensor input_43_cast_fp16 = reshape(shape = var_548, x = var_547_cast_fp16)[name = string("input_43_cast_fp16")]; + tensor encoder_layers_0_self_attn_linear_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16390528))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17439168))))[name = string("encoder_layers_0_self_attn_linear_out_weight_to_fp16_quantized")]; + tensor encoder_layers_0_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_0_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17441280)))]; + tensor linear_7_cast_fp16 = linear(bias = encoder_layers_0_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_0_self_attn_linear_out_weight_to_fp16_quantized, x = input_43_cast_fp16)[name = string("linear_7_cast_fp16")]; + tensor input_47_cast_fp16 = add(x = input_37_cast_fp16, y = linear_7_cast_fp16)[name = string("input_47_cast_fp16")]; + tensor x_17_axes_0 = const()[name = string("x_17_axes_0"), val = tensor([-1])]; + tensor encoder_layers_0_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_0_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17443392)))]; + tensor encoder_layers_0_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_0_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17445504)))]; + tensor x_17_cast_fp16 = layer_norm(axes = x_17_axes_0, beta = encoder_layers_0_norm_conv_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_0_norm_conv_weight_to_fp16, x = input_47_cast_fp16)[name = string("x_17_cast_fp16")]; + tensor input_49_perm_0 = const()[name = string("input_49_perm_0"), val = tensor([0, 2, 1])]; + string input_51_pad_type_0 = const()[name = string("input_51_pad_type_0"), val = string("valid")]; + tensor input_51_strides_0 = const()[name = string("input_51_strides_0"), val = tensor([1])]; + tensor input_51_pad_0 = const()[name = string("input_51_pad_0"), val = tensor([0, 0])]; + tensor input_51_dilations_0 = const()[name = string("input_51_dilations_0"), val = tensor([1])]; + int32 input_51_groups_0 = const()[name = string("input_51_groups_0"), val = int32(1)]; + tensor encoder_layers_0_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17447616))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19544832))))[name = string("encoder_layers_0_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_49_cast_fp16 = transpose(perm = input_49_perm_0, x = x_17_cast_fp16)[name = string("transpose_357")]; + tensor input_51_cast_fp16 = conv(dilations = input_51_dilations_0, groups = input_51_groups_0, pad = input_51_pad_0, pad_type = input_51_pad_type_0, strides = input_51_strides_0, weight = encoder_layers_0_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_49_cast_fp16)[name = string("input_51_cast_fp16")]; + int32 x_19_split_num_splits_0 = const()[name = string("x_19_split_num_splits_0"), val = int32(2)]; + int32 x_19_split_axis_0 = const()[name = string("x_19_split_axis_0"), val = int32(1)]; + tensor x_19_split_cast_fp16_0, tensor x_19_split_cast_fp16_1 = split(axis = x_19_split_axis_0, num_splits = x_19_split_num_splits_0, x = input_51_cast_fp16)[name = string("x_19_split_cast_fp16")]; + tensor x_19_split_1_sigmoid_cast_fp16 = sigmoid(x = x_19_split_cast_fp16_1)[name = string("x_19_split_1_sigmoid_cast_fp16")]; + tensor x_19_cast_fp16 = mul(x = x_19_split_cast_fp16_0, y = x_19_split_1_sigmoid_cast_fp16)[name = string("x_19_cast_fp16")]; + tensor var_574_axes_0 = const()[name = string("op_574_axes_0"), val = tensor([1])]; + tensor var_574 = expand_dims(axes = var_574_axes_0, x = pad_mask)[name = string("op_574")]; + tensor input_53_cast_fp16 = select(a = var_43_to_fp16, b = x_19_cast_fp16, cond = var_574)[name = string("input_53_cast_fp16")]; + bool new_x_3_interleave_0 = const()[name = string("new_x_3_interleave_0"), val = bool(false)]; + tensor new_x_3_cast_fp16 = concat(axis = var_58, interleave = new_x_3_interleave_0, values = (cache_3_cast_fp16, input_53_cast_fp16))[name = string("new_x_3_cast_fp16")]; + tensor var_587_begin_0 = const()[name = string("op_587_begin_0"), val = tensor([0, 0, 56])]; + tensor var_587_end_0 = const()[name = string("op_587_end_0"), val = tensor([1, 1024, 64])]; + tensor var_587_end_mask_0 = const()[name = string("op_587_end_mask_0"), val = tensor([true, true, true])]; + tensor var_587_cast_fp16 = slice_by_index(begin = var_587_begin_0, end = var_587_end_0, end_mask = var_587_end_mask_0, x = new_x_3_cast_fp16)[name = string("op_587_cast_fp16")]; + string x_21_pad_type_0 = const()[name = string("x_21_pad_type_0"), val = string("valid")]; + int32 x_21_groups_0 = const()[name = string("x_21_groups_0"), val = int32(1024)]; + tensor x_21_strides_0 = const()[name = string("x_21_strides_0"), val = tensor([1])]; + tensor x_21_pad_0 = const()[name = string("x_21_pad_0"), val = tensor([0, 0])]; + tensor x_21_dilations_0 = const()[name = string("x_21_dilations_0"), val = tensor([1])]; + tensor encoder_layers_0_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19548992))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19558272))))[name = string("encoder_layers_0_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_21_cast_fp16 = conv(dilations = x_21_dilations_0, groups = x_21_groups_0, pad = x_21_pad_0, pad_type = x_21_pad_type_0, strides = x_21_strides_0, weight = encoder_layers_0_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_3_cast_fp16)[name = string("x_21_cast_fp16")]; + tensor input_55_perm_0 = const()[name = string("input_55_perm_0"), val = tensor([0, 2, 1])]; + tensor x_23_axes_0 = const()[name = string("x_23_axes_0"), val = tensor([-1])]; + tensor encoder_layers_0_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_0_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19560384)))]; + tensor encoder_layers_0_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_0_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19562496)))]; + tensor input_55_cast_fp16 = transpose(perm = input_55_perm_0, x = x_21_cast_fp16)[name = string("transpose_356")]; + tensor x_23_cast_fp16 = layer_norm(axes = x_23_axes_0, beta = encoder_layers_0_conv_batch_norm_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_0_conv_batch_norm_weight_to_fp16, x = input_55_cast_fp16)[name = string("x_23_cast_fp16")]; + tensor input_57_perm_0 = const()[name = string("input_57_perm_0"), val = tensor([0, 2, 1])]; + tensor input_57_cast_fp16 = transpose(perm = input_57_perm_0, x = x_23_cast_fp16)[name = string("transpose_355")]; + tensor input_59_cast_fp16 = silu(x = input_57_cast_fp16)[name = string("input_59_cast_fp16")]; + string x_25_pad_type_0 = const()[name = string("x_25_pad_type_0"), val = string("valid")]; + tensor x_25_strides_0 = const()[name = string("x_25_strides_0"), val = tensor([1])]; + tensor x_25_pad_0 = const()[name = string("x_25_pad_0"), val = tensor([0, 0])]; + tensor x_25_dilations_0 = const()[name = string("x_25_dilations_0"), val = tensor([1])]; + int32 x_25_groups_0 = const()[name = string("x_25_groups_0"), val = int32(1)]; + tensor encoder_layers_0_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19564608))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20613248))))[name = string("encoder_layers_0_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_25_cast_fp16 = conv(dilations = x_25_dilations_0, groups = x_25_groups_0, pad = x_25_pad_0, pad_type = x_25_pad_type_0, strides = x_25_strides_0, weight = encoder_layers_0_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_59_cast_fp16)[name = string("x_25_cast_fp16")]; + tensor input_61_perm_0 = const()[name = string("input_61_perm_0"), val = tensor([0, 2, 1])]; + tensor input_61_cast_fp16 = transpose(perm = input_61_perm_0, x = x_25_cast_fp16)[name = string("transpose_354")]; + tensor input_63_cast_fp16 = add(x = input_47_cast_fp16, y = input_61_cast_fp16)[name = string("input_63_cast_fp16")]; + tensor input_65_axes_0 = const()[name = string("input_65_axes_0"), val = tensor([-1])]; + tensor encoder_layers_0_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_0_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20615360)))]; + tensor encoder_layers_0_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_0_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20617472)))]; + tensor input_65_cast_fp16 = layer_norm(axes = input_65_axes_0, beta = encoder_layers_0_norm_feed_forward2_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_0_norm_feed_forward2_weight_to_fp16, x = input_63_cast_fp16)[name = string("input_65_cast_fp16")]; + tensor encoder_layers_0_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20619584))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24813952))))[name = string("encoder_layers_0_feed_forward2_linear1_weight_to_fp16_quantized")]; + tensor encoder_layers_0_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_0_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24822208)))]; + tensor linear_8_cast_fp16 = linear(bias = encoder_layers_0_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_0_feed_forward2_linear1_weight_to_fp16_quantized, x = input_65_cast_fp16)[name = string("linear_8_cast_fp16")]; + tensor input_69_cast_fp16 = silu(x = linear_8_cast_fp16)[name = string("input_69_cast_fp16")]; + tensor encoder_layers_0_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24830464))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29024832))))[name = string("encoder_layers_0_feed_forward2_linear2_weight_to_fp16_quantized")]; + tensor encoder_layers_0_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_0_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29026944)))]; + tensor linear_9_cast_fp16 = linear(bias = encoder_layers_0_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_0_feed_forward2_linear2_weight_to_fp16_quantized, x = input_69_cast_fp16)[name = string("linear_9_cast_fp16")]; + fp16 var_630_to_fp16 = const()[name = string("op_630_to_fp16"), val = fp16(0x1p-1)]; + tensor var_631_cast_fp16 = mul(x = linear_9_cast_fp16, y = var_630_to_fp16)[name = string("op_631_cast_fp16")]; + tensor input_75_cast_fp16 = add(x = input_63_cast_fp16, y = var_631_cast_fp16)[name = string("input_75_cast_fp16")]; + tensor input_77_axes_0 = const()[name = string("input_77_axes_0"), val = tensor([-1])]; + tensor encoder_layers_0_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_0_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29029056)))]; + tensor encoder_layers_0_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_0_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29031168)))]; + tensor input_77_cast_fp16 = layer_norm(axes = input_77_axes_0, beta = encoder_layers_0_norm_out_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_0_norm_out_weight_to_fp16, x = input_75_cast_fp16)[name = string("input_77_cast_fp16")]; + tensor cache_5_begin_0 = const()[name = string("cache_5_begin_0"), val = tensor([1, 0, 0, 0])]; + tensor cache_5_end_0 = const()[name = string("cache_5_end_0"), val = tensor([2, 1, 42, 1024])]; + tensor cache_5_end_mask_0 = const()[name = string("cache_5_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_5_squeeze_mask_0 = const()[name = string("cache_5_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_5_cast_fp16 = slice_by_index(begin = cache_5_begin_0, end = cache_5_end_0, end_mask = cache_5_end_mask_0, squeeze_mask = cache_5_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_5_cast_fp16")]; + tensor cache_7_begin_0 = const()[name = string("cache_7_begin_0"), val = tensor([1, 0, 0, 0])]; + tensor cache_7_end_0 = const()[name = string("cache_7_end_0"), val = tensor([2, 1, 1024, 8])]; + tensor cache_7_end_mask_0 = const()[name = string("cache_7_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_7_squeeze_mask_0 = const()[name = string("cache_7_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_7_cast_fp16 = slice_by_index(begin = cache_7_begin_0, end = cache_7_end_0, end_mask = cache_7_end_mask_0, squeeze_mask = cache_7_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_7_cast_fp16")]; + tensor input_79_axes_0 = const()[name = string("input_79_axes_0"), val = tensor([-1])]; + tensor encoder_layers_1_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_1_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29033280)))]; + tensor encoder_layers_1_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_1_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29035392)))]; + tensor input_79_cast_fp16 = layer_norm(axes = input_79_axes_0, beta = encoder_layers_1_norm_feed_forward1_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_1_norm_feed_forward1_weight_to_fp16, x = input_77_cast_fp16)[name = string("input_79_cast_fp16")]; + tensor encoder_layers_1_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29037504))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33231872))))[name = string("encoder_layers_1_feed_forward1_linear1_weight_to_fp16_quantized")]; + tensor encoder_layers_1_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_1_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33240128)))]; + tensor linear_10_cast_fp16 = linear(bias = encoder_layers_1_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_1_feed_forward1_linear1_weight_to_fp16_quantized, x = input_79_cast_fp16)[name = string("linear_10_cast_fp16")]; + tensor input_83_cast_fp16 = silu(x = linear_10_cast_fp16)[name = string("input_83_cast_fp16")]; + tensor encoder_layers_1_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33248384))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37442752))))[name = string("encoder_layers_1_feed_forward1_linear2_weight_to_fp16_quantized")]; + tensor encoder_layers_1_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_1_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37444864)))]; + tensor linear_11_cast_fp16 = linear(bias = encoder_layers_1_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_1_feed_forward1_linear2_weight_to_fp16_quantized, x = input_83_cast_fp16)[name = string("linear_11_cast_fp16")]; + fp16 var_667_to_fp16 = const()[name = string("op_667_to_fp16"), val = fp16(0x1p-1)]; + tensor var_668_cast_fp16 = mul(x = linear_11_cast_fp16, y = var_667_to_fp16)[name = string("op_668_cast_fp16")]; + tensor input_89_cast_fp16 = add(x = input_77_cast_fp16, y = var_668_cast_fp16)[name = string("input_89_cast_fp16")]; + tensor key_3_axes_0 = const()[name = string("key_3_axes_0"), val = tensor([-1])]; + tensor encoder_layers_1_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_1_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37446976)))]; + tensor encoder_layers_1_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_1_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37449088)))]; + tensor key_3_cast_fp16 = layer_norm(axes = key_3_axes_0, beta = encoder_layers_1_norm_self_att_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_1_norm_self_att_weight_to_fp16, x = input_89_cast_fp16)[name = string("key_3_cast_fp16")]; + bool input_91_interleave_0 = const()[name = string("input_91_interleave_0"), val = bool(false)]; + tensor input_91_cast_fp16 = concat(axis = var_67, interleave = input_91_interleave_0, values = (cache_5_cast_fp16, key_3_cast_fp16))[name = string("input_91_cast_fp16")]; + bool var_696_interleave_0 = const()[name = string("op_696_interleave_0"), val = bool(false)]; + tensor var_696_cast_fp16 = concat(axis = var_67, interleave = var_696_interleave_0, values = key_3_cast_fp16)[name = string("op_696_cast_fp16")]; + tensor encoder_layers_1_self_attn_linear_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37451200))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38499840))))[name = string("encoder_layers_1_self_attn_linear_q_weight_to_fp16_quantized")]; + tensor encoder_layers_1_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_1_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38501952)))]; + tensor linear_12_cast_fp16 = linear(bias = encoder_layers_1_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_1_self_attn_linear_q_weight_to_fp16_quantized, x = key_3_cast_fp16)[name = string("linear_12_cast_fp16")]; + tensor var_701 = const()[name = string("op_701"), val = tensor([1, -1, 8, 128])]; + tensor q_7_cast_fp16 = reshape(shape = var_701, x = linear_12_cast_fp16)[name = string("q_7_cast_fp16")]; + tensor encoder_layers_1_self_attn_linear_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38504064))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39552704))))[name = string("encoder_layers_1_self_attn_linear_k_weight_to_fp16_quantized")]; + tensor encoder_layers_1_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_1_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39554816)))]; + tensor linear_13_cast_fp16 = linear(bias = encoder_layers_1_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_1_self_attn_linear_k_weight_to_fp16_quantized, x = input_91_cast_fp16)[name = string("linear_13_cast_fp16")]; + tensor var_706 = const()[name = string("op_706"), val = tensor([1, -1, 8, 128])]; + tensor k_5_cast_fp16 = reshape(shape = var_706, x = linear_13_cast_fp16)[name = string("k_5_cast_fp16")]; + tensor encoder_layers_1_self_attn_linear_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39556928))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40605568))))[name = string("encoder_layers_1_self_attn_linear_v_weight_to_fp16_quantized")]; + tensor encoder_layers_1_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_1_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40607680)))]; + tensor linear_14_cast_fp16 = linear(bias = encoder_layers_1_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_1_self_attn_linear_v_weight_to_fp16_quantized, x = input_91_cast_fp16)[name = string("linear_14_cast_fp16")]; + tensor var_711 = const()[name = string("op_711"), val = tensor([1, -1, 8, 128])]; + tensor v_3_cast_fp16 = reshape(shape = var_711, x = linear_14_cast_fp16)[name = string("v_3_cast_fp16")]; + tensor value_11_perm_0 = const()[name = string("value_11_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_1_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_1_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40609792)))]; + tensor var_724_cast_fp16 = add(x = q_7_cast_fp16, y = encoder_layers_1_self_attn_pos_bias_u_to_fp16)[name = string("op_724_cast_fp16")]; + tensor encoder_layers_1_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_1_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40611904)))]; + tensor var_726_cast_fp16 = add(x = q_7_cast_fp16, y = encoder_layers_1_self_attn_pos_bias_v_to_fp16)[name = string("op_726_cast_fp16")]; + tensor q_with_bias_v_3_perm_0 = const()[name = string("q_with_bias_v_3_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_33_transpose_x_0 = const()[name = string("x_33_transpose_x_0"), val = bool(false)]; + bool x_33_transpose_y_0 = const()[name = string("x_33_transpose_y_0"), val = bool(false)]; + tensor op_728_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40614016))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40813760))))[name = string("op_728_to_fp16_quantized")]; + tensor q_with_bias_v_3_cast_fp16 = transpose(perm = q_with_bias_v_3_perm_0, x = var_726_cast_fp16)[name = string("transpose_353")]; + tensor x_33_cast_fp16 = matmul(transpose_x = x_33_transpose_x_0, transpose_y = x_33_transpose_y_0, x = q_with_bias_v_3_cast_fp16, y = op_728_to_fp16_quantized)[name = string("x_33_cast_fp16")]; + tensor x_35_pad_0 = const()[name = string("x_35_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_35_mode_0 = const()[name = string("x_35_mode_0"), val = string("constant")]; + fp16 const_92_to_fp16 = const()[name = string("const_92_to_fp16"), val = fp16(0x0p+0)]; + tensor x_35_cast_fp16 = pad(constant_val = const_92_to_fp16, mode = x_35_mode_0, pad = x_35_pad_0, x = x_33_cast_fp16)[name = string("x_35_cast_fp16")]; + tensor var_736 = const()[name = string("op_736"), val = tensor([1, 8, -1, 56])]; + tensor x_37_cast_fp16 = reshape(shape = var_736, x = x_35_cast_fp16)[name = string("x_37_cast_fp16")]; + tensor var_740_begin_0 = const()[name = string("op_740_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_740_end_0 = const()[name = string("op_740_end_0"), val = tensor([1, 8, 196, 56])]; + tensor var_740_end_mask_0 = const()[name = string("op_740_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_740_cast_fp16 = slice_by_index(begin = var_740_begin_0, end = var_740_end_0, end_mask = var_740_end_mask_0, x = x_37_cast_fp16)[name = string("op_740_cast_fp16")]; + tensor var_741 = const()[name = string("op_741"), val = tensor([1, 8, 56, 195])]; + tensor matrix_bd_5_cast_fp16 = reshape(shape = var_741, x = var_740_cast_fp16)[name = string("matrix_bd_5_cast_fp16")]; + bool matrix_ac_3_transpose_x_0 = const()[name = string("matrix_ac_3_transpose_x_0"), val = bool(false)]; + bool matrix_ac_3_transpose_y_0 = const()[name = string("matrix_ac_3_transpose_y_0"), val = bool(false)]; + tensor transpose_98_perm_0 = const()[name = string("transpose_98_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_99_perm_0 = const()[name = string("transpose_99_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_99 = transpose(perm = transpose_99_perm_0, x = k_5_cast_fp16)[name = string("transpose_351")]; + tensor transpose_98 = transpose(perm = transpose_98_perm_0, x = var_724_cast_fp16)[name = string("transpose_352")]; + tensor matrix_ac_3_cast_fp16 = matmul(transpose_x = matrix_ac_3_transpose_x_0, transpose_y = matrix_ac_3_transpose_y_0, x = transpose_98, y = transpose_99)[name = string("matrix_ac_3_cast_fp16")]; + tensor matrix_bd_7_begin_0 = const()[name = string("matrix_bd_7_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_7_end_0 = const()[name = string("matrix_bd_7_end_0"), val = tensor([1, 8, 56, 98])]; + tensor matrix_bd_7_end_mask_0 = const()[name = string("matrix_bd_7_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_7_cast_fp16 = slice_by_index(begin = matrix_bd_7_begin_0, end = matrix_bd_7_end_0, end_mask = matrix_bd_7_end_mask_0, x = matrix_bd_5_cast_fp16)[name = string("matrix_bd_7_cast_fp16")]; + tensor var_750_cast_fp16 = add(x = matrix_ac_3_cast_fp16, y = matrix_bd_7_cast_fp16)[name = string("op_750_cast_fp16")]; + fp16 _inversed_scores_5_y_0_to_fp16 = const()[name = string("_inversed_scores_5_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_5_cast_fp16 = mul(x = var_750_cast_fp16, y = _inversed_scores_5_y_0_to_fp16)[name = string("_inversed_scores_5_cast_fp16")]; + tensor scores_7_cast_fp16 = select(a = var_44_to_fp16, b = _inversed_scores_5_cast_fp16, cond = mask_11)[name = string("scores_7_cast_fp16")]; + tensor var_756_cast_fp16 = softmax(axis = var_58, x = scores_7_cast_fp16)[name = string("op_756_cast_fp16")]; + tensor input_93_cast_fp16 = select(a = var_43_to_fp16, b = var_756_cast_fp16, cond = mask_11)[name = string("input_93_cast_fp16")]; + bool x_39_transpose_x_0 = const()[name = string("x_39_transpose_x_0"), val = bool(false)]; + bool x_39_transpose_y_0 = const()[name = string("x_39_transpose_y_0"), val = bool(false)]; + tensor value_11_cast_fp16 = transpose(perm = value_11_perm_0, x = v_3_cast_fp16)[name = string("transpose_350")]; + tensor x_39_cast_fp16 = matmul(transpose_x = x_39_transpose_x_0, transpose_y = x_39_transpose_y_0, x = input_93_cast_fp16, y = value_11_cast_fp16)[name = string("x_39_cast_fp16")]; + tensor var_760_perm_0 = const()[name = string("op_760_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_761 = const()[name = string("op_761"), val = tensor([1, -1, 1024])]; + tensor var_760_cast_fp16 = transpose(perm = var_760_perm_0, x = x_39_cast_fp16)[name = string("transpose_349")]; + tensor input_95_cast_fp16 = reshape(shape = var_761, x = var_760_cast_fp16)[name = string("input_95_cast_fp16")]; + tensor encoder_layers_1_self_attn_linear_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40814272))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41862912))))[name = string("encoder_layers_1_self_attn_linear_out_weight_to_fp16_quantized")]; + tensor encoder_layers_1_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_1_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41865024)))]; + tensor linear_16_cast_fp16 = linear(bias = encoder_layers_1_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_1_self_attn_linear_out_weight_to_fp16_quantized, x = input_95_cast_fp16)[name = string("linear_16_cast_fp16")]; + tensor input_99_cast_fp16 = add(x = input_89_cast_fp16, y = linear_16_cast_fp16)[name = string("input_99_cast_fp16")]; + tensor x_43_axes_0 = const()[name = string("x_43_axes_0"), val = tensor([-1])]; + tensor encoder_layers_1_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_1_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41867136)))]; + tensor encoder_layers_1_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_1_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41869248)))]; + tensor x_43_cast_fp16 = layer_norm(axes = x_43_axes_0, beta = encoder_layers_1_norm_conv_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_1_norm_conv_weight_to_fp16, x = input_99_cast_fp16)[name = string("x_43_cast_fp16")]; + tensor input_101_perm_0 = const()[name = string("input_101_perm_0"), val = tensor([0, 2, 1])]; + string input_103_pad_type_0 = const()[name = string("input_103_pad_type_0"), val = string("valid")]; + tensor input_103_strides_0 = const()[name = string("input_103_strides_0"), val = tensor([1])]; + tensor input_103_pad_0 = const()[name = string("input_103_pad_0"), val = tensor([0, 0])]; + tensor input_103_dilations_0 = const()[name = string("input_103_dilations_0"), val = tensor([1])]; + int32 input_103_groups_0 = const()[name = string("input_103_groups_0"), val = int32(1)]; + tensor encoder_layers_1_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41871360))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43968576))))[name = string("encoder_layers_1_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_101_cast_fp16 = transpose(perm = input_101_perm_0, x = x_43_cast_fp16)[name = string("transpose_348")]; + tensor input_103_cast_fp16 = conv(dilations = input_103_dilations_0, groups = input_103_groups_0, pad = input_103_pad_0, pad_type = input_103_pad_type_0, strides = input_103_strides_0, weight = encoder_layers_1_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_101_cast_fp16)[name = string("input_103_cast_fp16")]; + int32 x_45_split_num_splits_0 = const()[name = string("x_45_split_num_splits_0"), val = int32(2)]; + int32 x_45_split_axis_0 = const()[name = string("x_45_split_axis_0"), val = int32(1)]; + tensor x_45_split_cast_fp16_0, tensor x_45_split_cast_fp16_1 = split(axis = x_45_split_axis_0, num_splits = x_45_split_num_splits_0, x = input_103_cast_fp16)[name = string("x_45_split_cast_fp16")]; + tensor x_45_split_1_sigmoid_cast_fp16 = sigmoid(x = x_45_split_cast_fp16_1)[name = string("x_45_split_1_sigmoid_cast_fp16")]; + tensor x_45_cast_fp16 = mul(x = x_45_split_cast_fp16_0, y = x_45_split_1_sigmoid_cast_fp16)[name = string("x_45_cast_fp16")]; + tensor input_105_cast_fp16 = select(a = var_43_to_fp16, b = x_45_cast_fp16, cond = var_574)[name = string("input_105_cast_fp16")]; + bool new_x_7_interleave_0 = const()[name = string("new_x_7_interleave_0"), val = bool(false)]; + tensor new_x_7_cast_fp16 = concat(axis = var_58, interleave = new_x_7_interleave_0, values = (cache_7_cast_fp16, input_105_cast_fp16))[name = string("new_x_7_cast_fp16")]; + tensor var_800_begin_0 = const()[name = string("op_800_begin_0"), val = tensor([0, 0, 56])]; + tensor var_800_end_0 = const()[name = string("op_800_end_0"), val = tensor([1, 1024, 64])]; + tensor var_800_end_mask_0 = const()[name = string("op_800_end_mask_0"), val = tensor([true, true, true])]; + tensor var_800_cast_fp16 = slice_by_index(begin = var_800_begin_0, end = var_800_end_0, end_mask = var_800_end_mask_0, x = new_x_7_cast_fp16)[name = string("op_800_cast_fp16")]; + string x_47_pad_type_0 = const()[name = string("x_47_pad_type_0"), val = string("valid")]; + int32 x_47_groups_0 = const()[name = string("x_47_groups_0"), val = int32(1024)]; + tensor x_47_strides_0 = const()[name = string("x_47_strides_0"), val = tensor([1])]; + tensor x_47_pad_0 = const()[name = string("x_47_pad_0"), val = tensor([0, 0])]; + tensor x_47_dilations_0 = const()[name = string("x_47_dilations_0"), val = tensor([1])]; + tensor encoder_layers_1_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43972736))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43982016))))[name = string("encoder_layers_1_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_47_cast_fp16 = conv(dilations = x_47_dilations_0, groups = x_47_groups_0, pad = x_47_pad_0, pad_type = x_47_pad_type_0, strides = x_47_strides_0, weight = encoder_layers_1_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_7_cast_fp16)[name = string("x_47_cast_fp16")]; + tensor input_107_perm_0 = const()[name = string("input_107_perm_0"), val = tensor([0, 2, 1])]; + tensor x_49_axes_0 = const()[name = string("x_49_axes_0"), val = tensor([-1])]; + tensor encoder_layers_1_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_1_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43984128)))]; + tensor encoder_layers_1_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_1_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43986240)))]; + tensor input_107_cast_fp16 = transpose(perm = input_107_perm_0, x = x_47_cast_fp16)[name = string("transpose_347")]; + tensor x_49_cast_fp16 = layer_norm(axes = x_49_axes_0, beta = encoder_layers_1_conv_batch_norm_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_1_conv_batch_norm_weight_to_fp16, x = input_107_cast_fp16)[name = string("x_49_cast_fp16")]; + tensor input_109_perm_0 = const()[name = string("input_109_perm_0"), val = tensor([0, 2, 1])]; + tensor input_109_cast_fp16 = transpose(perm = input_109_perm_0, x = x_49_cast_fp16)[name = string("transpose_346")]; + tensor input_111_cast_fp16 = silu(x = input_109_cast_fp16)[name = string("input_111_cast_fp16")]; + string x_51_pad_type_0 = const()[name = string("x_51_pad_type_0"), val = string("valid")]; + tensor x_51_strides_0 = const()[name = string("x_51_strides_0"), val = tensor([1])]; + tensor x_51_pad_0 = const()[name = string("x_51_pad_0"), val = tensor([0, 0])]; + tensor x_51_dilations_0 = const()[name = string("x_51_dilations_0"), val = tensor([1])]; + int32 x_51_groups_0 = const()[name = string("x_51_groups_0"), val = int32(1)]; + tensor encoder_layers_1_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43988352))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45036992))))[name = string("encoder_layers_1_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_51_cast_fp16 = conv(dilations = x_51_dilations_0, groups = x_51_groups_0, pad = x_51_pad_0, pad_type = x_51_pad_type_0, strides = x_51_strides_0, weight = encoder_layers_1_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_111_cast_fp16)[name = string("x_51_cast_fp16")]; + tensor input_113_perm_0 = const()[name = string("input_113_perm_0"), val = tensor([0, 2, 1])]; + tensor input_113_cast_fp16 = transpose(perm = input_113_perm_0, x = x_51_cast_fp16)[name = string("transpose_345")]; + tensor input_115_cast_fp16 = add(x = input_99_cast_fp16, y = input_113_cast_fp16)[name = string("input_115_cast_fp16")]; + tensor input_117_axes_0 = const()[name = string("input_117_axes_0"), val = tensor([-1])]; + tensor encoder_layers_1_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_1_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45039104)))]; + tensor encoder_layers_1_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_1_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45041216)))]; + tensor input_117_cast_fp16 = layer_norm(axes = input_117_axes_0, beta = encoder_layers_1_norm_feed_forward2_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_1_norm_feed_forward2_weight_to_fp16, x = input_115_cast_fp16)[name = string("input_117_cast_fp16")]; + tensor encoder_layers_1_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45043328))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49237696))))[name = string("encoder_layers_1_feed_forward2_linear1_weight_to_fp16_quantized")]; + tensor encoder_layers_1_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_1_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49245952)))]; + tensor linear_17_cast_fp16 = linear(bias = encoder_layers_1_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_1_feed_forward2_linear1_weight_to_fp16_quantized, x = input_117_cast_fp16)[name = string("linear_17_cast_fp16")]; + tensor input_121_cast_fp16 = silu(x = linear_17_cast_fp16)[name = string("input_121_cast_fp16")]; + tensor encoder_layers_1_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49254208))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53448576))))[name = string("encoder_layers_1_feed_forward2_linear2_weight_to_fp16_quantized")]; + tensor encoder_layers_1_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_1_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53450688)))]; + tensor linear_18_cast_fp16 = linear(bias = encoder_layers_1_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_1_feed_forward2_linear2_weight_to_fp16_quantized, x = input_121_cast_fp16)[name = string("linear_18_cast_fp16")]; + fp16 var_843_to_fp16 = const()[name = string("op_843_to_fp16"), val = fp16(0x1p-1)]; + tensor var_844_cast_fp16 = mul(x = linear_18_cast_fp16, y = var_843_to_fp16)[name = string("op_844_cast_fp16")]; + tensor input_127_cast_fp16 = add(x = input_115_cast_fp16, y = var_844_cast_fp16)[name = string("input_127_cast_fp16")]; + tensor input_129_axes_0 = const()[name = string("input_129_axes_0"), val = tensor([-1])]; + tensor encoder_layers_1_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_1_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53452800)))]; + tensor encoder_layers_1_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_1_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53454912)))]; + tensor input_129_cast_fp16 = layer_norm(axes = input_129_axes_0, beta = encoder_layers_1_norm_out_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_1_norm_out_weight_to_fp16, x = input_127_cast_fp16)[name = string("input_129_cast_fp16")]; + tensor cache_9_begin_0 = const()[name = string("cache_9_begin_0"), val = tensor([2, 0, 0, 0])]; + tensor cache_9_end_0 = const()[name = string("cache_9_end_0"), val = tensor([3, 1, 42, 1024])]; + tensor cache_9_end_mask_0 = const()[name = string("cache_9_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_9_squeeze_mask_0 = const()[name = string("cache_9_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_9_cast_fp16 = slice_by_index(begin = cache_9_begin_0, end = cache_9_end_0, end_mask = cache_9_end_mask_0, squeeze_mask = cache_9_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_9_cast_fp16")]; + tensor cache_11_begin_0 = const()[name = string("cache_11_begin_0"), val = tensor([2, 0, 0, 0])]; + tensor cache_11_end_0 = const()[name = string("cache_11_end_0"), val = tensor([3, 1, 1024, 8])]; + tensor cache_11_end_mask_0 = const()[name = string("cache_11_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_11_squeeze_mask_0 = const()[name = string("cache_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_11_cast_fp16 = slice_by_index(begin = cache_11_begin_0, end = cache_11_end_0, end_mask = cache_11_end_mask_0, squeeze_mask = cache_11_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_11_cast_fp16")]; + tensor input_131_axes_0 = const()[name = string("input_131_axes_0"), val = tensor([-1])]; + tensor encoder_layers_2_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_2_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53457024)))]; + tensor encoder_layers_2_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_2_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53459136)))]; + tensor input_131_cast_fp16 = layer_norm(axes = input_131_axes_0, beta = encoder_layers_2_norm_feed_forward1_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_2_norm_feed_forward1_weight_to_fp16, x = input_129_cast_fp16)[name = string("input_131_cast_fp16")]; + tensor encoder_layers_2_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53461248))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57655616))))[name = string("encoder_layers_2_feed_forward1_linear1_weight_to_fp16_quantized")]; + tensor encoder_layers_2_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_2_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57663872)))]; + tensor linear_19_cast_fp16 = linear(bias = encoder_layers_2_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_2_feed_forward1_linear1_weight_to_fp16_quantized, x = input_131_cast_fp16)[name = string("linear_19_cast_fp16")]; + tensor input_135_cast_fp16 = silu(x = linear_19_cast_fp16)[name = string("input_135_cast_fp16")]; + tensor encoder_layers_2_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57672128))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61866496))))[name = string("encoder_layers_2_feed_forward1_linear2_weight_to_fp16_quantized")]; + tensor encoder_layers_2_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_2_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61868608)))]; + tensor linear_20_cast_fp16 = linear(bias = encoder_layers_2_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_2_feed_forward1_linear2_weight_to_fp16_quantized, x = input_135_cast_fp16)[name = string("linear_20_cast_fp16")]; + fp16 var_880_to_fp16 = const()[name = string("op_880_to_fp16"), val = fp16(0x1p-1)]; + tensor var_881_cast_fp16 = mul(x = linear_20_cast_fp16, y = var_880_to_fp16)[name = string("op_881_cast_fp16")]; + tensor input_141_cast_fp16 = add(x = input_129_cast_fp16, y = var_881_cast_fp16)[name = string("input_141_cast_fp16")]; + tensor key_5_axes_0 = const()[name = string("key_5_axes_0"), val = tensor([-1])]; + tensor encoder_layers_2_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_2_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61870720)))]; + tensor encoder_layers_2_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_2_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61872832)))]; + tensor key_5_cast_fp16 = layer_norm(axes = key_5_axes_0, beta = encoder_layers_2_norm_self_att_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_2_norm_self_att_weight_to_fp16, x = input_141_cast_fp16)[name = string("key_5_cast_fp16")]; + bool input_143_interleave_0 = const()[name = string("input_143_interleave_0"), val = bool(false)]; + tensor input_143_cast_fp16 = concat(axis = var_67, interleave = input_143_interleave_0, values = (cache_9_cast_fp16, key_5_cast_fp16))[name = string("input_143_cast_fp16")]; + bool var_909_interleave_0 = const()[name = string("op_909_interleave_0"), val = bool(false)]; + tensor var_909_cast_fp16 = concat(axis = var_67, interleave = var_909_interleave_0, values = key_5_cast_fp16)[name = string("op_909_cast_fp16")]; + tensor encoder_layers_2_self_attn_linear_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61874944))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62923584))))[name = string("encoder_layers_2_self_attn_linear_q_weight_to_fp16_quantized")]; + tensor encoder_layers_2_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_2_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62925696)))]; + tensor linear_21_cast_fp16 = linear(bias = encoder_layers_2_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_2_self_attn_linear_q_weight_to_fp16_quantized, x = key_5_cast_fp16)[name = string("linear_21_cast_fp16")]; + tensor var_914 = const()[name = string("op_914"), val = tensor([1, -1, 8, 128])]; + tensor q_13_cast_fp16 = reshape(shape = var_914, x = linear_21_cast_fp16)[name = string("q_13_cast_fp16")]; + tensor encoder_layers_2_self_attn_linear_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62927808))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(63976448))))[name = string("encoder_layers_2_self_attn_linear_k_weight_to_fp16_quantized")]; + tensor encoder_layers_2_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_2_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(63978560)))]; + tensor linear_22_cast_fp16 = linear(bias = encoder_layers_2_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_2_self_attn_linear_k_weight_to_fp16_quantized, x = input_143_cast_fp16)[name = string("linear_22_cast_fp16")]; + tensor var_919 = const()[name = string("op_919"), val = tensor([1, -1, 8, 128])]; + tensor k_9_cast_fp16 = reshape(shape = var_919, x = linear_22_cast_fp16)[name = string("k_9_cast_fp16")]; + tensor encoder_layers_2_self_attn_linear_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(63980672))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65029312))))[name = string("encoder_layers_2_self_attn_linear_v_weight_to_fp16_quantized")]; + tensor encoder_layers_2_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_2_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65031424)))]; + tensor linear_23_cast_fp16 = linear(bias = encoder_layers_2_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_2_self_attn_linear_v_weight_to_fp16_quantized, x = input_143_cast_fp16)[name = string("linear_23_cast_fp16")]; + tensor var_924 = const()[name = string("op_924"), val = tensor([1, -1, 8, 128])]; + tensor v_5_cast_fp16 = reshape(shape = var_924, x = linear_23_cast_fp16)[name = string("v_5_cast_fp16")]; + tensor value_13_perm_0 = const()[name = string("value_13_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_2_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_2_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65033536)))]; + tensor var_937_cast_fp16 = add(x = q_13_cast_fp16, y = encoder_layers_2_self_attn_pos_bias_u_to_fp16)[name = string("op_937_cast_fp16")]; + tensor encoder_layers_2_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_2_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65035648)))]; + tensor var_939_cast_fp16 = add(x = q_13_cast_fp16, y = encoder_layers_2_self_attn_pos_bias_v_to_fp16)[name = string("op_939_cast_fp16")]; + tensor q_with_bias_v_5_perm_0 = const()[name = string("q_with_bias_v_5_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_59_transpose_x_0 = const()[name = string("x_59_transpose_x_0"), val = bool(false)]; + bool x_59_transpose_y_0 = const()[name = string("x_59_transpose_y_0"), val = bool(false)]; + tensor op_941_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65037760))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65237504))))[name = string("op_941_to_fp16_quantized")]; + tensor q_with_bias_v_5_cast_fp16 = transpose(perm = q_with_bias_v_5_perm_0, x = var_939_cast_fp16)[name = string("transpose_344")]; + tensor x_59_cast_fp16 = matmul(transpose_x = x_59_transpose_x_0, transpose_y = x_59_transpose_y_0, x = q_with_bias_v_5_cast_fp16, y = op_941_to_fp16_quantized)[name = string("x_59_cast_fp16")]; + tensor x_61_pad_0 = const()[name = string("x_61_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_61_mode_0 = const()[name = string("x_61_mode_0"), val = string("constant")]; + fp16 const_105_to_fp16 = const()[name = string("const_105_to_fp16"), val = fp16(0x0p+0)]; + tensor x_61_cast_fp16 = pad(constant_val = const_105_to_fp16, mode = x_61_mode_0, pad = x_61_pad_0, x = x_59_cast_fp16)[name = string("x_61_cast_fp16")]; + tensor var_949 = const()[name = string("op_949"), val = tensor([1, 8, -1, 56])]; + tensor x_63_cast_fp16 = reshape(shape = var_949, x = x_61_cast_fp16)[name = string("x_63_cast_fp16")]; + tensor var_953_begin_0 = const()[name = string("op_953_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_953_end_0 = const()[name = string("op_953_end_0"), val = tensor([1, 8, 196, 56])]; + tensor var_953_end_mask_0 = const()[name = string("op_953_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_953_cast_fp16 = slice_by_index(begin = var_953_begin_0, end = var_953_end_0, end_mask = var_953_end_mask_0, x = x_63_cast_fp16)[name = string("op_953_cast_fp16")]; + tensor var_954 = const()[name = string("op_954"), val = tensor([1, 8, 56, 195])]; + tensor matrix_bd_9_cast_fp16 = reshape(shape = var_954, x = var_953_cast_fp16)[name = string("matrix_bd_9_cast_fp16")]; + bool matrix_ac_5_transpose_x_0 = const()[name = string("matrix_ac_5_transpose_x_0"), val = bool(false)]; + bool matrix_ac_5_transpose_y_0 = const()[name = string("matrix_ac_5_transpose_y_0"), val = bool(false)]; + tensor transpose_100_perm_0 = const()[name = string("transpose_100_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_101_perm_0 = const()[name = string("transpose_101_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_101 = transpose(perm = transpose_101_perm_0, x = k_9_cast_fp16)[name = string("transpose_342")]; + tensor transpose_100 = transpose(perm = transpose_100_perm_0, x = var_937_cast_fp16)[name = string("transpose_343")]; + tensor matrix_ac_5_cast_fp16 = matmul(transpose_x = matrix_ac_5_transpose_x_0, transpose_y = matrix_ac_5_transpose_y_0, x = transpose_100, y = transpose_101)[name = string("matrix_ac_5_cast_fp16")]; + tensor matrix_bd_11_begin_0 = const()[name = string("matrix_bd_11_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_11_end_0 = const()[name = string("matrix_bd_11_end_0"), val = tensor([1, 8, 56, 98])]; + tensor matrix_bd_11_end_mask_0 = const()[name = string("matrix_bd_11_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_11_cast_fp16 = slice_by_index(begin = matrix_bd_11_begin_0, end = matrix_bd_11_end_0, end_mask = matrix_bd_11_end_mask_0, x = matrix_bd_9_cast_fp16)[name = string("matrix_bd_11_cast_fp16")]; + tensor var_963_cast_fp16 = add(x = matrix_ac_5_cast_fp16, y = matrix_bd_11_cast_fp16)[name = string("op_963_cast_fp16")]; + fp16 _inversed_scores_9_y_0_to_fp16 = const()[name = string("_inversed_scores_9_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_9_cast_fp16 = mul(x = var_963_cast_fp16, y = _inversed_scores_9_y_0_to_fp16)[name = string("_inversed_scores_9_cast_fp16")]; + tensor scores_11_cast_fp16 = select(a = var_44_to_fp16, b = _inversed_scores_9_cast_fp16, cond = mask_11)[name = string("scores_11_cast_fp16")]; + tensor var_969_cast_fp16 = softmax(axis = var_58, x = scores_11_cast_fp16)[name = string("op_969_cast_fp16")]; + tensor input_145_cast_fp16 = select(a = var_43_to_fp16, b = var_969_cast_fp16, cond = mask_11)[name = string("input_145_cast_fp16")]; + bool x_65_transpose_x_0 = const()[name = string("x_65_transpose_x_0"), val = bool(false)]; + bool x_65_transpose_y_0 = const()[name = string("x_65_transpose_y_0"), val = bool(false)]; + tensor value_13_cast_fp16 = transpose(perm = value_13_perm_0, x = v_5_cast_fp16)[name = string("transpose_341")]; + tensor x_65_cast_fp16 = matmul(transpose_x = x_65_transpose_x_0, transpose_y = x_65_transpose_y_0, x = input_145_cast_fp16, y = value_13_cast_fp16)[name = string("x_65_cast_fp16")]; + tensor var_973_perm_0 = const()[name = string("op_973_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_974 = const()[name = string("op_974"), val = tensor([1, -1, 1024])]; + tensor var_973_cast_fp16 = transpose(perm = var_973_perm_0, x = x_65_cast_fp16)[name = string("transpose_340")]; + tensor input_147_cast_fp16 = reshape(shape = var_974, x = var_973_cast_fp16)[name = string("input_147_cast_fp16")]; + tensor encoder_layers_2_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65238016))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66024512))))[name = string("encoder_layers_2_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_2_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_2_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66024704)))]; + tensor linear_25_cast_fp16 = linear(bias = encoder_layers_2_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_2_self_attn_linear_out_weight_to_fp16_palettized, x = input_147_cast_fp16)[name = string("linear_25_cast_fp16")]; + tensor input_151_cast_fp16 = add(x = input_141_cast_fp16, y = linear_25_cast_fp16)[name = string("input_151_cast_fp16")]; + tensor x_69_axes_0 = const()[name = string("x_69_axes_0"), val = tensor([-1])]; + tensor encoder_layers_2_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_2_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66026816)))]; + tensor encoder_layers_2_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_2_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66028928)))]; + tensor x_69_cast_fp16 = layer_norm(axes = x_69_axes_0, beta = encoder_layers_2_norm_conv_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_2_norm_conv_weight_to_fp16, x = input_151_cast_fp16)[name = string("x_69_cast_fp16")]; + tensor input_153_perm_0 = const()[name = string("input_153_perm_0"), val = tensor([0, 2, 1])]; + string input_155_pad_type_0 = const()[name = string("input_155_pad_type_0"), val = string("valid")]; + tensor input_155_strides_0 = const()[name = string("input_155_strides_0"), val = tensor([1])]; + tensor input_155_pad_0 = const()[name = string("input_155_pad_0"), val = tensor([0, 0])]; + tensor input_155_dilations_0 = const()[name = string("input_155_dilations_0"), val = tensor([1])]; + int32 input_155_groups_0 = const()[name = string("input_155_groups_0"), val = int32(1)]; + tensor encoder_layers_2_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66031040))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68128256))))[name = string("encoder_layers_2_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_153_cast_fp16 = transpose(perm = input_153_perm_0, x = x_69_cast_fp16)[name = string("transpose_339")]; + tensor input_155_cast_fp16 = conv(dilations = input_155_dilations_0, groups = input_155_groups_0, pad = input_155_pad_0, pad_type = input_155_pad_type_0, strides = input_155_strides_0, weight = encoder_layers_2_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_153_cast_fp16)[name = string("input_155_cast_fp16")]; + int32 x_71_split_num_splits_0 = const()[name = string("x_71_split_num_splits_0"), val = int32(2)]; + int32 x_71_split_axis_0 = const()[name = string("x_71_split_axis_0"), val = int32(1)]; + tensor x_71_split_cast_fp16_0, tensor x_71_split_cast_fp16_1 = split(axis = x_71_split_axis_0, num_splits = x_71_split_num_splits_0, x = input_155_cast_fp16)[name = string("x_71_split_cast_fp16")]; + tensor x_71_split_1_sigmoid_cast_fp16 = sigmoid(x = x_71_split_cast_fp16_1)[name = string("x_71_split_1_sigmoid_cast_fp16")]; + tensor x_71_cast_fp16 = mul(x = x_71_split_cast_fp16_0, y = x_71_split_1_sigmoid_cast_fp16)[name = string("x_71_cast_fp16")]; + tensor input_157_cast_fp16 = select(a = var_43_to_fp16, b = x_71_cast_fp16, cond = var_574)[name = string("input_157_cast_fp16")]; + bool new_x_11_interleave_0 = const()[name = string("new_x_11_interleave_0"), val = bool(false)]; + tensor new_x_11_cast_fp16 = concat(axis = var_58, interleave = new_x_11_interleave_0, values = (cache_11_cast_fp16, input_157_cast_fp16))[name = string("new_x_11_cast_fp16")]; + tensor var_1013_begin_0 = const()[name = string("op_1013_begin_0"), val = tensor([0, 0, 56])]; + tensor var_1013_end_0 = const()[name = string("op_1013_end_0"), val = tensor([1, 1024, 64])]; + tensor var_1013_end_mask_0 = const()[name = string("op_1013_end_mask_0"), val = tensor([true, true, true])]; + tensor var_1013_cast_fp16 = slice_by_index(begin = var_1013_begin_0, end = var_1013_end_0, end_mask = var_1013_end_mask_0, x = new_x_11_cast_fp16)[name = string("op_1013_cast_fp16")]; + string x_73_pad_type_0 = const()[name = string("x_73_pad_type_0"), val = string("valid")]; + int32 x_73_groups_0 = const()[name = string("x_73_groups_0"), val = int32(1024)]; + tensor x_73_strides_0 = const()[name = string("x_73_strides_0"), val = tensor([1])]; + tensor x_73_pad_0 = const()[name = string("x_73_pad_0"), val = tensor([0, 0])]; + tensor x_73_dilations_0 = const()[name = string("x_73_dilations_0"), val = tensor([1])]; + tensor encoder_layers_2_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68132416))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68141696))))[name = string("encoder_layers_2_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_73_cast_fp16 = conv(dilations = x_73_dilations_0, groups = x_73_groups_0, pad = x_73_pad_0, pad_type = x_73_pad_type_0, strides = x_73_strides_0, weight = encoder_layers_2_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_11_cast_fp16)[name = string("x_73_cast_fp16")]; + tensor input_159_perm_0 = const()[name = string("input_159_perm_0"), val = tensor([0, 2, 1])]; + tensor x_75_axes_0 = const()[name = string("x_75_axes_0"), val = tensor([-1])]; + tensor encoder_layers_2_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_2_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68143808)))]; + tensor encoder_layers_2_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_2_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68145920)))]; + tensor input_159_cast_fp16 = transpose(perm = input_159_perm_0, x = x_73_cast_fp16)[name = string("transpose_338")]; + tensor x_75_cast_fp16 = layer_norm(axes = x_75_axes_0, beta = encoder_layers_2_conv_batch_norm_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_2_conv_batch_norm_weight_to_fp16, x = input_159_cast_fp16)[name = string("x_75_cast_fp16")]; + tensor input_161_perm_0 = const()[name = string("input_161_perm_0"), val = tensor([0, 2, 1])]; + tensor input_161_cast_fp16 = transpose(perm = input_161_perm_0, x = x_75_cast_fp16)[name = string("transpose_337")]; + tensor input_163_cast_fp16 = silu(x = input_161_cast_fp16)[name = string("input_163_cast_fp16")]; + string x_77_pad_type_0 = const()[name = string("x_77_pad_type_0"), val = string("valid")]; + tensor x_77_strides_0 = const()[name = string("x_77_strides_0"), val = tensor([1])]; + tensor x_77_pad_0 = const()[name = string("x_77_pad_0"), val = tensor([0, 0])]; + tensor x_77_dilations_0 = const()[name = string("x_77_dilations_0"), val = tensor([1])]; + int32 x_77_groups_0 = const()[name = string("x_77_groups_0"), val = int32(1)]; + tensor encoder_layers_2_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68148032))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69196672))))[name = string("encoder_layers_2_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_77_cast_fp16 = conv(dilations = x_77_dilations_0, groups = x_77_groups_0, pad = x_77_pad_0, pad_type = x_77_pad_type_0, strides = x_77_strides_0, weight = encoder_layers_2_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_163_cast_fp16)[name = string("x_77_cast_fp16")]; + tensor input_165_perm_0 = const()[name = string("input_165_perm_0"), val = tensor([0, 2, 1])]; + tensor input_165_cast_fp16 = transpose(perm = input_165_perm_0, x = x_77_cast_fp16)[name = string("transpose_336")]; + tensor input_167_cast_fp16 = add(x = input_151_cast_fp16, y = input_165_cast_fp16)[name = string("input_167_cast_fp16")]; + tensor input_169_axes_0 = const()[name = string("input_169_axes_0"), val = tensor([-1])]; + tensor encoder_layers_2_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_2_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69198784)))]; + tensor encoder_layers_2_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_2_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69200896)))]; + tensor input_169_cast_fp16 = layer_norm(axes = input_169_axes_0, beta = encoder_layers_2_norm_feed_forward2_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_2_norm_feed_forward2_weight_to_fp16, x = input_167_cast_fp16)[name = string("input_169_cast_fp16")]; + tensor encoder_layers_2_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69203008))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72348800))))[name = string("encoder_layers_2_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_2_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_2_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72348992)))]; + tensor linear_26_cast_fp16 = linear(bias = encoder_layers_2_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_2_feed_forward2_linear1_weight_to_fp16_palettized, x = input_169_cast_fp16)[name = string("linear_26_cast_fp16")]; + tensor input_173_cast_fp16 = silu(x = linear_26_cast_fp16)[name = string("input_173_cast_fp16")]; + tensor encoder_layers_2_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72357248))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75503040))))[name = string("encoder_layers_2_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_2_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_2_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75503232)))]; + tensor linear_27_cast_fp16 = linear(bias = encoder_layers_2_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_2_feed_forward2_linear2_weight_to_fp16_palettized, x = input_173_cast_fp16)[name = string("linear_27_cast_fp16")]; + fp16 var_1056_to_fp16 = const()[name = string("op_1056_to_fp16"), val = fp16(0x1p-1)]; + tensor var_1057_cast_fp16 = mul(x = linear_27_cast_fp16, y = var_1056_to_fp16)[name = string("op_1057_cast_fp16")]; + tensor input_179_cast_fp16 = add(x = input_167_cast_fp16, y = var_1057_cast_fp16)[name = string("input_179_cast_fp16")]; + tensor input_181_axes_0 = const()[name = string("input_181_axes_0"), val = tensor([-1])]; + tensor encoder_layers_2_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_2_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75505344)))]; + tensor encoder_layers_2_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_2_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75507456)))]; + tensor input_181_cast_fp16 = layer_norm(axes = input_181_axes_0, beta = encoder_layers_2_norm_out_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_2_norm_out_weight_to_fp16, x = input_179_cast_fp16)[name = string("input_181_cast_fp16")]; + tensor cache_13_begin_0 = const()[name = string("cache_13_begin_0"), val = tensor([3, 0, 0, 0])]; + tensor cache_13_end_0 = const()[name = string("cache_13_end_0"), val = tensor([4, 1, 42, 1024])]; + tensor cache_13_end_mask_0 = const()[name = string("cache_13_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_13_squeeze_mask_0 = const()[name = string("cache_13_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_13_cast_fp16 = slice_by_index(begin = cache_13_begin_0, end = cache_13_end_0, end_mask = cache_13_end_mask_0, squeeze_mask = cache_13_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_13_cast_fp16")]; + tensor cache_15_begin_0 = const()[name = string("cache_15_begin_0"), val = tensor([3, 0, 0, 0])]; + tensor cache_15_end_0 = const()[name = string("cache_15_end_0"), val = tensor([4, 1, 1024, 8])]; + tensor cache_15_end_mask_0 = const()[name = string("cache_15_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_15_squeeze_mask_0 = const()[name = string("cache_15_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_15_cast_fp16 = slice_by_index(begin = cache_15_begin_0, end = cache_15_end_0, end_mask = cache_15_end_mask_0, squeeze_mask = cache_15_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_15_cast_fp16")]; + tensor input_183_axes_0 = const()[name = string("input_183_axes_0"), val = tensor([-1])]; + tensor encoder_layers_3_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_3_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75509568)))]; + tensor encoder_layers_3_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_3_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75511680)))]; + tensor input_183_cast_fp16 = layer_norm(axes = input_183_axes_0, beta = encoder_layers_3_norm_feed_forward1_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_3_norm_feed_forward1_weight_to_fp16, x = input_181_cast_fp16)[name = string("input_183_cast_fp16")]; + tensor encoder_layers_3_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75513792))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78659584))))[name = string("encoder_layers_3_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_3_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_3_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78659776)))]; + tensor linear_28_cast_fp16 = linear(bias = encoder_layers_3_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_3_feed_forward1_linear1_weight_to_fp16_palettized, x = input_183_cast_fp16)[name = string("linear_28_cast_fp16")]; + tensor input_187_cast_fp16 = silu(x = linear_28_cast_fp16)[name = string("input_187_cast_fp16")]; + tensor encoder_layers_3_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78668032))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81813824))))[name = string("encoder_layers_3_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_3_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_3_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81814016)))]; + tensor linear_29_cast_fp16 = linear(bias = encoder_layers_3_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_3_feed_forward1_linear2_weight_to_fp16_palettized, x = input_187_cast_fp16)[name = string("linear_29_cast_fp16")]; + fp16 var_1093_to_fp16 = const()[name = string("op_1093_to_fp16"), val = fp16(0x1p-1)]; + tensor var_1094_cast_fp16 = mul(x = linear_29_cast_fp16, y = var_1093_to_fp16)[name = string("op_1094_cast_fp16")]; + tensor input_193_cast_fp16 = add(x = input_181_cast_fp16, y = var_1094_cast_fp16)[name = string("input_193_cast_fp16")]; + tensor key_7_axes_0 = const()[name = string("key_7_axes_0"), val = tensor([-1])]; + tensor encoder_layers_3_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_3_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81816128)))]; + tensor encoder_layers_3_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_3_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81818240)))]; + tensor key_7_cast_fp16 = layer_norm(axes = key_7_axes_0, beta = encoder_layers_3_norm_self_att_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_3_norm_self_att_weight_to_fp16, x = input_193_cast_fp16)[name = string("key_7_cast_fp16")]; + bool input_195_interleave_0 = const()[name = string("input_195_interleave_0"), val = bool(false)]; + tensor input_195_cast_fp16 = concat(axis = var_67, interleave = input_195_interleave_0, values = (cache_13_cast_fp16, key_7_cast_fp16))[name = string("input_195_cast_fp16")]; + bool var_1122_interleave_0 = const()[name = string("op_1122_interleave_0"), val = bool(false)]; + tensor var_1122_cast_fp16 = concat(axis = var_67, interleave = var_1122_interleave_0, values = key_7_cast_fp16)[name = string("op_1122_cast_fp16")]; + tensor encoder_layers_3_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81820352))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82606848))))[name = string("encoder_layers_3_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_3_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_3_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82607040)))]; + tensor linear_30_cast_fp16 = linear(bias = encoder_layers_3_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_3_self_attn_linear_q_weight_to_fp16_palettized, x = key_7_cast_fp16)[name = string("linear_30_cast_fp16")]; + tensor var_1127 = const()[name = string("op_1127"), val = tensor([1, -1, 8, 128])]; + tensor q_19_cast_fp16 = reshape(shape = var_1127, x = linear_30_cast_fp16)[name = string("q_19_cast_fp16")]; + tensor encoder_layers_3_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82609152))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83395648))))[name = string("encoder_layers_3_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_3_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_3_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83395840)))]; + tensor linear_31_cast_fp16 = linear(bias = encoder_layers_3_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_3_self_attn_linear_k_weight_to_fp16_palettized, x = input_195_cast_fp16)[name = string("linear_31_cast_fp16")]; + tensor var_1132 = const()[name = string("op_1132"), val = tensor([1, -1, 8, 128])]; + tensor k_13_cast_fp16 = reshape(shape = var_1132, x = linear_31_cast_fp16)[name = string("k_13_cast_fp16")]; + tensor encoder_layers_3_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83397952))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84184448))))[name = string("encoder_layers_3_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_3_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_3_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84184640)))]; + tensor linear_32_cast_fp16 = linear(bias = encoder_layers_3_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_3_self_attn_linear_v_weight_to_fp16_palettized, x = input_195_cast_fp16)[name = string("linear_32_cast_fp16")]; + tensor var_1137 = const()[name = string("op_1137"), val = tensor([1, -1, 8, 128])]; + tensor v_7_cast_fp16 = reshape(shape = var_1137, x = linear_32_cast_fp16)[name = string("v_7_cast_fp16")]; + tensor value_15_perm_0 = const()[name = string("value_15_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_3_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_3_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84186752)))]; + tensor var_1150_cast_fp16 = add(x = q_19_cast_fp16, y = encoder_layers_3_self_attn_pos_bias_u_to_fp16)[name = string("op_1150_cast_fp16")]; + tensor encoder_layers_3_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_3_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84188864)))]; + tensor var_1152_cast_fp16 = add(x = q_19_cast_fp16, y = encoder_layers_3_self_attn_pos_bias_v_to_fp16)[name = string("op_1152_cast_fp16")]; + tensor q_with_bias_v_7_perm_0 = const()[name = string("q_with_bias_v_7_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_85_transpose_x_0 = const()[name = string("x_85_transpose_x_0"), val = bool(false)]; + bool x_85_transpose_y_0 = const()[name = string("x_85_transpose_y_0"), val = bool(false)]; + tensor op_1154_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84190976))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84390720))))[name = string("op_1154_to_fp16_quantized")]; + tensor q_with_bias_v_7_cast_fp16 = transpose(perm = q_with_bias_v_7_perm_0, x = var_1152_cast_fp16)[name = string("transpose_335")]; + tensor x_85_cast_fp16 = matmul(transpose_x = x_85_transpose_x_0, transpose_y = x_85_transpose_y_0, x = q_with_bias_v_7_cast_fp16, y = op_1154_to_fp16_quantized)[name = string("x_85_cast_fp16")]; + tensor x_87_pad_0 = const()[name = string("x_87_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_87_mode_0 = const()[name = string("x_87_mode_0"), val = string("constant")]; + fp16 const_118_to_fp16 = const()[name = string("const_118_to_fp16"), val = fp16(0x0p+0)]; + tensor x_87_cast_fp16 = pad(constant_val = const_118_to_fp16, mode = x_87_mode_0, pad = x_87_pad_0, x = x_85_cast_fp16)[name = string("x_87_cast_fp16")]; + tensor var_1162 = const()[name = string("op_1162"), val = tensor([1, 8, -1, 56])]; + tensor x_89_cast_fp16 = reshape(shape = var_1162, x = x_87_cast_fp16)[name = string("x_89_cast_fp16")]; + tensor var_1166_begin_0 = const()[name = string("op_1166_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_1166_end_0 = const()[name = string("op_1166_end_0"), val = tensor([1, 8, 196, 56])]; + tensor var_1166_end_mask_0 = const()[name = string("op_1166_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_1166_cast_fp16 = slice_by_index(begin = var_1166_begin_0, end = var_1166_end_0, end_mask = var_1166_end_mask_0, x = x_89_cast_fp16)[name = string("op_1166_cast_fp16")]; + tensor var_1167 = const()[name = string("op_1167"), val = tensor([1, 8, 56, 195])]; + tensor matrix_bd_13_cast_fp16 = reshape(shape = var_1167, x = var_1166_cast_fp16)[name = string("matrix_bd_13_cast_fp16")]; + bool matrix_ac_7_transpose_x_0 = const()[name = string("matrix_ac_7_transpose_x_0"), val = bool(false)]; + bool matrix_ac_7_transpose_y_0 = const()[name = string("matrix_ac_7_transpose_y_0"), val = bool(false)]; + tensor transpose_102_perm_0 = const()[name = string("transpose_102_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_103_perm_0 = const()[name = string("transpose_103_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_103 = transpose(perm = transpose_103_perm_0, x = k_13_cast_fp16)[name = string("transpose_333")]; + tensor transpose_102 = transpose(perm = transpose_102_perm_0, x = var_1150_cast_fp16)[name = string("transpose_334")]; + tensor matrix_ac_7_cast_fp16 = matmul(transpose_x = matrix_ac_7_transpose_x_0, transpose_y = matrix_ac_7_transpose_y_0, x = transpose_102, y = transpose_103)[name = string("matrix_ac_7_cast_fp16")]; + tensor matrix_bd_15_begin_0 = const()[name = string("matrix_bd_15_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_15_end_0 = const()[name = string("matrix_bd_15_end_0"), val = tensor([1, 8, 56, 98])]; + tensor matrix_bd_15_end_mask_0 = const()[name = string("matrix_bd_15_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_15_cast_fp16 = slice_by_index(begin = matrix_bd_15_begin_0, end = matrix_bd_15_end_0, end_mask = matrix_bd_15_end_mask_0, x = matrix_bd_13_cast_fp16)[name = string("matrix_bd_15_cast_fp16")]; + tensor var_1176_cast_fp16 = add(x = matrix_ac_7_cast_fp16, y = matrix_bd_15_cast_fp16)[name = string("op_1176_cast_fp16")]; + fp16 _inversed_scores_13_y_0_to_fp16 = const()[name = string("_inversed_scores_13_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_13_cast_fp16 = mul(x = var_1176_cast_fp16, y = _inversed_scores_13_y_0_to_fp16)[name = string("_inversed_scores_13_cast_fp16")]; + tensor scores_15_cast_fp16 = select(a = var_44_to_fp16, b = _inversed_scores_13_cast_fp16, cond = mask_11)[name = string("scores_15_cast_fp16")]; + tensor var_1182_cast_fp16 = softmax(axis = var_58, x = scores_15_cast_fp16)[name = string("op_1182_cast_fp16")]; + tensor input_197_cast_fp16 = select(a = var_43_to_fp16, b = var_1182_cast_fp16, cond = mask_11)[name = string("input_197_cast_fp16")]; + bool x_91_transpose_x_0 = const()[name = string("x_91_transpose_x_0"), val = bool(false)]; + bool x_91_transpose_y_0 = const()[name = string("x_91_transpose_y_0"), val = bool(false)]; + tensor value_15_cast_fp16 = transpose(perm = value_15_perm_0, x = v_7_cast_fp16)[name = string("transpose_332")]; + tensor x_91_cast_fp16 = matmul(transpose_x = x_91_transpose_x_0, transpose_y = x_91_transpose_y_0, x = input_197_cast_fp16, y = value_15_cast_fp16)[name = string("x_91_cast_fp16")]; + tensor var_1186_perm_0 = const()[name = string("op_1186_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1187 = const()[name = string("op_1187"), val = tensor([1, -1, 1024])]; + tensor var_1186_cast_fp16 = transpose(perm = var_1186_perm_0, x = x_91_cast_fp16)[name = string("transpose_331")]; + tensor input_199_cast_fp16 = reshape(shape = var_1187, x = var_1186_cast_fp16)[name = string("input_199_cast_fp16")]; + tensor encoder_layers_3_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84391232))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85177728))))[name = string("encoder_layers_3_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_3_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_3_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85177920)))]; + tensor linear_34_cast_fp16 = linear(bias = encoder_layers_3_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_3_self_attn_linear_out_weight_to_fp16_palettized, x = input_199_cast_fp16)[name = string("linear_34_cast_fp16")]; + tensor input_203_cast_fp16 = add(x = input_193_cast_fp16, y = linear_34_cast_fp16)[name = string("input_203_cast_fp16")]; + tensor x_95_axes_0 = const()[name = string("x_95_axes_0"), val = tensor([-1])]; + tensor encoder_layers_3_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_3_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85180032)))]; + tensor encoder_layers_3_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_3_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85182144)))]; + tensor x_95_cast_fp16 = layer_norm(axes = x_95_axes_0, beta = encoder_layers_3_norm_conv_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_3_norm_conv_weight_to_fp16, x = input_203_cast_fp16)[name = string("x_95_cast_fp16")]; + tensor input_205_perm_0 = const()[name = string("input_205_perm_0"), val = tensor([0, 2, 1])]; + string input_207_pad_type_0 = const()[name = string("input_207_pad_type_0"), val = string("valid")]; + tensor input_207_strides_0 = const()[name = string("input_207_strides_0"), val = tensor([1])]; + tensor input_207_pad_0 = const()[name = string("input_207_pad_0"), val = tensor([0, 0])]; + tensor input_207_dilations_0 = const()[name = string("input_207_dilations_0"), val = tensor([1])]; + int32 input_207_groups_0 = const()[name = string("input_207_groups_0"), val = int32(1)]; + tensor encoder_layers_3_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85184256))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87281472))))[name = string("encoder_layers_3_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_205_cast_fp16 = transpose(perm = input_205_perm_0, x = x_95_cast_fp16)[name = string("transpose_330")]; + tensor input_207_cast_fp16 = conv(dilations = input_207_dilations_0, groups = input_207_groups_0, pad = input_207_pad_0, pad_type = input_207_pad_type_0, strides = input_207_strides_0, weight = encoder_layers_3_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_205_cast_fp16)[name = string("input_207_cast_fp16")]; + int32 x_97_split_num_splits_0 = const()[name = string("x_97_split_num_splits_0"), val = int32(2)]; + int32 x_97_split_axis_0 = const()[name = string("x_97_split_axis_0"), val = int32(1)]; + tensor x_97_split_cast_fp16_0, tensor x_97_split_cast_fp16_1 = split(axis = x_97_split_axis_0, num_splits = x_97_split_num_splits_0, x = input_207_cast_fp16)[name = string("x_97_split_cast_fp16")]; + tensor x_97_split_1_sigmoid_cast_fp16 = sigmoid(x = x_97_split_cast_fp16_1)[name = string("x_97_split_1_sigmoid_cast_fp16")]; + tensor x_97_cast_fp16 = mul(x = x_97_split_cast_fp16_0, y = x_97_split_1_sigmoid_cast_fp16)[name = string("x_97_cast_fp16")]; + tensor input_209_cast_fp16 = select(a = var_43_to_fp16, b = x_97_cast_fp16, cond = var_574)[name = string("input_209_cast_fp16")]; + bool new_x_15_interleave_0 = const()[name = string("new_x_15_interleave_0"), val = bool(false)]; + tensor new_x_15_cast_fp16 = concat(axis = var_58, interleave = new_x_15_interleave_0, values = (cache_15_cast_fp16, input_209_cast_fp16))[name = string("new_x_15_cast_fp16")]; + tensor var_1226_begin_0 = const()[name = string("op_1226_begin_0"), val = tensor([0, 0, 56])]; + tensor var_1226_end_0 = const()[name = string("op_1226_end_0"), val = tensor([1, 1024, 64])]; + tensor var_1226_end_mask_0 = const()[name = string("op_1226_end_mask_0"), val = tensor([true, true, true])]; + tensor var_1226_cast_fp16 = slice_by_index(begin = var_1226_begin_0, end = var_1226_end_0, end_mask = var_1226_end_mask_0, x = new_x_15_cast_fp16)[name = string("op_1226_cast_fp16")]; + string x_99_pad_type_0 = const()[name = string("x_99_pad_type_0"), val = string("valid")]; + int32 x_99_groups_0 = const()[name = string("x_99_groups_0"), val = int32(1024)]; + tensor x_99_strides_0 = const()[name = string("x_99_strides_0"), val = tensor([1])]; + tensor x_99_pad_0 = const()[name = string("x_99_pad_0"), val = tensor([0, 0])]; + tensor x_99_dilations_0 = const()[name = string("x_99_dilations_0"), val = tensor([1])]; + tensor encoder_layers_3_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87285632))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87294912))))[name = string("encoder_layers_3_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_99_cast_fp16 = conv(dilations = x_99_dilations_0, groups = x_99_groups_0, pad = x_99_pad_0, pad_type = x_99_pad_type_0, strides = x_99_strides_0, weight = encoder_layers_3_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_15_cast_fp16)[name = string("x_99_cast_fp16")]; + tensor input_211_perm_0 = const()[name = string("input_211_perm_0"), val = tensor([0, 2, 1])]; + tensor x_101_axes_0 = const()[name = string("x_101_axes_0"), val = tensor([-1])]; + tensor encoder_layers_3_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_3_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87297024)))]; + tensor encoder_layers_3_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_3_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87299136)))]; + tensor input_211_cast_fp16 = transpose(perm = input_211_perm_0, x = x_99_cast_fp16)[name = string("transpose_329")]; + tensor x_101_cast_fp16 = layer_norm(axes = x_101_axes_0, beta = encoder_layers_3_conv_batch_norm_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_3_conv_batch_norm_weight_to_fp16, x = input_211_cast_fp16)[name = string("x_101_cast_fp16")]; + tensor input_213_perm_0 = const()[name = string("input_213_perm_0"), val = tensor([0, 2, 1])]; + tensor input_213_cast_fp16 = transpose(perm = input_213_perm_0, x = x_101_cast_fp16)[name = string("transpose_328")]; + tensor input_215_cast_fp16 = silu(x = input_213_cast_fp16)[name = string("input_215_cast_fp16")]; + string x_103_pad_type_0 = const()[name = string("x_103_pad_type_0"), val = string("valid")]; + tensor x_103_strides_0 = const()[name = string("x_103_strides_0"), val = tensor([1])]; + tensor x_103_pad_0 = const()[name = string("x_103_pad_0"), val = tensor([0, 0])]; + tensor x_103_dilations_0 = const()[name = string("x_103_dilations_0"), val = tensor([1])]; + int32 x_103_groups_0 = const()[name = string("x_103_groups_0"), val = int32(1)]; + tensor encoder_layers_3_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87301248))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88349888))))[name = string("encoder_layers_3_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_103_cast_fp16 = conv(dilations = x_103_dilations_0, groups = x_103_groups_0, pad = x_103_pad_0, pad_type = x_103_pad_type_0, strides = x_103_strides_0, weight = encoder_layers_3_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_215_cast_fp16)[name = string("x_103_cast_fp16")]; + tensor input_217_perm_0 = const()[name = string("input_217_perm_0"), val = tensor([0, 2, 1])]; + tensor input_217_cast_fp16 = transpose(perm = input_217_perm_0, x = x_103_cast_fp16)[name = string("transpose_327")]; + tensor input_219_cast_fp16 = add(x = input_203_cast_fp16, y = input_217_cast_fp16)[name = string("input_219_cast_fp16")]; + tensor input_221_axes_0 = const()[name = string("input_221_axes_0"), val = tensor([-1])]; + tensor encoder_layers_3_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_3_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88352000)))]; + tensor encoder_layers_3_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_3_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88354112)))]; + tensor input_221_cast_fp16 = layer_norm(axes = input_221_axes_0, beta = encoder_layers_3_norm_feed_forward2_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_3_norm_feed_forward2_weight_to_fp16, x = input_219_cast_fp16)[name = string("input_221_cast_fp16")]; + tensor encoder_layers_3_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88356224))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(91502016))))[name = string("encoder_layers_3_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_3_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_3_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(91502208)))]; + tensor linear_35_cast_fp16 = linear(bias = encoder_layers_3_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_3_feed_forward2_linear1_weight_to_fp16_palettized, x = input_221_cast_fp16)[name = string("linear_35_cast_fp16")]; + tensor input_225_cast_fp16 = silu(x = linear_35_cast_fp16)[name = string("input_225_cast_fp16")]; + tensor encoder_layers_3_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(91510464))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94656256))))[name = string("encoder_layers_3_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_3_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_3_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94656448)))]; + tensor linear_36_cast_fp16 = linear(bias = encoder_layers_3_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_3_feed_forward2_linear2_weight_to_fp16_palettized, x = input_225_cast_fp16)[name = string("linear_36_cast_fp16")]; + fp16 var_1269_to_fp16 = const()[name = string("op_1269_to_fp16"), val = fp16(0x1p-1)]; + tensor var_1270_cast_fp16 = mul(x = linear_36_cast_fp16, y = var_1269_to_fp16)[name = string("op_1270_cast_fp16")]; + tensor input_231_cast_fp16 = add(x = input_219_cast_fp16, y = var_1270_cast_fp16)[name = string("input_231_cast_fp16")]; + tensor input_233_axes_0 = const()[name = string("input_233_axes_0"), val = tensor([-1])]; + tensor encoder_layers_3_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_3_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94658560)))]; + tensor encoder_layers_3_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_3_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94660672)))]; + tensor input_233_cast_fp16 = layer_norm(axes = input_233_axes_0, beta = encoder_layers_3_norm_out_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_3_norm_out_weight_to_fp16, x = input_231_cast_fp16)[name = string("input_233_cast_fp16")]; + tensor cache_17_begin_0 = const()[name = string("cache_17_begin_0"), val = tensor([4, 0, 0, 0])]; + tensor cache_17_end_0 = const()[name = string("cache_17_end_0"), val = tensor([5, 1, 42, 1024])]; + tensor cache_17_end_mask_0 = const()[name = string("cache_17_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_17_squeeze_mask_0 = const()[name = string("cache_17_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_17_cast_fp16 = slice_by_index(begin = cache_17_begin_0, end = cache_17_end_0, end_mask = cache_17_end_mask_0, squeeze_mask = cache_17_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_17_cast_fp16")]; + tensor cache_19_begin_0 = const()[name = string("cache_19_begin_0"), val = tensor([4, 0, 0, 0])]; + tensor cache_19_end_0 = const()[name = string("cache_19_end_0"), val = tensor([5, 1, 1024, 8])]; + tensor cache_19_end_mask_0 = const()[name = string("cache_19_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_19_squeeze_mask_0 = const()[name = string("cache_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_19_cast_fp16 = slice_by_index(begin = cache_19_begin_0, end = cache_19_end_0, end_mask = cache_19_end_mask_0, squeeze_mask = cache_19_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_19_cast_fp16")]; + tensor input_235_axes_0 = const()[name = string("input_235_axes_0"), val = tensor([-1])]; + tensor encoder_layers_4_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_4_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94662784)))]; + tensor encoder_layers_4_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_4_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94664896)))]; + tensor input_235_cast_fp16 = layer_norm(axes = input_235_axes_0, beta = encoder_layers_4_norm_feed_forward1_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_4_norm_feed_forward1_weight_to_fp16, x = input_233_cast_fp16)[name = string("input_235_cast_fp16")]; + tensor encoder_layers_4_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94667008))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97812800))))[name = string("encoder_layers_4_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_4_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_4_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97812992)))]; + tensor linear_37_cast_fp16 = linear(bias = encoder_layers_4_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_4_feed_forward1_linear1_weight_to_fp16_palettized, x = input_235_cast_fp16)[name = string("linear_37_cast_fp16")]; + tensor input_239_cast_fp16 = silu(x = linear_37_cast_fp16)[name = string("input_239_cast_fp16")]; + tensor encoder_layers_4_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97821248))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100967040))))[name = string("encoder_layers_4_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_4_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_4_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100967232)))]; + tensor linear_38_cast_fp16 = linear(bias = encoder_layers_4_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_4_feed_forward1_linear2_weight_to_fp16_palettized, x = input_239_cast_fp16)[name = string("linear_38_cast_fp16")]; + fp16 var_1306_to_fp16 = const()[name = string("op_1306_to_fp16"), val = fp16(0x1p-1)]; + tensor var_1307_cast_fp16 = mul(x = linear_38_cast_fp16, y = var_1306_to_fp16)[name = string("op_1307_cast_fp16")]; + tensor input_245_cast_fp16 = add(x = input_233_cast_fp16, y = var_1307_cast_fp16)[name = string("input_245_cast_fp16")]; + tensor key_9_axes_0 = const()[name = string("key_9_axes_0"), val = tensor([-1])]; + tensor encoder_layers_4_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_4_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100969344)))]; + tensor encoder_layers_4_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_4_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100971456)))]; + tensor key_9_cast_fp16 = layer_norm(axes = key_9_axes_0, beta = encoder_layers_4_norm_self_att_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_4_norm_self_att_weight_to_fp16, x = input_245_cast_fp16)[name = string("key_9_cast_fp16")]; + bool input_247_interleave_0 = const()[name = string("input_247_interleave_0"), val = bool(false)]; + tensor input_247_cast_fp16 = concat(axis = var_67, interleave = input_247_interleave_0, values = (cache_17_cast_fp16, key_9_cast_fp16))[name = string("input_247_cast_fp16")]; + bool var_1335_interleave_0 = const()[name = string("op_1335_interleave_0"), val = bool(false)]; + tensor var_1335_cast_fp16 = concat(axis = var_67, interleave = var_1335_interleave_0, values = key_9_cast_fp16)[name = string("op_1335_cast_fp16")]; + tensor encoder_layers_4_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100973568))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(101760064))))[name = string("encoder_layers_4_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_4_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_4_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(101760256)))]; + tensor linear_39_cast_fp16 = linear(bias = encoder_layers_4_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_4_self_attn_linear_q_weight_to_fp16_palettized, x = key_9_cast_fp16)[name = string("linear_39_cast_fp16")]; + tensor var_1340 = const()[name = string("op_1340"), val = tensor([1, -1, 8, 128])]; + tensor q_25_cast_fp16 = reshape(shape = var_1340, x = linear_39_cast_fp16)[name = string("q_25_cast_fp16")]; + tensor encoder_layers_4_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(101762368))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102548864))))[name = string("encoder_layers_4_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_4_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_4_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102549056)))]; + tensor linear_40_cast_fp16 = linear(bias = encoder_layers_4_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_4_self_attn_linear_k_weight_to_fp16_palettized, x = input_247_cast_fp16)[name = string("linear_40_cast_fp16")]; + tensor var_1345 = const()[name = string("op_1345"), val = tensor([1, -1, 8, 128])]; + tensor k_17_cast_fp16 = reshape(shape = var_1345, x = linear_40_cast_fp16)[name = string("k_17_cast_fp16")]; + tensor encoder_layers_4_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102551168))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103337664))))[name = string("encoder_layers_4_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_4_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_4_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103337856)))]; + tensor linear_41_cast_fp16 = linear(bias = encoder_layers_4_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_4_self_attn_linear_v_weight_to_fp16_palettized, x = input_247_cast_fp16)[name = string("linear_41_cast_fp16")]; + tensor var_1350 = const()[name = string("op_1350"), val = tensor([1, -1, 8, 128])]; + tensor v_9_cast_fp16 = reshape(shape = var_1350, x = linear_41_cast_fp16)[name = string("v_9_cast_fp16")]; + tensor value_17_perm_0 = const()[name = string("value_17_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_4_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_4_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103339968)))]; + tensor var_1363_cast_fp16 = add(x = q_25_cast_fp16, y = encoder_layers_4_self_attn_pos_bias_u_to_fp16)[name = string("op_1363_cast_fp16")]; + tensor encoder_layers_4_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_4_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103342080)))]; + tensor var_1365_cast_fp16 = add(x = q_25_cast_fp16, y = encoder_layers_4_self_attn_pos_bias_v_to_fp16)[name = string("op_1365_cast_fp16")]; + tensor q_with_bias_v_9_perm_0 = const()[name = string("q_with_bias_v_9_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_111_transpose_x_0 = const()[name = string("x_111_transpose_x_0"), val = bool(false)]; + bool x_111_transpose_y_0 = const()[name = string("x_111_transpose_y_0"), val = bool(false)]; + tensor op_1367_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103344192))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103543936))))[name = string("op_1367_to_fp16_quantized")]; + tensor q_with_bias_v_9_cast_fp16 = transpose(perm = q_with_bias_v_9_perm_0, x = var_1365_cast_fp16)[name = string("transpose_326")]; + tensor x_111_cast_fp16 = matmul(transpose_x = x_111_transpose_x_0, transpose_y = x_111_transpose_y_0, x = q_with_bias_v_9_cast_fp16, y = op_1367_to_fp16_quantized)[name = string("x_111_cast_fp16")]; + tensor x_113_pad_0 = const()[name = string("x_113_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_113_mode_0 = const()[name = string("x_113_mode_0"), val = string("constant")]; + fp16 const_131_to_fp16 = const()[name = string("const_131_to_fp16"), val = fp16(0x0p+0)]; + tensor x_113_cast_fp16 = pad(constant_val = const_131_to_fp16, mode = x_113_mode_0, pad = x_113_pad_0, x = x_111_cast_fp16)[name = string("x_113_cast_fp16")]; + tensor var_1375 = const()[name = string("op_1375"), val = tensor([1, 8, -1, 56])]; + tensor x_115_cast_fp16 = reshape(shape = var_1375, x = x_113_cast_fp16)[name = string("x_115_cast_fp16")]; + tensor var_1379_begin_0 = const()[name = string("op_1379_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_1379_end_0 = const()[name = string("op_1379_end_0"), val = tensor([1, 8, 196, 56])]; + tensor var_1379_end_mask_0 = const()[name = string("op_1379_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_1379_cast_fp16 = slice_by_index(begin = var_1379_begin_0, end = var_1379_end_0, end_mask = var_1379_end_mask_0, x = x_115_cast_fp16)[name = string("op_1379_cast_fp16")]; + tensor var_1380 = const()[name = string("op_1380"), val = tensor([1, 8, 56, 195])]; + tensor matrix_bd_17_cast_fp16 = reshape(shape = var_1380, x = var_1379_cast_fp16)[name = string("matrix_bd_17_cast_fp16")]; + bool matrix_ac_9_transpose_x_0 = const()[name = string("matrix_ac_9_transpose_x_0"), val = bool(false)]; + bool matrix_ac_9_transpose_y_0 = const()[name = string("matrix_ac_9_transpose_y_0"), val = bool(false)]; + tensor transpose_104_perm_0 = const()[name = string("transpose_104_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_105_perm_0 = const()[name = string("transpose_105_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_105 = transpose(perm = transpose_105_perm_0, x = k_17_cast_fp16)[name = string("transpose_324")]; + tensor transpose_104 = transpose(perm = transpose_104_perm_0, x = var_1363_cast_fp16)[name = string("transpose_325")]; + tensor matrix_ac_9_cast_fp16 = matmul(transpose_x = matrix_ac_9_transpose_x_0, transpose_y = matrix_ac_9_transpose_y_0, x = transpose_104, y = transpose_105)[name = string("matrix_ac_9_cast_fp16")]; + tensor matrix_bd_19_begin_0 = const()[name = string("matrix_bd_19_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_19_end_0 = const()[name = string("matrix_bd_19_end_0"), val = tensor([1, 8, 56, 98])]; + tensor matrix_bd_19_end_mask_0 = const()[name = string("matrix_bd_19_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_19_cast_fp16 = slice_by_index(begin = matrix_bd_19_begin_0, end = matrix_bd_19_end_0, end_mask = matrix_bd_19_end_mask_0, x = matrix_bd_17_cast_fp16)[name = string("matrix_bd_19_cast_fp16")]; + tensor var_1389_cast_fp16 = add(x = matrix_ac_9_cast_fp16, y = matrix_bd_19_cast_fp16)[name = string("op_1389_cast_fp16")]; + fp16 _inversed_scores_17_y_0_to_fp16 = const()[name = string("_inversed_scores_17_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_17_cast_fp16 = mul(x = var_1389_cast_fp16, y = _inversed_scores_17_y_0_to_fp16)[name = string("_inversed_scores_17_cast_fp16")]; + tensor scores_19_cast_fp16 = select(a = var_44_to_fp16, b = _inversed_scores_17_cast_fp16, cond = mask_11)[name = string("scores_19_cast_fp16")]; + tensor var_1395_cast_fp16 = softmax(axis = var_58, x = scores_19_cast_fp16)[name = string("op_1395_cast_fp16")]; + tensor input_249_cast_fp16 = select(a = var_43_to_fp16, b = var_1395_cast_fp16, cond = mask_11)[name = string("input_249_cast_fp16")]; + bool x_117_transpose_x_0 = const()[name = string("x_117_transpose_x_0"), val = bool(false)]; + bool x_117_transpose_y_0 = const()[name = string("x_117_transpose_y_0"), val = bool(false)]; + tensor value_17_cast_fp16 = transpose(perm = value_17_perm_0, x = v_9_cast_fp16)[name = string("transpose_323")]; + tensor x_117_cast_fp16 = matmul(transpose_x = x_117_transpose_x_0, transpose_y = x_117_transpose_y_0, x = input_249_cast_fp16, y = value_17_cast_fp16)[name = string("x_117_cast_fp16")]; + tensor var_1399_perm_0 = const()[name = string("op_1399_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1400 = const()[name = string("op_1400"), val = tensor([1, -1, 1024])]; + tensor var_1399_cast_fp16 = transpose(perm = var_1399_perm_0, x = x_117_cast_fp16)[name = string("transpose_322")]; + tensor input_251_cast_fp16 = reshape(shape = var_1400, x = var_1399_cast_fp16)[name = string("input_251_cast_fp16")]; + tensor encoder_layers_4_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103544448))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104330944))))[name = string("encoder_layers_4_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_4_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_4_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104331136)))]; + tensor linear_43_cast_fp16 = linear(bias = encoder_layers_4_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_4_self_attn_linear_out_weight_to_fp16_palettized, x = input_251_cast_fp16)[name = string("linear_43_cast_fp16")]; + tensor input_255_cast_fp16 = add(x = input_245_cast_fp16, y = linear_43_cast_fp16)[name = string("input_255_cast_fp16")]; + tensor x_121_axes_0 = const()[name = string("x_121_axes_0"), val = tensor([-1])]; + tensor encoder_layers_4_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_4_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104333248)))]; + tensor encoder_layers_4_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_4_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104335360)))]; + tensor x_121_cast_fp16 = layer_norm(axes = x_121_axes_0, beta = encoder_layers_4_norm_conv_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_4_norm_conv_weight_to_fp16, x = input_255_cast_fp16)[name = string("x_121_cast_fp16")]; + tensor input_257_perm_0 = const()[name = string("input_257_perm_0"), val = tensor([0, 2, 1])]; + string input_259_pad_type_0 = const()[name = string("input_259_pad_type_0"), val = string("valid")]; + tensor input_259_strides_0 = const()[name = string("input_259_strides_0"), val = tensor([1])]; + tensor input_259_pad_0 = const()[name = string("input_259_pad_0"), val = tensor([0, 0])]; + tensor input_259_dilations_0 = const()[name = string("input_259_dilations_0"), val = tensor([1])]; + int32 input_259_groups_0 = const()[name = string("input_259_groups_0"), val = int32(1)]; + tensor encoder_layers_4_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104337472))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106434688))))[name = string("encoder_layers_4_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_257_cast_fp16 = transpose(perm = input_257_perm_0, x = x_121_cast_fp16)[name = string("transpose_321")]; + tensor input_259_cast_fp16 = conv(dilations = input_259_dilations_0, groups = input_259_groups_0, pad = input_259_pad_0, pad_type = input_259_pad_type_0, strides = input_259_strides_0, weight = encoder_layers_4_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_257_cast_fp16)[name = string("input_259_cast_fp16")]; + int32 x_123_split_num_splits_0 = const()[name = string("x_123_split_num_splits_0"), val = int32(2)]; + int32 x_123_split_axis_0 = const()[name = string("x_123_split_axis_0"), val = int32(1)]; + tensor x_123_split_cast_fp16_0, tensor x_123_split_cast_fp16_1 = split(axis = x_123_split_axis_0, num_splits = x_123_split_num_splits_0, x = input_259_cast_fp16)[name = string("x_123_split_cast_fp16")]; + tensor x_123_split_1_sigmoid_cast_fp16 = sigmoid(x = x_123_split_cast_fp16_1)[name = string("x_123_split_1_sigmoid_cast_fp16")]; + tensor x_123_cast_fp16 = mul(x = x_123_split_cast_fp16_0, y = x_123_split_1_sigmoid_cast_fp16)[name = string("x_123_cast_fp16")]; + tensor input_261_cast_fp16 = select(a = var_43_to_fp16, b = x_123_cast_fp16, cond = var_574)[name = string("input_261_cast_fp16")]; + bool new_x_19_interleave_0 = const()[name = string("new_x_19_interleave_0"), val = bool(false)]; + tensor new_x_19_cast_fp16 = concat(axis = var_58, interleave = new_x_19_interleave_0, values = (cache_19_cast_fp16, input_261_cast_fp16))[name = string("new_x_19_cast_fp16")]; + tensor var_1439_begin_0 = const()[name = string("op_1439_begin_0"), val = tensor([0, 0, 56])]; + tensor var_1439_end_0 = const()[name = string("op_1439_end_0"), val = tensor([1, 1024, 64])]; + tensor var_1439_end_mask_0 = const()[name = string("op_1439_end_mask_0"), val = tensor([true, true, true])]; + tensor var_1439_cast_fp16 = slice_by_index(begin = var_1439_begin_0, end = var_1439_end_0, end_mask = var_1439_end_mask_0, x = new_x_19_cast_fp16)[name = string("op_1439_cast_fp16")]; + string x_125_pad_type_0 = const()[name = string("x_125_pad_type_0"), val = string("valid")]; + int32 x_125_groups_0 = const()[name = string("x_125_groups_0"), val = int32(1024)]; + tensor x_125_strides_0 = const()[name = string("x_125_strides_0"), val = tensor([1])]; + tensor x_125_pad_0 = const()[name = string("x_125_pad_0"), val = tensor([0, 0])]; + tensor x_125_dilations_0 = const()[name = string("x_125_dilations_0"), val = tensor([1])]; + tensor encoder_layers_4_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106438848))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106448128))))[name = string("encoder_layers_4_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_125_cast_fp16 = conv(dilations = x_125_dilations_0, groups = x_125_groups_0, pad = x_125_pad_0, pad_type = x_125_pad_type_0, strides = x_125_strides_0, weight = encoder_layers_4_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_19_cast_fp16)[name = string("x_125_cast_fp16")]; + tensor input_263_perm_0 = const()[name = string("input_263_perm_0"), val = tensor([0, 2, 1])]; + tensor x_127_axes_0 = const()[name = string("x_127_axes_0"), val = tensor([-1])]; + tensor encoder_layers_4_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_4_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106450240)))]; + tensor encoder_layers_4_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_4_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106452352)))]; + tensor input_263_cast_fp16 = transpose(perm = input_263_perm_0, x = x_125_cast_fp16)[name = string("transpose_320")]; + tensor x_127_cast_fp16 = layer_norm(axes = x_127_axes_0, beta = encoder_layers_4_conv_batch_norm_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_4_conv_batch_norm_weight_to_fp16, x = input_263_cast_fp16)[name = string("x_127_cast_fp16")]; + tensor input_265_perm_0 = const()[name = string("input_265_perm_0"), val = tensor([0, 2, 1])]; + tensor input_265_cast_fp16 = transpose(perm = input_265_perm_0, x = x_127_cast_fp16)[name = string("transpose_319")]; + tensor input_267_cast_fp16 = silu(x = input_265_cast_fp16)[name = string("input_267_cast_fp16")]; + string x_129_pad_type_0 = const()[name = string("x_129_pad_type_0"), val = string("valid")]; + tensor x_129_strides_0 = const()[name = string("x_129_strides_0"), val = tensor([1])]; + tensor x_129_pad_0 = const()[name = string("x_129_pad_0"), val = tensor([0, 0])]; + tensor x_129_dilations_0 = const()[name = string("x_129_dilations_0"), val = tensor([1])]; + int32 x_129_groups_0 = const()[name = string("x_129_groups_0"), val = int32(1)]; + tensor encoder_layers_4_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106454464))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107503104))))[name = string("encoder_layers_4_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_129_cast_fp16 = conv(dilations = x_129_dilations_0, groups = x_129_groups_0, pad = x_129_pad_0, pad_type = x_129_pad_type_0, strides = x_129_strides_0, weight = encoder_layers_4_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_267_cast_fp16)[name = string("x_129_cast_fp16")]; + tensor input_269_perm_0 = const()[name = string("input_269_perm_0"), val = tensor([0, 2, 1])]; + tensor input_269_cast_fp16 = transpose(perm = input_269_perm_0, x = x_129_cast_fp16)[name = string("transpose_318")]; + tensor input_271_cast_fp16 = add(x = input_255_cast_fp16, y = input_269_cast_fp16)[name = string("input_271_cast_fp16")]; + tensor input_273_axes_0 = const()[name = string("input_273_axes_0"), val = tensor([-1])]; + tensor encoder_layers_4_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_4_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107505216)))]; + tensor encoder_layers_4_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_4_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107507328)))]; + tensor input_273_cast_fp16 = layer_norm(axes = input_273_axes_0, beta = encoder_layers_4_norm_feed_forward2_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_4_norm_feed_forward2_weight_to_fp16, x = input_271_cast_fp16)[name = string("input_273_cast_fp16")]; + tensor encoder_layers_4_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107509440))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110655232))))[name = string("encoder_layers_4_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_4_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_4_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110655424)))]; + tensor linear_44_cast_fp16 = linear(bias = encoder_layers_4_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_4_feed_forward2_linear1_weight_to_fp16_palettized, x = input_273_cast_fp16)[name = string("linear_44_cast_fp16")]; + tensor input_277_cast_fp16 = silu(x = linear_44_cast_fp16)[name = string("input_277_cast_fp16")]; + tensor encoder_layers_4_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110663680))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113809472))))[name = string("encoder_layers_4_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_4_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_4_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113809664)))]; + tensor linear_45_cast_fp16 = linear(bias = encoder_layers_4_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_4_feed_forward2_linear2_weight_to_fp16_palettized, x = input_277_cast_fp16)[name = string("linear_45_cast_fp16")]; + fp16 var_1482_to_fp16 = const()[name = string("op_1482_to_fp16"), val = fp16(0x1p-1)]; + tensor var_1483_cast_fp16 = mul(x = linear_45_cast_fp16, y = var_1482_to_fp16)[name = string("op_1483_cast_fp16")]; + tensor input_283_cast_fp16 = add(x = input_271_cast_fp16, y = var_1483_cast_fp16)[name = string("input_283_cast_fp16")]; + tensor input_285_axes_0 = const()[name = string("input_285_axes_0"), val = tensor([-1])]; + tensor encoder_layers_4_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_4_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113811776)))]; + tensor encoder_layers_4_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_4_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113813888)))]; + tensor input_285_cast_fp16 = layer_norm(axes = input_285_axes_0, beta = encoder_layers_4_norm_out_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_4_norm_out_weight_to_fp16, x = input_283_cast_fp16)[name = string("input_285_cast_fp16")]; + tensor cache_21_begin_0 = const()[name = string("cache_21_begin_0"), val = tensor([5, 0, 0, 0])]; + tensor cache_21_end_0 = const()[name = string("cache_21_end_0"), val = tensor([6, 1, 42, 1024])]; + tensor cache_21_end_mask_0 = const()[name = string("cache_21_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_21_squeeze_mask_0 = const()[name = string("cache_21_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_21_cast_fp16 = slice_by_index(begin = cache_21_begin_0, end = cache_21_end_0, end_mask = cache_21_end_mask_0, squeeze_mask = cache_21_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_21_cast_fp16")]; + tensor cache_23_begin_0 = const()[name = string("cache_23_begin_0"), val = tensor([5, 0, 0, 0])]; + tensor cache_23_end_0 = const()[name = string("cache_23_end_0"), val = tensor([6, 1, 1024, 8])]; + tensor cache_23_end_mask_0 = const()[name = string("cache_23_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_23_squeeze_mask_0 = const()[name = string("cache_23_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_23_cast_fp16 = slice_by_index(begin = cache_23_begin_0, end = cache_23_end_0, end_mask = cache_23_end_mask_0, squeeze_mask = cache_23_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_23_cast_fp16")]; + tensor input_287_axes_0 = const()[name = string("input_287_axes_0"), val = tensor([-1])]; + tensor encoder_layers_5_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_5_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113816000)))]; + tensor encoder_layers_5_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_5_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113818112)))]; + tensor input_287_cast_fp16 = layer_norm(axes = input_287_axes_0, beta = encoder_layers_5_norm_feed_forward1_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_5_norm_feed_forward1_weight_to_fp16, x = input_285_cast_fp16)[name = string("input_287_cast_fp16")]; + tensor encoder_layers_5_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113820224))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116966016))))[name = string("encoder_layers_5_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_5_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_5_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116966208)))]; + tensor linear_46_cast_fp16 = linear(bias = encoder_layers_5_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_5_feed_forward1_linear1_weight_to_fp16_palettized, x = input_287_cast_fp16)[name = string("linear_46_cast_fp16")]; + tensor input_291_cast_fp16 = silu(x = linear_46_cast_fp16)[name = string("input_291_cast_fp16")]; + tensor encoder_layers_5_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116974464))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120120256))))[name = string("encoder_layers_5_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_5_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_5_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120120448)))]; + tensor linear_47_cast_fp16 = linear(bias = encoder_layers_5_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_5_feed_forward1_linear2_weight_to_fp16_palettized, x = input_291_cast_fp16)[name = string("linear_47_cast_fp16")]; + fp16 var_1519_to_fp16 = const()[name = string("op_1519_to_fp16"), val = fp16(0x1p-1)]; + tensor var_1520_cast_fp16 = mul(x = linear_47_cast_fp16, y = var_1519_to_fp16)[name = string("op_1520_cast_fp16")]; + tensor input_297_cast_fp16 = add(x = input_285_cast_fp16, y = var_1520_cast_fp16)[name = string("input_297_cast_fp16")]; + tensor key_11_axes_0 = const()[name = string("key_11_axes_0"), val = tensor([-1])]; + tensor encoder_layers_5_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_5_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120122560)))]; + tensor encoder_layers_5_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_5_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120124672)))]; + tensor key_11_cast_fp16 = layer_norm(axes = key_11_axes_0, beta = encoder_layers_5_norm_self_att_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_5_norm_self_att_weight_to_fp16, x = input_297_cast_fp16)[name = string("key_11_cast_fp16")]; + bool input_299_interleave_0 = const()[name = string("input_299_interleave_0"), val = bool(false)]; + tensor input_299_cast_fp16 = concat(axis = var_67, interleave = input_299_interleave_0, values = (cache_21_cast_fp16, key_11_cast_fp16))[name = string("input_299_cast_fp16")]; + bool var_1548_interleave_0 = const()[name = string("op_1548_interleave_0"), val = bool(false)]; + tensor var_1548_cast_fp16 = concat(axis = var_67, interleave = var_1548_interleave_0, values = key_11_cast_fp16)[name = string("op_1548_cast_fp16")]; + tensor encoder_layers_5_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120126784))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120913280))))[name = string("encoder_layers_5_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_5_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_5_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120913472)))]; + tensor linear_48_cast_fp16 = linear(bias = encoder_layers_5_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_5_self_attn_linear_q_weight_to_fp16_palettized, x = key_11_cast_fp16)[name = string("linear_48_cast_fp16")]; + tensor var_1553 = const()[name = string("op_1553"), val = tensor([1, -1, 8, 128])]; + tensor q_31_cast_fp16 = reshape(shape = var_1553, x = linear_48_cast_fp16)[name = string("q_31_cast_fp16")]; + tensor encoder_layers_5_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120915584))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121702080))))[name = string("encoder_layers_5_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_5_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_5_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121702272)))]; + tensor linear_49_cast_fp16 = linear(bias = encoder_layers_5_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_5_self_attn_linear_k_weight_to_fp16_palettized, x = input_299_cast_fp16)[name = string("linear_49_cast_fp16")]; + tensor var_1558 = const()[name = string("op_1558"), val = tensor([1, -1, 8, 128])]; + tensor k_21_cast_fp16 = reshape(shape = var_1558, x = linear_49_cast_fp16)[name = string("k_21_cast_fp16")]; + tensor encoder_layers_5_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121704384))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122490880))))[name = string("encoder_layers_5_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_5_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_5_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122491072)))]; + tensor linear_50_cast_fp16 = linear(bias = encoder_layers_5_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_5_self_attn_linear_v_weight_to_fp16_palettized, x = input_299_cast_fp16)[name = string("linear_50_cast_fp16")]; + tensor var_1563 = const()[name = string("op_1563"), val = tensor([1, -1, 8, 128])]; + tensor v_11_cast_fp16 = reshape(shape = var_1563, x = linear_50_cast_fp16)[name = string("v_11_cast_fp16")]; + tensor value_19_perm_0 = const()[name = string("value_19_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_5_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_5_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122493184)))]; + tensor var_1576_cast_fp16 = add(x = q_31_cast_fp16, y = encoder_layers_5_self_attn_pos_bias_u_to_fp16)[name = string("op_1576_cast_fp16")]; + tensor encoder_layers_5_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_5_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122495296)))]; + tensor var_1578_cast_fp16 = add(x = q_31_cast_fp16, y = encoder_layers_5_self_attn_pos_bias_v_to_fp16)[name = string("op_1578_cast_fp16")]; + tensor q_with_bias_v_11_perm_0 = const()[name = string("q_with_bias_v_11_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_137_transpose_x_0 = const()[name = string("x_137_transpose_x_0"), val = bool(false)]; + bool x_137_transpose_y_0 = const()[name = string("x_137_transpose_y_0"), val = bool(false)]; + tensor op_1580_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122497408))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122697152))))[name = string("op_1580_to_fp16_quantized")]; + tensor q_with_bias_v_11_cast_fp16 = transpose(perm = q_with_bias_v_11_perm_0, x = var_1578_cast_fp16)[name = string("transpose_317")]; + tensor x_137_cast_fp16 = matmul(transpose_x = x_137_transpose_x_0, transpose_y = x_137_transpose_y_0, x = q_with_bias_v_11_cast_fp16, y = op_1580_to_fp16_quantized)[name = string("x_137_cast_fp16")]; + tensor x_139_pad_0 = const()[name = string("x_139_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_139_mode_0 = const()[name = string("x_139_mode_0"), val = string("constant")]; + fp16 const_144_to_fp16 = const()[name = string("const_144_to_fp16"), val = fp16(0x0p+0)]; + tensor x_139_cast_fp16 = pad(constant_val = const_144_to_fp16, mode = x_139_mode_0, pad = x_139_pad_0, x = x_137_cast_fp16)[name = string("x_139_cast_fp16")]; + tensor var_1588 = const()[name = string("op_1588"), val = tensor([1, 8, -1, 56])]; + tensor x_141_cast_fp16 = reshape(shape = var_1588, x = x_139_cast_fp16)[name = string("x_141_cast_fp16")]; + tensor var_1592_begin_0 = const()[name = string("op_1592_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_1592_end_0 = const()[name = string("op_1592_end_0"), val = tensor([1, 8, 196, 56])]; + tensor var_1592_end_mask_0 = const()[name = string("op_1592_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_1592_cast_fp16 = slice_by_index(begin = var_1592_begin_0, end = var_1592_end_0, end_mask = var_1592_end_mask_0, x = x_141_cast_fp16)[name = string("op_1592_cast_fp16")]; + tensor var_1593 = const()[name = string("op_1593"), val = tensor([1, 8, 56, 195])]; + tensor matrix_bd_21_cast_fp16 = reshape(shape = var_1593, x = var_1592_cast_fp16)[name = string("matrix_bd_21_cast_fp16")]; + bool matrix_ac_11_transpose_x_0 = const()[name = string("matrix_ac_11_transpose_x_0"), val = bool(false)]; + bool matrix_ac_11_transpose_y_0 = const()[name = string("matrix_ac_11_transpose_y_0"), val = bool(false)]; + tensor transpose_106_perm_0 = const()[name = string("transpose_106_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_107_perm_0 = const()[name = string("transpose_107_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_107 = transpose(perm = transpose_107_perm_0, x = k_21_cast_fp16)[name = string("transpose_315")]; + tensor transpose_106 = transpose(perm = transpose_106_perm_0, x = var_1576_cast_fp16)[name = string("transpose_316")]; + tensor matrix_ac_11_cast_fp16 = matmul(transpose_x = matrix_ac_11_transpose_x_0, transpose_y = matrix_ac_11_transpose_y_0, x = transpose_106, y = transpose_107)[name = string("matrix_ac_11_cast_fp16")]; + tensor matrix_bd_23_begin_0 = const()[name = string("matrix_bd_23_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_23_end_0 = const()[name = string("matrix_bd_23_end_0"), val = tensor([1, 8, 56, 98])]; + tensor matrix_bd_23_end_mask_0 = const()[name = string("matrix_bd_23_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_23_cast_fp16 = slice_by_index(begin = matrix_bd_23_begin_0, end = matrix_bd_23_end_0, end_mask = matrix_bd_23_end_mask_0, x = matrix_bd_21_cast_fp16)[name = string("matrix_bd_23_cast_fp16")]; + tensor var_1602_cast_fp16 = add(x = matrix_ac_11_cast_fp16, y = matrix_bd_23_cast_fp16)[name = string("op_1602_cast_fp16")]; + fp16 _inversed_scores_21_y_0_to_fp16 = const()[name = string("_inversed_scores_21_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_21_cast_fp16 = mul(x = var_1602_cast_fp16, y = _inversed_scores_21_y_0_to_fp16)[name = string("_inversed_scores_21_cast_fp16")]; + tensor scores_23_cast_fp16 = select(a = var_44_to_fp16, b = _inversed_scores_21_cast_fp16, cond = mask_11)[name = string("scores_23_cast_fp16")]; + tensor var_1608_cast_fp16 = softmax(axis = var_58, x = scores_23_cast_fp16)[name = string("op_1608_cast_fp16")]; + tensor input_301_cast_fp16 = select(a = var_43_to_fp16, b = var_1608_cast_fp16, cond = mask_11)[name = string("input_301_cast_fp16")]; + bool x_143_transpose_x_0 = const()[name = string("x_143_transpose_x_0"), val = bool(false)]; + bool x_143_transpose_y_0 = const()[name = string("x_143_transpose_y_0"), val = bool(false)]; + tensor value_19_cast_fp16 = transpose(perm = value_19_perm_0, x = v_11_cast_fp16)[name = string("transpose_314")]; + tensor x_143_cast_fp16 = matmul(transpose_x = x_143_transpose_x_0, transpose_y = x_143_transpose_y_0, x = input_301_cast_fp16, y = value_19_cast_fp16)[name = string("x_143_cast_fp16")]; + tensor var_1612_perm_0 = const()[name = string("op_1612_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1613 = const()[name = string("op_1613"), val = tensor([1, -1, 1024])]; + tensor var_1612_cast_fp16 = transpose(perm = var_1612_perm_0, x = x_143_cast_fp16)[name = string("transpose_313")]; + tensor input_303_cast_fp16 = reshape(shape = var_1613, x = var_1612_cast_fp16)[name = string("input_303_cast_fp16")]; + tensor encoder_layers_5_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122697664))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123484160))))[name = string("encoder_layers_5_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_5_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_5_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123484352)))]; + tensor linear_52_cast_fp16 = linear(bias = encoder_layers_5_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_5_self_attn_linear_out_weight_to_fp16_palettized, x = input_303_cast_fp16)[name = string("linear_52_cast_fp16")]; + tensor input_307_cast_fp16 = add(x = input_297_cast_fp16, y = linear_52_cast_fp16)[name = string("input_307_cast_fp16")]; + tensor x_147_axes_0 = const()[name = string("x_147_axes_0"), val = tensor([-1])]; + tensor encoder_layers_5_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_5_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123486464)))]; + tensor encoder_layers_5_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_5_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123488576)))]; + tensor x_147_cast_fp16 = layer_norm(axes = x_147_axes_0, beta = encoder_layers_5_norm_conv_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_5_norm_conv_weight_to_fp16, x = input_307_cast_fp16)[name = string("x_147_cast_fp16")]; + tensor input_309_perm_0 = const()[name = string("input_309_perm_0"), val = tensor([0, 2, 1])]; + string input_311_pad_type_0 = const()[name = string("input_311_pad_type_0"), val = string("valid")]; + tensor input_311_strides_0 = const()[name = string("input_311_strides_0"), val = tensor([1])]; + tensor input_311_pad_0 = const()[name = string("input_311_pad_0"), val = tensor([0, 0])]; + tensor input_311_dilations_0 = const()[name = string("input_311_dilations_0"), val = tensor([1])]; + int32 input_311_groups_0 = const()[name = string("input_311_groups_0"), val = int32(1)]; + tensor encoder_layers_5_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123490688))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125587904))))[name = string("encoder_layers_5_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_309_cast_fp16 = transpose(perm = input_309_perm_0, x = x_147_cast_fp16)[name = string("transpose_312")]; + tensor input_311_cast_fp16 = conv(dilations = input_311_dilations_0, groups = input_311_groups_0, pad = input_311_pad_0, pad_type = input_311_pad_type_0, strides = input_311_strides_0, weight = encoder_layers_5_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_309_cast_fp16)[name = string("input_311_cast_fp16")]; + int32 x_149_split_num_splits_0 = const()[name = string("x_149_split_num_splits_0"), val = int32(2)]; + int32 x_149_split_axis_0 = const()[name = string("x_149_split_axis_0"), val = int32(1)]; + tensor x_149_split_cast_fp16_0, tensor x_149_split_cast_fp16_1 = split(axis = x_149_split_axis_0, num_splits = x_149_split_num_splits_0, x = input_311_cast_fp16)[name = string("x_149_split_cast_fp16")]; + tensor x_149_split_1_sigmoid_cast_fp16 = sigmoid(x = x_149_split_cast_fp16_1)[name = string("x_149_split_1_sigmoid_cast_fp16")]; + tensor x_149_cast_fp16 = mul(x = x_149_split_cast_fp16_0, y = x_149_split_1_sigmoid_cast_fp16)[name = string("x_149_cast_fp16")]; + tensor input_313_cast_fp16 = select(a = var_43_to_fp16, b = x_149_cast_fp16, cond = var_574)[name = string("input_313_cast_fp16")]; + bool new_x_23_interleave_0 = const()[name = string("new_x_23_interleave_0"), val = bool(false)]; + tensor new_x_23_cast_fp16 = concat(axis = var_58, interleave = new_x_23_interleave_0, values = (cache_23_cast_fp16, input_313_cast_fp16))[name = string("new_x_23_cast_fp16")]; + tensor var_1652_begin_0 = const()[name = string("op_1652_begin_0"), val = tensor([0, 0, 56])]; + tensor var_1652_end_0 = const()[name = string("op_1652_end_0"), val = tensor([1, 1024, 64])]; + tensor var_1652_end_mask_0 = const()[name = string("op_1652_end_mask_0"), val = tensor([true, true, true])]; + tensor var_1652_cast_fp16 = slice_by_index(begin = var_1652_begin_0, end = var_1652_end_0, end_mask = var_1652_end_mask_0, x = new_x_23_cast_fp16)[name = string("op_1652_cast_fp16")]; + string x_151_pad_type_0 = const()[name = string("x_151_pad_type_0"), val = string("valid")]; + int32 x_151_groups_0 = const()[name = string("x_151_groups_0"), val = int32(1024)]; + tensor x_151_strides_0 = const()[name = string("x_151_strides_0"), val = tensor([1])]; + tensor x_151_pad_0 = const()[name = string("x_151_pad_0"), val = tensor([0, 0])]; + tensor x_151_dilations_0 = const()[name = string("x_151_dilations_0"), val = tensor([1])]; + tensor encoder_layers_5_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125592064))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125601344))))[name = string("encoder_layers_5_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_151_cast_fp16 = conv(dilations = x_151_dilations_0, groups = x_151_groups_0, pad = x_151_pad_0, pad_type = x_151_pad_type_0, strides = x_151_strides_0, weight = encoder_layers_5_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_23_cast_fp16)[name = string("x_151_cast_fp16")]; + tensor input_315_perm_0 = const()[name = string("input_315_perm_0"), val = tensor([0, 2, 1])]; + tensor x_153_axes_0 = const()[name = string("x_153_axes_0"), val = tensor([-1])]; + tensor encoder_layers_5_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_5_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125603456)))]; + tensor encoder_layers_5_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_5_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125605568)))]; + tensor input_315_cast_fp16 = transpose(perm = input_315_perm_0, x = x_151_cast_fp16)[name = string("transpose_311")]; + tensor x_153_cast_fp16 = layer_norm(axes = x_153_axes_0, beta = encoder_layers_5_conv_batch_norm_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_5_conv_batch_norm_weight_to_fp16, x = input_315_cast_fp16)[name = string("x_153_cast_fp16")]; + tensor input_317_perm_0 = const()[name = string("input_317_perm_0"), val = tensor([0, 2, 1])]; + tensor input_317_cast_fp16 = transpose(perm = input_317_perm_0, x = x_153_cast_fp16)[name = string("transpose_310")]; + tensor input_319_cast_fp16 = silu(x = input_317_cast_fp16)[name = string("input_319_cast_fp16")]; + string x_155_pad_type_0 = const()[name = string("x_155_pad_type_0"), val = string("valid")]; + tensor x_155_strides_0 = const()[name = string("x_155_strides_0"), val = tensor([1])]; + tensor x_155_pad_0 = const()[name = string("x_155_pad_0"), val = tensor([0, 0])]; + tensor x_155_dilations_0 = const()[name = string("x_155_dilations_0"), val = tensor([1])]; + int32 x_155_groups_0 = const()[name = string("x_155_groups_0"), val = int32(1)]; + tensor encoder_layers_5_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125607680))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126656320))))[name = string("encoder_layers_5_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_155_cast_fp16 = conv(dilations = x_155_dilations_0, groups = x_155_groups_0, pad = x_155_pad_0, pad_type = x_155_pad_type_0, strides = x_155_strides_0, weight = encoder_layers_5_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_319_cast_fp16)[name = string("x_155_cast_fp16")]; + tensor input_321_perm_0 = const()[name = string("input_321_perm_0"), val = tensor([0, 2, 1])]; + tensor input_321_cast_fp16 = transpose(perm = input_321_perm_0, x = x_155_cast_fp16)[name = string("transpose_309")]; + tensor input_323_cast_fp16 = add(x = input_307_cast_fp16, y = input_321_cast_fp16)[name = string("input_323_cast_fp16")]; + tensor input_325_axes_0 = const()[name = string("input_325_axes_0"), val = tensor([-1])]; + tensor encoder_layers_5_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_5_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126658432)))]; + tensor encoder_layers_5_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_5_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126660544)))]; + tensor input_325_cast_fp16 = layer_norm(axes = input_325_axes_0, beta = encoder_layers_5_norm_feed_forward2_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_5_norm_feed_forward2_weight_to_fp16, x = input_323_cast_fp16)[name = string("input_325_cast_fp16")]; + tensor encoder_layers_5_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126662656))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(129808448))))[name = string("encoder_layers_5_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_5_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_5_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(129808640)))]; + tensor linear_53_cast_fp16 = linear(bias = encoder_layers_5_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_5_feed_forward2_linear1_weight_to_fp16_palettized, x = input_325_cast_fp16)[name = string("linear_53_cast_fp16")]; + tensor input_329_cast_fp16 = silu(x = linear_53_cast_fp16)[name = string("input_329_cast_fp16")]; + tensor encoder_layers_5_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(129816896))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132962688))))[name = string("encoder_layers_5_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_5_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_5_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132962880)))]; + tensor linear_54_cast_fp16 = linear(bias = encoder_layers_5_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_5_feed_forward2_linear2_weight_to_fp16_palettized, x = input_329_cast_fp16)[name = string("linear_54_cast_fp16")]; + fp16 var_1695_to_fp16 = const()[name = string("op_1695_to_fp16"), val = fp16(0x1p-1)]; + tensor var_1696_cast_fp16 = mul(x = linear_54_cast_fp16, y = var_1695_to_fp16)[name = string("op_1696_cast_fp16")]; + tensor input_335_cast_fp16 = add(x = input_323_cast_fp16, y = var_1696_cast_fp16)[name = string("input_335_cast_fp16")]; + tensor input_337_axes_0 = const()[name = string("input_337_axes_0"), val = tensor([-1])]; + tensor encoder_layers_5_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_5_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132964992)))]; + tensor encoder_layers_5_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_5_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132967104)))]; + tensor input_337_cast_fp16 = layer_norm(axes = input_337_axes_0, beta = encoder_layers_5_norm_out_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_5_norm_out_weight_to_fp16, x = input_335_cast_fp16)[name = string("input_337_cast_fp16")]; + tensor cache_25_begin_0 = const()[name = string("cache_25_begin_0"), val = tensor([6, 0, 0, 0])]; + tensor cache_25_end_0 = const()[name = string("cache_25_end_0"), val = tensor([7, 1, 42, 1024])]; + tensor cache_25_end_mask_0 = const()[name = string("cache_25_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_25_squeeze_mask_0 = const()[name = string("cache_25_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_25_cast_fp16 = slice_by_index(begin = cache_25_begin_0, end = cache_25_end_0, end_mask = cache_25_end_mask_0, squeeze_mask = cache_25_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_25_cast_fp16")]; + tensor cache_27_begin_0 = const()[name = string("cache_27_begin_0"), val = tensor([6, 0, 0, 0])]; + tensor cache_27_end_0 = const()[name = string("cache_27_end_0"), val = tensor([7, 1, 1024, 8])]; + tensor cache_27_end_mask_0 = const()[name = string("cache_27_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_27_squeeze_mask_0 = const()[name = string("cache_27_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_27_cast_fp16 = slice_by_index(begin = cache_27_begin_0, end = cache_27_end_0, end_mask = cache_27_end_mask_0, squeeze_mask = cache_27_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_27_cast_fp16")]; + tensor input_339_axes_0 = const()[name = string("input_339_axes_0"), val = tensor([-1])]; + tensor encoder_layers_6_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_6_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132969216)))]; + tensor encoder_layers_6_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_6_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132971328)))]; + tensor input_339_cast_fp16 = layer_norm(axes = input_339_axes_0, beta = encoder_layers_6_norm_feed_forward1_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_6_norm_feed_forward1_weight_to_fp16, x = input_337_cast_fp16)[name = string("input_339_cast_fp16")]; + tensor encoder_layers_6_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132973440))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136119232))))[name = string("encoder_layers_6_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_6_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_6_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136119424)))]; + tensor linear_55_cast_fp16 = linear(bias = encoder_layers_6_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_6_feed_forward1_linear1_weight_to_fp16_palettized, x = input_339_cast_fp16)[name = string("linear_55_cast_fp16")]; + tensor input_343_cast_fp16 = silu(x = linear_55_cast_fp16)[name = string("input_343_cast_fp16")]; + tensor encoder_layers_6_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136127680))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(139273472))))[name = string("encoder_layers_6_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_6_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_6_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(139273664)))]; + tensor linear_56_cast_fp16 = linear(bias = encoder_layers_6_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_6_feed_forward1_linear2_weight_to_fp16_palettized, x = input_343_cast_fp16)[name = string("linear_56_cast_fp16")]; + fp16 var_1732_to_fp16 = const()[name = string("op_1732_to_fp16"), val = fp16(0x1p-1)]; + tensor var_1733_cast_fp16 = mul(x = linear_56_cast_fp16, y = var_1732_to_fp16)[name = string("op_1733_cast_fp16")]; + tensor input_349_cast_fp16 = add(x = input_337_cast_fp16, y = var_1733_cast_fp16)[name = string("input_349_cast_fp16")]; + tensor key_13_axes_0 = const()[name = string("key_13_axes_0"), val = tensor([-1])]; + tensor encoder_layers_6_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_6_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(139275776)))]; + tensor encoder_layers_6_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_6_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(139277888)))]; + tensor key_13_cast_fp16 = layer_norm(axes = key_13_axes_0, beta = encoder_layers_6_norm_self_att_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_6_norm_self_att_weight_to_fp16, x = input_349_cast_fp16)[name = string("key_13_cast_fp16")]; + bool input_351_interleave_0 = const()[name = string("input_351_interleave_0"), val = bool(false)]; + tensor input_351_cast_fp16 = concat(axis = var_67, interleave = input_351_interleave_0, values = (cache_25_cast_fp16, key_13_cast_fp16))[name = string("input_351_cast_fp16")]; + bool var_1761_interleave_0 = const()[name = string("op_1761_interleave_0"), val = bool(false)]; + tensor var_1761_cast_fp16 = concat(axis = var_67, interleave = var_1761_interleave_0, values = key_13_cast_fp16)[name = string("op_1761_cast_fp16")]; + tensor encoder_layers_6_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(139280000))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140066496))))[name = string("encoder_layers_6_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_6_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_6_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140066688)))]; + tensor linear_57_cast_fp16 = linear(bias = encoder_layers_6_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_6_self_attn_linear_q_weight_to_fp16_palettized, x = key_13_cast_fp16)[name = string("linear_57_cast_fp16")]; + tensor var_1766 = const()[name = string("op_1766"), val = tensor([1, -1, 8, 128])]; + tensor q_37_cast_fp16 = reshape(shape = var_1766, x = linear_57_cast_fp16)[name = string("q_37_cast_fp16")]; + tensor encoder_layers_6_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140068800))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140855296))))[name = string("encoder_layers_6_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_6_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_6_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140855488)))]; + tensor linear_58_cast_fp16 = linear(bias = encoder_layers_6_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_6_self_attn_linear_k_weight_to_fp16_palettized, x = input_351_cast_fp16)[name = string("linear_58_cast_fp16")]; + tensor var_1771 = const()[name = string("op_1771"), val = tensor([1, -1, 8, 128])]; + tensor k_25_cast_fp16 = reshape(shape = var_1771, x = linear_58_cast_fp16)[name = string("k_25_cast_fp16")]; + tensor encoder_layers_6_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140857600))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141644096))))[name = string("encoder_layers_6_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_6_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_6_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141644288)))]; + tensor linear_59_cast_fp16 = linear(bias = encoder_layers_6_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_6_self_attn_linear_v_weight_to_fp16_palettized, x = input_351_cast_fp16)[name = string("linear_59_cast_fp16")]; + tensor var_1776 = const()[name = string("op_1776"), val = tensor([1, -1, 8, 128])]; + tensor v_13_cast_fp16 = reshape(shape = var_1776, x = linear_59_cast_fp16)[name = string("v_13_cast_fp16")]; + tensor value_21_perm_0 = const()[name = string("value_21_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_6_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_6_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141646400)))]; + tensor var_1789_cast_fp16 = add(x = q_37_cast_fp16, y = encoder_layers_6_self_attn_pos_bias_u_to_fp16)[name = string("op_1789_cast_fp16")]; + tensor encoder_layers_6_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_6_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141648512)))]; + tensor var_1791_cast_fp16 = add(x = q_37_cast_fp16, y = encoder_layers_6_self_attn_pos_bias_v_to_fp16)[name = string("op_1791_cast_fp16")]; + tensor q_with_bias_v_13_perm_0 = const()[name = string("q_with_bias_v_13_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_163_transpose_x_0 = const()[name = string("x_163_transpose_x_0"), val = bool(false)]; + bool x_163_transpose_y_0 = const()[name = string("x_163_transpose_y_0"), val = bool(false)]; + tensor op_1793_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141650624))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141850368))))[name = string("op_1793_to_fp16_quantized")]; + tensor q_with_bias_v_13_cast_fp16 = transpose(perm = q_with_bias_v_13_perm_0, x = var_1791_cast_fp16)[name = string("transpose_308")]; + tensor x_163_cast_fp16 = matmul(transpose_x = x_163_transpose_x_0, transpose_y = x_163_transpose_y_0, x = q_with_bias_v_13_cast_fp16, y = op_1793_to_fp16_quantized)[name = string("x_163_cast_fp16")]; + tensor x_165_pad_0 = const()[name = string("x_165_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_165_mode_0 = const()[name = string("x_165_mode_0"), val = string("constant")]; + fp16 const_157_to_fp16 = const()[name = string("const_157_to_fp16"), val = fp16(0x0p+0)]; + tensor x_165_cast_fp16 = pad(constant_val = const_157_to_fp16, mode = x_165_mode_0, pad = x_165_pad_0, x = x_163_cast_fp16)[name = string("x_165_cast_fp16")]; + tensor var_1801 = const()[name = string("op_1801"), val = tensor([1, 8, -1, 56])]; + tensor x_167_cast_fp16 = reshape(shape = var_1801, x = x_165_cast_fp16)[name = string("x_167_cast_fp16")]; + tensor var_1805_begin_0 = const()[name = string("op_1805_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_1805_end_0 = const()[name = string("op_1805_end_0"), val = tensor([1, 8, 196, 56])]; + tensor var_1805_end_mask_0 = const()[name = string("op_1805_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_1805_cast_fp16 = slice_by_index(begin = var_1805_begin_0, end = var_1805_end_0, end_mask = var_1805_end_mask_0, x = x_167_cast_fp16)[name = string("op_1805_cast_fp16")]; + tensor var_1806 = const()[name = string("op_1806"), val = tensor([1, 8, 56, 195])]; + tensor matrix_bd_25_cast_fp16 = reshape(shape = var_1806, x = var_1805_cast_fp16)[name = string("matrix_bd_25_cast_fp16")]; + bool matrix_ac_13_transpose_x_0 = const()[name = string("matrix_ac_13_transpose_x_0"), val = bool(false)]; + bool matrix_ac_13_transpose_y_0 = const()[name = string("matrix_ac_13_transpose_y_0"), val = bool(false)]; + tensor transpose_108_perm_0 = const()[name = string("transpose_108_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_109_perm_0 = const()[name = string("transpose_109_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_109 = transpose(perm = transpose_109_perm_0, x = k_25_cast_fp16)[name = string("transpose_306")]; + tensor transpose_108 = transpose(perm = transpose_108_perm_0, x = var_1789_cast_fp16)[name = string("transpose_307")]; + tensor matrix_ac_13_cast_fp16 = matmul(transpose_x = matrix_ac_13_transpose_x_0, transpose_y = matrix_ac_13_transpose_y_0, x = transpose_108, y = transpose_109)[name = string("matrix_ac_13_cast_fp16")]; + tensor matrix_bd_27_begin_0 = const()[name = string("matrix_bd_27_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_27_end_0 = const()[name = string("matrix_bd_27_end_0"), val = tensor([1, 8, 56, 98])]; + tensor matrix_bd_27_end_mask_0 = const()[name = string("matrix_bd_27_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_27_cast_fp16 = slice_by_index(begin = matrix_bd_27_begin_0, end = matrix_bd_27_end_0, end_mask = matrix_bd_27_end_mask_0, x = matrix_bd_25_cast_fp16)[name = string("matrix_bd_27_cast_fp16")]; + tensor var_1815_cast_fp16 = add(x = matrix_ac_13_cast_fp16, y = matrix_bd_27_cast_fp16)[name = string("op_1815_cast_fp16")]; + fp16 _inversed_scores_25_y_0_to_fp16 = const()[name = string("_inversed_scores_25_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_25_cast_fp16 = mul(x = var_1815_cast_fp16, y = _inversed_scores_25_y_0_to_fp16)[name = string("_inversed_scores_25_cast_fp16")]; + tensor scores_27_cast_fp16 = select(a = var_44_to_fp16, b = _inversed_scores_25_cast_fp16, cond = mask_11)[name = string("scores_27_cast_fp16")]; + tensor var_1821_cast_fp16 = softmax(axis = var_58, x = scores_27_cast_fp16)[name = string("op_1821_cast_fp16")]; + tensor input_353_cast_fp16 = select(a = var_43_to_fp16, b = var_1821_cast_fp16, cond = mask_11)[name = string("input_353_cast_fp16")]; + bool x_169_transpose_x_0 = const()[name = string("x_169_transpose_x_0"), val = bool(false)]; + bool x_169_transpose_y_0 = const()[name = string("x_169_transpose_y_0"), val = bool(false)]; + tensor value_21_cast_fp16 = transpose(perm = value_21_perm_0, x = v_13_cast_fp16)[name = string("transpose_305")]; + tensor x_169_cast_fp16 = matmul(transpose_x = x_169_transpose_x_0, transpose_y = x_169_transpose_y_0, x = input_353_cast_fp16, y = value_21_cast_fp16)[name = string("x_169_cast_fp16")]; + tensor var_1825_perm_0 = const()[name = string("op_1825_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1826 = const()[name = string("op_1826"), val = tensor([1, -1, 1024])]; + tensor var_1825_cast_fp16 = transpose(perm = var_1825_perm_0, x = x_169_cast_fp16)[name = string("transpose_304")]; + tensor input_355_cast_fp16 = reshape(shape = var_1826, x = var_1825_cast_fp16)[name = string("input_355_cast_fp16")]; + tensor encoder_layers_6_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141850880))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142637376))))[name = string("encoder_layers_6_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_6_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_6_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142637568)))]; + tensor linear_61_cast_fp16 = linear(bias = encoder_layers_6_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_6_self_attn_linear_out_weight_to_fp16_palettized, x = input_355_cast_fp16)[name = string("linear_61_cast_fp16")]; + tensor input_359_cast_fp16 = add(x = input_349_cast_fp16, y = linear_61_cast_fp16)[name = string("input_359_cast_fp16")]; + tensor x_173_axes_0 = const()[name = string("x_173_axes_0"), val = tensor([-1])]; + tensor encoder_layers_6_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_6_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142639680)))]; + tensor encoder_layers_6_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_6_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142641792)))]; + tensor x_173_cast_fp16 = layer_norm(axes = x_173_axes_0, beta = encoder_layers_6_norm_conv_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_6_norm_conv_weight_to_fp16, x = input_359_cast_fp16)[name = string("x_173_cast_fp16")]; + tensor input_361_perm_0 = const()[name = string("input_361_perm_0"), val = tensor([0, 2, 1])]; + string input_363_pad_type_0 = const()[name = string("input_363_pad_type_0"), val = string("valid")]; + tensor input_363_strides_0 = const()[name = string("input_363_strides_0"), val = tensor([1])]; + tensor input_363_pad_0 = const()[name = string("input_363_pad_0"), val = tensor([0, 0])]; + tensor input_363_dilations_0 = const()[name = string("input_363_dilations_0"), val = tensor([1])]; + int32 input_363_groups_0 = const()[name = string("input_363_groups_0"), val = int32(1)]; + tensor encoder_layers_6_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142643904))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144741120))))[name = string("encoder_layers_6_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_361_cast_fp16 = transpose(perm = input_361_perm_0, x = x_173_cast_fp16)[name = string("transpose_303")]; + tensor input_363_cast_fp16 = conv(dilations = input_363_dilations_0, groups = input_363_groups_0, pad = input_363_pad_0, pad_type = input_363_pad_type_0, strides = input_363_strides_0, weight = encoder_layers_6_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_361_cast_fp16)[name = string("input_363_cast_fp16")]; + int32 x_175_split_num_splits_0 = const()[name = string("x_175_split_num_splits_0"), val = int32(2)]; + int32 x_175_split_axis_0 = const()[name = string("x_175_split_axis_0"), val = int32(1)]; + tensor x_175_split_cast_fp16_0, tensor x_175_split_cast_fp16_1 = split(axis = x_175_split_axis_0, num_splits = x_175_split_num_splits_0, x = input_363_cast_fp16)[name = string("x_175_split_cast_fp16")]; + tensor x_175_split_1_sigmoid_cast_fp16 = sigmoid(x = x_175_split_cast_fp16_1)[name = string("x_175_split_1_sigmoid_cast_fp16")]; + tensor x_175_cast_fp16 = mul(x = x_175_split_cast_fp16_0, y = x_175_split_1_sigmoid_cast_fp16)[name = string("x_175_cast_fp16")]; + tensor input_365_cast_fp16 = select(a = var_43_to_fp16, b = x_175_cast_fp16, cond = var_574)[name = string("input_365_cast_fp16")]; + bool new_x_27_interleave_0 = const()[name = string("new_x_27_interleave_0"), val = bool(false)]; + tensor new_x_27_cast_fp16 = concat(axis = var_58, interleave = new_x_27_interleave_0, values = (cache_27_cast_fp16, input_365_cast_fp16))[name = string("new_x_27_cast_fp16")]; + tensor var_1865_begin_0 = const()[name = string("op_1865_begin_0"), val = tensor([0, 0, 56])]; + tensor var_1865_end_0 = const()[name = string("op_1865_end_0"), val = tensor([1, 1024, 64])]; + tensor var_1865_end_mask_0 = const()[name = string("op_1865_end_mask_0"), val = tensor([true, true, true])]; + tensor var_1865_cast_fp16 = slice_by_index(begin = var_1865_begin_0, end = var_1865_end_0, end_mask = var_1865_end_mask_0, x = new_x_27_cast_fp16)[name = string("op_1865_cast_fp16")]; + string x_177_pad_type_0 = const()[name = string("x_177_pad_type_0"), val = string("valid")]; + int32 x_177_groups_0 = const()[name = string("x_177_groups_0"), val = int32(1024)]; + tensor x_177_strides_0 = const()[name = string("x_177_strides_0"), val = tensor([1])]; + tensor x_177_pad_0 = const()[name = string("x_177_pad_0"), val = tensor([0, 0])]; + tensor x_177_dilations_0 = const()[name = string("x_177_dilations_0"), val = tensor([1])]; + tensor encoder_layers_6_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144745280))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144754560))))[name = string("encoder_layers_6_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_177_cast_fp16 = conv(dilations = x_177_dilations_0, groups = x_177_groups_0, pad = x_177_pad_0, pad_type = x_177_pad_type_0, strides = x_177_strides_0, weight = encoder_layers_6_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_27_cast_fp16)[name = string("x_177_cast_fp16")]; + tensor input_367_perm_0 = const()[name = string("input_367_perm_0"), val = tensor([0, 2, 1])]; + tensor x_179_axes_0 = const()[name = string("x_179_axes_0"), val = tensor([-1])]; + tensor encoder_layers_6_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_6_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144756672)))]; + tensor encoder_layers_6_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_6_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144758784)))]; + tensor input_367_cast_fp16 = transpose(perm = input_367_perm_0, x = x_177_cast_fp16)[name = string("transpose_302")]; + tensor x_179_cast_fp16 = layer_norm(axes = x_179_axes_0, beta = encoder_layers_6_conv_batch_norm_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_6_conv_batch_norm_weight_to_fp16, x = input_367_cast_fp16)[name = string("x_179_cast_fp16")]; + tensor input_369_perm_0 = const()[name = string("input_369_perm_0"), val = tensor([0, 2, 1])]; + tensor input_369_cast_fp16 = transpose(perm = input_369_perm_0, x = x_179_cast_fp16)[name = string("transpose_301")]; + tensor input_371_cast_fp16 = silu(x = input_369_cast_fp16)[name = string("input_371_cast_fp16")]; + string x_181_pad_type_0 = const()[name = string("x_181_pad_type_0"), val = string("valid")]; + tensor x_181_strides_0 = const()[name = string("x_181_strides_0"), val = tensor([1])]; + tensor x_181_pad_0 = const()[name = string("x_181_pad_0"), val = tensor([0, 0])]; + tensor x_181_dilations_0 = const()[name = string("x_181_dilations_0"), val = tensor([1])]; + int32 x_181_groups_0 = const()[name = string("x_181_groups_0"), val = int32(1)]; + tensor encoder_layers_6_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144760896))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(145809536))))[name = string("encoder_layers_6_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_181_cast_fp16 = conv(dilations = x_181_dilations_0, groups = x_181_groups_0, pad = x_181_pad_0, pad_type = x_181_pad_type_0, strides = x_181_strides_0, weight = encoder_layers_6_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_371_cast_fp16)[name = string("x_181_cast_fp16")]; + tensor input_373_perm_0 = const()[name = string("input_373_perm_0"), val = tensor([0, 2, 1])]; + tensor input_373_cast_fp16 = transpose(perm = input_373_perm_0, x = x_181_cast_fp16)[name = string("transpose_300")]; + tensor input_375_cast_fp16 = add(x = input_359_cast_fp16, y = input_373_cast_fp16)[name = string("input_375_cast_fp16")]; + tensor input_377_axes_0 = const()[name = string("input_377_axes_0"), val = tensor([-1])]; + tensor encoder_layers_6_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_6_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(145811648)))]; + tensor encoder_layers_6_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_6_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(145813760)))]; + tensor input_377_cast_fp16 = layer_norm(axes = input_377_axes_0, beta = encoder_layers_6_norm_feed_forward2_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_6_norm_feed_forward2_weight_to_fp16, x = input_375_cast_fp16)[name = string("input_377_cast_fp16")]; + tensor encoder_layers_6_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(145815872))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(148961664))))[name = string("encoder_layers_6_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_6_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_6_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(148961856)))]; + tensor linear_62_cast_fp16 = linear(bias = encoder_layers_6_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_6_feed_forward2_linear1_weight_to_fp16_palettized, x = input_377_cast_fp16)[name = string("linear_62_cast_fp16")]; + tensor input_381_cast_fp16 = silu(x = linear_62_cast_fp16)[name = string("input_381_cast_fp16")]; + tensor encoder_layers_6_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(148970112))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152115904))))[name = string("encoder_layers_6_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_6_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_6_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152116096)))]; + tensor linear_63_cast_fp16 = linear(bias = encoder_layers_6_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_6_feed_forward2_linear2_weight_to_fp16_palettized, x = input_381_cast_fp16)[name = string("linear_63_cast_fp16")]; + fp16 var_1908_to_fp16 = const()[name = string("op_1908_to_fp16"), val = fp16(0x1p-1)]; + tensor var_1909_cast_fp16 = mul(x = linear_63_cast_fp16, y = var_1908_to_fp16)[name = string("op_1909_cast_fp16")]; + tensor input_387_cast_fp16 = add(x = input_375_cast_fp16, y = var_1909_cast_fp16)[name = string("input_387_cast_fp16")]; + tensor input_389_axes_0 = const()[name = string("input_389_axes_0"), val = tensor([-1])]; + tensor encoder_layers_6_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_6_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152118208)))]; + tensor encoder_layers_6_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_6_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152120320)))]; + tensor input_389_cast_fp16 = layer_norm(axes = input_389_axes_0, beta = encoder_layers_6_norm_out_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_6_norm_out_weight_to_fp16, x = input_387_cast_fp16)[name = string("input_389_cast_fp16")]; + tensor cache_29_begin_0 = const()[name = string("cache_29_begin_0"), val = tensor([7, 0, 0, 0])]; + tensor cache_29_end_0 = const()[name = string("cache_29_end_0"), val = tensor([8, 1, 42, 1024])]; + tensor cache_29_end_mask_0 = const()[name = string("cache_29_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_29_squeeze_mask_0 = const()[name = string("cache_29_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_29_cast_fp16 = slice_by_index(begin = cache_29_begin_0, end = cache_29_end_0, end_mask = cache_29_end_mask_0, squeeze_mask = cache_29_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_29_cast_fp16")]; + tensor cache_31_begin_0 = const()[name = string("cache_31_begin_0"), val = tensor([7, 0, 0, 0])]; + tensor cache_31_end_0 = const()[name = string("cache_31_end_0"), val = tensor([8, 1, 1024, 8])]; + tensor cache_31_end_mask_0 = const()[name = string("cache_31_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_31_squeeze_mask_0 = const()[name = string("cache_31_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_31_cast_fp16 = slice_by_index(begin = cache_31_begin_0, end = cache_31_end_0, end_mask = cache_31_end_mask_0, squeeze_mask = cache_31_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_31_cast_fp16")]; + tensor input_391_axes_0 = const()[name = string("input_391_axes_0"), val = tensor([-1])]; + tensor encoder_layers_7_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_7_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152122432)))]; + tensor encoder_layers_7_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_7_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152124544)))]; + tensor input_391_cast_fp16 = layer_norm(axes = input_391_axes_0, beta = encoder_layers_7_norm_feed_forward1_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_7_norm_feed_forward1_weight_to_fp16, x = input_389_cast_fp16)[name = string("input_391_cast_fp16")]; + tensor encoder_layers_7_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152126656))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(155272448))))[name = string("encoder_layers_7_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_7_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_7_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(155272640)))]; + tensor linear_64_cast_fp16 = linear(bias = encoder_layers_7_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_7_feed_forward1_linear1_weight_to_fp16_palettized, x = input_391_cast_fp16)[name = string("linear_64_cast_fp16")]; + tensor input_395_cast_fp16 = silu(x = linear_64_cast_fp16)[name = string("input_395_cast_fp16")]; + tensor encoder_layers_7_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(155280896))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158426688))))[name = string("encoder_layers_7_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_7_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_7_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158426880)))]; + tensor linear_65_cast_fp16 = linear(bias = encoder_layers_7_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_7_feed_forward1_linear2_weight_to_fp16_palettized, x = input_395_cast_fp16)[name = string("linear_65_cast_fp16")]; + fp16 var_1945_to_fp16 = const()[name = string("op_1945_to_fp16"), val = fp16(0x1p-1)]; + tensor var_1946_cast_fp16 = mul(x = linear_65_cast_fp16, y = var_1945_to_fp16)[name = string("op_1946_cast_fp16")]; + tensor input_401_cast_fp16 = add(x = input_389_cast_fp16, y = var_1946_cast_fp16)[name = string("input_401_cast_fp16")]; + tensor key_15_axes_0 = const()[name = string("key_15_axes_0"), val = tensor([-1])]; + tensor encoder_layers_7_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_7_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158428992)))]; + tensor encoder_layers_7_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_7_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158431104)))]; + tensor key_15_cast_fp16 = layer_norm(axes = key_15_axes_0, beta = encoder_layers_7_norm_self_att_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_7_norm_self_att_weight_to_fp16, x = input_401_cast_fp16)[name = string("key_15_cast_fp16")]; + bool input_403_interleave_0 = const()[name = string("input_403_interleave_0"), val = bool(false)]; + tensor input_403_cast_fp16 = concat(axis = var_67, interleave = input_403_interleave_0, values = (cache_29_cast_fp16, key_15_cast_fp16))[name = string("input_403_cast_fp16")]; + bool var_1974_interleave_0 = const()[name = string("op_1974_interleave_0"), val = bool(false)]; + tensor var_1974_cast_fp16 = concat(axis = var_67, interleave = var_1974_interleave_0, values = key_15_cast_fp16)[name = string("op_1974_cast_fp16")]; + tensor encoder_layers_7_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158433216))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(159219712))))[name = string("encoder_layers_7_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_7_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_7_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(159219904)))]; + tensor linear_66_cast_fp16 = linear(bias = encoder_layers_7_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_7_self_attn_linear_q_weight_to_fp16_palettized, x = key_15_cast_fp16)[name = string("linear_66_cast_fp16")]; + tensor var_1979 = const()[name = string("op_1979"), val = tensor([1, -1, 8, 128])]; + tensor q_43_cast_fp16 = reshape(shape = var_1979, x = linear_66_cast_fp16)[name = string("q_43_cast_fp16")]; + tensor encoder_layers_7_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(159222016))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160008512))))[name = string("encoder_layers_7_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_7_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_7_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160008704)))]; + tensor linear_67_cast_fp16 = linear(bias = encoder_layers_7_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_7_self_attn_linear_k_weight_to_fp16_palettized, x = input_403_cast_fp16)[name = string("linear_67_cast_fp16")]; + tensor var_1984 = const()[name = string("op_1984"), val = tensor([1, -1, 8, 128])]; + tensor k_29_cast_fp16 = reshape(shape = var_1984, x = linear_67_cast_fp16)[name = string("k_29_cast_fp16")]; + tensor encoder_layers_7_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160010816))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160797312))))[name = string("encoder_layers_7_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_7_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_7_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160797504)))]; + tensor linear_68_cast_fp16 = linear(bias = encoder_layers_7_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_7_self_attn_linear_v_weight_to_fp16_palettized, x = input_403_cast_fp16)[name = string("linear_68_cast_fp16")]; + tensor var_1989 = const()[name = string("op_1989"), val = tensor([1, -1, 8, 128])]; + tensor v_15_cast_fp16 = reshape(shape = var_1989, x = linear_68_cast_fp16)[name = string("v_15_cast_fp16")]; + tensor value_23_perm_0 = const()[name = string("value_23_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_7_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_7_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160799616)))]; + tensor var_2002_cast_fp16 = add(x = q_43_cast_fp16, y = encoder_layers_7_self_attn_pos_bias_u_to_fp16)[name = string("op_2002_cast_fp16")]; + tensor encoder_layers_7_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_7_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160801728)))]; + tensor var_2004_cast_fp16 = add(x = q_43_cast_fp16, y = encoder_layers_7_self_attn_pos_bias_v_to_fp16)[name = string("op_2004_cast_fp16")]; + tensor q_with_bias_v_15_perm_0 = const()[name = string("q_with_bias_v_15_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_189_transpose_x_0 = const()[name = string("x_189_transpose_x_0"), val = bool(false)]; + bool x_189_transpose_y_0 = const()[name = string("x_189_transpose_y_0"), val = bool(false)]; + tensor op_2006_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160803840))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161003584))))[name = string("op_2006_to_fp16_quantized")]; + tensor q_with_bias_v_15_cast_fp16 = transpose(perm = q_with_bias_v_15_perm_0, x = var_2004_cast_fp16)[name = string("transpose_299")]; + tensor x_189_cast_fp16 = matmul(transpose_x = x_189_transpose_x_0, transpose_y = x_189_transpose_y_0, x = q_with_bias_v_15_cast_fp16, y = op_2006_to_fp16_quantized)[name = string("x_189_cast_fp16")]; + tensor x_191_pad_0 = const()[name = string("x_191_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_191_mode_0 = const()[name = string("x_191_mode_0"), val = string("constant")]; + fp16 const_170_to_fp16 = const()[name = string("const_170_to_fp16"), val = fp16(0x0p+0)]; + tensor x_191_cast_fp16 = pad(constant_val = const_170_to_fp16, mode = x_191_mode_0, pad = x_191_pad_0, x = x_189_cast_fp16)[name = string("x_191_cast_fp16")]; + tensor var_2014 = const()[name = string("op_2014"), val = tensor([1, 8, -1, 56])]; + tensor x_193_cast_fp16 = reshape(shape = var_2014, x = x_191_cast_fp16)[name = string("x_193_cast_fp16")]; + tensor var_2018_begin_0 = const()[name = string("op_2018_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_2018_end_0 = const()[name = string("op_2018_end_0"), val = tensor([1, 8, 196, 56])]; + tensor var_2018_end_mask_0 = const()[name = string("op_2018_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_2018_cast_fp16 = slice_by_index(begin = var_2018_begin_0, end = var_2018_end_0, end_mask = var_2018_end_mask_0, x = x_193_cast_fp16)[name = string("op_2018_cast_fp16")]; + tensor var_2019 = const()[name = string("op_2019"), val = tensor([1, 8, 56, 195])]; + tensor matrix_bd_29_cast_fp16 = reshape(shape = var_2019, x = var_2018_cast_fp16)[name = string("matrix_bd_29_cast_fp16")]; + bool matrix_ac_15_transpose_x_0 = const()[name = string("matrix_ac_15_transpose_x_0"), val = bool(false)]; + bool matrix_ac_15_transpose_y_0 = const()[name = string("matrix_ac_15_transpose_y_0"), val = bool(false)]; + tensor transpose_110_perm_0 = const()[name = string("transpose_110_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_111_perm_0 = const()[name = string("transpose_111_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_111 = transpose(perm = transpose_111_perm_0, x = k_29_cast_fp16)[name = string("transpose_297")]; + tensor transpose_110 = transpose(perm = transpose_110_perm_0, x = var_2002_cast_fp16)[name = string("transpose_298")]; + tensor matrix_ac_15_cast_fp16 = matmul(transpose_x = matrix_ac_15_transpose_x_0, transpose_y = matrix_ac_15_transpose_y_0, x = transpose_110, y = transpose_111)[name = string("matrix_ac_15_cast_fp16")]; + tensor matrix_bd_31_begin_0 = const()[name = string("matrix_bd_31_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_31_end_0 = const()[name = string("matrix_bd_31_end_0"), val = tensor([1, 8, 56, 98])]; + tensor matrix_bd_31_end_mask_0 = const()[name = string("matrix_bd_31_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_31_cast_fp16 = slice_by_index(begin = matrix_bd_31_begin_0, end = matrix_bd_31_end_0, end_mask = matrix_bd_31_end_mask_0, x = matrix_bd_29_cast_fp16)[name = string("matrix_bd_31_cast_fp16")]; + tensor var_2028_cast_fp16 = add(x = matrix_ac_15_cast_fp16, y = matrix_bd_31_cast_fp16)[name = string("op_2028_cast_fp16")]; + fp16 _inversed_scores_29_y_0_to_fp16 = const()[name = string("_inversed_scores_29_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_29_cast_fp16 = mul(x = var_2028_cast_fp16, y = _inversed_scores_29_y_0_to_fp16)[name = string("_inversed_scores_29_cast_fp16")]; + tensor scores_31_cast_fp16 = select(a = var_44_to_fp16, b = _inversed_scores_29_cast_fp16, cond = mask_11)[name = string("scores_31_cast_fp16")]; + tensor var_2034_cast_fp16 = softmax(axis = var_58, x = scores_31_cast_fp16)[name = string("op_2034_cast_fp16")]; + tensor input_405_cast_fp16 = select(a = var_43_to_fp16, b = var_2034_cast_fp16, cond = mask_11)[name = string("input_405_cast_fp16")]; + bool x_195_transpose_x_0 = const()[name = string("x_195_transpose_x_0"), val = bool(false)]; + bool x_195_transpose_y_0 = const()[name = string("x_195_transpose_y_0"), val = bool(false)]; + tensor value_23_cast_fp16 = transpose(perm = value_23_perm_0, x = v_15_cast_fp16)[name = string("transpose_296")]; + tensor x_195_cast_fp16 = matmul(transpose_x = x_195_transpose_x_0, transpose_y = x_195_transpose_y_0, x = input_405_cast_fp16, y = value_23_cast_fp16)[name = string("x_195_cast_fp16")]; + tensor var_2038_perm_0 = const()[name = string("op_2038_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_2039 = const()[name = string("op_2039"), val = tensor([1, -1, 1024])]; + tensor var_2038_cast_fp16 = transpose(perm = var_2038_perm_0, x = x_195_cast_fp16)[name = string("transpose_295")]; + tensor input_407_cast_fp16 = reshape(shape = var_2039, x = var_2038_cast_fp16)[name = string("input_407_cast_fp16")]; + tensor encoder_layers_7_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161004096))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161790592))))[name = string("encoder_layers_7_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_7_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_7_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161790784)))]; + tensor linear_70_cast_fp16 = linear(bias = encoder_layers_7_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_7_self_attn_linear_out_weight_to_fp16_palettized, x = input_407_cast_fp16)[name = string("linear_70_cast_fp16")]; + tensor input_411_cast_fp16 = add(x = input_401_cast_fp16, y = linear_70_cast_fp16)[name = string("input_411_cast_fp16")]; + tensor x_199_axes_0 = const()[name = string("x_199_axes_0"), val = tensor([-1])]; + tensor encoder_layers_7_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_7_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161792896)))]; + tensor encoder_layers_7_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_7_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161795008)))]; + tensor x_199_cast_fp16 = layer_norm(axes = x_199_axes_0, beta = encoder_layers_7_norm_conv_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_7_norm_conv_weight_to_fp16, x = input_411_cast_fp16)[name = string("x_199_cast_fp16")]; + tensor input_413_perm_0 = const()[name = string("input_413_perm_0"), val = tensor([0, 2, 1])]; + string input_415_pad_type_0 = const()[name = string("input_415_pad_type_0"), val = string("valid")]; + tensor input_415_strides_0 = const()[name = string("input_415_strides_0"), val = tensor([1])]; + tensor input_415_pad_0 = const()[name = string("input_415_pad_0"), val = tensor([0, 0])]; + tensor input_415_dilations_0 = const()[name = string("input_415_dilations_0"), val = tensor([1])]; + int32 input_415_groups_0 = const()[name = string("input_415_groups_0"), val = int32(1)]; + tensor encoder_layers_7_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161797120))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163894336))))[name = string("encoder_layers_7_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_413_cast_fp16 = transpose(perm = input_413_perm_0, x = x_199_cast_fp16)[name = string("transpose_294")]; + tensor input_415_cast_fp16 = conv(dilations = input_415_dilations_0, groups = input_415_groups_0, pad = input_415_pad_0, pad_type = input_415_pad_type_0, strides = input_415_strides_0, weight = encoder_layers_7_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_413_cast_fp16)[name = string("input_415_cast_fp16")]; + int32 x_201_split_num_splits_0 = const()[name = string("x_201_split_num_splits_0"), val = int32(2)]; + int32 x_201_split_axis_0 = const()[name = string("x_201_split_axis_0"), val = int32(1)]; + tensor x_201_split_cast_fp16_0, tensor x_201_split_cast_fp16_1 = split(axis = x_201_split_axis_0, num_splits = x_201_split_num_splits_0, x = input_415_cast_fp16)[name = string("x_201_split_cast_fp16")]; + tensor x_201_split_1_sigmoid_cast_fp16 = sigmoid(x = x_201_split_cast_fp16_1)[name = string("x_201_split_1_sigmoid_cast_fp16")]; + tensor x_201_cast_fp16 = mul(x = x_201_split_cast_fp16_0, y = x_201_split_1_sigmoid_cast_fp16)[name = string("x_201_cast_fp16")]; + tensor input_417_cast_fp16 = select(a = var_43_to_fp16, b = x_201_cast_fp16, cond = var_574)[name = string("input_417_cast_fp16")]; + bool new_x_31_interleave_0 = const()[name = string("new_x_31_interleave_0"), val = bool(false)]; + tensor new_x_31_cast_fp16 = concat(axis = var_58, interleave = new_x_31_interleave_0, values = (cache_31_cast_fp16, input_417_cast_fp16))[name = string("new_x_31_cast_fp16")]; + tensor var_2078_begin_0 = const()[name = string("op_2078_begin_0"), val = tensor([0, 0, 56])]; + tensor var_2078_end_0 = const()[name = string("op_2078_end_0"), val = tensor([1, 1024, 64])]; + tensor var_2078_end_mask_0 = const()[name = string("op_2078_end_mask_0"), val = tensor([true, true, true])]; + tensor var_2078_cast_fp16 = slice_by_index(begin = var_2078_begin_0, end = var_2078_end_0, end_mask = var_2078_end_mask_0, x = new_x_31_cast_fp16)[name = string("op_2078_cast_fp16")]; + string x_203_pad_type_0 = const()[name = string("x_203_pad_type_0"), val = string("valid")]; + int32 x_203_groups_0 = const()[name = string("x_203_groups_0"), val = int32(1024)]; + tensor x_203_strides_0 = const()[name = string("x_203_strides_0"), val = tensor([1])]; + tensor x_203_pad_0 = const()[name = string("x_203_pad_0"), val = tensor([0, 0])]; + tensor x_203_dilations_0 = const()[name = string("x_203_dilations_0"), val = tensor([1])]; + tensor encoder_layers_7_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163898496))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163907776))))[name = string("encoder_layers_7_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_203_cast_fp16 = conv(dilations = x_203_dilations_0, groups = x_203_groups_0, pad = x_203_pad_0, pad_type = x_203_pad_type_0, strides = x_203_strides_0, weight = encoder_layers_7_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_31_cast_fp16)[name = string("x_203_cast_fp16")]; + tensor input_419_perm_0 = const()[name = string("input_419_perm_0"), val = tensor([0, 2, 1])]; + tensor x_205_axes_0 = const()[name = string("x_205_axes_0"), val = tensor([-1])]; + tensor encoder_layers_7_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_7_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163909888)))]; + tensor encoder_layers_7_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_7_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163912000)))]; + tensor input_419_cast_fp16 = transpose(perm = input_419_perm_0, x = x_203_cast_fp16)[name = string("transpose_293")]; + tensor x_205_cast_fp16 = layer_norm(axes = x_205_axes_0, beta = encoder_layers_7_conv_batch_norm_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_7_conv_batch_norm_weight_to_fp16, x = input_419_cast_fp16)[name = string("x_205_cast_fp16")]; + tensor input_421_perm_0 = const()[name = string("input_421_perm_0"), val = tensor([0, 2, 1])]; + tensor input_421_cast_fp16 = transpose(perm = input_421_perm_0, x = x_205_cast_fp16)[name = string("transpose_292")]; + tensor input_423_cast_fp16 = silu(x = input_421_cast_fp16)[name = string("input_423_cast_fp16")]; + string x_207_pad_type_0 = const()[name = string("x_207_pad_type_0"), val = string("valid")]; + tensor x_207_strides_0 = const()[name = string("x_207_strides_0"), val = tensor([1])]; + tensor x_207_pad_0 = const()[name = string("x_207_pad_0"), val = tensor([0, 0])]; + tensor x_207_dilations_0 = const()[name = string("x_207_dilations_0"), val = tensor([1])]; + int32 x_207_groups_0 = const()[name = string("x_207_groups_0"), val = int32(1)]; + tensor encoder_layers_7_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163914112))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(164962752))))[name = string("encoder_layers_7_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_207_cast_fp16 = conv(dilations = x_207_dilations_0, groups = x_207_groups_0, pad = x_207_pad_0, pad_type = x_207_pad_type_0, strides = x_207_strides_0, weight = encoder_layers_7_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_423_cast_fp16)[name = string("x_207_cast_fp16")]; + tensor input_425_perm_0 = const()[name = string("input_425_perm_0"), val = tensor([0, 2, 1])]; + tensor input_425_cast_fp16 = transpose(perm = input_425_perm_0, x = x_207_cast_fp16)[name = string("transpose_291")]; + tensor input_427_cast_fp16 = add(x = input_411_cast_fp16, y = input_425_cast_fp16)[name = string("input_427_cast_fp16")]; + tensor input_429_axes_0 = const()[name = string("input_429_axes_0"), val = tensor([-1])]; + tensor encoder_layers_7_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_7_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(164964864)))]; + tensor encoder_layers_7_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_7_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(164966976)))]; + tensor input_429_cast_fp16 = layer_norm(axes = input_429_axes_0, beta = encoder_layers_7_norm_feed_forward2_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_7_norm_feed_forward2_weight_to_fp16, x = input_427_cast_fp16)[name = string("input_429_cast_fp16")]; + tensor encoder_layers_7_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(164969088))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(168114880))))[name = string("encoder_layers_7_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_7_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_7_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(168115072)))]; + tensor linear_71_cast_fp16 = linear(bias = encoder_layers_7_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_7_feed_forward2_linear1_weight_to_fp16_palettized, x = input_429_cast_fp16)[name = string("linear_71_cast_fp16")]; + tensor input_433_cast_fp16 = silu(x = linear_71_cast_fp16)[name = string("input_433_cast_fp16")]; + tensor encoder_layers_7_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(168123328))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(171269120))))[name = string("encoder_layers_7_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_7_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_7_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(171269312)))]; + tensor linear_72_cast_fp16 = linear(bias = encoder_layers_7_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_7_feed_forward2_linear2_weight_to_fp16_palettized, x = input_433_cast_fp16)[name = string("linear_72_cast_fp16")]; + fp16 var_2121_to_fp16 = const()[name = string("op_2121_to_fp16"), val = fp16(0x1p-1)]; + tensor var_2122_cast_fp16 = mul(x = linear_72_cast_fp16, y = var_2121_to_fp16)[name = string("op_2122_cast_fp16")]; + tensor input_439_cast_fp16 = add(x = input_427_cast_fp16, y = var_2122_cast_fp16)[name = string("input_439_cast_fp16")]; + tensor input_441_axes_0 = const()[name = string("input_441_axes_0"), val = tensor([-1])]; + tensor encoder_layers_7_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_7_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(171271424)))]; + tensor encoder_layers_7_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_7_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(171273536)))]; + tensor input_441_cast_fp16 = layer_norm(axes = input_441_axes_0, beta = encoder_layers_7_norm_out_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_7_norm_out_weight_to_fp16, x = input_439_cast_fp16)[name = string("input_441_cast_fp16")]; + tensor cache_33_begin_0 = const()[name = string("cache_33_begin_0"), val = tensor([8, 0, 0, 0])]; + tensor cache_33_end_0 = const()[name = string("cache_33_end_0"), val = tensor([9, 1, 42, 1024])]; + tensor cache_33_end_mask_0 = const()[name = string("cache_33_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_33_squeeze_mask_0 = const()[name = string("cache_33_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_33_cast_fp16 = slice_by_index(begin = cache_33_begin_0, end = cache_33_end_0, end_mask = cache_33_end_mask_0, squeeze_mask = cache_33_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_33_cast_fp16")]; + tensor cache_35_begin_0 = const()[name = string("cache_35_begin_0"), val = tensor([8, 0, 0, 0])]; + tensor cache_35_end_0 = const()[name = string("cache_35_end_0"), val = tensor([9, 1, 1024, 8])]; + tensor cache_35_end_mask_0 = const()[name = string("cache_35_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_35_squeeze_mask_0 = const()[name = string("cache_35_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_35_cast_fp16 = slice_by_index(begin = cache_35_begin_0, end = cache_35_end_0, end_mask = cache_35_end_mask_0, squeeze_mask = cache_35_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_35_cast_fp16")]; + tensor input_443_axes_0 = const()[name = string("input_443_axes_0"), val = tensor([-1])]; + tensor encoder_layers_8_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_8_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(171275648)))]; + tensor encoder_layers_8_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_8_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(171277760)))]; + tensor input_443_cast_fp16 = layer_norm(axes = input_443_axes_0, beta = encoder_layers_8_norm_feed_forward1_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_8_norm_feed_forward1_weight_to_fp16, x = input_441_cast_fp16)[name = string("input_443_cast_fp16")]; + tensor encoder_layers_8_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(171279872))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(174425664))))[name = string("encoder_layers_8_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_8_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_8_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(174425856)))]; + tensor linear_73_cast_fp16 = linear(bias = encoder_layers_8_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_8_feed_forward1_linear1_weight_to_fp16_palettized, x = input_443_cast_fp16)[name = string("linear_73_cast_fp16")]; + tensor input_447_cast_fp16 = silu(x = linear_73_cast_fp16)[name = string("input_447_cast_fp16")]; + tensor encoder_layers_8_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(174434112))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(177579904))))[name = string("encoder_layers_8_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_8_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_8_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(177580096)))]; + tensor linear_74_cast_fp16 = linear(bias = encoder_layers_8_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_8_feed_forward1_linear2_weight_to_fp16_palettized, x = input_447_cast_fp16)[name = string("linear_74_cast_fp16")]; + fp16 var_2158_to_fp16 = const()[name = string("op_2158_to_fp16"), val = fp16(0x1p-1)]; + tensor var_2159_cast_fp16 = mul(x = linear_74_cast_fp16, y = var_2158_to_fp16)[name = string("op_2159_cast_fp16")]; + tensor input_453_cast_fp16 = add(x = input_441_cast_fp16, y = var_2159_cast_fp16)[name = string("input_453_cast_fp16")]; + tensor key_17_axes_0 = const()[name = string("key_17_axes_0"), val = tensor([-1])]; + tensor encoder_layers_8_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_8_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(177582208)))]; + tensor encoder_layers_8_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_8_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(177584320)))]; + tensor key_17_cast_fp16 = layer_norm(axes = key_17_axes_0, beta = encoder_layers_8_norm_self_att_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_8_norm_self_att_weight_to_fp16, x = input_453_cast_fp16)[name = string("key_17_cast_fp16")]; + bool input_455_interleave_0 = const()[name = string("input_455_interleave_0"), val = bool(false)]; + tensor input_455_cast_fp16 = concat(axis = var_67, interleave = input_455_interleave_0, values = (cache_33_cast_fp16, key_17_cast_fp16))[name = string("input_455_cast_fp16")]; + bool var_2187_interleave_0 = const()[name = string("op_2187_interleave_0"), val = bool(false)]; + tensor var_2187_cast_fp16 = concat(axis = var_67, interleave = var_2187_interleave_0, values = key_17_cast_fp16)[name = string("op_2187_cast_fp16")]; + tensor encoder_layers_8_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(177586432))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(178372928))))[name = string("encoder_layers_8_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_8_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_8_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(178373120)))]; + tensor linear_75_cast_fp16 = linear(bias = encoder_layers_8_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_8_self_attn_linear_q_weight_to_fp16_palettized, x = key_17_cast_fp16)[name = string("linear_75_cast_fp16")]; + tensor var_2192 = const()[name = string("op_2192"), val = tensor([1, -1, 8, 128])]; + tensor q_49_cast_fp16 = reshape(shape = var_2192, x = linear_75_cast_fp16)[name = string("q_49_cast_fp16")]; + tensor encoder_layers_8_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(178375232))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179161728))))[name = string("encoder_layers_8_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_8_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_8_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179161920)))]; + tensor linear_76_cast_fp16 = linear(bias = encoder_layers_8_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_8_self_attn_linear_k_weight_to_fp16_palettized, x = input_455_cast_fp16)[name = string("linear_76_cast_fp16")]; + tensor var_2197 = const()[name = string("op_2197"), val = tensor([1, -1, 8, 128])]; + tensor k_33_cast_fp16 = reshape(shape = var_2197, x = linear_76_cast_fp16)[name = string("k_33_cast_fp16")]; + tensor encoder_layers_8_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179164032))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179950528))))[name = string("encoder_layers_8_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_8_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_8_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179950720)))]; + tensor linear_77_cast_fp16 = linear(bias = encoder_layers_8_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_8_self_attn_linear_v_weight_to_fp16_palettized, x = input_455_cast_fp16)[name = string("linear_77_cast_fp16")]; + tensor var_2202 = const()[name = string("op_2202"), val = tensor([1, -1, 8, 128])]; + tensor v_17_cast_fp16 = reshape(shape = var_2202, x = linear_77_cast_fp16)[name = string("v_17_cast_fp16")]; + tensor value_25_perm_0 = const()[name = string("value_25_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_8_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_8_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179952832)))]; + tensor var_2215_cast_fp16 = add(x = q_49_cast_fp16, y = encoder_layers_8_self_attn_pos_bias_u_to_fp16)[name = string("op_2215_cast_fp16")]; + tensor encoder_layers_8_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_8_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179954944)))]; + tensor var_2217_cast_fp16 = add(x = q_49_cast_fp16, y = encoder_layers_8_self_attn_pos_bias_v_to_fp16)[name = string("op_2217_cast_fp16")]; + tensor q_with_bias_v_17_perm_0 = const()[name = string("q_with_bias_v_17_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_215_transpose_x_0 = const()[name = string("x_215_transpose_x_0"), val = bool(false)]; + bool x_215_transpose_y_0 = const()[name = string("x_215_transpose_y_0"), val = bool(false)]; + tensor op_2219_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179957056))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(180156800))))[name = string("op_2219_to_fp16_quantized")]; + tensor q_with_bias_v_17_cast_fp16 = transpose(perm = q_with_bias_v_17_perm_0, x = var_2217_cast_fp16)[name = string("transpose_290")]; + tensor x_215_cast_fp16 = matmul(transpose_x = x_215_transpose_x_0, transpose_y = x_215_transpose_y_0, x = q_with_bias_v_17_cast_fp16, y = op_2219_to_fp16_quantized)[name = string("x_215_cast_fp16")]; + tensor x_217_pad_0 = const()[name = string("x_217_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_217_mode_0 = const()[name = string("x_217_mode_0"), val = string("constant")]; + fp16 const_183_to_fp16 = const()[name = string("const_183_to_fp16"), val = fp16(0x0p+0)]; + tensor x_217_cast_fp16 = pad(constant_val = const_183_to_fp16, mode = x_217_mode_0, pad = x_217_pad_0, x = x_215_cast_fp16)[name = string("x_217_cast_fp16")]; + tensor var_2227 = const()[name = string("op_2227"), val = tensor([1, 8, -1, 56])]; + tensor x_219_cast_fp16 = reshape(shape = var_2227, x = x_217_cast_fp16)[name = string("x_219_cast_fp16")]; + tensor var_2231_begin_0 = const()[name = string("op_2231_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_2231_end_0 = const()[name = string("op_2231_end_0"), val = tensor([1, 8, 196, 56])]; + tensor var_2231_end_mask_0 = const()[name = string("op_2231_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_2231_cast_fp16 = slice_by_index(begin = var_2231_begin_0, end = var_2231_end_0, end_mask = var_2231_end_mask_0, x = x_219_cast_fp16)[name = string("op_2231_cast_fp16")]; + tensor var_2232 = const()[name = string("op_2232"), val = tensor([1, 8, 56, 195])]; + tensor matrix_bd_33_cast_fp16 = reshape(shape = var_2232, x = var_2231_cast_fp16)[name = string("matrix_bd_33_cast_fp16")]; + bool matrix_ac_17_transpose_x_0 = const()[name = string("matrix_ac_17_transpose_x_0"), val = bool(false)]; + bool matrix_ac_17_transpose_y_0 = const()[name = string("matrix_ac_17_transpose_y_0"), val = bool(false)]; + tensor transpose_112_perm_0 = const()[name = string("transpose_112_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_113_perm_0 = const()[name = string("transpose_113_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_113 = transpose(perm = transpose_113_perm_0, x = k_33_cast_fp16)[name = string("transpose_288")]; + tensor transpose_112 = transpose(perm = transpose_112_perm_0, x = var_2215_cast_fp16)[name = string("transpose_289")]; + tensor matrix_ac_17_cast_fp16 = matmul(transpose_x = matrix_ac_17_transpose_x_0, transpose_y = matrix_ac_17_transpose_y_0, x = transpose_112, y = transpose_113)[name = string("matrix_ac_17_cast_fp16")]; + tensor matrix_bd_35_begin_0 = const()[name = string("matrix_bd_35_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_35_end_0 = const()[name = string("matrix_bd_35_end_0"), val = tensor([1, 8, 56, 98])]; + tensor matrix_bd_35_end_mask_0 = const()[name = string("matrix_bd_35_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_35_cast_fp16 = slice_by_index(begin = matrix_bd_35_begin_0, end = matrix_bd_35_end_0, end_mask = matrix_bd_35_end_mask_0, x = matrix_bd_33_cast_fp16)[name = string("matrix_bd_35_cast_fp16")]; + tensor var_2241_cast_fp16 = add(x = matrix_ac_17_cast_fp16, y = matrix_bd_35_cast_fp16)[name = string("op_2241_cast_fp16")]; + fp16 _inversed_scores_33_y_0_to_fp16 = const()[name = string("_inversed_scores_33_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_33_cast_fp16 = mul(x = var_2241_cast_fp16, y = _inversed_scores_33_y_0_to_fp16)[name = string("_inversed_scores_33_cast_fp16")]; + tensor scores_35_cast_fp16 = select(a = var_44_to_fp16, b = _inversed_scores_33_cast_fp16, cond = mask_11)[name = string("scores_35_cast_fp16")]; + tensor var_2247_cast_fp16 = softmax(axis = var_58, x = scores_35_cast_fp16)[name = string("op_2247_cast_fp16")]; + tensor input_457_cast_fp16 = select(a = var_43_to_fp16, b = var_2247_cast_fp16, cond = mask_11)[name = string("input_457_cast_fp16")]; + bool x_221_transpose_x_0 = const()[name = string("x_221_transpose_x_0"), val = bool(false)]; + bool x_221_transpose_y_0 = const()[name = string("x_221_transpose_y_0"), val = bool(false)]; + tensor value_25_cast_fp16 = transpose(perm = value_25_perm_0, x = v_17_cast_fp16)[name = string("transpose_287")]; + tensor x_221_cast_fp16 = matmul(transpose_x = x_221_transpose_x_0, transpose_y = x_221_transpose_y_0, x = input_457_cast_fp16, y = value_25_cast_fp16)[name = string("x_221_cast_fp16")]; + tensor var_2251_perm_0 = const()[name = string("op_2251_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_2252 = const()[name = string("op_2252"), val = tensor([1, -1, 1024])]; + tensor var_2251_cast_fp16 = transpose(perm = var_2251_perm_0, x = x_221_cast_fp16)[name = string("transpose_286")]; + tensor input_459_cast_fp16 = reshape(shape = var_2252, x = var_2251_cast_fp16)[name = string("input_459_cast_fp16")]; + tensor encoder_layers_8_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(180157312))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(180943808))))[name = string("encoder_layers_8_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_8_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_8_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(180944000)))]; + tensor linear_79_cast_fp16 = linear(bias = encoder_layers_8_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_8_self_attn_linear_out_weight_to_fp16_palettized, x = input_459_cast_fp16)[name = string("linear_79_cast_fp16")]; + tensor input_463_cast_fp16 = add(x = input_453_cast_fp16, y = linear_79_cast_fp16)[name = string("input_463_cast_fp16")]; + tensor x_225_axes_0 = const()[name = string("x_225_axes_0"), val = tensor([-1])]; + tensor encoder_layers_8_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_8_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(180946112)))]; + tensor encoder_layers_8_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_8_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(180948224)))]; + tensor x_225_cast_fp16 = layer_norm(axes = x_225_axes_0, beta = encoder_layers_8_norm_conv_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_8_norm_conv_weight_to_fp16, x = input_463_cast_fp16)[name = string("x_225_cast_fp16")]; + tensor input_465_perm_0 = const()[name = string("input_465_perm_0"), val = tensor([0, 2, 1])]; + string input_467_pad_type_0 = const()[name = string("input_467_pad_type_0"), val = string("valid")]; + tensor input_467_strides_0 = const()[name = string("input_467_strides_0"), val = tensor([1])]; + tensor input_467_pad_0 = const()[name = string("input_467_pad_0"), val = tensor([0, 0])]; + tensor input_467_dilations_0 = const()[name = string("input_467_dilations_0"), val = tensor([1])]; + int32 input_467_groups_0 = const()[name = string("input_467_groups_0"), val = int32(1)]; + tensor encoder_layers_8_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(180950336))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183047552))))[name = string("encoder_layers_8_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_465_cast_fp16 = transpose(perm = input_465_perm_0, x = x_225_cast_fp16)[name = string("transpose_285")]; + tensor input_467_cast_fp16 = conv(dilations = input_467_dilations_0, groups = input_467_groups_0, pad = input_467_pad_0, pad_type = input_467_pad_type_0, strides = input_467_strides_0, weight = encoder_layers_8_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_465_cast_fp16)[name = string("input_467_cast_fp16")]; + int32 x_227_split_num_splits_0 = const()[name = string("x_227_split_num_splits_0"), val = int32(2)]; + int32 x_227_split_axis_0 = const()[name = string("x_227_split_axis_0"), val = int32(1)]; + tensor x_227_split_cast_fp16_0, tensor x_227_split_cast_fp16_1 = split(axis = x_227_split_axis_0, num_splits = x_227_split_num_splits_0, x = input_467_cast_fp16)[name = string("x_227_split_cast_fp16")]; + tensor x_227_split_1_sigmoid_cast_fp16 = sigmoid(x = x_227_split_cast_fp16_1)[name = string("x_227_split_1_sigmoid_cast_fp16")]; + tensor x_227_cast_fp16 = mul(x = x_227_split_cast_fp16_0, y = x_227_split_1_sigmoid_cast_fp16)[name = string("x_227_cast_fp16")]; + tensor input_469_cast_fp16 = select(a = var_43_to_fp16, b = x_227_cast_fp16, cond = var_574)[name = string("input_469_cast_fp16")]; + bool new_x_35_interleave_0 = const()[name = string("new_x_35_interleave_0"), val = bool(false)]; + tensor new_x_35_cast_fp16 = concat(axis = var_58, interleave = new_x_35_interleave_0, values = (cache_35_cast_fp16, input_469_cast_fp16))[name = string("new_x_35_cast_fp16")]; + tensor var_2291_begin_0 = const()[name = string("op_2291_begin_0"), val = tensor([0, 0, 56])]; + tensor var_2291_end_0 = const()[name = string("op_2291_end_0"), val = tensor([1, 1024, 64])]; + tensor var_2291_end_mask_0 = const()[name = string("op_2291_end_mask_0"), val = tensor([true, true, true])]; + tensor var_2291_cast_fp16 = slice_by_index(begin = var_2291_begin_0, end = var_2291_end_0, end_mask = var_2291_end_mask_0, x = new_x_35_cast_fp16)[name = string("op_2291_cast_fp16")]; + string x_229_pad_type_0 = const()[name = string("x_229_pad_type_0"), val = string("valid")]; + int32 x_229_groups_0 = const()[name = string("x_229_groups_0"), val = int32(1024)]; + tensor x_229_strides_0 = const()[name = string("x_229_strides_0"), val = tensor([1])]; + tensor x_229_pad_0 = const()[name = string("x_229_pad_0"), val = tensor([0, 0])]; + tensor x_229_dilations_0 = const()[name = string("x_229_dilations_0"), val = tensor([1])]; + tensor encoder_layers_8_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183051712))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183060992))))[name = string("encoder_layers_8_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_229_cast_fp16 = conv(dilations = x_229_dilations_0, groups = x_229_groups_0, pad = x_229_pad_0, pad_type = x_229_pad_type_0, strides = x_229_strides_0, weight = encoder_layers_8_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_35_cast_fp16)[name = string("x_229_cast_fp16")]; + tensor input_471_perm_0 = const()[name = string("input_471_perm_0"), val = tensor([0, 2, 1])]; + tensor x_231_axes_0 = const()[name = string("x_231_axes_0"), val = tensor([-1])]; + tensor encoder_layers_8_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_8_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183063104)))]; + tensor encoder_layers_8_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_8_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183065216)))]; + tensor input_471_cast_fp16 = transpose(perm = input_471_perm_0, x = x_229_cast_fp16)[name = string("transpose_284")]; + tensor x_231_cast_fp16 = layer_norm(axes = x_231_axes_0, beta = encoder_layers_8_conv_batch_norm_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_8_conv_batch_norm_weight_to_fp16, x = input_471_cast_fp16)[name = string("x_231_cast_fp16")]; + tensor input_473_perm_0 = const()[name = string("input_473_perm_0"), val = tensor([0, 2, 1])]; + tensor input_473_cast_fp16 = transpose(perm = input_473_perm_0, x = x_231_cast_fp16)[name = string("transpose_283")]; + tensor input_475_cast_fp16 = silu(x = input_473_cast_fp16)[name = string("input_475_cast_fp16")]; + string x_233_pad_type_0 = const()[name = string("x_233_pad_type_0"), val = string("valid")]; + tensor x_233_strides_0 = const()[name = string("x_233_strides_0"), val = tensor([1])]; + tensor x_233_pad_0 = const()[name = string("x_233_pad_0"), val = tensor([0, 0])]; + tensor x_233_dilations_0 = const()[name = string("x_233_dilations_0"), val = tensor([1])]; + int32 x_233_groups_0 = const()[name = string("x_233_groups_0"), val = int32(1)]; + tensor encoder_layers_8_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183067328))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(184115968))))[name = string("encoder_layers_8_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_233_cast_fp16 = conv(dilations = x_233_dilations_0, groups = x_233_groups_0, pad = x_233_pad_0, pad_type = x_233_pad_type_0, strides = x_233_strides_0, weight = encoder_layers_8_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_475_cast_fp16)[name = string("x_233_cast_fp16")]; + tensor input_477_perm_0 = const()[name = string("input_477_perm_0"), val = tensor([0, 2, 1])]; + tensor input_477_cast_fp16 = transpose(perm = input_477_perm_0, x = x_233_cast_fp16)[name = string("transpose_282")]; + tensor input_479_cast_fp16 = add(x = input_463_cast_fp16, y = input_477_cast_fp16)[name = string("input_479_cast_fp16")]; + tensor input_481_axes_0 = const()[name = string("input_481_axes_0"), val = tensor([-1])]; + tensor encoder_layers_8_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_8_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(184118080)))]; + tensor encoder_layers_8_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_8_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(184120192)))]; + tensor input_481_cast_fp16 = layer_norm(axes = input_481_axes_0, beta = encoder_layers_8_norm_feed_forward2_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_8_norm_feed_forward2_weight_to_fp16, x = input_479_cast_fp16)[name = string("input_481_cast_fp16")]; + tensor encoder_layers_8_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(184122304))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(187268096))))[name = string("encoder_layers_8_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_8_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_8_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(187268288)))]; + tensor linear_80_cast_fp16 = linear(bias = encoder_layers_8_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_8_feed_forward2_linear1_weight_to_fp16_palettized, x = input_481_cast_fp16)[name = string("linear_80_cast_fp16")]; + tensor input_485_cast_fp16 = silu(x = linear_80_cast_fp16)[name = string("input_485_cast_fp16")]; + tensor encoder_layers_8_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(187276544))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(190422336))))[name = string("encoder_layers_8_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_8_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_8_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(190422528)))]; + tensor linear_81_cast_fp16 = linear(bias = encoder_layers_8_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_8_feed_forward2_linear2_weight_to_fp16_palettized, x = input_485_cast_fp16)[name = string("linear_81_cast_fp16")]; + fp16 var_2334_to_fp16 = const()[name = string("op_2334_to_fp16"), val = fp16(0x1p-1)]; + tensor var_2335_cast_fp16 = mul(x = linear_81_cast_fp16, y = var_2334_to_fp16)[name = string("op_2335_cast_fp16")]; + tensor input_491_cast_fp16 = add(x = input_479_cast_fp16, y = var_2335_cast_fp16)[name = string("input_491_cast_fp16")]; + tensor input_493_axes_0 = const()[name = string("input_493_axes_0"), val = tensor([-1])]; + tensor encoder_layers_8_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_8_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(190424640)))]; + tensor encoder_layers_8_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_8_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(190426752)))]; + tensor input_493_cast_fp16 = layer_norm(axes = input_493_axes_0, beta = encoder_layers_8_norm_out_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_8_norm_out_weight_to_fp16, x = input_491_cast_fp16)[name = string("input_493_cast_fp16")]; + tensor cache_37_begin_0 = const()[name = string("cache_37_begin_0"), val = tensor([9, 0, 0, 0])]; + tensor cache_37_end_0 = const()[name = string("cache_37_end_0"), val = tensor([10, 1, 42, 1024])]; + tensor cache_37_end_mask_0 = const()[name = string("cache_37_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_37_squeeze_mask_0 = const()[name = string("cache_37_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_37_cast_fp16 = slice_by_index(begin = cache_37_begin_0, end = cache_37_end_0, end_mask = cache_37_end_mask_0, squeeze_mask = cache_37_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_37_cast_fp16")]; + tensor cache_39_begin_0 = const()[name = string("cache_39_begin_0"), val = tensor([9, 0, 0, 0])]; + tensor cache_39_end_0 = const()[name = string("cache_39_end_0"), val = tensor([10, 1, 1024, 8])]; + tensor cache_39_end_mask_0 = const()[name = string("cache_39_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_39_squeeze_mask_0 = const()[name = string("cache_39_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_39_cast_fp16 = slice_by_index(begin = cache_39_begin_0, end = cache_39_end_0, end_mask = cache_39_end_mask_0, squeeze_mask = cache_39_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_39_cast_fp16")]; + tensor input_495_axes_0 = const()[name = string("input_495_axes_0"), val = tensor([-1])]; + tensor encoder_layers_9_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_9_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(190428864)))]; + tensor encoder_layers_9_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_9_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(190430976)))]; + tensor input_495_cast_fp16 = layer_norm(axes = input_495_axes_0, beta = encoder_layers_9_norm_feed_forward1_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_9_norm_feed_forward1_weight_to_fp16, x = input_493_cast_fp16)[name = string("input_495_cast_fp16")]; + tensor encoder_layers_9_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(190433088))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(193578880))))[name = string("encoder_layers_9_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_9_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_9_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(193579072)))]; + tensor linear_82_cast_fp16 = linear(bias = encoder_layers_9_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_9_feed_forward1_linear1_weight_to_fp16_palettized, x = input_495_cast_fp16)[name = string("linear_82_cast_fp16")]; + tensor input_499_cast_fp16 = silu(x = linear_82_cast_fp16)[name = string("input_499_cast_fp16")]; + tensor encoder_layers_9_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(193587328))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196733120))))[name = string("encoder_layers_9_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_9_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_9_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196733312)))]; + tensor linear_83_cast_fp16 = linear(bias = encoder_layers_9_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_9_feed_forward1_linear2_weight_to_fp16_palettized, x = input_499_cast_fp16)[name = string("linear_83_cast_fp16")]; + fp16 var_2371_to_fp16 = const()[name = string("op_2371_to_fp16"), val = fp16(0x1p-1)]; + tensor var_2372_cast_fp16 = mul(x = linear_83_cast_fp16, y = var_2371_to_fp16)[name = string("op_2372_cast_fp16")]; + tensor input_505_cast_fp16 = add(x = input_493_cast_fp16, y = var_2372_cast_fp16)[name = string("input_505_cast_fp16")]; + tensor key_19_axes_0 = const()[name = string("key_19_axes_0"), val = tensor([-1])]; + tensor encoder_layers_9_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_9_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196735424)))]; + tensor encoder_layers_9_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_9_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196737536)))]; + tensor key_19_cast_fp16 = layer_norm(axes = key_19_axes_0, beta = encoder_layers_9_norm_self_att_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_9_norm_self_att_weight_to_fp16, x = input_505_cast_fp16)[name = string("key_19_cast_fp16")]; + bool input_507_interleave_0 = const()[name = string("input_507_interleave_0"), val = bool(false)]; + tensor input_507_cast_fp16 = concat(axis = var_67, interleave = input_507_interleave_0, values = (cache_37_cast_fp16, key_19_cast_fp16))[name = string("input_507_cast_fp16")]; + bool var_2400_interleave_0 = const()[name = string("op_2400_interleave_0"), val = bool(false)]; + tensor var_2400_cast_fp16 = concat(axis = var_67, interleave = var_2400_interleave_0, values = key_19_cast_fp16)[name = string("op_2400_cast_fp16")]; + tensor encoder_layers_9_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196739648))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197526144))))[name = string("encoder_layers_9_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_9_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_9_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197526336)))]; + tensor linear_84_cast_fp16 = linear(bias = encoder_layers_9_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_9_self_attn_linear_q_weight_to_fp16_palettized, x = key_19_cast_fp16)[name = string("linear_84_cast_fp16")]; + tensor var_2405 = const()[name = string("op_2405"), val = tensor([1, -1, 8, 128])]; + tensor q_55_cast_fp16 = reshape(shape = var_2405, x = linear_84_cast_fp16)[name = string("q_55_cast_fp16")]; + tensor encoder_layers_9_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197528448))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198314944))))[name = string("encoder_layers_9_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_9_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_9_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198315136)))]; + tensor linear_85_cast_fp16 = linear(bias = encoder_layers_9_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_9_self_attn_linear_k_weight_to_fp16_palettized, x = input_507_cast_fp16)[name = string("linear_85_cast_fp16")]; + tensor var_2410 = const()[name = string("op_2410"), val = tensor([1, -1, 8, 128])]; + tensor k_37_cast_fp16 = reshape(shape = var_2410, x = linear_85_cast_fp16)[name = string("k_37_cast_fp16")]; + tensor encoder_layers_9_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198317248))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199103744))))[name = string("encoder_layers_9_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_9_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_9_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199103936)))]; + tensor linear_86_cast_fp16 = linear(bias = encoder_layers_9_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_9_self_attn_linear_v_weight_to_fp16_palettized, x = input_507_cast_fp16)[name = string("linear_86_cast_fp16")]; + tensor var_2415 = const()[name = string("op_2415"), val = tensor([1, -1, 8, 128])]; + tensor v_19_cast_fp16 = reshape(shape = var_2415, x = linear_86_cast_fp16)[name = string("v_19_cast_fp16")]; + tensor value_27_perm_0 = const()[name = string("value_27_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_9_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_9_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199106048)))]; + tensor var_2428_cast_fp16 = add(x = q_55_cast_fp16, y = encoder_layers_9_self_attn_pos_bias_u_to_fp16)[name = string("op_2428_cast_fp16")]; + tensor encoder_layers_9_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_9_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199108160)))]; + tensor var_2430_cast_fp16 = add(x = q_55_cast_fp16, y = encoder_layers_9_self_attn_pos_bias_v_to_fp16)[name = string("op_2430_cast_fp16")]; + tensor q_with_bias_v_19_perm_0 = const()[name = string("q_with_bias_v_19_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_241_transpose_x_0 = const()[name = string("x_241_transpose_x_0"), val = bool(false)]; + bool x_241_transpose_y_0 = const()[name = string("x_241_transpose_y_0"), val = bool(false)]; + tensor op_2432_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199110272))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199310016))))[name = string("op_2432_to_fp16_quantized")]; + tensor q_with_bias_v_19_cast_fp16 = transpose(perm = q_with_bias_v_19_perm_0, x = var_2430_cast_fp16)[name = string("transpose_281")]; + tensor x_241_cast_fp16 = matmul(transpose_x = x_241_transpose_x_0, transpose_y = x_241_transpose_y_0, x = q_with_bias_v_19_cast_fp16, y = op_2432_to_fp16_quantized)[name = string("x_241_cast_fp16")]; + tensor x_243_pad_0 = const()[name = string("x_243_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_243_mode_0 = const()[name = string("x_243_mode_0"), val = string("constant")]; + fp16 const_196_to_fp16 = const()[name = string("const_196_to_fp16"), val = fp16(0x0p+0)]; + tensor x_243_cast_fp16 = pad(constant_val = const_196_to_fp16, mode = x_243_mode_0, pad = x_243_pad_0, x = x_241_cast_fp16)[name = string("x_243_cast_fp16")]; + tensor var_2440 = const()[name = string("op_2440"), val = tensor([1, 8, -1, 56])]; + tensor x_245_cast_fp16 = reshape(shape = var_2440, x = x_243_cast_fp16)[name = string("x_245_cast_fp16")]; + tensor var_2444_begin_0 = const()[name = string("op_2444_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_2444_end_0 = const()[name = string("op_2444_end_0"), val = tensor([1, 8, 196, 56])]; + tensor var_2444_end_mask_0 = const()[name = string("op_2444_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_2444_cast_fp16 = slice_by_index(begin = var_2444_begin_0, end = var_2444_end_0, end_mask = var_2444_end_mask_0, x = x_245_cast_fp16)[name = string("op_2444_cast_fp16")]; + tensor var_2445 = const()[name = string("op_2445"), val = tensor([1, 8, 56, 195])]; + tensor matrix_bd_37_cast_fp16 = reshape(shape = var_2445, x = var_2444_cast_fp16)[name = string("matrix_bd_37_cast_fp16")]; + bool matrix_ac_19_transpose_x_0 = const()[name = string("matrix_ac_19_transpose_x_0"), val = bool(false)]; + bool matrix_ac_19_transpose_y_0 = const()[name = string("matrix_ac_19_transpose_y_0"), val = bool(false)]; + tensor transpose_114_perm_0 = const()[name = string("transpose_114_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_115_perm_0 = const()[name = string("transpose_115_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_115 = transpose(perm = transpose_115_perm_0, x = k_37_cast_fp16)[name = string("transpose_279")]; + tensor transpose_114 = transpose(perm = transpose_114_perm_0, x = var_2428_cast_fp16)[name = string("transpose_280")]; + tensor matrix_ac_19_cast_fp16 = matmul(transpose_x = matrix_ac_19_transpose_x_0, transpose_y = matrix_ac_19_transpose_y_0, x = transpose_114, y = transpose_115)[name = string("matrix_ac_19_cast_fp16")]; + tensor matrix_bd_39_begin_0 = const()[name = string("matrix_bd_39_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_39_end_0 = const()[name = string("matrix_bd_39_end_0"), val = tensor([1, 8, 56, 98])]; + tensor matrix_bd_39_end_mask_0 = const()[name = string("matrix_bd_39_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_39_cast_fp16 = slice_by_index(begin = matrix_bd_39_begin_0, end = matrix_bd_39_end_0, end_mask = matrix_bd_39_end_mask_0, x = matrix_bd_37_cast_fp16)[name = string("matrix_bd_39_cast_fp16")]; + tensor var_2454_cast_fp16 = add(x = matrix_ac_19_cast_fp16, y = matrix_bd_39_cast_fp16)[name = string("op_2454_cast_fp16")]; + fp16 _inversed_scores_37_y_0_to_fp16 = const()[name = string("_inversed_scores_37_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_37_cast_fp16 = mul(x = var_2454_cast_fp16, y = _inversed_scores_37_y_0_to_fp16)[name = string("_inversed_scores_37_cast_fp16")]; + tensor scores_39_cast_fp16 = select(a = var_44_to_fp16, b = _inversed_scores_37_cast_fp16, cond = mask_11)[name = string("scores_39_cast_fp16")]; + tensor var_2460_cast_fp16 = softmax(axis = var_58, x = scores_39_cast_fp16)[name = string("op_2460_cast_fp16")]; + tensor input_509_cast_fp16 = select(a = var_43_to_fp16, b = var_2460_cast_fp16, cond = mask_11)[name = string("input_509_cast_fp16")]; + bool x_247_transpose_x_0 = const()[name = string("x_247_transpose_x_0"), val = bool(false)]; + bool x_247_transpose_y_0 = const()[name = string("x_247_transpose_y_0"), val = bool(false)]; + tensor value_27_cast_fp16 = transpose(perm = value_27_perm_0, x = v_19_cast_fp16)[name = string("transpose_278")]; + tensor x_247_cast_fp16 = matmul(transpose_x = x_247_transpose_x_0, transpose_y = x_247_transpose_y_0, x = input_509_cast_fp16, y = value_27_cast_fp16)[name = string("x_247_cast_fp16")]; + tensor var_2464_perm_0 = const()[name = string("op_2464_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_2465 = const()[name = string("op_2465"), val = tensor([1, -1, 1024])]; + tensor var_2464_cast_fp16 = transpose(perm = var_2464_perm_0, x = x_247_cast_fp16)[name = string("transpose_277")]; + tensor input_511_cast_fp16 = reshape(shape = var_2465, x = var_2464_cast_fp16)[name = string("input_511_cast_fp16")]; + tensor encoder_layers_9_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199310528))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(200097024))))[name = string("encoder_layers_9_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_9_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_9_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(200097216)))]; + tensor linear_88_cast_fp16 = linear(bias = encoder_layers_9_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_9_self_attn_linear_out_weight_to_fp16_palettized, x = input_511_cast_fp16)[name = string("linear_88_cast_fp16")]; + tensor input_515_cast_fp16 = add(x = input_505_cast_fp16, y = linear_88_cast_fp16)[name = string("input_515_cast_fp16")]; + tensor x_251_axes_0 = const()[name = string("x_251_axes_0"), val = tensor([-1])]; + tensor encoder_layers_9_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_9_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(200099328)))]; + tensor encoder_layers_9_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_9_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(200101440)))]; + tensor x_251_cast_fp16 = layer_norm(axes = x_251_axes_0, beta = encoder_layers_9_norm_conv_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_9_norm_conv_weight_to_fp16, x = input_515_cast_fp16)[name = string("x_251_cast_fp16")]; + tensor input_517_perm_0 = const()[name = string("input_517_perm_0"), val = tensor([0, 2, 1])]; + string input_519_pad_type_0 = const()[name = string("input_519_pad_type_0"), val = string("valid")]; + tensor input_519_strides_0 = const()[name = string("input_519_strides_0"), val = tensor([1])]; + tensor input_519_pad_0 = const()[name = string("input_519_pad_0"), val = tensor([0, 0])]; + tensor input_519_dilations_0 = const()[name = string("input_519_dilations_0"), val = tensor([1])]; + int32 input_519_groups_0 = const()[name = string("input_519_groups_0"), val = int32(1)]; + tensor encoder_layers_9_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(200103552))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202200768))))[name = string("encoder_layers_9_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_517_cast_fp16 = transpose(perm = input_517_perm_0, x = x_251_cast_fp16)[name = string("transpose_276")]; + tensor input_519_cast_fp16 = conv(dilations = input_519_dilations_0, groups = input_519_groups_0, pad = input_519_pad_0, pad_type = input_519_pad_type_0, strides = input_519_strides_0, weight = encoder_layers_9_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_517_cast_fp16)[name = string("input_519_cast_fp16")]; + int32 x_253_split_num_splits_0 = const()[name = string("x_253_split_num_splits_0"), val = int32(2)]; + int32 x_253_split_axis_0 = const()[name = string("x_253_split_axis_0"), val = int32(1)]; + tensor x_253_split_cast_fp16_0, tensor x_253_split_cast_fp16_1 = split(axis = x_253_split_axis_0, num_splits = x_253_split_num_splits_0, x = input_519_cast_fp16)[name = string("x_253_split_cast_fp16")]; + tensor x_253_split_1_sigmoid_cast_fp16 = sigmoid(x = x_253_split_cast_fp16_1)[name = string("x_253_split_1_sigmoid_cast_fp16")]; + tensor x_253_cast_fp16 = mul(x = x_253_split_cast_fp16_0, y = x_253_split_1_sigmoid_cast_fp16)[name = string("x_253_cast_fp16")]; + tensor input_521_cast_fp16 = select(a = var_43_to_fp16, b = x_253_cast_fp16, cond = var_574)[name = string("input_521_cast_fp16")]; + bool new_x_39_interleave_0 = const()[name = string("new_x_39_interleave_0"), val = bool(false)]; + tensor new_x_39_cast_fp16 = concat(axis = var_58, interleave = new_x_39_interleave_0, values = (cache_39_cast_fp16, input_521_cast_fp16))[name = string("new_x_39_cast_fp16")]; + tensor var_2504_begin_0 = const()[name = string("op_2504_begin_0"), val = tensor([0, 0, 56])]; + tensor var_2504_end_0 = const()[name = string("op_2504_end_0"), val = tensor([1, 1024, 64])]; + tensor var_2504_end_mask_0 = const()[name = string("op_2504_end_mask_0"), val = tensor([true, true, true])]; + tensor var_2504_cast_fp16 = slice_by_index(begin = var_2504_begin_0, end = var_2504_end_0, end_mask = var_2504_end_mask_0, x = new_x_39_cast_fp16)[name = string("op_2504_cast_fp16")]; + string x_255_pad_type_0 = const()[name = string("x_255_pad_type_0"), val = string("valid")]; + int32 x_255_groups_0 = const()[name = string("x_255_groups_0"), val = int32(1024)]; + tensor x_255_strides_0 = const()[name = string("x_255_strides_0"), val = tensor([1])]; + tensor x_255_pad_0 = const()[name = string("x_255_pad_0"), val = tensor([0, 0])]; + tensor x_255_dilations_0 = const()[name = string("x_255_dilations_0"), val = tensor([1])]; + tensor encoder_layers_9_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202204928))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202214208))))[name = string("encoder_layers_9_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_255_cast_fp16 = conv(dilations = x_255_dilations_0, groups = x_255_groups_0, pad = x_255_pad_0, pad_type = x_255_pad_type_0, strides = x_255_strides_0, weight = encoder_layers_9_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_39_cast_fp16)[name = string("x_255_cast_fp16")]; + tensor input_523_perm_0 = const()[name = string("input_523_perm_0"), val = tensor([0, 2, 1])]; + tensor x_257_axes_0 = const()[name = string("x_257_axes_0"), val = tensor([-1])]; + tensor encoder_layers_9_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_9_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202216320)))]; + tensor encoder_layers_9_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_9_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202218432)))]; + tensor input_523_cast_fp16 = transpose(perm = input_523_perm_0, x = x_255_cast_fp16)[name = string("transpose_275")]; + tensor x_257_cast_fp16 = layer_norm(axes = x_257_axes_0, beta = encoder_layers_9_conv_batch_norm_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_9_conv_batch_norm_weight_to_fp16, x = input_523_cast_fp16)[name = string("x_257_cast_fp16")]; + tensor input_525_perm_0 = const()[name = string("input_525_perm_0"), val = tensor([0, 2, 1])]; + tensor input_525_cast_fp16 = transpose(perm = input_525_perm_0, x = x_257_cast_fp16)[name = string("transpose_274")]; + tensor input_527_cast_fp16 = silu(x = input_525_cast_fp16)[name = string("input_527_cast_fp16")]; + string x_259_pad_type_0 = const()[name = string("x_259_pad_type_0"), val = string("valid")]; + tensor x_259_strides_0 = const()[name = string("x_259_strides_0"), val = tensor([1])]; + tensor x_259_pad_0 = const()[name = string("x_259_pad_0"), val = tensor([0, 0])]; + tensor x_259_dilations_0 = const()[name = string("x_259_dilations_0"), val = tensor([1])]; + int32 x_259_groups_0 = const()[name = string("x_259_groups_0"), val = int32(1)]; + tensor encoder_layers_9_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202220544))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(203269184))))[name = string("encoder_layers_9_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_259_cast_fp16 = conv(dilations = x_259_dilations_0, groups = x_259_groups_0, pad = x_259_pad_0, pad_type = x_259_pad_type_0, strides = x_259_strides_0, weight = encoder_layers_9_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_527_cast_fp16)[name = string("x_259_cast_fp16")]; + tensor input_529_perm_0 = const()[name = string("input_529_perm_0"), val = tensor([0, 2, 1])]; + tensor input_529_cast_fp16 = transpose(perm = input_529_perm_0, x = x_259_cast_fp16)[name = string("transpose_273")]; + tensor input_531_cast_fp16 = add(x = input_515_cast_fp16, y = input_529_cast_fp16)[name = string("input_531_cast_fp16")]; + tensor input_533_axes_0 = const()[name = string("input_533_axes_0"), val = tensor([-1])]; + tensor encoder_layers_9_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_9_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(203271296)))]; + tensor encoder_layers_9_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_9_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(203273408)))]; + tensor input_533_cast_fp16 = layer_norm(axes = input_533_axes_0, beta = encoder_layers_9_norm_feed_forward2_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_9_norm_feed_forward2_weight_to_fp16, x = input_531_cast_fp16)[name = string("input_533_cast_fp16")]; + tensor encoder_layers_9_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(203275520))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(206421312))))[name = string("encoder_layers_9_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_9_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_9_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(206421504)))]; + tensor linear_89_cast_fp16 = linear(bias = encoder_layers_9_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_9_feed_forward2_linear1_weight_to_fp16_palettized, x = input_533_cast_fp16)[name = string("linear_89_cast_fp16")]; + tensor input_537_cast_fp16 = silu(x = linear_89_cast_fp16)[name = string("input_537_cast_fp16")]; + tensor encoder_layers_9_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(206429760))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(209575552))))[name = string("encoder_layers_9_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_9_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_9_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(209575744)))]; + tensor linear_90_cast_fp16 = linear(bias = encoder_layers_9_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_9_feed_forward2_linear2_weight_to_fp16_palettized, x = input_537_cast_fp16)[name = string("linear_90_cast_fp16")]; + fp16 var_2547_to_fp16 = const()[name = string("op_2547_to_fp16"), val = fp16(0x1p-1)]; + tensor var_2548_cast_fp16 = mul(x = linear_90_cast_fp16, y = var_2547_to_fp16)[name = string("op_2548_cast_fp16")]; + tensor input_543_cast_fp16 = add(x = input_531_cast_fp16, y = var_2548_cast_fp16)[name = string("input_543_cast_fp16")]; + tensor input_545_axes_0 = const()[name = string("input_545_axes_0"), val = tensor([-1])]; + tensor encoder_layers_9_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_9_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(209577856)))]; + tensor encoder_layers_9_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_9_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(209579968)))]; + tensor input_545_cast_fp16 = layer_norm(axes = input_545_axes_0, beta = encoder_layers_9_norm_out_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_9_norm_out_weight_to_fp16, x = input_543_cast_fp16)[name = string("input_545_cast_fp16")]; + tensor cache_41_begin_0 = const()[name = string("cache_41_begin_0"), val = tensor([10, 0, 0, 0])]; + tensor cache_41_end_0 = const()[name = string("cache_41_end_0"), val = tensor([11, 1, 42, 1024])]; + tensor cache_41_end_mask_0 = const()[name = string("cache_41_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_41_squeeze_mask_0 = const()[name = string("cache_41_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_41_cast_fp16 = slice_by_index(begin = cache_41_begin_0, end = cache_41_end_0, end_mask = cache_41_end_mask_0, squeeze_mask = cache_41_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_41_cast_fp16")]; + tensor cache_43_begin_0 = const()[name = string("cache_43_begin_0"), val = tensor([10, 0, 0, 0])]; + tensor cache_43_end_0 = const()[name = string("cache_43_end_0"), val = tensor([11, 1, 1024, 8])]; + tensor cache_43_end_mask_0 = const()[name = string("cache_43_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_43_squeeze_mask_0 = const()[name = string("cache_43_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_43_cast_fp16 = slice_by_index(begin = cache_43_begin_0, end = cache_43_end_0, end_mask = cache_43_end_mask_0, squeeze_mask = cache_43_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_43_cast_fp16")]; + tensor input_547_axes_0 = const()[name = string("input_547_axes_0"), val = tensor([-1])]; + tensor encoder_layers_10_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_10_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(209582080)))]; + tensor encoder_layers_10_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_10_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(209584192)))]; + tensor input_547_cast_fp16 = layer_norm(axes = input_547_axes_0, beta = encoder_layers_10_norm_feed_forward1_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_10_norm_feed_forward1_weight_to_fp16, x = input_545_cast_fp16)[name = string("input_547_cast_fp16")]; + tensor encoder_layers_10_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(209586304))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212732096))))[name = string("encoder_layers_10_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_10_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_10_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212732288)))]; + tensor linear_91_cast_fp16 = linear(bias = encoder_layers_10_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_10_feed_forward1_linear1_weight_to_fp16_palettized, x = input_547_cast_fp16)[name = string("linear_91_cast_fp16")]; + tensor input_551_cast_fp16 = silu(x = linear_91_cast_fp16)[name = string("input_551_cast_fp16")]; + tensor encoder_layers_10_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212740544))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215886336))))[name = string("encoder_layers_10_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_10_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_10_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215886528)))]; + tensor linear_92_cast_fp16 = linear(bias = encoder_layers_10_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_10_feed_forward1_linear2_weight_to_fp16_palettized, x = input_551_cast_fp16)[name = string("linear_92_cast_fp16")]; + fp16 var_2584_to_fp16 = const()[name = string("op_2584_to_fp16"), val = fp16(0x1p-1)]; + tensor var_2585_cast_fp16 = mul(x = linear_92_cast_fp16, y = var_2584_to_fp16)[name = string("op_2585_cast_fp16")]; + tensor input_557_cast_fp16 = add(x = input_545_cast_fp16, y = var_2585_cast_fp16)[name = string("input_557_cast_fp16")]; + tensor key_21_axes_0 = const()[name = string("key_21_axes_0"), val = tensor([-1])]; + tensor encoder_layers_10_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_10_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215888640)))]; + tensor encoder_layers_10_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_10_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215890752)))]; + tensor key_21_cast_fp16 = layer_norm(axes = key_21_axes_0, beta = encoder_layers_10_norm_self_att_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_10_norm_self_att_weight_to_fp16, x = input_557_cast_fp16)[name = string("key_21_cast_fp16")]; + bool input_559_interleave_0 = const()[name = string("input_559_interleave_0"), val = bool(false)]; + tensor input_559_cast_fp16 = concat(axis = var_67, interleave = input_559_interleave_0, values = (cache_41_cast_fp16, key_21_cast_fp16))[name = string("input_559_cast_fp16")]; + bool var_2613_interleave_0 = const()[name = string("op_2613_interleave_0"), val = bool(false)]; + tensor var_2613_cast_fp16 = concat(axis = var_67, interleave = var_2613_interleave_0, values = key_21_cast_fp16)[name = string("op_2613_cast_fp16")]; + tensor encoder_layers_10_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215892864))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216679360))))[name = string("encoder_layers_10_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_10_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_10_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216679552)))]; + tensor linear_93_cast_fp16 = linear(bias = encoder_layers_10_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_10_self_attn_linear_q_weight_to_fp16_palettized, x = key_21_cast_fp16)[name = string("linear_93_cast_fp16")]; + tensor var_2618 = const()[name = string("op_2618"), val = tensor([1, -1, 8, 128])]; + tensor q_61_cast_fp16 = reshape(shape = var_2618, x = linear_93_cast_fp16)[name = string("q_61_cast_fp16")]; + tensor encoder_layers_10_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216681664))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217468160))))[name = string("encoder_layers_10_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_10_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_10_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217468352)))]; + tensor linear_94_cast_fp16 = linear(bias = encoder_layers_10_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_10_self_attn_linear_k_weight_to_fp16_palettized, x = input_559_cast_fp16)[name = string("linear_94_cast_fp16")]; + tensor var_2623 = const()[name = string("op_2623"), val = tensor([1, -1, 8, 128])]; + tensor k_41_cast_fp16 = reshape(shape = var_2623, x = linear_94_cast_fp16)[name = string("k_41_cast_fp16")]; + tensor encoder_layers_10_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217470464))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218256960))))[name = string("encoder_layers_10_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_10_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_10_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218257152)))]; + tensor linear_95_cast_fp16 = linear(bias = encoder_layers_10_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_10_self_attn_linear_v_weight_to_fp16_palettized, x = input_559_cast_fp16)[name = string("linear_95_cast_fp16")]; + tensor var_2628 = const()[name = string("op_2628"), val = tensor([1, -1, 8, 128])]; + tensor v_21_cast_fp16 = reshape(shape = var_2628, x = linear_95_cast_fp16)[name = string("v_21_cast_fp16")]; + tensor value_29_perm_0 = const()[name = string("value_29_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_10_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_10_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218259264)))]; + tensor var_2641_cast_fp16 = add(x = q_61_cast_fp16, y = encoder_layers_10_self_attn_pos_bias_u_to_fp16)[name = string("op_2641_cast_fp16")]; + tensor encoder_layers_10_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_10_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218261376)))]; + tensor var_2643_cast_fp16 = add(x = q_61_cast_fp16, y = encoder_layers_10_self_attn_pos_bias_v_to_fp16)[name = string("op_2643_cast_fp16")]; + tensor q_with_bias_v_21_perm_0 = const()[name = string("q_with_bias_v_21_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_267_transpose_x_0 = const()[name = string("x_267_transpose_x_0"), val = bool(false)]; + bool x_267_transpose_y_0 = const()[name = string("x_267_transpose_y_0"), val = bool(false)]; + tensor op_2645_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218263488))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218463232))))[name = string("op_2645_to_fp16_quantized")]; + tensor q_with_bias_v_21_cast_fp16 = transpose(perm = q_with_bias_v_21_perm_0, x = var_2643_cast_fp16)[name = string("transpose_272")]; + tensor x_267_cast_fp16 = matmul(transpose_x = x_267_transpose_x_0, transpose_y = x_267_transpose_y_0, x = q_with_bias_v_21_cast_fp16, y = op_2645_to_fp16_quantized)[name = string("x_267_cast_fp16")]; + tensor x_269_pad_0 = const()[name = string("x_269_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_269_mode_0 = const()[name = string("x_269_mode_0"), val = string("constant")]; + fp16 const_209_to_fp16 = const()[name = string("const_209_to_fp16"), val = fp16(0x0p+0)]; + tensor x_269_cast_fp16 = pad(constant_val = const_209_to_fp16, mode = x_269_mode_0, pad = x_269_pad_0, x = x_267_cast_fp16)[name = string("x_269_cast_fp16")]; + tensor var_2653 = const()[name = string("op_2653"), val = tensor([1, 8, -1, 56])]; + tensor x_271_cast_fp16 = reshape(shape = var_2653, x = x_269_cast_fp16)[name = string("x_271_cast_fp16")]; + tensor var_2657_begin_0 = const()[name = string("op_2657_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_2657_end_0 = const()[name = string("op_2657_end_0"), val = tensor([1, 8, 196, 56])]; + tensor var_2657_end_mask_0 = const()[name = string("op_2657_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_2657_cast_fp16 = slice_by_index(begin = var_2657_begin_0, end = var_2657_end_0, end_mask = var_2657_end_mask_0, x = x_271_cast_fp16)[name = string("op_2657_cast_fp16")]; + tensor var_2658 = const()[name = string("op_2658"), val = tensor([1, 8, 56, 195])]; + tensor matrix_bd_41_cast_fp16 = reshape(shape = var_2658, x = var_2657_cast_fp16)[name = string("matrix_bd_41_cast_fp16")]; + bool matrix_ac_21_transpose_x_0 = const()[name = string("matrix_ac_21_transpose_x_0"), val = bool(false)]; + bool matrix_ac_21_transpose_y_0 = const()[name = string("matrix_ac_21_transpose_y_0"), val = bool(false)]; + tensor transpose_116_perm_0 = const()[name = string("transpose_116_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_117_perm_0 = const()[name = string("transpose_117_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_117 = transpose(perm = transpose_117_perm_0, x = k_41_cast_fp16)[name = string("transpose_270")]; + tensor transpose_116 = transpose(perm = transpose_116_perm_0, x = var_2641_cast_fp16)[name = string("transpose_271")]; + tensor matrix_ac_21_cast_fp16 = matmul(transpose_x = matrix_ac_21_transpose_x_0, transpose_y = matrix_ac_21_transpose_y_0, x = transpose_116, y = transpose_117)[name = string("matrix_ac_21_cast_fp16")]; + tensor matrix_bd_43_begin_0 = const()[name = string("matrix_bd_43_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_43_end_0 = const()[name = string("matrix_bd_43_end_0"), val = tensor([1, 8, 56, 98])]; + tensor matrix_bd_43_end_mask_0 = const()[name = string("matrix_bd_43_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_43_cast_fp16 = slice_by_index(begin = matrix_bd_43_begin_0, end = matrix_bd_43_end_0, end_mask = matrix_bd_43_end_mask_0, x = matrix_bd_41_cast_fp16)[name = string("matrix_bd_43_cast_fp16")]; + tensor var_2667_cast_fp16 = add(x = matrix_ac_21_cast_fp16, y = matrix_bd_43_cast_fp16)[name = string("op_2667_cast_fp16")]; + fp16 _inversed_scores_41_y_0_to_fp16 = const()[name = string("_inversed_scores_41_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_41_cast_fp16 = mul(x = var_2667_cast_fp16, y = _inversed_scores_41_y_0_to_fp16)[name = string("_inversed_scores_41_cast_fp16")]; + tensor scores_43_cast_fp16 = select(a = var_44_to_fp16, b = _inversed_scores_41_cast_fp16, cond = mask_11)[name = string("scores_43_cast_fp16")]; + tensor var_2673_cast_fp16 = softmax(axis = var_58, x = scores_43_cast_fp16)[name = string("op_2673_cast_fp16")]; + tensor input_561_cast_fp16 = select(a = var_43_to_fp16, b = var_2673_cast_fp16, cond = mask_11)[name = string("input_561_cast_fp16")]; + bool x_273_transpose_x_0 = const()[name = string("x_273_transpose_x_0"), val = bool(false)]; + bool x_273_transpose_y_0 = const()[name = string("x_273_transpose_y_0"), val = bool(false)]; + tensor value_29_cast_fp16 = transpose(perm = value_29_perm_0, x = v_21_cast_fp16)[name = string("transpose_269")]; + tensor x_273_cast_fp16 = matmul(transpose_x = x_273_transpose_x_0, transpose_y = x_273_transpose_y_0, x = input_561_cast_fp16, y = value_29_cast_fp16)[name = string("x_273_cast_fp16")]; + tensor var_2677_perm_0 = const()[name = string("op_2677_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_2678 = const()[name = string("op_2678"), val = tensor([1, -1, 1024])]; + tensor var_2677_cast_fp16 = transpose(perm = var_2677_perm_0, x = x_273_cast_fp16)[name = string("transpose_268")]; + tensor input_563_cast_fp16 = reshape(shape = var_2678, x = var_2677_cast_fp16)[name = string("input_563_cast_fp16")]; + tensor encoder_layers_10_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218463744))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(219250240))))[name = string("encoder_layers_10_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_10_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_10_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(219250432)))]; + tensor linear_97_cast_fp16 = linear(bias = encoder_layers_10_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_10_self_attn_linear_out_weight_to_fp16_palettized, x = input_563_cast_fp16)[name = string("linear_97_cast_fp16")]; + tensor input_567_cast_fp16 = add(x = input_557_cast_fp16, y = linear_97_cast_fp16)[name = string("input_567_cast_fp16")]; + tensor x_277_axes_0 = const()[name = string("x_277_axes_0"), val = tensor([-1])]; + tensor encoder_layers_10_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_10_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(219252544)))]; + tensor encoder_layers_10_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_10_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(219254656)))]; + tensor x_277_cast_fp16 = layer_norm(axes = x_277_axes_0, beta = encoder_layers_10_norm_conv_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_10_norm_conv_weight_to_fp16, x = input_567_cast_fp16)[name = string("x_277_cast_fp16")]; + tensor input_569_perm_0 = const()[name = string("input_569_perm_0"), val = tensor([0, 2, 1])]; + string input_571_pad_type_0 = const()[name = string("input_571_pad_type_0"), val = string("valid")]; + tensor input_571_strides_0 = const()[name = string("input_571_strides_0"), val = tensor([1])]; + tensor input_571_pad_0 = const()[name = string("input_571_pad_0"), val = tensor([0, 0])]; + tensor input_571_dilations_0 = const()[name = string("input_571_dilations_0"), val = tensor([1])]; + int32 input_571_groups_0 = const()[name = string("input_571_groups_0"), val = int32(1)]; + tensor encoder_layers_10_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(219256768))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221353984))))[name = string("encoder_layers_10_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_569_cast_fp16 = transpose(perm = input_569_perm_0, x = x_277_cast_fp16)[name = string("transpose_267")]; + tensor input_571_cast_fp16 = conv(dilations = input_571_dilations_0, groups = input_571_groups_0, pad = input_571_pad_0, pad_type = input_571_pad_type_0, strides = input_571_strides_0, weight = encoder_layers_10_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_569_cast_fp16)[name = string("input_571_cast_fp16")]; + int32 x_279_split_num_splits_0 = const()[name = string("x_279_split_num_splits_0"), val = int32(2)]; + int32 x_279_split_axis_0 = const()[name = string("x_279_split_axis_0"), val = int32(1)]; + tensor x_279_split_cast_fp16_0, tensor x_279_split_cast_fp16_1 = split(axis = x_279_split_axis_0, num_splits = x_279_split_num_splits_0, x = input_571_cast_fp16)[name = string("x_279_split_cast_fp16")]; + tensor x_279_split_1_sigmoid_cast_fp16 = sigmoid(x = x_279_split_cast_fp16_1)[name = string("x_279_split_1_sigmoid_cast_fp16")]; + tensor x_279_cast_fp16 = mul(x = x_279_split_cast_fp16_0, y = x_279_split_1_sigmoid_cast_fp16)[name = string("x_279_cast_fp16")]; + tensor input_573_cast_fp16 = select(a = var_43_to_fp16, b = x_279_cast_fp16, cond = var_574)[name = string("input_573_cast_fp16")]; + bool new_x_43_interleave_0 = const()[name = string("new_x_43_interleave_0"), val = bool(false)]; + tensor new_x_43_cast_fp16 = concat(axis = var_58, interleave = new_x_43_interleave_0, values = (cache_43_cast_fp16, input_573_cast_fp16))[name = string("new_x_43_cast_fp16")]; + tensor var_2717_begin_0 = const()[name = string("op_2717_begin_0"), val = tensor([0, 0, 56])]; + tensor var_2717_end_0 = const()[name = string("op_2717_end_0"), val = tensor([1, 1024, 64])]; + tensor var_2717_end_mask_0 = const()[name = string("op_2717_end_mask_0"), val = tensor([true, true, true])]; + tensor var_2717_cast_fp16 = slice_by_index(begin = var_2717_begin_0, end = var_2717_end_0, end_mask = var_2717_end_mask_0, x = new_x_43_cast_fp16)[name = string("op_2717_cast_fp16")]; + string x_281_pad_type_0 = const()[name = string("x_281_pad_type_0"), val = string("valid")]; + int32 x_281_groups_0 = const()[name = string("x_281_groups_0"), val = int32(1024)]; + tensor x_281_strides_0 = const()[name = string("x_281_strides_0"), val = tensor([1])]; + tensor x_281_pad_0 = const()[name = string("x_281_pad_0"), val = tensor([0, 0])]; + tensor x_281_dilations_0 = const()[name = string("x_281_dilations_0"), val = tensor([1])]; + tensor encoder_layers_10_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221358144))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221367424))))[name = string("encoder_layers_10_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_281_cast_fp16 = conv(dilations = x_281_dilations_0, groups = x_281_groups_0, pad = x_281_pad_0, pad_type = x_281_pad_type_0, strides = x_281_strides_0, weight = encoder_layers_10_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_43_cast_fp16)[name = string("x_281_cast_fp16")]; + tensor input_575_perm_0 = const()[name = string("input_575_perm_0"), val = tensor([0, 2, 1])]; + tensor x_283_axes_0 = const()[name = string("x_283_axes_0"), val = tensor([-1])]; + tensor encoder_layers_10_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_10_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221369536)))]; + tensor encoder_layers_10_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_10_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221371648)))]; + tensor input_575_cast_fp16 = transpose(perm = input_575_perm_0, x = x_281_cast_fp16)[name = string("transpose_266")]; + tensor x_283_cast_fp16 = layer_norm(axes = x_283_axes_0, beta = encoder_layers_10_conv_batch_norm_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_10_conv_batch_norm_weight_to_fp16, x = input_575_cast_fp16)[name = string("x_283_cast_fp16")]; + tensor input_577_perm_0 = const()[name = string("input_577_perm_0"), val = tensor([0, 2, 1])]; + tensor input_577_cast_fp16 = transpose(perm = input_577_perm_0, x = x_283_cast_fp16)[name = string("transpose_265")]; + tensor input_579_cast_fp16 = silu(x = input_577_cast_fp16)[name = string("input_579_cast_fp16")]; + string x_285_pad_type_0 = const()[name = string("x_285_pad_type_0"), val = string("valid")]; + tensor x_285_strides_0 = const()[name = string("x_285_strides_0"), val = tensor([1])]; + tensor x_285_pad_0 = const()[name = string("x_285_pad_0"), val = tensor([0, 0])]; + tensor x_285_dilations_0 = const()[name = string("x_285_dilations_0"), val = tensor([1])]; + int32 x_285_groups_0 = const()[name = string("x_285_groups_0"), val = int32(1)]; + tensor encoder_layers_10_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221373760))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(222422400))))[name = string("encoder_layers_10_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_285_cast_fp16 = conv(dilations = x_285_dilations_0, groups = x_285_groups_0, pad = x_285_pad_0, pad_type = x_285_pad_type_0, strides = x_285_strides_0, weight = encoder_layers_10_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_579_cast_fp16)[name = string("x_285_cast_fp16")]; + tensor input_581_perm_0 = const()[name = string("input_581_perm_0"), val = tensor([0, 2, 1])]; + tensor input_581_cast_fp16 = transpose(perm = input_581_perm_0, x = x_285_cast_fp16)[name = string("transpose_264")]; + tensor input_583_cast_fp16 = add(x = input_567_cast_fp16, y = input_581_cast_fp16)[name = string("input_583_cast_fp16")]; + tensor input_585_axes_0 = const()[name = string("input_585_axes_0"), val = tensor([-1])]; + tensor encoder_layers_10_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_10_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(222424512)))]; + tensor encoder_layers_10_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_10_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(222426624)))]; + tensor input_585_cast_fp16 = layer_norm(axes = input_585_axes_0, beta = encoder_layers_10_norm_feed_forward2_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_10_norm_feed_forward2_weight_to_fp16, x = input_583_cast_fp16)[name = string("input_585_cast_fp16")]; + tensor encoder_layers_10_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(222428736))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(225574528))))[name = string("encoder_layers_10_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_10_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_10_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(225574720)))]; + tensor linear_98_cast_fp16 = linear(bias = encoder_layers_10_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_10_feed_forward2_linear1_weight_to_fp16_palettized, x = input_585_cast_fp16)[name = string("linear_98_cast_fp16")]; + tensor input_589_cast_fp16 = silu(x = linear_98_cast_fp16)[name = string("input_589_cast_fp16")]; + tensor encoder_layers_10_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(225582976))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(228728768))))[name = string("encoder_layers_10_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_10_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_10_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(228728960)))]; + tensor linear_99_cast_fp16 = linear(bias = encoder_layers_10_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_10_feed_forward2_linear2_weight_to_fp16_palettized, x = input_589_cast_fp16)[name = string("linear_99_cast_fp16")]; + fp16 var_2760_to_fp16 = const()[name = string("op_2760_to_fp16"), val = fp16(0x1p-1)]; + tensor var_2761_cast_fp16 = mul(x = linear_99_cast_fp16, y = var_2760_to_fp16)[name = string("op_2761_cast_fp16")]; + tensor input_595_cast_fp16 = add(x = input_583_cast_fp16, y = var_2761_cast_fp16)[name = string("input_595_cast_fp16")]; + tensor input_597_axes_0 = const()[name = string("input_597_axes_0"), val = tensor([-1])]; + tensor encoder_layers_10_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_10_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(228731072)))]; + tensor encoder_layers_10_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_10_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(228733184)))]; + tensor input_597_cast_fp16 = layer_norm(axes = input_597_axes_0, beta = encoder_layers_10_norm_out_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_10_norm_out_weight_to_fp16, x = input_595_cast_fp16)[name = string("input_597_cast_fp16")]; + tensor cache_45_begin_0 = const()[name = string("cache_45_begin_0"), val = tensor([11, 0, 0, 0])]; + tensor cache_45_end_0 = const()[name = string("cache_45_end_0"), val = tensor([12, 1, 42, 1024])]; + tensor cache_45_end_mask_0 = const()[name = string("cache_45_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_45_squeeze_mask_0 = const()[name = string("cache_45_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_45_cast_fp16 = slice_by_index(begin = cache_45_begin_0, end = cache_45_end_0, end_mask = cache_45_end_mask_0, squeeze_mask = cache_45_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_45_cast_fp16")]; + tensor cache_47_begin_0 = const()[name = string("cache_47_begin_0"), val = tensor([11, 0, 0, 0])]; + tensor cache_47_end_0 = const()[name = string("cache_47_end_0"), val = tensor([12, 1, 1024, 8])]; + tensor cache_47_end_mask_0 = const()[name = string("cache_47_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_47_squeeze_mask_0 = const()[name = string("cache_47_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_47_cast_fp16 = slice_by_index(begin = cache_47_begin_0, end = cache_47_end_0, end_mask = cache_47_end_mask_0, squeeze_mask = cache_47_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_47_cast_fp16")]; + tensor input_599_axes_0 = const()[name = string("input_599_axes_0"), val = tensor([-1])]; + tensor encoder_layers_11_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_11_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(228735296)))]; + tensor encoder_layers_11_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_11_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(228737408)))]; + tensor input_599_cast_fp16 = layer_norm(axes = input_599_axes_0, beta = encoder_layers_11_norm_feed_forward1_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_11_norm_feed_forward1_weight_to_fp16, x = input_597_cast_fp16)[name = string("input_599_cast_fp16")]; + tensor encoder_layers_11_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(228739520))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(231885312))))[name = string("encoder_layers_11_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_11_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_11_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(231885504)))]; + tensor linear_100_cast_fp16 = linear(bias = encoder_layers_11_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_11_feed_forward1_linear1_weight_to_fp16_palettized, x = input_599_cast_fp16)[name = string("linear_100_cast_fp16")]; + tensor input_603_cast_fp16 = silu(x = linear_100_cast_fp16)[name = string("input_603_cast_fp16")]; + tensor encoder_layers_11_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(231893760))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(235039552))))[name = string("encoder_layers_11_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_11_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_11_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(235039744)))]; + tensor linear_101_cast_fp16 = linear(bias = encoder_layers_11_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_11_feed_forward1_linear2_weight_to_fp16_palettized, x = input_603_cast_fp16)[name = string("linear_101_cast_fp16")]; + fp16 var_2797_to_fp16 = const()[name = string("op_2797_to_fp16"), val = fp16(0x1p-1)]; + tensor var_2798_cast_fp16 = mul(x = linear_101_cast_fp16, y = var_2797_to_fp16)[name = string("op_2798_cast_fp16")]; + tensor input_609_cast_fp16 = add(x = input_597_cast_fp16, y = var_2798_cast_fp16)[name = string("input_609_cast_fp16")]; + tensor key_23_axes_0 = const()[name = string("key_23_axes_0"), val = tensor([-1])]; + tensor encoder_layers_11_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_11_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(235041856)))]; + tensor encoder_layers_11_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_11_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(235043968)))]; + tensor key_23_cast_fp16 = layer_norm(axes = key_23_axes_0, beta = encoder_layers_11_norm_self_att_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_11_norm_self_att_weight_to_fp16, x = input_609_cast_fp16)[name = string("key_23_cast_fp16")]; + bool input_611_interleave_0 = const()[name = string("input_611_interleave_0"), val = bool(false)]; + tensor input_611_cast_fp16 = concat(axis = var_67, interleave = input_611_interleave_0, values = (cache_45_cast_fp16, key_23_cast_fp16))[name = string("input_611_cast_fp16")]; + bool var_2826_interleave_0 = const()[name = string("op_2826_interleave_0"), val = bool(false)]; + tensor var_2826_cast_fp16 = concat(axis = var_67, interleave = var_2826_interleave_0, values = key_23_cast_fp16)[name = string("op_2826_cast_fp16")]; + tensor encoder_layers_11_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(235046080))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(235832576))))[name = string("encoder_layers_11_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_11_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_11_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(235832768)))]; + tensor linear_102_cast_fp16 = linear(bias = encoder_layers_11_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_11_self_attn_linear_q_weight_to_fp16_palettized, x = key_23_cast_fp16)[name = string("linear_102_cast_fp16")]; + tensor var_2831 = const()[name = string("op_2831"), val = tensor([1, -1, 8, 128])]; + tensor q_67_cast_fp16 = reshape(shape = var_2831, x = linear_102_cast_fp16)[name = string("q_67_cast_fp16")]; + tensor encoder_layers_11_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(235834880))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236621376))))[name = string("encoder_layers_11_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_11_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_11_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236621568)))]; + tensor linear_103_cast_fp16 = linear(bias = encoder_layers_11_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_11_self_attn_linear_k_weight_to_fp16_palettized, x = input_611_cast_fp16)[name = string("linear_103_cast_fp16")]; + tensor var_2836 = const()[name = string("op_2836"), val = tensor([1, -1, 8, 128])]; + tensor k_45_cast_fp16 = reshape(shape = var_2836, x = linear_103_cast_fp16)[name = string("k_45_cast_fp16")]; + tensor encoder_layers_11_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236623680))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(237410176))))[name = string("encoder_layers_11_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_11_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_11_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(237410368)))]; + tensor linear_104_cast_fp16 = linear(bias = encoder_layers_11_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_11_self_attn_linear_v_weight_to_fp16_palettized, x = input_611_cast_fp16)[name = string("linear_104_cast_fp16")]; + tensor var_2841 = const()[name = string("op_2841"), val = tensor([1, -1, 8, 128])]; + tensor v_23_cast_fp16 = reshape(shape = var_2841, x = linear_104_cast_fp16)[name = string("v_23_cast_fp16")]; + tensor value_31_perm_0 = const()[name = string("value_31_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_11_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_11_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(237412480)))]; + tensor var_2854_cast_fp16 = add(x = q_67_cast_fp16, y = encoder_layers_11_self_attn_pos_bias_u_to_fp16)[name = string("op_2854_cast_fp16")]; + tensor encoder_layers_11_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_11_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(237414592)))]; + tensor var_2856_cast_fp16 = add(x = q_67_cast_fp16, y = encoder_layers_11_self_attn_pos_bias_v_to_fp16)[name = string("op_2856_cast_fp16")]; + tensor q_with_bias_v_23_perm_0 = const()[name = string("q_with_bias_v_23_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_293_transpose_x_0 = const()[name = string("x_293_transpose_x_0"), val = bool(false)]; + bool x_293_transpose_y_0 = const()[name = string("x_293_transpose_y_0"), val = bool(false)]; + tensor op_2858_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(237416704))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(237616448))))[name = string("op_2858_to_fp16_quantized")]; + tensor q_with_bias_v_23_cast_fp16 = transpose(perm = q_with_bias_v_23_perm_0, x = var_2856_cast_fp16)[name = string("transpose_263")]; + tensor x_293_cast_fp16 = matmul(transpose_x = x_293_transpose_x_0, transpose_y = x_293_transpose_y_0, x = q_with_bias_v_23_cast_fp16, y = op_2858_to_fp16_quantized)[name = string("x_293_cast_fp16")]; + tensor x_295_pad_0 = const()[name = string("x_295_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_295_mode_0 = const()[name = string("x_295_mode_0"), val = string("constant")]; + fp16 const_222_to_fp16 = const()[name = string("const_222_to_fp16"), val = fp16(0x0p+0)]; + tensor x_295_cast_fp16 = pad(constant_val = const_222_to_fp16, mode = x_295_mode_0, pad = x_295_pad_0, x = x_293_cast_fp16)[name = string("x_295_cast_fp16")]; + tensor var_2866 = const()[name = string("op_2866"), val = tensor([1, 8, -1, 56])]; + tensor x_297_cast_fp16 = reshape(shape = var_2866, x = x_295_cast_fp16)[name = string("x_297_cast_fp16")]; + tensor var_2870_begin_0 = const()[name = string("op_2870_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_2870_end_0 = const()[name = string("op_2870_end_0"), val = tensor([1, 8, 196, 56])]; + tensor var_2870_end_mask_0 = const()[name = string("op_2870_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_2870_cast_fp16 = slice_by_index(begin = var_2870_begin_0, end = var_2870_end_0, end_mask = var_2870_end_mask_0, x = x_297_cast_fp16)[name = string("op_2870_cast_fp16")]; + tensor var_2871 = const()[name = string("op_2871"), val = tensor([1, 8, 56, 195])]; + tensor matrix_bd_45_cast_fp16 = reshape(shape = var_2871, x = var_2870_cast_fp16)[name = string("matrix_bd_45_cast_fp16")]; + bool matrix_ac_23_transpose_x_0 = const()[name = string("matrix_ac_23_transpose_x_0"), val = bool(false)]; + bool matrix_ac_23_transpose_y_0 = const()[name = string("matrix_ac_23_transpose_y_0"), val = bool(false)]; + tensor transpose_118_perm_0 = const()[name = string("transpose_118_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_119_perm_0 = const()[name = string("transpose_119_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_119 = transpose(perm = transpose_119_perm_0, x = k_45_cast_fp16)[name = string("transpose_261")]; + tensor transpose_118 = transpose(perm = transpose_118_perm_0, x = var_2854_cast_fp16)[name = string("transpose_262")]; + tensor matrix_ac_23_cast_fp16 = matmul(transpose_x = matrix_ac_23_transpose_x_0, transpose_y = matrix_ac_23_transpose_y_0, x = transpose_118, y = transpose_119)[name = string("matrix_ac_23_cast_fp16")]; + tensor matrix_bd_47_begin_0 = const()[name = string("matrix_bd_47_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_47_end_0 = const()[name = string("matrix_bd_47_end_0"), val = tensor([1, 8, 56, 98])]; + tensor matrix_bd_47_end_mask_0 = const()[name = string("matrix_bd_47_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_47_cast_fp16 = slice_by_index(begin = matrix_bd_47_begin_0, end = matrix_bd_47_end_0, end_mask = matrix_bd_47_end_mask_0, x = matrix_bd_45_cast_fp16)[name = string("matrix_bd_47_cast_fp16")]; + tensor var_2880_cast_fp16 = add(x = matrix_ac_23_cast_fp16, y = matrix_bd_47_cast_fp16)[name = string("op_2880_cast_fp16")]; + fp16 _inversed_scores_45_y_0_to_fp16 = const()[name = string("_inversed_scores_45_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_45_cast_fp16 = mul(x = var_2880_cast_fp16, y = _inversed_scores_45_y_0_to_fp16)[name = string("_inversed_scores_45_cast_fp16")]; + tensor scores_47_cast_fp16 = select(a = var_44_to_fp16, b = _inversed_scores_45_cast_fp16, cond = mask_11)[name = string("scores_47_cast_fp16")]; + tensor var_2886_cast_fp16 = softmax(axis = var_58, x = scores_47_cast_fp16)[name = string("op_2886_cast_fp16")]; + tensor input_613_cast_fp16 = select(a = var_43_to_fp16, b = var_2886_cast_fp16, cond = mask_11)[name = string("input_613_cast_fp16")]; + bool x_299_transpose_x_0 = const()[name = string("x_299_transpose_x_0"), val = bool(false)]; + bool x_299_transpose_y_0 = const()[name = string("x_299_transpose_y_0"), val = bool(false)]; + tensor value_31_cast_fp16 = transpose(perm = value_31_perm_0, x = v_23_cast_fp16)[name = string("transpose_260")]; + tensor x_299_cast_fp16 = matmul(transpose_x = x_299_transpose_x_0, transpose_y = x_299_transpose_y_0, x = input_613_cast_fp16, y = value_31_cast_fp16)[name = string("x_299_cast_fp16")]; + tensor var_2890_perm_0 = const()[name = string("op_2890_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_2891 = const()[name = string("op_2891"), val = tensor([1, -1, 1024])]; + tensor var_2890_cast_fp16 = transpose(perm = var_2890_perm_0, x = x_299_cast_fp16)[name = string("transpose_259")]; + tensor input_615_cast_fp16 = reshape(shape = var_2891, x = var_2890_cast_fp16)[name = string("input_615_cast_fp16")]; + tensor encoder_layers_11_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(237616960))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(238403456))))[name = string("encoder_layers_11_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_11_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_11_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(238403648)))]; + tensor linear_106_cast_fp16 = linear(bias = encoder_layers_11_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_11_self_attn_linear_out_weight_to_fp16_palettized, x = input_615_cast_fp16)[name = string("linear_106_cast_fp16")]; + tensor input_619_cast_fp16 = add(x = input_609_cast_fp16, y = linear_106_cast_fp16)[name = string("input_619_cast_fp16")]; + tensor x_303_axes_0 = const()[name = string("x_303_axes_0"), val = tensor([-1])]; + tensor encoder_layers_11_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_11_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(238405760)))]; + tensor encoder_layers_11_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_11_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(238407872)))]; + tensor x_303_cast_fp16 = layer_norm(axes = x_303_axes_0, beta = encoder_layers_11_norm_conv_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_11_norm_conv_weight_to_fp16, x = input_619_cast_fp16)[name = string("x_303_cast_fp16")]; + tensor input_621_perm_0 = const()[name = string("input_621_perm_0"), val = tensor([0, 2, 1])]; + string input_623_pad_type_0 = const()[name = string("input_623_pad_type_0"), val = string("valid")]; + tensor input_623_strides_0 = const()[name = string("input_623_strides_0"), val = tensor([1])]; + tensor input_623_pad_0 = const()[name = string("input_623_pad_0"), val = tensor([0, 0])]; + tensor input_623_dilations_0 = const()[name = string("input_623_dilations_0"), val = tensor([1])]; + int32 input_623_groups_0 = const()[name = string("input_623_groups_0"), val = int32(1)]; + tensor encoder_layers_11_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(238409984))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240507200))))[name = string("encoder_layers_11_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_621_cast_fp16 = transpose(perm = input_621_perm_0, x = x_303_cast_fp16)[name = string("transpose_258")]; + tensor input_623_cast_fp16 = conv(dilations = input_623_dilations_0, groups = input_623_groups_0, pad = input_623_pad_0, pad_type = input_623_pad_type_0, strides = input_623_strides_0, weight = encoder_layers_11_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_621_cast_fp16)[name = string("input_623_cast_fp16")]; + int32 x_305_split_num_splits_0 = const()[name = string("x_305_split_num_splits_0"), val = int32(2)]; + int32 x_305_split_axis_0 = const()[name = string("x_305_split_axis_0"), val = int32(1)]; + tensor x_305_split_cast_fp16_0, tensor x_305_split_cast_fp16_1 = split(axis = x_305_split_axis_0, num_splits = x_305_split_num_splits_0, x = input_623_cast_fp16)[name = string("x_305_split_cast_fp16")]; + tensor x_305_split_1_sigmoid_cast_fp16 = sigmoid(x = x_305_split_cast_fp16_1)[name = string("x_305_split_1_sigmoid_cast_fp16")]; + tensor x_305_cast_fp16 = mul(x = x_305_split_cast_fp16_0, y = x_305_split_1_sigmoid_cast_fp16)[name = string("x_305_cast_fp16")]; + tensor input_625_cast_fp16 = select(a = var_43_to_fp16, b = x_305_cast_fp16, cond = var_574)[name = string("input_625_cast_fp16")]; + bool new_x_47_interleave_0 = const()[name = string("new_x_47_interleave_0"), val = bool(false)]; + tensor new_x_47_cast_fp16 = concat(axis = var_58, interleave = new_x_47_interleave_0, values = (cache_47_cast_fp16, input_625_cast_fp16))[name = string("new_x_47_cast_fp16")]; + tensor var_2930_begin_0 = const()[name = string("op_2930_begin_0"), val = tensor([0, 0, 56])]; + tensor var_2930_end_0 = const()[name = string("op_2930_end_0"), val = tensor([1, 1024, 64])]; + tensor var_2930_end_mask_0 = const()[name = string("op_2930_end_mask_0"), val = tensor([true, true, true])]; + tensor var_2930_cast_fp16 = slice_by_index(begin = var_2930_begin_0, end = var_2930_end_0, end_mask = var_2930_end_mask_0, x = new_x_47_cast_fp16)[name = string("op_2930_cast_fp16")]; + string x_307_pad_type_0 = const()[name = string("x_307_pad_type_0"), val = string("valid")]; + int32 x_307_groups_0 = const()[name = string("x_307_groups_0"), val = int32(1024)]; + tensor x_307_strides_0 = const()[name = string("x_307_strides_0"), val = tensor([1])]; + tensor x_307_pad_0 = const()[name = string("x_307_pad_0"), val = tensor([0, 0])]; + tensor x_307_dilations_0 = const()[name = string("x_307_dilations_0"), val = tensor([1])]; + tensor encoder_layers_11_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240511360))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240520640))))[name = string("encoder_layers_11_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_307_cast_fp16 = conv(dilations = x_307_dilations_0, groups = x_307_groups_0, pad = x_307_pad_0, pad_type = x_307_pad_type_0, strides = x_307_strides_0, weight = encoder_layers_11_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_47_cast_fp16)[name = string("x_307_cast_fp16")]; + tensor input_627_perm_0 = const()[name = string("input_627_perm_0"), val = tensor([0, 2, 1])]; + tensor x_309_axes_0 = const()[name = string("x_309_axes_0"), val = tensor([-1])]; + tensor encoder_layers_11_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_11_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240522752)))]; + tensor encoder_layers_11_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_11_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240524864)))]; + tensor input_627_cast_fp16 = transpose(perm = input_627_perm_0, x = x_307_cast_fp16)[name = string("transpose_257")]; + tensor x_309_cast_fp16 = layer_norm(axes = x_309_axes_0, beta = encoder_layers_11_conv_batch_norm_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_11_conv_batch_norm_weight_to_fp16, x = input_627_cast_fp16)[name = string("x_309_cast_fp16")]; + tensor input_629_perm_0 = const()[name = string("input_629_perm_0"), val = tensor([0, 2, 1])]; + tensor input_629_cast_fp16 = transpose(perm = input_629_perm_0, x = x_309_cast_fp16)[name = string("transpose_256")]; + tensor input_631_cast_fp16 = silu(x = input_629_cast_fp16)[name = string("input_631_cast_fp16")]; + string x_311_pad_type_0 = const()[name = string("x_311_pad_type_0"), val = string("valid")]; + tensor x_311_strides_0 = const()[name = string("x_311_strides_0"), val = tensor([1])]; + tensor x_311_pad_0 = const()[name = string("x_311_pad_0"), val = tensor([0, 0])]; + tensor x_311_dilations_0 = const()[name = string("x_311_dilations_0"), val = tensor([1])]; + int32 x_311_groups_0 = const()[name = string("x_311_groups_0"), val = int32(1)]; + tensor encoder_layers_11_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240526976))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241575616))))[name = string("encoder_layers_11_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_311_cast_fp16 = conv(dilations = x_311_dilations_0, groups = x_311_groups_0, pad = x_311_pad_0, pad_type = x_311_pad_type_0, strides = x_311_strides_0, weight = encoder_layers_11_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_631_cast_fp16)[name = string("x_311_cast_fp16")]; + tensor input_633_perm_0 = const()[name = string("input_633_perm_0"), val = tensor([0, 2, 1])]; + tensor input_633_cast_fp16 = transpose(perm = input_633_perm_0, x = x_311_cast_fp16)[name = string("transpose_255")]; + tensor input_635_cast_fp16 = add(x = input_619_cast_fp16, y = input_633_cast_fp16)[name = string("input_635_cast_fp16")]; + tensor input_637_axes_0 = const()[name = string("input_637_axes_0"), val = tensor([-1])]; + tensor encoder_layers_11_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_11_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241577728)))]; + tensor encoder_layers_11_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_11_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241579840)))]; + tensor input_637_cast_fp16 = layer_norm(axes = input_637_axes_0, beta = encoder_layers_11_norm_feed_forward2_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_11_norm_feed_forward2_weight_to_fp16, x = input_635_cast_fp16)[name = string("input_637_cast_fp16")]; + tensor encoder_layers_11_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241581952))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(244727744))))[name = string("encoder_layers_11_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_11_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_11_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(244727936)))]; + tensor linear_107_cast_fp16 = linear(bias = encoder_layers_11_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_11_feed_forward2_linear1_weight_to_fp16_palettized, x = input_637_cast_fp16)[name = string("linear_107_cast_fp16")]; + tensor input_641_cast_fp16 = silu(x = linear_107_cast_fp16)[name = string("input_641_cast_fp16")]; + tensor encoder_layers_11_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(244736192))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(247881984))))[name = string("encoder_layers_11_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_11_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_11_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(247882176)))]; + tensor linear_108_cast_fp16 = linear(bias = encoder_layers_11_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_11_feed_forward2_linear2_weight_to_fp16_palettized, x = input_641_cast_fp16)[name = string("linear_108_cast_fp16")]; + fp16 var_2973_to_fp16 = const()[name = string("op_2973_to_fp16"), val = fp16(0x1p-1)]; + tensor var_2974_cast_fp16 = mul(x = linear_108_cast_fp16, y = var_2973_to_fp16)[name = string("op_2974_cast_fp16")]; + tensor input_647_cast_fp16 = add(x = input_635_cast_fp16, y = var_2974_cast_fp16)[name = string("input_647_cast_fp16")]; + tensor input_649_axes_0 = const()[name = string("input_649_axes_0"), val = tensor([-1])]; + tensor encoder_layers_11_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_11_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(247884288)))]; + tensor encoder_layers_11_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_11_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(247886400)))]; + tensor input_649_cast_fp16 = layer_norm(axes = input_649_axes_0, beta = encoder_layers_11_norm_out_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_11_norm_out_weight_to_fp16, x = input_647_cast_fp16)[name = string("input_649_cast_fp16")]; + tensor cache_49_begin_0 = const()[name = string("cache_49_begin_0"), val = tensor([12, 0, 0, 0])]; + tensor cache_49_end_0 = const()[name = string("cache_49_end_0"), val = tensor([13, 1, 42, 1024])]; + tensor cache_49_end_mask_0 = const()[name = string("cache_49_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_49_squeeze_mask_0 = const()[name = string("cache_49_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_49_cast_fp16 = slice_by_index(begin = cache_49_begin_0, end = cache_49_end_0, end_mask = cache_49_end_mask_0, squeeze_mask = cache_49_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_49_cast_fp16")]; + tensor cache_51_begin_0 = const()[name = string("cache_51_begin_0"), val = tensor([12, 0, 0, 0])]; + tensor cache_51_end_0 = const()[name = string("cache_51_end_0"), val = tensor([13, 1, 1024, 8])]; + tensor cache_51_end_mask_0 = const()[name = string("cache_51_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_51_squeeze_mask_0 = const()[name = string("cache_51_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_51_cast_fp16 = slice_by_index(begin = cache_51_begin_0, end = cache_51_end_0, end_mask = cache_51_end_mask_0, squeeze_mask = cache_51_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_51_cast_fp16")]; + tensor input_651_axes_0 = const()[name = string("input_651_axes_0"), val = tensor([-1])]; + tensor encoder_layers_12_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_12_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(247888512)))]; + tensor encoder_layers_12_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_12_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(247890624)))]; + tensor input_651_cast_fp16 = layer_norm(axes = input_651_axes_0, beta = encoder_layers_12_norm_feed_forward1_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_12_norm_feed_forward1_weight_to_fp16, x = input_649_cast_fp16)[name = string("input_651_cast_fp16")]; + tensor encoder_layers_12_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(247892736))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(251038528))))[name = string("encoder_layers_12_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_12_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_12_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(251038720)))]; + tensor linear_109_cast_fp16 = linear(bias = encoder_layers_12_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_12_feed_forward1_linear1_weight_to_fp16_palettized, x = input_651_cast_fp16)[name = string("linear_109_cast_fp16")]; + tensor input_655_cast_fp16 = silu(x = linear_109_cast_fp16)[name = string("input_655_cast_fp16")]; + tensor encoder_layers_12_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(251046976))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(254192768))))[name = string("encoder_layers_12_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_12_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_12_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(254192960)))]; + tensor linear_110_cast_fp16 = linear(bias = encoder_layers_12_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_12_feed_forward1_linear2_weight_to_fp16_palettized, x = input_655_cast_fp16)[name = string("linear_110_cast_fp16")]; + fp16 var_3010_to_fp16 = const()[name = string("op_3010_to_fp16"), val = fp16(0x1p-1)]; + tensor var_3011_cast_fp16 = mul(x = linear_110_cast_fp16, y = var_3010_to_fp16)[name = string("op_3011_cast_fp16")]; + tensor input_661_cast_fp16 = add(x = input_649_cast_fp16, y = var_3011_cast_fp16)[name = string("input_661_cast_fp16")]; + tensor key_25_axes_0 = const()[name = string("key_25_axes_0"), val = tensor([-1])]; + tensor encoder_layers_12_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_12_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(254195072)))]; + tensor encoder_layers_12_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_12_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(254197184)))]; + tensor key_25_cast_fp16 = layer_norm(axes = key_25_axes_0, beta = encoder_layers_12_norm_self_att_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_12_norm_self_att_weight_to_fp16, x = input_661_cast_fp16)[name = string("key_25_cast_fp16")]; + bool input_663_interleave_0 = const()[name = string("input_663_interleave_0"), val = bool(false)]; + tensor input_663_cast_fp16 = concat(axis = var_67, interleave = input_663_interleave_0, values = (cache_49_cast_fp16, key_25_cast_fp16))[name = string("input_663_cast_fp16")]; + bool var_3039_interleave_0 = const()[name = string("op_3039_interleave_0"), val = bool(false)]; + tensor var_3039_cast_fp16 = concat(axis = var_67, interleave = var_3039_interleave_0, values = key_25_cast_fp16)[name = string("op_3039_cast_fp16")]; + tensor encoder_layers_12_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(254199296))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(254985792))))[name = string("encoder_layers_12_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_12_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_12_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(254985984)))]; + tensor linear_111_cast_fp16 = linear(bias = encoder_layers_12_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_12_self_attn_linear_q_weight_to_fp16_palettized, x = key_25_cast_fp16)[name = string("linear_111_cast_fp16")]; + tensor var_3044 = const()[name = string("op_3044"), val = tensor([1, -1, 8, 128])]; + tensor q_73_cast_fp16 = reshape(shape = var_3044, x = linear_111_cast_fp16)[name = string("q_73_cast_fp16")]; + tensor encoder_layers_12_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(254988096))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255774592))))[name = string("encoder_layers_12_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_12_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_12_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255774784)))]; + tensor linear_112_cast_fp16 = linear(bias = encoder_layers_12_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_12_self_attn_linear_k_weight_to_fp16_palettized, x = input_663_cast_fp16)[name = string("linear_112_cast_fp16")]; + tensor var_3049 = const()[name = string("op_3049"), val = tensor([1, -1, 8, 128])]; + tensor k_49_cast_fp16 = reshape(shape = var_3049, x = linear_112_cast_fp16)[name = string("k_49_cast_fp16")]; + tensor encoder_layers_12_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255776896))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256563392))))[name = string("encoder_layers_12_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_12_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_12_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256563584)))]; + tensor linear_113_cast_fp16 = linear(bias = encoder_layers_12_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_12_self_attn_linear_v_weight_to_fp16_palettized, x = input_663_cast_fp16)[name = string("linear_113_cast_fp16")]; + tensor var_3054 = const()[name = string("op_3054"), val = tensor([1, -1, 8, 128])]; + tensor v_25_cast_fp16 = reshape(shape = var_3054, x = linear_113_cast_fp16)[name = string("v_25_cast_fp16")]; + tensor value_33_perm_0 = const()[name = string("value_33_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_12_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_12_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256565696)))]; + tensor var_3067_cast_fp16 = add(x = q_73_cast_fp16, y = encoder_layers_12_self_attn_pos_bias_u_to_fp16)[name = string("op_3067_cast_fp16")]; + tensor encoder_layers_12_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_12_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256567808)))]; + tensor var_3069_cast_fp16 = add(x = q_73_cast_fp16, y = encoder_layers_12_self_attn_pos_bias_v_to_fp16)[name = string("op_3069_cast_fp16")]; + tensor q_with_bias_v_25_perm_0 = const()[name = string("q_with_bias_v_25_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_319_transpose_x_0 = const()[name = string("x_319_transpose_x_0"), val = bool(false)]; + bool x_319_transpose_y_0 = const()[name = string("x_319_transpose_y_0"), val = bool(false)]; + tensor op_3071_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256569920))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256769664))))[name = string("op_3071_to_fp16_quantized")]; + tensor q_with_bias_v_25_cast_fp16 = transpose(perm = q_with_bias_v_25_perm_0, x = var_3069_cast_fp16)[name = string("transpose_254")]; + tensor x_319_cast_fp16 = matmul(transpose_x = x_319_transpose_x_0, transpose_y = x_319_transpose_y_0, x = q_with_bias_v_25_cast_fp16, y = op_3071_to_fp16_quantized)[name = string("x_319_cast_fp16")]; + tensor x_321_pad_0 = const()[name = string("x_321_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_321_mode_0 = const()[name = string("x_321_mode_0"), val = string("constant")]; + fp16 const_235_to_fp16 = const()[name = string("const_235_to_fp16"), val = fp16(0x0p+0)]; + tensor x_321_cast_fp16 = pad(constant_val = const_235_to_fp16, mode = x_321_mode_0, pad = x_321_pad_0, x = x_319_cast_fp16)[name = string("x_321_cast_fp16")]; + tensor var_3079 = const()[name = string("op_3079"), val = tensor([1, 8, -1, 56])]; + tensor x_323_cast_fp16 = reshape(shape = var_3079, x = x_321_cast_fp16)[name = string("x_323_cast_fp16")]; + tensor var_3083_begin_0 = const()[name = string("op_3083_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_3083_end_0 = const()[name = string("op_3083_end_0"), val = tensor([1, 8, 196, 56])]; + tensor var_3083_end_mask_0 = const()[name = string("op_3083_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_3083_cast_fp16 = slice_by_index(begin = var_3083_begin_0, end = var_3083_end_0, end_mask = var_3083_end_mask_0, x = x_323_cast_fp16)[name = string("op_3083_cast_fp16")]; + tensor var_3084 = const()[name = string("op_3084"), val = tensor([1, 8, 56, 195])]; + tensor matrix_bd_49_cast_fp16 = reshape(shape = var_3084, x = var_3083_cast_fp16)[name = string("matrix_bd_49_cast_fp16")]; + bool matrix_ac_25_transpose_x_0 = const()[name = string("matrix_ac_25_transpose_x_0"), val = bool(false)]; + bool matrix_ac_25_transpose_y_0 = const()[name = string("matrix_ac_25_transpose_y_0"), val = bool(false)]; + tensor transpose_120_perm_0 = const()[name = string("transpose_120_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_121_perm_0 = const()[name = string("transpose_121_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_121 = transpose(perm = transpose_121_perm_0, x = k_49_cast_fp16)[name = string("transpose_252")]; + tensor transpose_120 = transpose(perm = transpose_120_perm_0, x = var_3067_cast_fp16)[name = string("transpose_253")]; + tensor matrix_ac_25_cast_fp16 = matmul(transpose_x = matrix_ac_25_transpose_x_0, transpose_y = matrix_ac_25_transpose_y_0, x = transpose_120, y = transpose_121)[name = string("matrix_ac_25_cast_fp16")]; + tensor matrix_bd_51_begin_0 = const()[name = string("matrix_bd_51_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_51_end_0 = const()[name = string("matrix_bd_51_end_0"), val = tensor([1, 8, 56, 98])]; + tensor matrix_bd_51_end_mask_0 = const()[name = string("matrix_bd_51_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_51_cast_fp16 = slice_by_index(begin = matrix_bd_51_begin_0, end = matrix_bd_51_end_0, end_mask = matrix_bd_51_end_mask_0, x = matrix_bd_49_cast_fp16)[name = string("matrix_bd_51_cast_fp16")]; + tensor var_3093_cast_fp16 = add(x = matrix_ac_25_cast_fp16, y = matrix_bd_51_cast_fp16)[name = string("op_3093_cast_fp16")]; + fp16 _inversed_scores_49_y_0_to_fp16 = const()[name = string("_inversed_scores_49_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_49_cast_fp16 = mul(x = var_3093_cast_fp16, y = _inversed_scores_49_y_0_to_fp16)[name = string("_inversed_scores_49_cast_fp16")]; + tensor scores_51_cast_fp16 = select(a = var_44_to_fp16, b = _inversed_scores_49_cast_fp16, cond = mask_11)[name = string("scores_51_cast_fp16")]; + tensor var_3099_cast_fp16 = softmax(axis = var_58, x = scores_51_cast_fp16)[name = string("op_3099_cast_fp16")]; + tensor input_665_cast_fp16 = select(a = var_43_to_fp16, b = var_3099_cast_fp16, cond = mask_11)[name = string("input_665_cast_fp16")]; + bool x_325_transpose_x_0 = const()[name = string("x_325_transpose_x_0"), val = bool(false)]; + bool x_325_transpose_y_0 = const()[name = string("x_325_transpose_y_0"), val = bool(false)]; + tensor value_33_cast_fp16 = transpose(perm = value_33_perm_0, x = v_25_cast_fp16)[name = string("transpose_251")]; + tensor x_325_cast_fp16 = matmul(transpose_x = x_325_transpose_x_0, transpose_y = x_325_transpose_y_0, x = input_665_cast_fp16, y = value_33_cast_fp16)[name = string("x_325_cast_fp16")]; + tensor var_3103_perm_0 = const()[name = string("op_3103_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_3104 = const()[name = string("op_3104"), val = tensor([1, -1, 1024])]; + tensor var_3103_cast_fp16 = transpose(perm = var_3103_perm_0, x = x_325_cast_fp16)[name = string("transpose_250")]; + tensor input_667_cast_fp16 = reshape(shape = var_3104, x = var_3103_cast_fp16)[name = string("input_667_cast_fp16")]; + tensor encoder_layers_12_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256770176))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257556672))))[name = string("encoder_layers_12_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_12_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_12_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257556864)))]; + tensor linear_115_cast_fp16 = linear(bias = encoder_layers_12_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_12_self_attn_linear_out_weight_to_fp16_palettized, x = input_667_cast_fp16)[name = string("linear_115_cast_fp16")]; + tensor input_671_cast_fp16 = add(x = input_661_cast_fp16, y = linear_115_cast_fp16)[name = string("input_671_cast_fp16")]; + tensor x_329_axes_0 = const()[name = string("x_329_axes_0"), val = tensor([-1])]; + tensor encoder_layers_12_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_12_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257558976)))]; + tensor encoder_layers_12_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_12_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257561088)))]; + tensor x_329_cast_fp16 = layer_norm(axes = x_329_axes_0, beta = encoder_layers_12_norm_conv_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_12_norm_conv_weight_to_fp16, x = input_671_cast_fp16)[name = string("x_329_cast_fp16")]; + tensor input_673_perm_0 = const()[name = string("input_673_perm_0"), val = tensor([0, 2, 1])]; + string input_675_pad_type_0 = const()[name = string("input_675_pad_type_0"), val = string("valid")]; + tensor input_675_strides_0 = const()[name = string("input_675_strides_0"), val = tensor([1])]; + tensor input_675_pad_0 = const()[name = string("input_675_pad_0"), val = tensor([0, 0])]; + tensor input_675_dilations_0 = const()[name = string("input_675_dilations_0"), val = tensor([1])]; + int32 input_675_groups_0 = const()[name = string("input_675_groups_0"), val = int32(1)]; + tensor encoder_layers_12_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257563200))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259660416))))[name = string("encoder_layers_12_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_673_cast_fp16 = transpose(perm = input_673_perm_0, x = x_329_cast_fp16)[name = string("transpose_249")]; + tensor input_675_cast_fp16 = conv(dilations = input_675_dilations_0, groups = input_675_groups_0, pad = input_675_pad_0, pad_type = input_675_pad_type_0, strides = input_675_strides_0, weight = encoder_layers_12_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_673_cast_fp16)[name = string("input_675_cast_fp16")]; + int32 x_331_split_num_splits_0 = const()[name = string("x_331_split_num_splits_0"), val = int32(2)]; + int32 x_331_split_axis_0 = const()[name = string("x_331_split_axis_0"), val = int32(1)]; + tensor x_331_split_cast_fp16_0, tensor x_331_split_cast_fp16_1 = split(axis = x_331_split_axis_0, num_splits = x_331_split_num_splits_0, x = input_675_cast_fp16)[name = string("x_331_split_cast_fp16")]; + tensor x_331_split_1_sigmoid_cast_fp16 = sigmoid(x = x_331_split_cast_fp16_1)[name = string("x_331_split_1_sigmoid_cast_fp16")]; + tensor x_331_cast_fp16 = mul(x = x_331_split_cast_fp16_0, y = x_331_split_1_sigmoid_cast_fp16)[name = string("x_331_cast_fp16")]; + tensor input_677_cast_fp16 = select(a = var_43_to_fp16, b = x_331_cast_fp16, cond = var_574)[name = string("input_677_cast_fp16")]; + bool new_x_51_interleave_0 = const()[name = string("new_x_51_interleave_0"), val = bool(false)]; + tensor new_x_51_cast_fp16 = concat(axis = var_58, interleave = new_x_51_interleave_0, values = (cache_51_cast_fp16, input_677_cast_fp16))[name = string("new_x_51_cast_fp16")]; + tensor var_3143_begin_0 = const()[name = string("op_3143_begin_0"), val = tensor([0, 0, 56])]; + tensor var_3143_end_0 = const()[name = string("op_3143_end_0"), val = tensor([1, 1024, 64])]; + tensor var_3143_end_mask_0 = const()[name = string("op_3143_end_mask_0"), val = tensor([true, true, true])]; + tensor var_3143_cast_fp16 = slice_by_index(begin = var_3143_begin_0, end = var_3143_end_0, end_mask = var_3143_end_mask_0, x = new_x_51_cast_fp16)[name = string("op_3143_cast_fp16")]; + string x_333_pad_type_0 = const()[name = string("x_333_pad_type_0"), val = string("valid")]; + int32 x_333_groups_0 = const()[name = string("x_333_groups_0"), val = int32(1024)]; + tensor x_333_strides_0 = const()[name = string("x_333_strides_0"), val = tensor([1])]; + tensor x_333_pad_0 = const()[name = string("x_333_pad_0"), val = tensor([0, 0])]; + tensor x_333_dilations_0 = const()[name = string("x_333_dilations_0"), val = tensor([1])]; + tensor encoder_layers_12_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259664576))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259673856))))[name = string("encoder_layers_12_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_333_cast_fp16 = conv(dilations = x_333_dilations_0, groups = x_333_groups_0, pad = x_333_pad_0, pad_type = x_333_pad_type_0, strides = x_333_strides_0, weight = encoder_layers_12_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_51_cast_fp16)[name = string("x_333_cast_fp16")]; + tensor input_679_perm_0 = const()[name = string("input_679_perm_0"), val = tensor([0, 2, 1])]; + tensor x_335_axes_0 = const()[name = string("x_335_axes_0"), val = tensor([-1])]; + tensor encoder_layers_12_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_12_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259675968)))]; + tensor encoder_layers_12_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_12_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259678080)))]; + tensor input_679_cast_fp16 = transpose(perm = input_679_perm_0, x = x_333_cast_fp16)[name = string("transpose_248")]; + tensor x_335_cast_fp16 = layer_norm(axes = x_335_axes_0, beta = encoder_layers_12_conv_batch_norm_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_12_conv_batch_norm_weight_to_fp16, x = input_679_cast_fp16)[name = string("x_335_cast_fp16")]; + tensor input_681_perm_0 = const()[name = string("input_681_perm_0"), val = tensor([0, 2, 1])]; + tensor input_681_cast_fp16 = transpose(perm = input_681_perm_0, x = x_335_cast_fp16)[name = string("transpose_247")]; + tensor input_683_cast_fp16 = silu(x = input_681_cast_fp16)[name = string("input_683_cast_fp16")]; + string x_337_pad_type_0 = const()[name = string("x_337_pad_type_0"), val = string("valid")]; + tensor x_337_strides_0 = const()[name = string("x_337_strides_0"), val = tensor([1])]; + tensor x_337_pad_0 = const()[name = string("x_337_pad_0"), val = tensor([0, 0])]; + tensor x_337_dilations_0 = const()[name = string("x_337_dilations_0"), val = tensor([1])]; + int32 x_337_groups_0 = const()[name = string("x_337_groups_0"), val = int32(1)]; + tensor encoder_layers_12_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259680192))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(260728832))))[name = string("encoder_layers_12_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_337_cast_fp16 = conv(dilations = x_337_dilations_0, groups = x_337_groups_0, pad = x_337_pad_0, pad_type = x_337_pad_type_0, strides = x_337_strides_0, weight = encoder_layers_12_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_683_cast_fp16)[name = string("x_337_cast_fp16")]; + tensor input_685_perm_0 = const()[name = string("input_685_perm_0"), val = tensor([0, 2, 1])]; + tensor input_685_cast_fp16 = transpose(perm = input_685_perm_0, x = x_337_cast_fp16)[name = string("transpose_246")]; + tensor input_687_cast_fp16 = add(x = input_671_cast_fp16, y = input_685_cast_fp16)[name = string("input_687_cast_fp16")]; + tensor input_689_axes_0 = const()[name = string("input_689_axes_0"), val = tensor([-1])]; + tensor encoder_layers_12_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_12_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(260730944)))]; + tensor encoder_layers_12_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_12_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(260733056)))]; + tensor input_689_cast_fp16 = layer_norm(axes = input_689_axes_0, beta = encoder_layers_12_norm_feed_forward2_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_12_norm_feed_forward2_weight_to_fp16, x = input_687_cast_fp16)[name = string("input_689_cast_fp16")]; + tensor encoder_layers_12_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(260735168))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(263880960))))[name = string("encoder_layers_12_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_12_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_12_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(263881152)))]; + tensor linear_116_cast_fp16 = linear(bias = encoder_layers_12_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_12_feed_forward2_linear1_weight_to_fp16_palettized, x = input_689_cast_fp16)[name = string("linear_116_cast_fp16")]; + tensor input_693_cast_fp16 = silu(x = linear_116_cast_fp16)[name = string("input_693_cast_fp16")]; + tensor encoder_layers_12_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(263889408))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(267035200))))[name = string("encoder_layers_12_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_12_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_12_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(267035392)))]; + tensor linear_117_cast_fp16 = linear(bias = encoder_layers_12_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_12_feed_forward2_linear2_weight_to_fp16_palettized, x = input_693_cast_fp16)[name = string("linear_117_cast_fp16")]; + fp16 var_3186_to_fp16 = const()[name = string("op_3186_to_fp16"), val = fp16(0x1p-1)]; + tensor var_3187_cast_fp16 = mul(x = linear_117_cast_fp16, y = var_3186_to_fp16)[name = string("op_3187_cast_fp16")]; + tensor input_699_cast_fp16 = add(x = input_687_cast_fp16, y = var_3187_cast_fp16)[name = string("input_699_cast_fp16")]; + tensor input_701_axes_0 = const()[name = string("input_701_axes_0"), val = tensor([-1])]; + tensor encoder_layers_12_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_12_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(267037504)))]; + tensor encoder_layers_12_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_12_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(267039616)))]; + tensor input_701_cast_fp16 = layer_norm(axes = input_701_axes_0, beta = encoder_layers_12_norm_out_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_12_norm_out_weight_to_fp16, x = input_699_cast_fp16)[name = string("input_701_cast_fp16")]; + tensor cache_53_begin_0 = const()[name = string("cache_53_begin_0"), val = tensor([13, 0, 0, 0])]; + tensor cache_53_end_0 = const()[name = string("cache_53_end_0"), val = tensor([14, 1, 42, 1024])]; + tensor cache_53_end_mask_0 = const()[name = string("cache_53_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_53_squeeze_mask_0 = const()[name = string("cache_53_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_53_cast_fp16 = slice_by_index(begin = cache_53_begin_0, end = cache_53_end_0, end_mask = cache_53_end_mask_0, squeeze_mask = cache_53_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_53_cast_fp16")]; + tensor cache_55_begin_0 = const()[name = string("cache_55_begin_0"), val = tensor([13, 0, 0, 0])]; + tensor cache_55_end_0 = const()[name = string("cache_55_end_0"), val = tensor([14, 1, 1024, 8])]; + tensor cache_55_end_mask_0 = const()[name = string("cache_55_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_55_squeeze_mask_0 = const()[name = string("cache_55_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_55_cast_fp16 = slice_by_index(begin = cache_55_begin_0, end = cache_55_end_0, end_mask = cache_55_end_mask_0, squeeze_mask = cache_55_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_55_cast_fp16")]; + tensor input_703_axes_0 = const()[name = string("input_703_axes_0"), val = tensor([-1])]; + tensor encoder_layers_13_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_13_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(267041728)))]; + tensor encoder_layers_13_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_13_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(267043840)))]; + tensor input_703_cast_fp16 = layer_norm(axes = input_703_axes_0, beta = encoder_layers_13_norm_feed_forward1_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_13_norm_feed_forward1_weight_to_fp16, x = input_701_cast_fp16)[name = string("input_703_cast_fp16")]; + tensor encoder_layers_13_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(267045952))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(270191744))))[name = string("encoder_layers_13_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_13_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_13_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(270191936)))]; + tensor linear_118_cast_fp16 = linear(bias = encoder_layers_13_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_13_feed_forward1_linear1_weight_to_fp16_palettized, x = input_703_cast_fp16)[name = string("linear_118_cast_fp16")]; + tensor input_707_cast_fp16 = silu(x = linear_118_cast_fp16)[name = string("input_707_cast_fp16")]; + tensor encoder_layers_13_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(270200192))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(273345984))))[name = string("encoder_layers_13_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_13_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_13_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(273346176)))]; + tensor linear_119_cast_fp16 = linear(bias = encoder_layers_13_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_13_feed_forward1_linear2_weight_to_fp16_palettized, x = input_707_cast_fp16)[name = string("linear_119_cast_fp16")]; + fp16 var_3223_to_fp16 = const()[name = string("op_3223_to_fp16"), val = fp16(0x1p-1)]; + tensor var_3224_cast_fp16 = mul(x = linear_119_cast_fp16, y = var_3223_to_fp16)[name = string("op_3224_cast_fp16")]; + tensor input_713_cast_fp16 = add(x = input_701_cast_fp16, y = var_3224_cast_fp16)[name = string("input_713_cast_fp16")]; + tensor key_27_axes_0 = const()[name = string("key_27_axes_0"), val = tensor([-1])]; + tensor encoder_layers_13_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_13_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(273348288)))]; + tensor encoder_layers_13_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_13_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(273350400)))]; + tensor key_27_cast_fp16 = layer_norm(axes = key_27_axes_0, beta = encoder_layers_13_norm_self_att_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_13_norm_self_att_weight_to_fp16, x = input_713_cast_fp16)[name = string("key_27_cast_fp16")]; + bool input_715_interleave_0 = const()[name = string("input_715_interleave_0"), val = bool(false)]; + tensor input_715_cast_fp16 = concat(axis = var_67, interleave = input_715_interleave_0, values = (cache_53_cast_fp16, key_27_cast_fp16))[name = string("input_715_cast_fp16")]; + bool var_3252_interleave_0 = const()[name = string("op_3252_interleave_0"), val = bool(false)]; + tensor var_3252_cast_fp16 = concat(axis = var_67, interleave = var_3252_interleave_0, values = key_27_cast_fp16)[name = string("op_3252_cast_fp16")]; + tensor encoder_layers_13_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(273352512))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274139008))))[name = string("encoder_layers_13_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_13_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_13_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274139200)))]; + tensor linear_120_cast_fp16 = linear(bias = encoder_layers_13_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_13_self_attn_linear_q_weight_to_fp16_palettized, x = key_27_cast_fp16)[name = string("linear_120_cast_fp16")]; + tensor var_3257 = const()[name = string("op_3257"), val = tensor([1, -1, 8, 128])]; + tensor q_79_cast_fp16 = reshape(shape = var_3257, x = linear_120_cast_fp16)[name = string("q_79_cast_fp16")]; + tensor encoder_layers_13_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274141312))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274927808))))[name = string("encoder_layers_13_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_13_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_13_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274928000)))]; + tensor linear_121_cast_fp16 = linear(bias = encoder_layers_13_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_13_self_attn_linear_k_weight_to_fp16_palettized, x = input_715_cast_fp16)[name = string("linear_121_cast_fp16")]; + tensor var_3262 = const()[name = string("op_3262"), val = tensor([1, -1, 8, 128])]; + tensor k_53_cast_fp16 = reshape(shape = var_3262, x = linear_121_cast_fp16)[name = string("k_53_cast_fp16")]; + tensor encoder_layers_13_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274930112))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275716608))))[name = string("encoder_layers_13_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_13_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_13_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275716800)))]; + tensor linear_122_cast_fp16 = linear(bias = encoder_layers_13_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_13_self_attn_linear_v_weight_to_fp16_palettized, x = input_715_cast_fp16)[name = string("linear_122_cast_fp16")]; + tensor var_3267 = const()[name = string("op_3267"), val = tensor([1, -1, 8, 128])]; + tensor v_27_cast_fp16 = reshape(shape = var_3267, x = linear_122_cast_fp16)[name = string("v_27_cast_fp16")]; + tensor value_35_perm_0 = const()[name = string("value_35_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_13_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_13_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275718912)))]; + tensor var_3280_cast_fp16 = add(x = q_79_cast_fp16, y = encoder_layers_13_self_attn_pos_bias_u_to_fp16)[name = string("op_3280_cast_fp16")]; + tensor encoder_layers_13_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_13_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275721024)))]; + tensor var_3282_cast_fp16 = add(x = q_79_cast_fp16, y = encoder_layers_13_self_attn_pos_bias_v_to_fp16)[name = string("op_3282_cast_fp16")]; + tensor q_with_bias_v_27_perm_0 = const()[name = string("q_with_bias_v_27_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_345_transpose_x_0 = const()[name = string("x_345_transpose_x_0"), val = bool(false)]; + bool x_345_transpose_y_0 = const()[name = string("x_345_transpose_y_0"), val = bool(false)]; + tensor op_3284_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275723136))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275922880))))[name = string("op_3284_to_fp16_quantized")]; + tensor q_with_bias_v_27_cast_fp16 = transpose(perm = q_with_bias_v_27_perm_0, x = var_3282_cast_fp16)[name = string("transpose_245")]; + tensor x_345_cast_fp16 = matmul(transpose_x = x_345_transpose_x_0, transpose_y = x_345_transpose_y_0, x = q_with_bias_v_27_cast_fp16, y = op_3284_to_fp16_quantized)[name = string("x_345_cast_fp16")]; + tensor x_347_pad_0 = const()[name = string("x_347_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_347_mode_0 = const()[name = string("x_347_mode_0"), val = string("constant")]; + fp16 const_248_to_fp16 = const()[name = string("const_248_to_fp16"), val = fp16(0x0p+0)]; + tensor x_347_cast_fp16 = pad(constant_val = const_248_to_fp16, mode = x_347_mode_0, pad = x_347_pad_0, x = x_345_cast_fp16)[name = string("x_347_cast_fp16")]; + tensor var_3292 = const()[name = string("op_3292"), val = tensor([1, 8, -1, 56])]; + tensor x_349_cast_fp16 = reshape(shape = var_3292, x = x_347_cast_fp16)[name = string("x_349_cast_fp16")]; + tensor var_3296_begin_0 = const()[name = string("op_3296_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_3296_end_0 = const()[name = string("op_3296_end_0"), val = tensor([1, 8, 196, 56])]; + tensor var_3296_end_mask_0 = const()[name = string("op_3296_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_3296_cast_fp16 = slice_by_index(begin = var_3296_begin_0, end = var_3296_end_0, end_mask = var_3296_end_mask_0, x = x_349_cast_fp16)[name = string("op_3296_cast_fp16")]; + tensor var_3297 = const()[name = string("op_3297"), val = tensor([1, 8, 56, 195])]; + tensor matrix_bd_53_cast_fp16 = reshape(shape = var_3297, x = var_3296_cast_fp16)[name = string("matrix_bd_53_cast_fp16")]; + bool matrix_ac_27_transpose_x_0 = const()[name = string("matrix_ac_27_transpose_x_0"), val = bool(false)]; + bool matrix_ac_27_transpose_y_0 = const()[name = string("matrix_ac_27_transpose_y_0"), val = bool(false)]; + tensor transpose_122_perm_0 = const()[name = string("transpose_122_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_123_perm_0 = const()[name = string("transpose_123_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_123 = transpose(perm = transpose_123_perm_0, x = k_53_cast_fp16)[name = string("transpose_243")]; + tensor transpose_122 = transpose(perm = transpose_122_perm_0, x = var_3280_cast_fp16)[name = string("transpose_244")]; + tensor matrix_ac_27_cast_fp16 = matmul(transpose_x = matrix_ac_27_transpose_x_0, transpose_y = matrix_ac_27_transpose_y_0, x = transpose_122, y = transpose_123)[name = string("matrix_ac_27_cast_fp16")]; + tensor matrix_bd_55_begin_0 = const()[name = string("matrix_bd_55_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_55_end_0 = const()[name = string("matrix_bd_55_end_0"), val = tensor([1, 8, 56, 98])]; + tensor matrix_bd_55_end_mask_0 = const()[name = string("matrix_bd_55_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_55_cast_fp16 = slice_by_index(begin = matrix_bd_55_begin_0, end = matrix_bd_55_end_0, end_mask = matrix_bd_55_end_mask_0, x = matrix_bd_53_cast_fp16)[name = string("matrix_bd_55_cast_fp16")]; + tensor var_3306_cast_fp16 = add(x = matrix_ac_27_cast_fp16, y = matrix_bd_55_cast_fp16)[name = string("op_3306_cast_fp16")]; + fp16 _inversed_scores_53_y_0_to_fp16 = const()[name = string("_inversed_scores_53_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_53_cast_fp16 = mul(x = var_3306_cast_fp16, y = _inversed_scores_53_y_0_to_fp16)[name = string("_inversed_scores_53_cast_fp16")]; + tensor scores_55_cast_fp16 = select(a = var_44_to_fp16, b = _inversed_scores_53_cast_fp16, cond = mask_11)[name = string("scores_55_cast_fp16")]; + tensor var_3312_cast_fp16 = softmax(axis = var_58, x = scores_55_cast_fp16)[name = string("op_3312_cast_fp16")]; + tensor input_717_cast_fp16 = select(a = var_43_to_fp16, b = var_3312_cast_fp16, cond = mask_11)[name = string("input_717_cast_fp16")]; + bool x_351_transpose_x_0 = const()[name = string("x_351_transpose_x_0"), val = bool(false)]; + bool x_351_transpose_y_0 = const()[name = string("x_351_transpose_y_0"), val = bool(false)]; + tensor value_35_cast_fp16 = transpose(perm = value_35_perm_0, x = v_27_cast_fp16)[name = string("transpose_242")]; + tensor x_351_cast_fp16 = matmul(transpose_x = x_351_transpose_x_0, transpose_y = x_351_transpose_y_0, x = input_717_cast_fp16, y = value_35_cast_fp16)[name = string("x_351_cast_fp16")]; + tensor var_3316_perm_0 = const()[name = string("op_3316_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_3317 = const()[name = string("op_3317"), val = tensor([1, -1, 1024])]; + tensor var_3316_cast_fp16 = transpose(perm = var_3316_perm_0, x = x_351_cast_fp16)[name = string("transpose_241")]; + tensor input_719_cast_fp16 = reshape(shape = var_3317, x = var_3316_cast_fp16)[name = string("input_719_cast_fp16")]; + tensor encoder_layers_13_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275923392))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(276709888))))[name = string("encoder_layers_13_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_13_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_13_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(276710080)))]; + tensor linear_124_cast_fp16 = linear(bias = encoder_layers_13_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_13_self_attn_linear_out_weight_to_fp16_palettized, x = input_719_cast_fp16)[name = string("linear_124_cast_fp16")]; + tensor input_723_cast_fp16 = add(x = input_713_cast_fp16, y = linear_124_cast_fp16)[name = string("input_723_cast_fp16")]; + tensor x_355_axes_0 = const()[name = string("x_355_axes_0"), val = tensor([-1])]; + tensor encoder_layers_13_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_13_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(276712192)))]; + tensor encoder_layers_13_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_13_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(276714304)))]; + tensor x_355_cast_fp16 = layer_norm(axes = x_355_axes_0, beta = encoder_layers_13_norm_conv_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_13_norm_conv_weight_to_fp16, x = input_723_cast_fp16)[name = string("x_355_cast_fp16")]; + tensor input_725_perm_0 = const()[name = string("input_725_perm_0"), val = tensor([0, 2, 1])]; + string input_727_pad_type_0 = const()[name = string("input_727_pad_type_0"), val = string("valid")]; + tensor input_727_strides_0 = const()[name = string("input_727_strides_0"), val = tensor([1])]; + tensor input_727_pad_0 = const()[name = string("input_727_pad_0"), val = tensor([0, 0])]; + tensor input_727_dilations_0 = const()[name = string("input_727_dilations_0"), val = tensor([1])]; + int32 input_727_groups_0 = const()[name = string("input_727_groups_0"), val = int32(1)]; + tensor encoder_layers_13_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(276716416))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278813632))))[name = string("encoder_layers_13_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_725_cast_fp16 = transpose(perm = input_725_perm_0, x = x_355_cast_fp16)[name = string("transpose_240")]; + tensor input_727_cast_fp16 = conv(dilations = input_727_dilations_0, groups = input_727_groups_0, pad = input_727_pad_0, pad_type = input_727_pad_type_0, strides = input_727_strides_0, weight = encoder_layers_13_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_725_cast_fp16)[name = string("input_727_cast_fp16")]; + int32 x_357_split_num_splits_0 = const()[name = string("x_357_split_num_splits_0"), val = int32(2)]; + int32 x_357_split_axis_0 = const()[name = string("x_357_split_axis_0"), val = int32(1)]; + tensor x_357_split_cast_fp16_0, tensor x_357_split_cast_fp16_1 = split(axis = x_357_split_axis_0, num_splits = x_357_split_num_splits_0, x = input_727_cast_fp16)[name = string("x_357_split_cast_fp16")]; + tensor x_357_split_1_sigmoid_cast_fp16 = sigmoid(x = x_357_split_cast_fp16_1)[name = string("x_357_split_1_sigmoid_cast_fp16")]; + tensor x_357_cast_fp16 = mul(x = x_357_split_cast_fp16_0, y = x_357_split_1_sigmoid_cast_fp16)[name = string("x_357_cast_fp16")]; + tensor input_729_cast_fp16 = select(a = var_43_to_fp16, b = x_357_cast_fp16, cond = var_574)[name = string("input_729_cast_fp16")]; + bool new_x_55_interleave_0 = const()[name = string("new_x_55_interleave_0"), val = bool(false)]; + tensor new_x_55_cast_fp16 = concat(axis = var_58, interleave = new_x_55_interleave_0, values = (cache_55_cast_fp16, input_729_cast_fp16))[name = string("new_x_55_cast_fp16")]; + tensor var_3356_begin_0 = const()[name = string("op_3356_begin_0"), val = tensor([0, 0, 56])]; + tensor var_3356_end_0 = const()[name = string("op_3356_end_0"), val = tensor([1, 1024, 64])]; + tensor var_3356_end_mask_0 = const()[name = string("op_3356_end_mask_0"), val = tensor([true, true, true])]; + tensor var_3356_cast_fp16 = slice_by_index(begin = var_3356_begin_0, end = var_3356_end_0, end_mask = var_3356_end_mask_0, x = new_x_55_cast_fp16)[name = string("op_3356_cast_fp16")]; + string x_359_pad_type_0 = const()[name = string("x_359_pad_type_0"), val = string("valid")]; + int32 x_359_groups_0 = const()[name = string("x_359_groups_0"), val = int32(1024)]; + tensor x_359_strides_0 = const()[name = string("x_359_strides_0"), val = tensor([1])]; + tensor x_359_pad_0 = const()[name = string("x_359_pad_0"), val = tensor([0, 0])]; + tensor x_359_dilations_0 = const()[name = string("x_359_dilations_0"), val = tensor([1])]; + tensor encoder_layers_13_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278817792))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278827072))))[name = string("encoder_layers_13_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_359_cast_fp16 = conv(dilations = x_359_dilations_0, groups = x_359_groups_0, pad = x_359_pad_0, pad_type = x_359_pad_type_0, strides = x_359_strides_0, weight = encoder_layers_13_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_55_cast_fp16)[name = string("x_359_cast_fp16")]; + tensor input_731_perm_0 = const()[name = string("input_731_perm_0"), val = tensor([0, 2, 1])]; + tensor x_361_axes_0 = const()[name = string("x_361_axes_0"), val = tensor([-1])]; + tensor encoder_layers_13_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_13_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278829184)))]; + tensor encoder_layers_13_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_13_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278831296)))]; + tensor input_731_cast_fp16 = transpose(perm = input_731_perm_0, x = x_359_cast_fp16)[name = string("transpose_239")]; + tensor x_361_cast_fp16 = layer_norm(axes = x_361_axes_0, beta = encoder_layers_13_conv_batch_norm_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_13_conv_batch_norm_weight_to_fp16, x = input_731_cast_fp16)[name = string("x_361_cast_fp16")]; + tensor input_733_perm_0 = const()[name = string("input_733_perm_0"), val = tensor([0, 2, 1])]; + tensor input_733_cast_fp16 = transpose(perm = input_733_perm_0, x = x_361_cast_fp16)[name = string("transpose_238")]; + tensor input_735_cast_fp16 = silu(x = input_733_cast_fp16)[name = string("input_735_cast_fp16")]; + string x_363_pad_type_0 = const()[name = string("x_363_pad_type_0"), val = string("valid")]; + tensor x_363_strides_0 = const()[name = string("x_363_strides_0"), val = tensor([1])]; + tensor x_363_pad_0 = const()[name = string("x_363_pad_0"), val = tensor([0, 0])]; + tensor x_363_dilations_0 = const()[name = string("x_363_dilations_0"), val = tensor([1])]; + int32 x_363_groups_0 = const()[name = string("x_363_groups_0"), val = int32(1)]; + tensor encoder_layers_13_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278833408))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(279882048))))[name = string("encoder_layers_13_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_363_cast_fp16 = conv(dilations = x_363_dilations_0, groups = x_363_groups_0, pad = x_363_pad_0, pad_type = x_363_pad_type_0, strides = x_363_strides_0, weight = encoder_layers_13_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_735_cast_fp16)[name = string("x_363_cast_fp16")]; + tensor input_737_perm_0 = const()[name = string("input_737_perm_0"), val = tensor([0, 2, 1])]; + tensor input_737_cast_fp16 = transpose(perm = input_737_perm_0, x = x_363_cast_fp16)[name = string("transpose_237")]; + tensor input_739_cast_fp16 = add(x = input_723_cast_fp16, y = input_737_cast_fp16)[name = string("input_739_cast_fp16")]; + tensor input_741_axes_0 = const()[name = string("input_741_axes_0"), val = tensor([-1])]; + tensor encoder_layers_13_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_13_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(279884160)))]; + tensor encoder_layers_13_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_13_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(279886272)))]; + tensor input_741_cast_fp16 = layer_norm(axes = input_741_axes_0, beta = encoder_layers_13_norm_feed_forward2_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_13_norm_feed_forward2_weight_to_fp16, x = input_739_cast_fp16)[name = string("input_741_cast_fp16")]; + tensor encoder_layers_13_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(279888384))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(283034176))))[name = string("encoder_layers_13_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_13_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_13_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(283034368)))]; + tensor linear_125_cast_fp16 = linear(bias = encoder_layers_13_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_13_feed_forward2_linear1_weight_to_fp16_palettized, x = input_741_cast_fp16)[name = string("linear_125_cast_fp16")]; + tensor input_745_cast_fp16 = silu(x = linear_125_cast_fp16)[name = string("input_745_cast_fp16")]; + tensor encoder_layers_13_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(283042624))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286188416))))[name = string("encoder_layers_13_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_13_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_13_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286188608)))]; + tensor linear_126_cast_fp16 = linear(bias = encoder_layers_13_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_13_feed_forward2_linear2_weight_to_fp16_palettized, x = input_745_cast_fp16)[name = string("linear_126_cast_fp16")]; + fp16 var_3399_to_fp16 = const()[name = string("op_3399_to_fp16"), val = fp16(0x1p-1)]; + tensor var_3400_cast_fp16 = mul(x = linear_126_cast_fp16, y = var_3399_to_fp16)[name = string("op_3400_cast_fp16")]; + tensor input_751_cast_fp16 = add(x = input_739_cast_fp16, y = var_3400_cast_fp16)[name = string("input_751_cast_fp16")]; + tensor input_753_axes_0 = const()[name = string("input_753_axes_0"), val = tensor([-1])]; + tensor encoder_layers_13_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_13_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286190720)))]; + tensor encoder_layers_13_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_13_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286192832)))]; + tensor input_753_cast_fp16 = layer_norm(axes = input_753_axes_0, beta = encoder_layers_13_norm_out_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_13_norm_out_weight_to_fp16, x = input_751_cast_fp16)[name = string("input_753_cast_fp16")]; + tensor cache_57_begin_0 = const()[name = string("cache_57_begin_0"), val = tensor([14, 0, 0, 0])]; + tensor cache_57_end_0 = const()[name = string("cache_57_end_0"), val = tensor([15, 1, 42, 1024])]; + tensor cache_57_end_mask_0 = const()[name = string("cache_57_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_57_squeeze_mask_0 = const()[name = string("cache_57_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_57_cast_fp16 = slice_by_index(begin = cache_57_begin_0, end = cache_57_end_0, end_mask = cache_57_end_mask_0, squeeze_mask = cache_57_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_57_cast_fp16")]; + tensor cache_59_begin_0 = const()[name = string("cache_59_begin_0"), val = tensor([14, 0, 0, 0])]; + tensor cache_59_end_0 = const()[name = string("cache_59_end_0"), val = tensor([15, 1, 1024, 8])]; + tensor cache_59_end_mask_0 = const()[name = string("cache_59_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_59_squeeze_mask_0 = const()[name = string("cache_59_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_59_cast_fp16 = slice_by_index(begin = cache_59_begin_0, end = cache_59_end_0, end_mask = cache_59_end_mask_0, squeeze_mask = cache_59_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_59_cast_fp16")]; + tensor input_755_axes_0 = const()[name = string("input_755_axes_0"), val = tensor([-1])]; + tensor encoder_layers_14_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_14_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286194944)))]; + tensor encoder_layers_14_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_14_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286197056)))]; + tensor input_755_cast_fp16 = layer_norm(axes = input_755_axes_0, beta = encoder_layers_14_norm_feed_forward1_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_14_norm_feed_forward1_weight_to_fp16, x = input_753_cast_fp16)[name = string("input_755_cast_fp16")]; + tensor encoder_layers_14_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286199168))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(289344960))))[name = string("encoder_layers_14_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_14_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_14_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(289345152)))]; + tensor linear_127_cast_fp16 = linear(bias = encoder_layers_14_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_14_feed_forward1_linear1_weight_to_fp16_palettized, x = input_755_cast_fp16)[name = string("linear_127_cast_fp16")]; + tensor input_759_cast_fp16 = silu(x = linear_127_cast_fp16)[name = string("input_759_cast_fp16")]; + tensor encoder_layers_14_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(289353408))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(292499200))))[name = string("encoder_layers_14_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_14_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_14_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(292499392)))]; + tensor linear_128_cast_fp16 = linear(bias = encoder_layers_14_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_14_feed_forward1_linear2_weight_to_fp16_palettized, x = input_759_cast_fp16)[name = string("linear_128_cast_fp16")]; + fp16 var_3436_to_fp16 = const()[name = string("op_3436_to_fp16"), val = fp16(0x1p-1)]; + tensor var_3437_cast_fp16 = mul(x = linear_128_cast_fp16, y = var_3436_to_fp16)[name = string("op_3437_cast_fp16")]; + tensor input_765_cast_fp16 = add(x = input_753_cast_fp16, y = var_3437_cast_fp16)[name = string("input_765_cast_fp16")]; + tensor key_29_axes_0 = const()[name = string("key_29_axes_0"), val = tensor([-1])]; + tensor encoder_layers_14_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_14_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(292501504)))]; + tensor encoder_layers_14_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_14_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(292503616)))]; + tensor key_29_cast_fp16 = layer_norm(axes = key_29_axes_0, beta = encoder_layers_14_norm_self_att_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_14_norm_self_att_weight_to_fp16, x = input_765_cast_fp16)[name = string("key_29_cast_fp16")]; + bool input_767_interleave_0 = const()[name = string("input_767_interleave_0"), val = bool(false)]; + tensor input_767_cast_fp16 = concat(axis = var_67, interleave = input_767_interleave_0, values = (cache_57_cast_fp16, key_29_cast_fp16))[name = string("input_767_cast_fp16")]; + bool var_3465_interleave_0 = const()[name = string("op_3465_interleave_0"), val = bool(false)]; + tensor var_3465_cast_fp16 = concat(axis = var_67, interleave = var_3465_interleave_0, values = key_29_cast_fp16)[name = string("op_3465_cast_fp16")]; + tensor encoder_layers_14_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(292505728))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293292224))))[name = string("encoder_layers_14_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_14_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_14_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293292416)))]; + tensor linear_129_cast_fp16 = linear(bias = encoder_layers_14_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_14_self_attn_linear_q_weight_to_fp16_palettized, x = key_29_cast_fp16)[name = string("linear_129_cast_fp16")]; + tensor var_3470 = const()[name = string("op_3470"), val = tensor([1, -1, 8, 128])]; + tensor q_85_cast_fp16 = reshape(shape = var_3470, x = linear_129_cast_fp16)[name = string("q_85_cast_fp16")]; + tensor encoder_layers_14_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293294528))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294081024))))[name = string("encoder_layers_14_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_14_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_14_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294081216)))]; + tensor linear_130_cast_fp16 = linear(bias = encoder_layers_14_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_14_self_attn_linear_k_weight_to_fp16_palettized, x = input_767_cast_fp16)[name = string("linear_130_cast_fp16")]; + tensor var_3475 = const()[name = string("op_3475"), val = tensor([1, -1, 8, 128])]; + tensor k_57_cast_fp16 = reshape(shape = var_3475, x = linear_130_cast_fp16)[name = string("k_57_cast_fp16")]; + tensor encoder_layers_14_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294083328))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294869824))))[name = string("encoder_layers_14_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_14_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_14_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294870016)))]; + tensor linear_131_cast_fp16 = linear(bias = encoder_layers_14_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_14_self_attn_linear_v_weight_to_fp16_palettized, x = input_767_cast_fp16)[name = string("linear_131_cast_fp16")]; + tensor var_3480 = const()[name = string("op_3480"), val = tensor([1, -1, 8, 128])]; + tensor v_29_cast_fp16 = reshape(shape = var_3480, x = linear_131_cast_fp16)[name = string("v_29_cast_fp16")]; + tensor value_37_perm_0 = const()[name = string("value_37_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_14_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_14_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294872128)))]; + tensor var_3493_cast_fp16 = add(x = q_85_cast_fp16, y = encoder_layers_14_self_attn_pos_bias_u_to_fp16)[name = string("op_3493_cast_fp16")]; + tensor encoder_layers_14_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_14_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294874240)))]; + tensor var_3495_cast_fp16 = add(x = q_85_cast_fp16, y = encoder_layers_14_self_attn_pos_bias_v_to_fp16)[name = string("op_3495_cast_fp16")]; + tensor q_with_bias_v_29_perm_0 = const()[name = string("q_with_bias_v_29_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_371_transpose_x_0 = const()[name = string("x_371_transpose_x_0"), val = bool(false)]; + bool x_371_transpose_y_0 = const()[name = string("x_371_transpose_y_0"), val = bool(false)]; + tensor op_3497_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294876352))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(295076096))))[name = string("op_3497_to_fp16_quantized")]; + tensor q_with_bias_v_29_cast_fp16 = transpose(perm = q_with_bias_v_29_perm_0, x = var_3495_cast_fp16)[name = string("transpose_236")]; + tensor x_371_cast_fp16 = matmul(transpose_x = x_371_transpose_x_0, transpose_y = x_371_transpose_y_0, x = q_with_bias_v_29_cast_fp16, y = op_3497_to_fp16_quantized)[name = string("x_371_cast_fp16")]; + tensor x_373_pad_0 = const()[name = string("x_373_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_373_mode_0 = const()[name = string("x_373_mode_0"), val = string("constant")]; + fp16 const_261_to_fp16 = const()[name = string("const_261_to_fp16"), val = fp16(0x0p+0)]; + tensor x_373_cast_fp16 = pad(constant_val = const_261_to_fp16, mode = x_373_mode_0, pad = x_373_pad_0, x = x_371_cast_fp16)[name = string("x_373_cast_fp16")]; + tensor var_3505 = const()[name = string("op_3505"), val = tensor([1, 8, -1, 56])]; + tensor x_375_cast_fp16 = reshape(shape = var_3505, x = x_373_cast_fp16)[name = string("x_375_cast_fp16")]; + tensor var_3509_begin_0 = const()[name = string("op_3509_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_3509_end_0 = const()[name = string("op_3509_end_0"), val = tensor([1, 8, 196, 56])]; + tensor var_3509_end_mask_0 = const()[name = string("op_3509_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_3509_cast_fp16 = slice_by_index(begin = var_3509_begin_0, end = var_3509_end_0, end_mask = var_3509_end_mask_0, x = x_375_cast_fp16)[name = string("op_3509_cast_fp16")]; + tensor var_3510 = const()[name = string("op_3510"), val = tensor([1, 8, 56, 195])]; + tensor matrix_bd_57_cast_fp16 = reshape(shape = var_3510, x = var_3509_cast_fp16)[name = string("matrix_bd_57_cast_fp16")]; + bool matrix_ac_29_transpose_x_0 = const()[name = string("matrix_ac_29_transpose_x_0"), val = bool(false)]; + bool matrix_ac_29_transpose_y_0 = const()[name = string("matrix_ac_29_transpose_y_0"), val = bool(false)]; + tensor transpose_124_perm_0 = const()[name = string("transpose_124_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_125_perm_0 = const()[name = string("transpose_125_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_125 = transpose(perm = transpose_125_perm_0, x = k_57_cast_fp16)[name = string("transpose_234")]; + tensor transpose_124 = transpose(perm = transpose_124_perm_0, x = var_3493_cast_fp16)[name = string("transpose_235")]; + tensor matrix_ac_29_cast_fp16 = matmul(transpose_x = matrix_ac_29_transpose_x_0, transpose_y = matrix_ac_29_transpose_y_0, x = transpose_124, y = transpose_125)[name = string("matrix_ac_29_cast_fp16")]; + tensor matrix_bd_59_begin_0 = const()[name = string("matrix_bd_59_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_59_end_0 = const()[name = string("matrix_bd_59_end_0"), val = tensor([1, 8, 56, 98])]; + tensor matrix_bd_59_end_mask_0 = const()[name = string("matrix_bd_59_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_59_cast_fp16 = slice_by_index(begin = matrix_bd_59_begin_0, end = matrix_bd_59_end_0, end_mask = matrix_bd_59_end_mask_0, x = matrix_bd_57_cast_fp16)[name = string("matrix_bd_59_cast_fp16")]; + tensor var_3519_cast_fp16 = add(x = matrix_ac_29_cast_fp16, y = matrix_bd_59_cast_fp16)[name = string("op_3519_cast_fp16")]; + fp16 _inversed_scores_57_y_0_to_fp16 = const()[name = string("_inversed_scores_57_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_57_cast_fp16 = mul(x = var_3519_cast_fp16, y = _inversed_scores_57_y_0_to_fp16)[name = string("_inversed_scores_57_cast_fp16")]; + tensor scores_59_cast_fp16 = select(a = var_44_to_fp16, b = _inversed_scores_57_cast_fp16, cond = mask_11)[name = string("scores_59_cast_fp16")]; + tensor var_3525_cast_fp16 = softmax(axis = var_58, x = scores_59_cast_fp16)[name = string("op_3525_cast_fp16")]; + tensor input_769_cast_fp16 = select(a = var_43_to_fp16, b = var_3525_cast_fp16, cond = mask_11)[name = string("input_769_cast_fp16")]; + bool x_377_transpose_x_0 = const()[name = string("x_377_transpose_x_0"), val = bool(false)]; + bool x_377_transpose_y_0 = const()[name = string("x_377_transpose_y_0"), val = bool(false)]; + tensor value_37_cast_fp16 = transpose(perm = value_37_perm_0, x = v_29_cast_fp16)[name = string("transpose_233")]; + tensor x_377_cast_fp16 = matmul(transpose_x = x_377_transpose_x_0, transpose_y = x_377_transpose_y_0, x = input_769_cast_fp16, y = value_37_cast_fp16)[name = string("x_377_cast_fp16")]; + tensor var_3529_perm_0 = const()[name = string("op_3529_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_3530 = const()[name = string("op_3530"), val = tensor([1, -1, 1024])]; + tensor var_3529_cast_fp16 = transpose(perm = var_3529_perm_0, x = x_377_cast_fp16)[name = string("transpose_232")]; + tensor input_771_cast_fp16 = reshape(shape = var_3530, x = var_3529_cast_fp16)[name = string("input_771_cast_fp16")]; + tensor encoder_layers_14_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(295076608))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(295863104))))[name = string("encoder_layers_14_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_14_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_14_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(295863296)))]; + tensor linear_133_cast_fp16 = linear(bias = encoder_layers_14_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_14_self_attn_linear_out_weight_to_fp16_palettized, x = input_771_cast_fp16)[name = string("linear_133_cast_fp16")]; + tensor input_775_cast_fp16 = add(x = input_765_cast_fp16, y = linear_133_cast_fp16)[name = string("input_775_cast_fp16")]; + tensor x_381_axes_0 = const()[name = string("x_381_axes_0"), val = tensor([-1])]; + tensor encoder_layers_14_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_14_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(295865408)))]; + tensor encoder_layers_14_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_14_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(295867520)))]; + tensor x_381_cast_fp16 = layer_norm(axes = x_381_axes_0, beta = encoder_layers_14_norm_conv_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_14_norm_conv_weight_to_fp16, x = input_775_cast_fp16)[name = string("x_381_cast_fp16")]; + tensor input_777_perm_0 = const()[name = string("input_777_perm_0"), val = tensor([0, 2, 1])]; + string input_779_pad_type_0 = const()[name = string("input_779_pad_type_0"), val = string("valid")]; + tensor input_779_strides_0 = const()[name = string("input_779_strides_0"), val = tensor([1])]; + tensor input_779_pad_0 = const()[name = string("input_779_pad_0"), val = tensor([0, 0])]; + tensor input_779_dilations_0 = const()[name = string("input_779_dilations_0"), val = tensor([1])]; + int32 input_779_groups_0 = const()[name = string("input_779_groups_0"), val = int32(1)]; + tensor encoder_layers_14_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(295869632))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297966848))))[name = string("encoder_layers_14_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_777_cast_fp16 = transpose(perm = input_777_perm_0, x = x_381_cast_fp16)[name = string("transpose_231")]; + tensor input_779_cast_fp16 = conv(dilations = input_779_dilations_0, groups = input_779_groups_0, pad = input_779_pad_0, pad_type = input_779_pad_type_0, strides = input_779_strides_0, weight = encoder_layers_14_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_777_cast_fp16)[name = string("input_779_cast_fp16")]; + int32 x_383_split_num_splits_0 = const()[name = string("x_383_split_num_splits_0"), val = int32(2)]; + int32 x_383_split_axis_0 = const()[name = string("x_383_split_axis_0"), val = int32(1)]; + tensor x_383_split_cast_fp16_0, tensor x_383_split_cast_fp16_1 = split(axis = x_383_split_axis_0, num_splits = x_383_split_num_splits_0, x = input_779_cast_fp16)[name = string("x_383_split_cast_fp16")]; + tensor x_383_split_1_sigmoid_cast_fp16 = sigmoid(x = x_383_split_cast_fp16_1)[name = string("x_383_split_1_sigmoid_cast_fp16")]; + tensor x_383_cast_fp16 = mul(x = x_383_split_cast_fp16_0, y = x_383_split_1_sigmoid_cast_fp16)[name = string("x_383_cast_fp16")]; + tensor input_781_cast_fp16 = select(a = var_43_to_fp16, b = x_383_cast_fp16, cond = var_574)[name = string("input_781_cast_fp16")]; + bool new_x_59_interleave_0 = const()[name = string("new_x_59_interleave_0"), val = bool(false)]; + tensor new_x_59_cast_fp16 = concat(axis = var_58, interleave = new_x_59_interleave_0, values = (cache_59_cast_fp16, input_781_cast_fp16))[name = string("new_x_59_cast_fp16")]; + tensor var_3569_begin_0 = const()[name = string("op_3569_begin_0"), val = tensor([0, 0, 56])]; + tensor var_3569_end_0 = const()[name = string("op_3569_end_0"), val = tensor([1, 1024, 64])]; + tensor var_3569_end_mask_0 = const()[name = string("op_3569_end_mask_0"), val = tensor([true, true, true])]; + tensor var_3569_cast_fp16 = slice_by_index(begin = var_3569_begin_0, end = var_3569_end_0, end_mask = var_3569_end_mask_0, x = new_x_59_cast_fp16)[name = string("op_3569_cast_fp16")]; + string x_385_pad_type_0 = const()[name = string("x_385_pad_type_0"), val = string("valid")]; + int32 x_385_groups_0 = const()[name = string("x_385_groups_0"), val = int32(1024)]; + tensor x_385_strides_0 = const()[name = string("x_385_strides_0"), val = tensor([1])]; + tensor x_385_pad_0 = const()[name = string("x_385_pad_0"), val = tensor([0, 0])]; + tensor x_385_dilations_0 = const()[name = string("x_385_dilations_0"), val = tensor([1])]; + tensor encoder_layers_14_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297971008))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297980288))))[name = string("encoder_layers_14_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_385_cast_fp16 = conv(dilations = x_385_dilations_0, groups = x_385_groups_0, pad = x_385_pad_0, pad_type = x_385_pad_type_0, strides = x_385_strides_0, weight = encoder_layers_14_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_59_cast_fp16)[name = string("x_385_cast_fp16")]; + tensor input_783_perm_0 = const()[name = string("input_783_perm_0"), val = tensor([0, 2, 1])]; + tensor x_387_axes_0 = const()[name = string("x_387_axes_0"), val = tensor([-1])]; + tensor encoder_layers_14_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_14_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297982400)))]; + tensor encoder_layers_14_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_14_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297984512)))]; + tensor input_783_cast_fp16 = transpose(perm = input_783_perm_0, x = x_385_cast_fp16)[name = string("transpose_230")]; + tensor x_387_cast_fp16 = layer_norm(axes = x_387_axes_0, beta = encoder_layers_14_conv_batch_norm_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_14_conv_batch_norm_weight_to_fp16, x = input_783_cast_fp16)[name = string("x_387_cast_fp16")]; + tensor input_785_perm_0 = const()[name = string("input_785_perm_0"), val = tensor([0, 2, 1])]; + tensor input_785_cast_fp16 = transpose(perm = input_785_perm_0, x = x_387_cast_fp16)[name = string("transpose_229")]; + tensor input_787_cast_fp16 = silu(x = input_785_cast_fp16)[name = string("input_787_cast_fp16")]; + string x_389_pad_type_0 = const()[name = string("x_389_pad_type_0"), val = string("valid")]; + tensor x_389_strides_0 = const()[name = string("x_389_strides_0"), val = tensor([1])]; + tensor x_389_pad_0 = const()[name = string("x_389_pad_0"), val = tensor([0, 0])]; + tensor x_389_dilations_0 = const()[name = string("x_389_dilations_0"), val = tensor([1])]; + int32 x_389_groups_0 = const()[name = string("x_389_groups_0"), val = int32(1)]; + tensor encoder_layers_14_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297986624))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(299035264))))[name = string("encoder_layers_14_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_389_cast_fp16 = conv(dilations = x_389_dilations_0, groups = x_389_groups_0, pad = x_389_pad_0, pad_type = x_389_pad_type_0, strides = x_389_strides_0, weight = encoder_layers_14_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_787_cast_fp16)[name = string("x_389_cast_fp16")]; + tensor input_789_perm_0 = const()[name = string("input_789_perm_0"), val = tensor([0, 2, 1])]; + tensor input_789_cast_fp16 = transpose(perm = input_789_perm_0, x = x_389_cast_fp16)[name = string("transpose_228")]; + tensor input_791_cast_fp16 = add(x = input_775_cast_fp16, y = input_789_cast_fp16)[name = string("input_791_cast_fp16")]; + tensor input_793_axes_0 = const()[name = string("input_793_axes_0"), val = tensor([-1])]; + tensor encoder_layers_14_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_14_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(299037376)))]; + tensor encoder_layers_14_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_14_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(299039488)))]; + tensor input_793_cast_fp16 = layer_norm(axes = input_793_axes_0, beta = encoder_layers_14_norm_feed_forward2_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_14_norm_feed_forward2_weight_to_fp16, x = input_791_cast_fp16)[name = string("input_793_cast_fp16")]; + tensor encoder_layers_14_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(299041600))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(302187392))))[name = string("encoder_layers_14_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_14_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_14_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(302187584)))]; + tensor linear_134_cast_fp16 = linear(bias = encoder_layers_14_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_14_feed_forward2_linear1_weight_to_fp16_palettized, x = input_793_cast_fp16)[name = string("linear_134_cast_fp16")]; + tensor input_797_cast_fp16 = silu(x = linear_134_cast_fp16)[name = string("input_797_cast_fp16")]; + tensor encoder_layers_14_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(302195840))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(305341632))))[name = string("encoder_layers_14_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_14_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_14_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(305341824)))]; + tensor linear_135_cast_fp16 = linear(bias = encoder_layers_14_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_14_feed_forward2_linear2_weight_to_fp16_palettized, x = input_797_cast_fp16)[name = string("linear_135_cast_fp16")]; + fp16 var_3612_to_fp16 = const()[name = string("op_3612_to_fp16"), val = fp16(0x1p-1)]; + tensor var_3613_cast_fp16 = mul(x = linear_135_cast_fp16, y = var_3612_to_fp16)[name = string("op_3613_cast_fp16")]; + tensor input_803_cast_fp16 = add(x = input_791_cast_fp16, y = var_3613_cast_fp16)[name = string("input_803_cast_fp16")]; + tensor input_805_axes_0 = const()[name = string("input_805_axes_0"), val = tensor([-1])]; + tensor encoder_layers_14_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_14_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(305343936)))]; + tensor encoder_layers_14_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_14_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(305346048)))]; + tensor input_805_cast_fp16 = layer_norm(axes = input_805_axes_0, beta = encoder_layers_14_norm_out_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_14_norm_out_weight_to_fp16, x = input_803_cast_fp16)[name = string("input_805_cast_fp16")]; + tensor cache_61_begin_0 = const()[name = string("cache_61_begin_0"), val = tensor([15, 0, 0, 0])]; + tensor cache_61_end_0 = const()[name = string("cache_61_end_0"), val = tensor([16, 1, 42, 1024])]; + tensor cache_61_end_mask_0 = const()[name = string("cache_61_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_61_squeeze_mask_0 = const()[name = string("cache_61_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_61_cast_fp16 = slice_by_index(begin = cache_61_begin_0, end = cache_61_end_0, end_mask = cache_61_end_mask_0, squeeze_mask = cache_61_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_61_cast_fp16")]; + tensor cache_63_begin_0 = const()[name = string("cache_63_begin_0"), val = tensor([15, 0, 0, 0])]; + tensor cache_63_end_0 = const()[name = string("cache_63_end_0"), val = tensor([16, 1, 1024, 8])]; + tensor cache_63_end_mask_0 = const()[name = string("cache_63_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_63_squeeze_mask_0 = const()[name = string("cache_63_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_63_cast_fp16 = slice_by_index(begin = cache_63_begin_0, end = cache_63_end_0, end_mask = cache_63_end_mask_0, squeeze_mask = cache_63_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_63_cast_fp16")]; + tensor input_807_axes_0 = const()[name = string("input_807_axes_0"), val = tensor([-1])]; + tensor encoder_layers_15_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_15_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(305348160)))]; + tensor encoder_layers_15_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_15_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(305350272)))]; + tensor input_807_cast_fp16 = layer_norm(axes = input_807_axes_0, beta = encoder_layers_15_norm_feed_forward1_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_15_norm_feed_forward1_weight_to_fp16, x = input_805_cast_fp16)[name = string("input_807_cast_fp16")]; + tensor encoder_layers_15_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(305352384))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308498176))))[name = string("encoder_layers_15_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_15_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_15_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308498368)))]; + tensor linear_136_cast_fp16 = linear(bias = encoder_layers_15_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_15_feed_forward1_linear1_weight_to_fp16_palettized, x = input_807_cast_fp16)[name = string("linear_136_cast_fp16")]; + tensor input_811_cast_fp16 = silu(x = linear_136_cast_fp16)[name = string("input_811_cast_fp16")]; + tensor encoder_layers_15_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308506624))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311652416))))[name = string("encoder_layers_15_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_15_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_15_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311652608)))]; + tensor linear_137_cast_fp16 = linear(bias = encoder_layers_15_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_15_feed_forward1_linear2_weight_to_fp16_palettized, x = input_811_cast_fp16)[name = string("linear_137_cast_fp16")]; + fp16 var_3649_to_fp16 = const()[name = string("op_3649_to_fp16"), val = fp16(0x1p-1)]; + tensor var_3650_cast_fp16 = mul(x = linear_137_cast_fp16, y = var_3649_to_fp16)[name = string("op_3650_cast_fp16")]; + tensor input_817_cast_fp16 = add(x = input_805_cast_fp16, y = var_3650_cast_fp16)[name = string("input_817_cast_fp16")]; + tensor key_31_axes_0 = const()[name = string("key_31_axes_0"), val = tensor([-1])]; + tensor encoder_layers_15_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_15_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311654720)))]; + tensor encoder_layers_15_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_15_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311656832)))]; + tensor key_31_cast_fp16 = layer_norm(axes = key_31_axes_0, beta = encoder_layers_15_norm_self_att_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_15_norm_self_att_weight_to_fp16, x = input_817_cast_fp16)[name = string("key_31_cast_fp16")]; + bool input_819_interleave_0 = const()[name = string("input_819_interleave_0"), val = bool(false)]; + tensor input_819_cast_fp16 = concat(axis = var_67, interleave = input_819_interleave_0, values = (cache_61_cast_fp16, key_31_cast_fp16))[name = string("input_819_cast_fp16")]; + bool var_3678_interleave_0 = const()[name = string("op_3678_interleave_0"), val = bool(false)]; + tensor var_3678_cast_fp16 = concat(axis = var_67, interleave = var_3678_interleave_0, values = key_31_cast_fp16)[name = string("op_3678_cast_fp16")]; + tensor encoder_layers_15_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311658944))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(312445440))))[name = string("encoder_layers_15_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_15_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_15_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(312445632)))]; + tensor linear_138_cast_fp16 = linear(bias = encoder_layers_15_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_15_self_attn_linear_q_weight_to_fp16_palettized, x = key_31_cast_fp16)[name = string("linear_138_cast_fp16")]; + tensor var_3683 = const()[name = string("op_3683"), val = tensor([1, -1, 8, 128])]; + tensor q_91_cast_fp16 = reshape(shape = var_3683, x = linear_138_cast_fp16)[name = string("q_91_cast_fp16")]; + tensor encoder_layers_15_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(312447744))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(313234240))))[name = string("encoder_layers_15_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_15_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_15_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(313234432)))]; + tensor linear_139_cast_fp16 = linear(bias = encoder_layers_15_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_15_self_attn_linear_k_weight_to_fp16_palettized, x = input_819_cast_fp16)[name = string("linear_139_cast_fp16")]; + tensor var_3688 = const()[name = string("op_3688"), val = tensor([1, -1, 8, 128])]; + tensor k_61_cast_fp16 = reshape(shape = var_3688, x = linear_139_cast_fp16)[name = string("k_61_cast_fp16")]; + tensor encoder_layers_15_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(313236544))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(314023040))))[name = string("encoder_layers_15_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_15_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_15_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(314023232)))]; + tensor linear_140_cast_fp16 = linear(bias = encoder_layers_15_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_15_self_attn_linear_v_weight_to_fp16_palettized, x = input_819_cast_fp16)[name = string("linear_140_cast_fp16")]; + tensor var_3693 = const()[name = string("op_3693"), val = tensor([1, -1, 8, 128])]; + tensor v_31_cast_fp16 = reshape(shape = var_3693, x = linear_140_cast_fp16)[name = string("v_31_cast_fp16")]; + tensor value_39_perm_0 = const()[name = string("value_39_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_15_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_15_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(314025344)))]; + tensor var_3706_cast_fp16 = add(x = q_91_cast_fp16, y = encoder_layers_15_self_attn_pos_bias_u_to_fp16)[name = string("op_3706_cast_fp16")]; + tensor encoder_layers_15_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_15_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(314027456)))]; + tensor var_3708_cast_fp16 = add(x = q_91_cast_fp16, y = encoder_layers_15_self_attn_pos_bias_v_to_fp16)[name = string("op_3708_cast_fp16")]; + tensor q_with_bias_v_31_perm_0 = const()[name = string("q_with_bias_v_31_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_397_transpose_x_0 = const()[name = string("x_397_transpose_x_0"), val = bool(false)]; + bool x_397_transpose_y_0 = const()[name = string("x_397_transpose_y_0"), val = bool(false)]; + tensor op_3710_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(314029568))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(314229312))))[name = string("op_3710_to_fp16_quantized")]; + tensor q_with_bias_v_31_cast_fp16 = transpose(perm = q_with_bias_v_31_perm_0, x = var_3708_cast_fp16)[name = string("transpose_227")]; + tensor x_397_cast_fp16 = matmul(transpose_x = x_397_transpose_x_0, transpose_y = x_397_transpose_y_0, x = q_with_bias_v_31_cast_fp16, y = op_3710_to_fp16_quantized)[name = string("x_397_cast_fp16")]; + tensor x_399_pad_0 = const()[name = string("x_399_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_399_mode_0 = const()[name = string("x_399_mode_0"), val = string("constant")]; + fp16 const_274_to_fp16 = const()[name = string("const_274_to_fp16"), val = fp16(0x0p+0)]; + tensor x_399_cast_fp16 = pad(constant_val = const_274_to_fp16, mode = x_399_mode_0, pad = x_399_pad_0, x = x_397_cast_fp16)[name = string("x_399_cast_fp16")]; + tensor var_3718 = const()[name = string("op_3718"), val = tensor([1, 8, -1, 56])]; + tensor x_401_cast_fp16 = reshape(shape = var_3718, x = x_399_cast_fp16)[name = string("x_401_cast_fp16")]; + tensor var_3722_begin_0 = const()[name = string("op_3722_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_3722_end_0 = const()[name = string("op_3722_end_0"), val = tensor([1, 8, 196, 56])]; + tensor var_3722_end_mask_0 = const()[name = string("op_3722_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_3722_cast_fp16 = slice_by_index(begin = var_3722_begin_0, end = var_3722_end_0, end_mask = var_3722_end_mask_0, x = x_401_cast_fp16)[name = string("op_3722_cast_fp16")]; + tensor var_3723 = const()[name = string("op_3723"), val = tensor([1, 8, 56, 195])]; + tensor matrix_bd_61_cast_fp16 = reshape(shape = var_3723, x = var_3722_cast_fp16)[name = string("matrix_bd_61_cast_fp16")]; + bool matrix_ac_31_transpose_x_0 = const()[name = string("matrix_ac_31_transpose_x_0"), val = bool(false)]; + bool matrix_ac_31_transpose_y_0 = const()[name = string("matrix_ac_31_transpose_y_0"), val = bool(false)]; + tensor transpose_126_perm_0 = const()[name = string("transpose_126_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_127_perm_0 = const()[name = string("transpose_127_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_127 = transpose(perm = transpose_127_perm_0, x = k_61_cast_fp16)[name = string("transpose_225")]; + tensor transpose_126 = transpose(perm = transpose_126_perm_0, x = var_3706_cast_fp16)[name = string("transpose_226")]; + tensor matrix_ac_31_cast_fp16 = matmul(transpose_x = matrix_ac_31_transpose_x_0, transpose_y = matrix_ac_31_transpose_y_0, x = transpose_126, y = transpose_127)[name = string("matrix_ac_31_cast_fp16")]; + tensor matrix_bd_63_begin_0 = const()[name = string("matrix_bd_63_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_63_end_0 = const()[name = string("matrix_bd_63_end_0"), val = tensor([1, 8, 56, 98])]; + tensor matrix_bd_63_end_mask_0 = const()[name = string("matrix_bd_63_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_63_cast_fp16 = slice_by_index(begin = matrix_bd_63_begin_0, end = matrix_bd_63_end_0, end_mask = matrix_bd_63_end_mask_0, x = matrix_bd_61_cast_fp16)[name = string("matrix_bd_63_cast_fp16")]; + tensor var_3732_cast_fp16 = add(x = matrix_ac_31_cast_fp16, y = matrix_bd_63_cast_fp16)[name = string("op_3732_cast_fp16")]; + fp16 _inversed_scores_61_y_0_to_fp16 = const()[name = string("_inversed_scores_61_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_61_cast_fp16 = mul(x = var_3732_cast_fp16, y = _inversed_scores_61_y_0_to_fp16)[name = string("_inversed_scores_61_cast_fp16")]; + tensor scores_63_cast_fp16 = select(a = var_44_to_fp16, b = _inversed_scores_61_cast_fp16, cond = mask_11)[name = string("scores_63_cast_fp16")]; + tensor var_3738_cast_fp16 = softmax(axis = var_58, x = scores_63_cast_fp16)[name = string("op_3738_cast_fp16")]; + tensor input_821_cast_fp16 = select(a = var_43_to_fp16, b = var_3738_cast_fp16, cond = mask_11)[name = string("input_821_cast_fp16")]; + bool x_403_transpose_x_0 = const()[name = string("x_403_transpose_x_0"), val = bool(false)]; + bool x_403_transpose_y_0 = const()[name = string("x_403_transpose_y_0"), val = bool(false)]; + tensor value_39_cast_fp16 = transpose(perm = value_39_perm_0, x = v_31_cast_fp16)[name = string("transpose_224")]; + tensor x_403_cast_fp16 = matmul(transpose_x = x_403_transpose_x_0, transpose_y = x_403_transpose_y_0, x = input_821_cast_fp16, y = value_39_cast_fp16)[name = string("x_403_cast_fp16")]; + tensor var_3742_perm_0 = const()[name = string("op_3742_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_3743 = const()[name = string("op_3743"), val = tensor([1, -1, 1024])]; + tensor var_3742_cast_fp16 = transpose(perm = var_3742_perm_0, x = x_403_cast_fp16)[name = string("transpose_223")]; + tensor input_823_cast_fp16 = reshape(shape = var_3743, x = var_3742_cast_fp16)[name = string("input_823_cast_fp16")]; + tensor encoder_layers_15_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(314229824))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315016320))))[name = string("encoder_layers_15_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_15_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_15_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315016512)))]; + tensor linear_142_cast_fp16 = linear(bias = encoder_layers_15_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_15_self_attn_linear_out_weight_to_fp16_palettized, x = input_823_cast_fp16)[name = string("linear_142_cast_fp16")]; + tensor input_827_cast_fp16 = add(x = input_817_cast_fp16, y = linear_142_cast_fp16)[name = string("input_827_cast_fp16")]; + tensor x_407_axes_0 = const()[name = string("x_407_axes_0"), val = tensor([-1])]; + tensor encoder_layers_15_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_15_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315018624)))]; + tensor encoder_layers_15_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_15_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315020736)))]; + tensor x_407_cast_fp16 = layer_norm(axes = x_407_axes_0, beta = encoder_layers_15_norm_conv_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_15_norm_conv_weight_to_fp16, x = input_827_cast_fp16)[name = string("x_407_cast_fp16")]; + tensor input_829_perm_0 = const()[name = string("input_829_perm_0"), val = tensor([0, 2, 1])]; + string input_831_pad_type_0 = const()[name = string("input_831_pad_type_0"), val = string("valid")]; + tensor input_831_strides_0 = const()[name = string("input_831_strides_0"), val = tensor([1])]; + tensor input_831_pad_0 = const()[name = string("input_831_pad_0"), val = tensor([0, 0])]; + tensor input_831_dilations_0 = const()[name = string("input_831_dilations_0"), val = tensor([1])]; + int32 input_831_groups_0 = const()[name = string("input_831_groups_0"), val = int32(1)]; + tensor encoder_layers_15_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315022848))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(317120064))))[name = string("encoder_layers_15_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_829_cast_fp16 = transpose(perm = input_829_perm_0, x = x_407_cast_fp16)[name = string("transpose_222")]; + tensor input_831_cast_fp16 = conv(dilations = input_831_dilations_0, groups = input_831_groups_0, pad = input_831_pad_0, pad_type = input_831_pad_type_0, strides = input_831_strides_0, weight = encoder_layers_15_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_829_cast_fp16)[name = string("input_831_cast_fp16")]; + int32 x_409_split_num_splits_0 = const()[name = string("x_409_split_num_splits_0"), val = int32(2)]; + int32 x_409_split_axis_0 = const()[name = string("x_409_split_axis_0"), val = int32(1)]; + tensor x_409_split_cast_fp16_0, tensor x_409_split_cast_fp16_1 = split(axis = x_409_split_axis_0, num_splits = x_409_split_num_splits_0, x = input_831_cast_fp16)[name = string("x_409_split_cast_fp16")]; + tensor x_409_split_1_sigmoid_cast_fp16 = sigmoid(x = x_409_split_cast_fp16_1)[name = string("x_409_split_1_sigmoid_cast_fp16")]; + tensor x_409_cast_fp16 = mul(x = x_409_split_cast_fp16_0, y = x_409_split_1_sigmoid_cast_fp16)[name = string("x_409_cast_fp16")]; + tensor input_833_cast_fp16 = select(a = var_43_to_fp16, b = x_409_cast_fp16, cond = var_574)[name = string("input_833_cast_fp16")]; + bool new_x_63_interleave_0 = const()[name = string("new_x_63_interleave_0"), val = bool(false)]; + tensor new_x_63_cast_fp16 = concat(axis = var_58, interleave = new_x_63_interleave_0, values = (cache_63_cast_fp16, input_833_cast_fp16))[name = string("new_x_63_cast_fp16")]; + tensor var_3782_begin_0 = const()[name = string("op_3782_begin_0"), val = tensor([0, 0, 56])]; + tensor var_3782_end_0 = const()[name = string("op_3782_end_0"), val = tensor([1, 1024, 64])]; + tensor var_3782_end_mask_0 = const()[name = string("op_3782_end_mask_0"), val = tensor([true, true, true])]; + tensor var_3782_cast_fp16 = slice_by_index(begin = var_3782_begin_0, end = var_3782_end_0, end_mask = var_3782_end_mask_0, x = new_x_63_cast_fp16)[name = string("op_3782_cast_fp16")]; + string x_411_pad_type_0 = const()[name = string("x_411_pad_type_0"), val = string("valid")]; + int32 x_411_groups_0 = const()[name = string("x_411_groups_0"), val = int32(1024)]; + tensor x_411_strides_0 = const()[name = string("x_411_strides_0"), val = tensor([1])]; + tensor x_411_pad_0 = const()[name = string("x_411_pad_0"), val = tensor([0, 0])]; + tensor x_411_dilations_0 = const()[name = string("x_411_dilations_0"), val = tensor([1])]; + tensor encoder_layers_15_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(317124224))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(317133504))))[name = string("encoder_layers_15_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_411_cast_fp16 = conv(dilations = x_411_dilations_0, groups = x_411_groups_0, pad = x_411_pad_0, pad_type = x_411_pad_type_0, strides = x_411_strides_0, weight = encoder_layers_15_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_63_cast_fp16)[name = string("x_411_cast_fp16")]; + tensor input_835_perm_0 = const()[name = string("input_835_perm_0"), val = tensor([0, 2, 1])]; + tensor x_413_axes_0 = const()[name = string("x_413_axes_0"), val = tensor([-1])]; + tensor encoder_layers_15_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_15_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(317135616)))]; + tensor encoder_layers_15_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_15_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(317137728)))]; + tensor input_835_cast_fp16 = transpose(perm = input_835_perm_0, x = x_411_cast_fp16)[name = string("transpose_221")]; + tensor x_413_cast_fp16 = layer_norm(axes = x_413_axes_0, beta = encoder_layers_15_conv_batch_norm_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_15_conv_batch_norm_weight_to_fp16, x = input_835_cast_fp16)[name = string("x_413_cast_fp16")]; + tensor input_837_perm_0 = const()[name = string("input_837_perm_0"), val = tensor([0, 2, 1])]; + tensor input_837_cast_fp16 = transpose(perm = input_837_perm_0, x = x_413_cast_fp16)[name = string("transpose_220")]; + tensor input_839_cast_fp16 = silu(x = input_837_cast_fp16)[name = string("input_839_cast_fp16")]; + string x_415_pad_type_0 = const()[name = string("x_415_pad_type_0"), val = string("valid")]; + tensor x_415_strides_0 = const()[name = string("x_415_strides_0"), val = tensor([1])]; + tensor x_415_pad_0 = const()[name = string("x_415_pad_0"), val = tensor([0, 0])]; + tensor x_415_dilations_0 = const()[name = string("x_415_dilations_0"), val = tensor([1])]; + int32 x_415_groups_0 = const()[name = string("x_415_groups_0"), val = int32(1)]; + tensor encoder_layers_15_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(317139840))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(318188480))))[name = string("encoder_layers_15_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_415_cast_fp16 = conv(dilations = x_415_dilations_0, groups = x_415_groups_0, pad = x_415_pad_0, pad_type = x_415_pad_type_0, strides = x_415_strides_0, weight = encoder_layers_15_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_839_cast_fp16)[name = string("x_415_cast_fp16")]; + tensor input_841_perm_0 = const()[name = string("input_841_perm_0"), val = tensor([0, 2, 1])]; + tensor input_841_cast_fp16 = transpose(perm = input_841_perm_0, x = x_415_cast_fp16)[name = string("transpose_219")]; + tensor input_843_cast_fp16 = add(x = input_827_cast_fp16, y = input_841_cast_fp16)[name = string("input_843_cast_fp16")]; + tensor input_845_axes_0 = const()[name = string("input_845_axes_0"), val = tensor([-1])]; + tensor encoder_layers_15_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_15_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(318190592)))]; + tensor encoder_layers_15_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_15_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(318192704)))]; + tensor input_845_cast_fp16 = layer_norm(axes = input_845_axes_0, beta = encoder_layers_15_norm_feed_forward2_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_15_norm_feed_forward2_weight_to_fp16, x = input_843_cast_fp16)[name = string("input_845_cast_fp16")]; + tensor encoder_layers_15_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(318194816))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(321340608))))[name = string("encoder_layers_15_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_15_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_15_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(321340800)))]; + tensor linear_143_cast_fp16 = linear(bias = encoder_layers_15_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_15_feed_forward2_linear1_weight_to_fp16_palettized, x = input_845_cast_fp16)[name = string("linear_143_cast_fp16")]; + tensor input_849_cast_fp16 = silu(x = linear_143_cast_fp16)[name = string("input_849_cast_fp16")]; + tensor encoder_layers_15_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(321349056))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(324494848))))[name = string("encoder_layers_15_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_15_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_15_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(324495040)))]; + tensor linear_144_cast_fp16 = linear(bias = encoder_layers_15_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_15_feed_forward2_linear2_weight_to_fp16_palettized, x = input_849_cast_fp16)[name = string("linear_144_cast_fp16")]; + fp16 var_3825_to_fp16 = const()[name = string("op_3825_to_fp16"), val = fp16(0x1p-1)]; + tensor var_3826_cast_fp16 = mul(x = linear_144_cast_fp16, y = var_3825_to_fp16)[name = string("op_3826_cast_fp16")]; + tensor input_855_cast_fp16 = add(x = input_843_cast_fp16, y = var_3826_cast_fp16)[name = string("input_855_cast_fp16")]; + tensor input_857_axes_0 = const()[name = string("input_857_axes_0"), val = tensor([-1])]; + tensor encoder_layers_15_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_15_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(324497152)))]; + tensor encoder_layers_15_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_15_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(324499264)))]; + tensor input_857_cast_fp16 = layer_norm(axes = input_857_axes_0, beta = encoder_layers_15_norm_out_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_15_norm_out_weight_to_fp16, x = input_855_cast_fp16)[name = string("input_857_cast_fp16")]; + tensor cache_65_begin_0 = const()[name = string("cache_65_begin_0"), val = tensor([16, 0, 0, 0])]; + tensor cache_65_end_0 = const()[name = string("cache_65_end_0"), val = tensor([17, 1, 42, 1024])]; + tensor cache_65_end_mask_0 = const()[name = string("cache_65_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_65_squeeze_mask_0 = const()[name = string("cache_65_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_65_cast_fp16 = slice_by_index(begin = cache_65_begin_0, end = cache_65_end_0, end_mask = cache_65_end_mask_0, squeeze_mask = cache_65_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_65_cast_fp16")]; + tensor cache_67_begin_0 = const()[name = string("cache_67_begin_0"), val = tensor([16, 0, 0, 0])]; + tensor cache_67_end_0 = const()[name = string("cache_67_end_0"), val = tensor([17, 1, 1024, 8])]; + tensor cache_67_end_mask_0 = const()[name = string("cache_67_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_67_squeeze_mask_0 = const()[name = string("cache_67_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_67_cast_fp16 = slice_by_index(begin = cache_67_begin_0, end = cache_67_end_0, end_mask = cache_67_end_mask_0, squeeze_mask = cache_67_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_67_cast_fp16")]; + tensor input_859_axes_0 = const()[name = string("input_859_axes_0"), val = tensor([-1])]; + tensor encoder_layers_16_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_16_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(324501376)))]; + tensor encoder_layers_16_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_16_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(324503488)))]; + tensor input_859_cast_fp16 = layer_norm(axes = input_859_axes_0, beta = encoder_layers_16_norm_feed_forward1_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_16_norm_feed_forward1_weight_to_fp16, x = input_857_cast_fp16)[name = string("input_859_cast_fp16")]; + tensor encoder_layers_16_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(324505600))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(327651392))))[name = string("encoder_layers_16_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_16_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_16_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(327651584)))]; + tensor linear_145_cast_fp16 = linear(bias = encoder_layers_16_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_16_feed_forward1_linear1_weight_to_fp16_palettized, x = input_859_cast_fp16)[name = string("linear_145_cast_fp16")]; + tensor input_863_cast_fp16 = silu(x = linear_145_cast_fp16)[name = string("input_863_cast_fp16")]; + tensor encoder_layers_16_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(327659840))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(330805632))))[name = string("encoder_layers_16_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_16_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_16_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(330805824)))]; + tensor linear_146_cast_fp16 = linear(bias = encoder_layers_16_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_16_feed_forward1_linear2_weight_to_fp16_palettized, x = input_863_cast_fp16)[name = string("linear_146_cast_fp16")]; + fp16 var_3862_to_fp16 = const()[name = string("op_3862_to_fp16"), val = fp16(0x1p-1)]; + tensor var_3863_cast_fp16 = mul(x = linear_146_cast_fp16, y = var_3862_to_fp16)[name = string("op_3863_cast_fp16")]; + tensor input_869_cast_fp16 = add(x = input_857_cast_fp16, y = var_3863_cast_fp16)[name = string("input_869_cast_fp16")]; + tensor key_33_axes_0 = const()[name = string("key_33_axes_0"), val = tensor([-1])]; + tensor encoder_layers_16_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_16_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(330807936)))]; + tensor encoder_layers_16_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_16_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(330810048)))]; + tensor key_33_cast_fp16 = layer_norm(axes = key_33_axes_0, beta = encoder_layers_16_norm_self_att_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_16_norm_self_att_weight_to_fp16, x = input_869_cast_fp16)[name = string("key_33_cast_fp16")]; + bool input_871_interleave_0 = const()[name = string("input_871_interleave_0"), val = bool(false)]; + tensor input_871_cast_fp16 = concat(axis = var_67, interleave = input_871_interleave_0, values = (cache_65_cast_fp16, key_33_cast_fp16))[name = string("input_871_cast_fp16")]; + bool var_3891_interleave_0 = const()[name = string("op_3891_interleave_0"), val = bool(false)]; + tensor var_3891_cast_fp16 = concat(axis = var_67, interleave = var_3891_interleave_0, values = key_33_cast_fp16)[name = string("op_3891_cast_fp16")]; + tensor encoder_layers_16_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(330812160))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331598656))))[name = string("encoder_layers_16_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_16_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_16_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331598848)))]; + tensor linear_147_cast_fp16 = linear(bias = encoder_layers_16_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_16_self_attn_linear_q_weight_to_fp16_palettized, x = key_33_cast_fp16)[name = string("linear_147_cast_fp16")]; + tensor var_3896 = const()[name = string("op_3896"), val = tensor([1, -1, 8, 128])]; + tensor q_97_cast_fp16 = reshape(shape = var_3896, x = linear_147_cast_fp16)[name = string("q_97_cast_fp16")]; + tensor encoder_layers_16_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331600960))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332387456))))[name = string("encoder_layers_16_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_16_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_16_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332387648)))]; + tensor linear_148_cast_fp16 = linear(bias = encoder_layers_16_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_16_self_attn_linear_k_weight_to_fp16_palettized, x = input_871_cast_fp16)[name = string("linear_148_cast_fp16")]; + tensor var_3901 = const()[name = string("op_3901"), val = tensor([1, -1, 8, 128])]; + tensor k_65_cast_fp16 = reshape(shape = var_3901, x = linear_148_cast_fp16)[name = string("k_65_cast_fp16")]; + tensor encoder_layers_16_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332389760))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(333176256))))[name = string("encoder_layers_16_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_16_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_16_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(333176448)))]; + tensor linear_149_cast_fp16 = linear(bias = encoder_layers_16_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_16_self_attn_linear_v_weight_to_fp16_palettized, x = input_871_cast_fp16)[name = string("linear_149_cast_fp16")]; + tensor var_3906 = const()[name = string("op_3906"), val = tensor([1, -1, 8, 128])]; + tensor v_33_cast_fp16 = reshape(shape = var_3906, x = linear_149_cast_fp16)[name = string("v_33_cast_fp16")]; + tensor value_41_perm_0 = const()[name = string("value_41_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_16_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_16_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(333178560)))]; + tensor var_3919_cast_fp16 = add(x = q_97_cast_fp16, y = encoder_layers_16_self_attn_pos_bias_u_to_fp16)[name = string("op_3919_cast_fp16")]; + tensor encoder_layers_16_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_16_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(333180672)))]; + tensor var_3921_cast_fp16 = add(x = q_97_cast_fp16, y = encoder_layers_16_self_attn_pos_bias_v_to_fp16)[name = string("op_3921_cast_fp16")]; + tensor q_with_bias_v_33_perm_0 = const()[name = string("q_with_bias_v_33_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_423_transpose_x_0 = const()[name = string("x_423_transpose_x_0"), val = bool(false)]; + bool x_423_transpose_y_0 = const()[name = string("x_423_transpose_y_0"), val = bool(false)]; + tensor op_3923_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(333182784))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(333382528))))[name = string("op_3923_to_fp16_quantized")]; + tensor q_with_bias_v_33_cast_fp16 = transpose(perm = q_with_bias_v_33_perm_0, x = var_3921_cast_fp16)[name = string("transpose_218")]; + tensor x_423_cast_fp16 = matmul(transpose_x = x_423_transpose_x_0, transpose_y = x_423_transpose_y_0, x = q_with_bias_v_33_cast_fp16, y = op_3923_to_fp16_quantized)[name = string("x_423_cast_fp16")]; + tensor x_425_pad_0 = const()[name = string("x_425_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_425_mode_0 = const()[name = string("x_425_mode_0"), val = string("constant")]; + fp16 const_287_to_fp16 = const()[name = string("const_287_to_fp16"), val = fp16(0x0p+0)]; + tensor x_425_cast_fp16 = pad(constant_val = const_287_to_fp16, mode = x_425_mode_0, pad = x_425_pad_0, x = x_423_cast_fp16)[name = string("x_425_cast_fp16")]; + tensor var_3931 = const()[name = string("op_3931"), val = tensor([1, 8, -1, 56])]; + tensor x_427_cast_fp16 = reshape(shape = var_3931, x = x_425_cast_fp16)[name = string("x_427_cast_fp16")]; + tensor var_3935_begin_0 = const()[name = string("op_3935_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_3935_end_0 = const()[name = string("op_3935_end_0"), val = tensor([1, 8, 196, 56])]; + tensor var_3935_end_mask_0 = const()[name = string("op_3935_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_3935_cast_fp16 = slice_by_index(begin = var_3935_begin_0, end = var_3935_end_0, end_mask = var_3935_end_mask_0, x = x_427_cast_fp16)[name = string("op_3935_cast_fp16")]; + tensor var_3936 = const()[name = string("op_3936"), val = tensor([1, 8, 56, 195])]; + tensor matrix_bd_65_cast_fp16 = reshape(shape = var_3936, x = var_3935_cast_fp16)[name = string("matrix_bd_65_cast_fp16")]; + bool matrix_ac_33_transpose_x_0 = const()[name = string("matrix_ac_33_transpose_x_0"), val = bool(false)]; + bool matrix_ac_33_transpose_y_0 = const()[name = string("matrix_ac_33_transpose_y_0"), val = bool(false)]; + tensor transpose_128_perm_0 = const()[name = string("transpose_128_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_129_perm_0 = const()[name = string("transpose_129_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_129 = transpose(perm = transpose_129_perm_0, x = k_65_cast_fp16)[name = string("transpose_216")]; + tensor transpose_128 = transpose(perm = transpose_128_perm_0, x = var_3919_cast_fp16)[name = string("transpose_217")]; + tensor matrix_ac_33_cast_fp16 = matmul(transpose_x = matrix_ac_33_transpose_x_0, transpose_y = matrix_ac_33_transpose_y_0, x = transpose_128, y = transpose_129)[name = string("matrix_ac_33_cast_fp16")]; + tensor matrix_bd_67_begin_0 = const()[name = string("matrix_bd_67_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_67_end_0 = const()[name = string("matrix_bd_67_end_0"), val = tensor([1, 8, 56, 98])]; + tensor matrix_bd_67_end_mask_0 = const()[name = string("matrix_bd_67_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_67_cast_fp16 = slice_by_index(begin = matrix_bd_67_begin_0, end = matrix_bd_67_end_0, end_mask = matrix_bd_67_end_mask_0, x = matrix_bd_65_cast_fp16)[name = string("matrix_bd_67_cast_fp16")]; + tensor var_3945_cast_fp16 = add(x = matrix_ac_33_cast_fp16, y = matrix_bd_67_cast_fp16)[name = string("op_3945_cast_fp16")]; + fp16 _inversed_scores_65_y_0_to_fp16 = const()[name = string("_inversed_scores_65_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_65_cast_fp16 = mul(x = var_3945_cast_fp16, y = _inversed_scores_65_y_0_to_fp16)[name = string("_inversed_scores_65_cast_fp16")]; + tensor scores_67_cast_fp16 = select(a = var_44_to_fp16, b = _inversed_scores_65_cast_fp16, cond = mask_11)[name = string("scores_67_cast_fp16")]; + tensor var_3951_cast_fp16 = softmax(axis = var_58, x = scores_67_cast_fp16)[name = string("op_3951_cast_fp16")]; + tensor input_873_cast_fp16 = select(a = var_43_to_fp16, b = var_3951_cast_fp16, cond = mask_11)[name = string("input_873_cast_fp16")]; + bool x_429_transpose_x_0 = const()[name = string("x_429_transpose_x_0"), val = bool(false)]; + bool x_429_transpose_y_0 = const()[name = string("x_429_transpose_y_0"), val = bool(false)]; + tensor value_41_cast_fp16 = transpose(perm = value_41_perm_0, x = v_33_cast_fp16)[name = string("transpose_215")]; + tensor x_429_cast_fp16 = matmul(transpose_x = x_429_transpose_x_0, transpose_y = x_429_transpose_y_0, x = input_873_cast_fp16, y = value_41_cast_fp16)[name = string("x_429_cast_fp16")]; + tensor var_3955_perm_0 = const()[name = string("op_3955_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_3956 = const()[name = string("op_3956"), val = tensor([1, -1, 1024])]; + tensor var_3955_cast_fp16 = transpose(perm = var_3955_perm_0, x = x_429_cast_fp16)[name = string("transpose_214")]; + tensor input_875_cast_fp16 = reshape(shape = var_3956, x = var_3955_cast_fp16)[name = string("input_875_cast_fp16")]; + tensor encoder_layers_16_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(333383040))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(334169536))))[name = string("encoder_layers_16_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_16_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_16_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(334169728)))]; + tensor linear_151_cast_fp16 = linear(bias = encoder_layers_16_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_16_self_attn_linear_out_weight_to_fp16_palettized, x = input_875_cast_fp16)[name = string("linear_151_cast_fp16")]; + tensor input_879_cast_fp16 = add(x = input_869_cast_fp16, y = linear_151_cast_fp16)[name = string("input_879_cast_fp16")]; + tensor x_433_axes_0 = const()[name = string("x_433_axes_0"), val = tensor([-1])]; + tensor encoder_layers_16_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_16_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(334171840)))]; + tensor encoder_layers_16_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_16_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(334173952)))]; + tensor x_433_cast_fp16 = layer_norm(axes = x_433_axes_0, beta = encoder_layers_16_norm_conv_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_16_norm_conv_weight_to_fp16, x = input_879_cast_fp16)[name = string("x_433_cast_fp16")]; + tensor input_881_perm_0 = const()[name = string("input_881_perm_0"), val = tensor([0, 2, 1])]; + string input_883_pad_type_0 = const()[name = string("input_883_pad_type_0"), val = string("valid")]; + tensor input_883_strides_0 = const()[name = string("input_883_strides_0"), val = tensor([1])]; + tensor input_883_pad_0 = const()[name = string("input_883_pad_0"), val = tensor([0, 0])]; + tensor input_883_dilations_0 = const()[name = string("input_883_dilations_0"), val = tensor([1])]; + int32 input_883_groups_0 = const()[name = string("input_883_groups_0"), val = int32(1)]; + tensor encoder_layers_16_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(334176064))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(336273280))))[name = string("encoder_layers_16_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_881_cast_fp16 = transpose(perm = input_881_perm_0, x = x_433_cast_fp16)[name = string("transpose_213")]; + tensor input_883_cast_fp16 = conv(dilations = input_883_dilations_0, groups = input_883_groups_0, pad = input_883_pad_0, pad_type = input_883_pad_type_0, strides = input_883_strides_0, weight = encoder_layers_16_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_881_cast_fp16)[name = string("input_883_cast_fp16")]; + int32 x_435_split_num_splits_0 = const()[name = string("x_435_split_num_splits_0"), val = int32(2)]; + int32 x_435_split_axis_0 = const()[name = string("x_435_split_axis_0"), val = int32(1)]; + tensor x_435_split_cast_fp16_0, tensor x_435_split_cast_fp16_1 = split(axis = x_435_split_axis_0, num_splits = x_435_split_num_splits_0, x = input_883_cast_fp16)[name = string("x_435_split_cast_fp16")]; + tensor x_435_split_1_sigmoid_cast_fp16 = sigmoid(x = x_435_split_cast_fp16_1)[name = string("x_435_split_1_sigmoid_cast_fp16")]; + tensor x_435_cast_fp16 = mul(x = x_435_split_cast_fp16_0, y = x_435_split_1_sigmoid_cast_fp16)[name = string("x_435_cast_fp16")]; + tensor input_885_cast_fp16 = select(a = var_43_to_fp16, b = x_435_cast_fp16, cond = var_574)[name = string("input_885_cast_fp16")]; + bool new_x_67_interleave_0 = const()[name = string("new_x_67_interleave_0"), val = bool(false)]; + tensor new_x_67_cast_fp16 = concat(axis = var_58, interleave = new_x_67_interleave_0, values = (cache_67_cast_fp16, input_885_cast_fp16))[name = string("new_x_67_cast_fp16")]; + tensor var_3995_begin_0 = const()[name = string("op_3995_begin_0"), val = tensor([0, 0, 56])]; + tensor var_3995_end_0 = const()[name = string("op_3995_end_0"), val = tensor([1, 1024, 64])]; + tensor var_3995_end_mask_0 = const()[name = string("op_3995_end_mask_0"), val = tensor([true, true, true])]; + tensor var_3995_cast_fp16 = slice_by_index(begin = var_3995_begin_0, end = var_3995_end_0, end_mask = var_3995_end_mask_0, x = new_x_67_cast_fp16)[name = string("op_3995_cast_fp16")]; + string x_437_pad_type_0 = const()[name = string("x_437_pad_type_0"), val = string("valid")]; + int32 x_437_groups_0 = const()[name = string("x_437_groups_0"), val = int32(1024)]; + tensor x_437_strides_0 = const()[name = string("x_437_strides_0"), val = tensor([1])]; + tensor x_437_pad_0 = const()[name = string("x_437_pad_0"), val = tensor([0, 0])]; + tensor x_437_dilations_0 = const()[name = string("x_437_dilations_0"), val = tensor([1])]; + tensor encoder_layers_16_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(336277440))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(336286720))))[name = string("encoder_layers_16_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_437_cast_fp16 = conv(dilations = x_437_dilations_0, groups = x_437_groups_0, pad = x_437_pad_0, pad_type = x_437_pad_type_0, strides = x_437_strides_0, weight = encoder_layers_16_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_67_cast_fp16)[name = string("x_437_cast_fp16")]; + tensor input_887_perm_0 = const()[name = string("input_887_perm_0"), val = tensor([0, 2, 1])]; + tensor x_439_axes_0 = const()[name = string("x_439_axes_0"), val = tensor([-1])]; + tensor encoder_layers_16_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_16_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(336288832)))]; + tensor encoder_layers_16_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_16_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(336290944)))]; + tensor input_887_cast_fp16 = transpose(perm = input_887_perm_0, x = x_437_cast_fp16)[name = string("transpose_212")]; + tensor x_439_cast_fp16 = layer_norm(axes = x_439_axes_0, beta = encoder_layers_16_conv_batch_norm_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_16_conv_batch_norm_weight_to_fp16, x = input_887_cast_fp16)[name = string("x_439_cast_fp16")]; + tensor input_889_perm_0 = const()[name = string("input_889_perm_0"), val = tensor([0, 2, 1])]; + tensor input_889_cast_fp16 = transpose(perm = input_889_perm_0, x = x_439_cast_fp16)[name = string("transpose_211")]; + tensor input_891_cast_fp16 = silu(x = input_889_cast_fp16)[name = string("input_891_cast_fp16")]; + string x_441_pad_type_0 = const()[name = string("x_441_pad_type_0"), val = string("valid")]; + tensor x_441_strides_0 = const()[name = string("x_441_strides_0"), val = tensor([1])]; + tensor x_441_pad_0 = const()[name = string("x_441_pad_0"), val = tensor([0, 0])]; + tensor x_441_dilations_0 = const()[name = string("x_441_dilations_0"), val = tensor([1])]; + int32 x_441_groups_0 = const()[name = string("x_441_groups_0"), val = int32(1)]; + tensor encoder_layers_16_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(336293056))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(337341696))))[name = string("encoder_layers_16_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_441_cast_fp16 = conv(dilations = x_441_dilations_0, groups = x_441_groups_0, pad = x_441_pad_0, pad_type = x_441_pad_type_0, strides = x_441_strides_0, weight = encoder_layers_16_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_891_cast_fp16)[name = string("x_441_cast_fp16")]; + tensor input_893_perm_0 = const()[name = string("input_893_perm_0"), val = tensor([0, 2, 1])]; + tensor input_893_cast_fp16 = transpose(perm = input_893_perm_0, x = x_441_cast_fp16)[name = string("transpose_210")]; + tensor input_895_cast_fp16 = add(x = input_879_cast_fp16, y = input_893_cast_fp16)[name = string("input_895_cast_fp16")]; + tensor input_897_axes_0 = const()[name = string("input_897_axes_0"), val = tensor([-1])]; + tensor encoder_layers_16_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_16_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(337343808)))]; + tensor encoder_layers_16_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_16_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(337345920)))]; + tensor input_897_cast_fp16 = layer_norm(axes = input_897_axes_0, beta = encoder_layers_16_norm_feed_forward2_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_16_norm_feed_forward2_weight_to_fp16, x = input_895_cast_fp16)[name = string("input_897_cast_fp16")]; + tensor encoder_layers_16_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(337348032))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(340493824))))[name = string("encoder_layers_16_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_16_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_16_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(340494016)))]; + tensor linear_152_cast_fp16 = linear(bias = encoder_layers_16_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_16_feed_forward2_linear1_weight_to_fp16_palettized, x = input_897_cast_fp16)[name = string("linear_152_cast_fp16")]; + tensor input_901_cast_fp16 = silu(x = linear_152_cast_fp16)[name = string("input_901_cast_fp16")]; + tensor encoder_layers_16_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(340502272))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(343648064))))[name = string("encoder_layers_16_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_16_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_16_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(343648256)))]; + tensor linear_153_cast_fp16 = linear(bias = encoder_layers_16_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_16_feed_forward2_linear2_weight_to_fp16_palettized, x = input_901_cast_fp16)[name = string("linear_153_cast_fp16")]; + fp16 var_4038_to_fp16 = const()[name = string("op_4038_to_fp16"), val = fp16(0x1p-1)]; + tensor var_4039_cast_fp16 = mul(x = linear_153_cast_fp16, y = var_4038_to_fp16)[name = string("op_4039_cast_fp16")]; + tensor input_907_cast_fp16 = add(x = input_895_cast_fp16, y = var_4039_cast_fp16)[name = string("input_907_cast_fp16")]; + tensor input_909_axes_0 = const()[name = string("input_909_axes_0"), val = tensor([-1])]; + tensor encoder_layers_16_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_16_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(343650368)))]; + tensor encoder_layers_16_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_16_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(343652480)))]; + tensor input_909_cast_fp16 = layer_norm(axes = input_909_axes_0, beta = encoder_layers_16_norm_out_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_16_norm_out_weight_to_fp16, x = input_907_cast_fp16)[name = string("input_909_cast_fp16")]; + tensor cache_69_begin_0 = const()[name = string("cache_69_begin_0"), val = tensor([17, 0, 0, 0])]; + tensor cache_69_end_0 = const()[name = string("cache_69_end_0"), val = tensor([18, 1, 42, 1024])]; + tensor cache_69_end_mask_0 = const()[name = string("cache_69_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_69_squeeze_mask_0 = const()[name = string("cache_69_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_69_cast_fp16 = slice_by_index(begin = cache_69_begin_0, end = cache_69_end_0, end_mask = cache_69_end_mask_0, squeeze_mask = cache_69_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_69_cast_fp16")]; + tensor cache_71_begin_0 = const()[name = string("cache_71_begin_0"), val = tensor([17, 0, 0, 0])]; + tensor cache_71_end_0 = const()[name = string("cache_71_end_0"), val = tensor([18, 1, 1024, 8])]; + tensor cache_71_end_mask_0 = const()[name = string("cache_71_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_71_squeeze_mask_0 = const()[name = string("cache_71_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_71_cast_fp16 = slice_by_index(begin = cache_71_begin_0, end = cache_71_end_0, end_mask = cache_71_end_mask_0, squeeze_mask = cache_71_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_71_cast_fp16")]; + tensor input_911_axes_0 = const()[name = string("input_911_axes_0"), val = tensor([-1])]; + tensor encoder_layers_17_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_17_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(343654592)))]; + tensor encoder_layers_17_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_17_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(343656704)))]; + tensor input_911_cast_fp16 = layer_norm(axes = input_911_axes_0, beta = encoder_layers_17_norm_feed_forward1_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_17_norm_feed_forward1_weight_to_fp16, x = input_909_cast_fp16)[name = string("input_911_cast_fp16")]; + tensor encoder_layers_17_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(343658816))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346804608))))[name = string("encoder_layers_17_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_17_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_17_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346804800)))]; + tensor linear_154_cast_fp16 = linear(bias = encoder_layers_17_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_17_feed_forward1_linear1_weight_to_fp16_palettized, x = input_911_cast_fp16)[name = string("linear_154_cast_fp16")]; + tensor input_915_cast_fp16 = silu(x = linear_154_cast_fp16)[name = string("input_915_cast_fp16")]; + tensor encoder_layers_17_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346813056))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(349958848))))[name = string("encoder_layers_17_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_17_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_17_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(349959040)))]; + tensor linear_155_cast_fp16 = linear(bias = encoder_layers_17_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_17_feed_forward1_linear2_weight_to_fp16_palettized, x = input_915_cast_fp16)[name = string("linear_155_cast_fp16")]; + fp16 var_4075_to_fp16 = const()[name = string("op_4075_to_fp16"), val = fp16(0x1p-1)]; + tensor var_4076_cast_fp16 = mul(x = linear_155_cast_fp16, y = var_4075_to_fp16)[name = string("op_4076_cast_fp16")]; + tensor input_921_cast_fp16 = add(x = input_909_cast_fp16, y = var_4076_cast_fp16)[name = string("input_921_cast_fp16")]; + tensor key_35_axes_0 = const()[name = string("key_35_axes_0"), val = tensor([-1])]; + tensor encoder_layers_17_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_17_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(349961152)))]; + tensor encoder_layers_17_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_17_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(349963264)))]; + tensor key_35_cast_fp16 = layer_norm(axes = key_35_axes_0, beta = encoder_layers_17_norm_self_att_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_17_norm_self_att_weight_to_fp16, x = input_921_cast_fp16)[name = string("key_35_cast_fp16")]; + bool input_923_interleave_0 = const()[name = string("input_923_interleave_0"), val = bool(false)]; + tensor input_923_cast_fp16 = concat(axis = var_67, interleave = input_923_interleave_0, values = (cache_69_cast_fp16, key_35_cast_fp16))[name = string("input_923_cast_fp16")]; + bool var_4104_interleave_0 = const()[name = string("op_4104_interleave_0"), val = bool(false)]; + tensor var_4104_cast_fp16 = concat(axis = var_67, interleave = var_4104_interleave_0, values = key_35_cast_fp16)[name = string("op_4104_cast_fp16")]; + tensor encoder_layers_17_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(349965376))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350751872))))[name = string("encoder_layers_17_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_17_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_17_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350752064)))]; + tensor linear_156_cast_fp16 = linear(bias = encoder_layers_17_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_17_self_attn_linear_q_weight_to_fp16_palettized, x = key_35_cast_fp16)[name = string("linear_156_cast_fp16")]; + tensor var_4109 = const()[name = string("op_4109"), val = tensor([1, -1, 8, 128])]; + tensor q_103_cast_fp16 = reshape(shape = var_4109, x = linear_156_cast_fp16)[name = string("q_103_cast_fp16")]; + tensor encoder_layers_17_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350754176))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351540672))))[name = string("encoder_layers_17_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_17_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_17_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351540864)))]; + tensor linear_157_cast_fp16 = linear(bias = encoder_layers_17_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_17_self_attn_linear_k_weight_to_fp16_palettized, x = input_923_cast_fp16)[name = string("linear_157_cast_fp16")]; + tensor var_4114 = const()[name = string("op_4114"), val = tensor([1, -1, 8, 128])]; + tensor k_69_cast_fp16 = reshape(shape = var_4114, x = linear_157_cast_fp16)[name = string("k_69_cast_fp16")]; + tensor encoder_layers_17_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351542976))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(352329472))))[name = string("encoder_layers_17_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_17_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_17_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(352329664)))]; + tensor linear_158_cast_fp16 = linear(bias = encoder_layers_17_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_17_self_attn_linear_v_weight_to_fp16_palettized, x = input_923_cast_fp16)[name = string("linear_158_cast_fp16")]; + tensor var_4119 = const()[name = string("op_4119"), val = tensor([1, -1, 8, 128])]; + tensor v_35_cast_fp16 = reshape(shape = var_4119, x = linear_158_cast_fp16)[name = string("v_35_cast_fp16")]; + tensor value_43_perm_0 = const()[name = string("value_43_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_17_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_17_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(352331776)))]; + tensor var_4132_cast_fp16 = add(x = q_103_cast_fp16, y = encoder_layers_17_self_attn_pos_bias_u_to_fp16)[name = string("op_4132_cast_fp16")]; + tensor encoder_layers_17_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_17_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(352333888)))]; + tensor var_4134_cast_fp16 = add(x = q_103_cast_fp16, y = encoder_layers_17_self_attn_pos_bias_v_to_fp16)[name = string("op_4134_cast_fp16")]; + tensor q_with_bias_v_35_perm_0 = const()[name = string("q_with_bias_v_35_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_449_transpose_x_0 = const()[name = string("x_449_transpose_x_0"), val = bool(false)]; + bool x_449_transpose_y_0 = const()[name = string("x_449_transpose_y_0"), val = bool(false)]; + tensor op_4136_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(352336000))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(352535744))))[name = string("op_4136_to_fp16_quantized")]; + tensor q_with_bias_v_35_cast_fp16 = transpose(perm = q_with_bias_v_35_perm_0, x = var_4134_cast_fp16)[name = string("transpose_209")]; + tensor x_449_cast_fp16 = matmul(transpose_x = x_449_transpose_x_0, transpose_y = x_449_transpose_y_0, x = q_with_bias_v_35_cast_fp16, y = op_4136_to_fp16_quantized)[name = string("x_449_cast_fp16")]; + tensor x_451_pad_0 = const()[name = string("x_451_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_451_mode_0 = const()[name = string("x_451_mode_0"), val = string("constant")]; + fp16 const_300_to_fp16 = const()[name = string("const_300_to_fp16"), val = fp16(0x0p+0)]; + tensor x_451_cast_fp16 = pad(constant_val = const_300_to_fp16, mode = x_451_mode_0, pad = x_451_pad_0, x = x_449_cast_fp16)[name = string("x_451_cast_fp16")]; + tensor var_4144 = const()[name = string("op_4144"), val = tensor([1, 8, -1, 56])]; + tensor x_453_cast_fp16 = reshape(shape = var_4144, x = x_451_cast_fp16)[name = string("x_453_cast_fp16")]; + tensor var_4148_begin_0 = const()[name = string("op_4148_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_4148_end_0 = const()[name = string("op_4148_end_0"), val = tensor([1, 8, 196, 56])]; + tensor var_4148_end_mask_0 = const()[name = string("op_4148_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_4148_cast_fp16 = slice_by_index(begin = var_4148_begin_0, end = var_4148_end_0, end_mask = var_4148_end_mask_0, x = x_453_cast_fp16)[name = string("op_4148_cast_fp16")]; + tensor var_4149 = const()[name = string("op_4149"), val = tensor([1, 8, 56, 195])]; + tensor matrix_bd_69_cast_fp16 = reshape(shape = var_4149, x = var_4148_cast_fp16)[name = string("matrix_bd_69_cast_fp16")]; + bool matrix_ac_35_transpose_x_0 = const()[name = string("matrix_ac_35_transpose_x_0"), val = bool(false)]; + bool matrix_ac_35_transpose_y_0 = const()[name = string("matrix_ac_35_transpose_y_0"), val = bool(false)]; + tensor transpose_130_perm_0 = const()[name = string("transpose_130_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_131_perm_0 = const()[name = string("transpose_131_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_131 = transpose(perm = transpose_131_perm_0, x = k_69_cast_fp16)[name = string("transpose_207")]; + tensor transpose_130 = transpose(perm = transpose_130_perm_0, x = var_4132_cast_fp16)[name = string("transpose_208")]; + tensor matrix_ac_35_cast_fp16 = matmul(transpose_x = matrix_ac_35_transpose_x_0, transpose_y = matrix_ac_35_transpose_y_0, x = transpose_130, y = transpose_131)[name = string("matrix_ac_35_cast_fp16")]; + tensor matrix_bd_71_begin_0 = const()[name = string("matrix_bd_71_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_71_end_0 = const()[name = string("matrix_bd_71_end_0"), val = tensor([1, 8, 56, 98])]; + tensor matrix_bd_71_end_mask_0 = const()[name = string("matrix_bd_71_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_71_cast_fp16 = slice_by_index(begin = matrix_bd_71_begin_0, end = matrix_bd_71_end_0, end_mask = matrix_bd_71_end_mask_0, x = matrix_bd_69_cast_fp16)[name = string("matrix_bd_71_cast_fp16")]; + tensor var_4158_cast_fp16 = add(x = matrix_ac_35_cast_fp16, y = matrix_bd_71_cast_fp16)[name = string("op_4158_cast_fp16")]; + fp16 _inversed_scores_69_y_0_to_fp16 = const()[name = string("_inversed_scores_69_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_69_cast_fp16 = mul(x = var_4158_cast_fp16, y = _inversed_scores_69_y_0_to_fp16)[name = string("_inversed_scores_69_cast_fp16")]; + tensor scores_71_cast_fp16 = select(a = var_44_to_fp16, b = _inversed_scores_69_cast_fp16, cond = mask_11)[name = string("scores_71_cast_fp16")]; + tensor var_4164_cast_fp16 = softmax(axis = var_58, x = scores_71_cast_fp16)[name = string("op_4164_cast_fp16")]; + tensor input_925_cast_fp16 = select(a = var_43_to_fp16, b = var_4164_cast_fp16, cond = mask_11)[name = string("input_925_cast_fp16")]; + bool x_455_transpose_x_0 = const()[name = string("x_455_transpose_x_0"), val = bool(false)]; + bool x_455_transpose_y_0 = const()[name = string("x_455_transpose_y_0"), val = bool(false)]; + tensor value_43_cast_fp16 = transpose(perm = value_43_perm_0, x = v_35_cast_fp16)[name = string("transpose_206")]; + tensor x_455_cast_fp16 = matmul(transpose_x = x_455_transpose_x_0, transpose_y = x_455_transpose_y_0, x = input_925_cast_fp16, y = value_43_cast_fp16)[name = string("x_455_cast_fp16")]; + tensor var_4168_perm_0 = const()[name = string("op_4168_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_4169 = const()[name = string("op_4169"), val = tensor([1, -1, 1024])]; + tensor var_4168_cast_fp16 = transpose(perm = var_4168_perm_0, x = x_455_cast_fp16)[name = string("transpose_205")]; + tensor input_927_cast_fp16 = reshape(shape = var_4169, x = var_4168_cast_fp16)[name = string("input_927_cast_fp16")]; + tensor encoder_layers_17_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(352536256))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(353322752))))[name = string("encoder_layers_17_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_17_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_17_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(353322944)))]; + tensor linear_160_cast_fp16 = linear(bias = encoder_layers_17_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_17_self_attn_linear_out_weight_to_fp16_palettized, x = input_927_cast_fp16)[name = string("linear_160_cast_fp16")]; + tensor input_931_cast_fp16 = add(x = input_921_cast_fp16, y = linear_160_cast_fp16)[name = string("input_931_cast_fp16")]; + tensor x_459_axes_0 = const()[name = string("x_459_axes_0"), val = tensor([-1])]; + tensor encoder_layers_17_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_17_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(353325056)))]; + tensor encoder_layers_17_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_17_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(353327168)))]; + tensor x_459_cast_fp16 = layer_norm(axes = x_459_axes_0, beta = encoder_layers_17_norm_conv_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_17_norm_conv_weight_to_fp16, x = input_931_cast_fp16)[name = string("x_459_cast_fp16")]; + tensor input_933_perm_0 = const()[name = string("input_933_perm_0"), val = tensor([0, 2, 1])]; + string input_935_pad_type_0 = const()[name = string("input_935_pad_type_0"), val = string("valid")]; + tensor input_935_strides_0 = const()[name = string("input_935_strides_0"), val = tensor([1])]; + tensor input_935_pad_0 = const()[name = string("input_935_pad_0"), val = tensor([0, 0])]; + tensor input_935_dilations_0 = const()[name = string("input_935_dilations_0"), val = tensor([1])]; + int32 input_935_groups_0 = const()[name = string("input_935_groups_0"), val = int32(1)]; + tensor encoder_layers_17_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(353329280))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(355426496))))[name = string("encoder_layers_17_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_933_cast_fp16 = transpose(perm = input_933_perm_0, x = x_459_cast_fp16)[name = string("transpose_204")]; + tensor input_935_cast_fp16 = conv(dilations = input_935_dilations_0, groups = input_935_groups_0, pad = input_935_pad_0, pad_type = input_935_pad_type_0, strides = input_935_strides_0, weight = encoder_layers_17_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_933_cast_fp16)[name = string("input_935_cast_fp16")]; + int32 x_461_split_num_splits_0 = const()[name = string("x_461_split_num_splits_0"), val = int32(2)]; + int32 x_461_split_axis_0 = const()[name = string("x_461_split_axis_0"), val = int32(1)]; + tensor x_461_split_cast_fp16_0, tensor x_461_split_cast_fp16_1 = split(axis = x_461_split_axis_0, num_splits = x_461_split_num_splits_0, x = input_935_cast_fp16)[name = string("x_461_split_cast_fp16")]; + tensor x_461_split_1_sigmoid_cast_fp16 = sigmoid(x = x_461_split_cast_fp16_1)[name = string("x_461_split_1_sigmoid_cast_fp16")]; + tensor x_461_cast_fp16 = mul(x = x_461_split_cast_fp16_0, y = x_461_split_1_sigmoid_cast_fp16)[name = string("x_461_cast_fp16")]; + tensor input_937_cast_fp16 = select(a = var_43_to_fp16, b = x_461_cast_fp16, cond = var_574)[name = string("input_937_cast_fp16")]; + bool new_x_71_interleave_0 = const()[name = string("new_x_71_interleave_0"), val = bool(false)]; + tensor new_x_71_cast_fp16 = concat(axis = var_58, interleave = new_x_71_interleave_0, values = (cache_71_cast_fp16, input_937_cast_fp16))[name = string("new_x_71_cast_fp16")]; + tensor var_4208_begin_0 = const()[name = string("op_4208_begin_0"), val = tensor([0, 0, 56])]; + tensor var_4208_end_0 = const()[name = string("op_4208_end_0"), val = tensor([1, 1024, 64])]; + tensor var_4208_end_mask_0 = const()[name = string("op_4208_end_mask_0"), val = tensor([true, true, true])]; + tensor var_4208_cast_fp16 = slice_by_index(begin = var_4208_begin_0, end = var_4208_end_0, end_mask = var_4208_end_mask_0, x = new_x_71_cast_fp16)[name = string("op_4208_cast_fp16")]; + string x_463_pad_type_0 = const()[name = string("x_463_pad_type_0"), val = string("valid")]; + int32 x_463_groups_0 = const()[name = string("x_463_groups_0"), val = int32(1024)]; + tensor x_463_strides_0 = const()[name = string("x_463_strides_0"), val = tensor([1])]; + tensor x_463_pad_0 = const()[name = string("x_463_pad_0"), val = tensor([0, 0])]; + tensor x_463_dilations_0 = const()[name = string("x_463_dilations_0"), val = tensor([1])]; + tensor encoder_layers_17_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(355430656))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(355439936))))[name = string("encoder_layers_17_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_463_cast_fp16 = conv(dilations = x_463_dilations_0, groups = x_463_groups_0, pad = x_463_pad_0, pad_type = x_463_pad_type_0, strides = x_463_strides_0, weight = encoder_layers_17_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_71_cast_fp16)[name = string("x_463_cast_fp16")]; + tensor input_939_perm_0 = const()[name = string("input_939_perm_0"), val = tensor([0, 2, 1])]; + tensor x_465_axes_0 = const()[name = string("x_465_axes_0"), val = tensor([-1])]; + tensor encoder_layers_17_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_17_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(355442048)))]; + tensor encoder_layers_17_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_17_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(355444160)))]; + tensor input_939_cast_fp16 = transpose(perm = input_939_perm_0, x = x_463_cast_fp16)[name = string("transpose_203")]; + tensor x_465_cast_fp16 = layer_norm(axes = x_465_axes_0, beta = encoder_layers_17_conv_batch_norm_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_17_conv_batch_norm_weight_to_fp16, x = input_939_cast_fp16)[name = string("x_465_cast_fp16")]; + tensor input_941_perm_0 = const()[name = string("input_941_perm_0"), val = tensor([0, 2, 1])]; + tensor input_941_cast_fp16 = transpose(perm = input_941_perm_0, x = x_465_cast_fp16)[name = string("transpose_202")]; + tensor input_943_cast_fp16 = silu(x = input_941_cast_fp16)[name = string("input_943_cast_fp16")]; + string x_467_pad_type_0 = const()[name = string("x_467_pad_type_0"), val = string("valid")]; + tensor x_467_strides_0 = const()[name = string("x_467_strides_0"), val = tensor([1])]; + tensor x_467_pad_0 = const()[name = string("x_467_pad_0"), val = tensor([0, 0])]; + tensor x_467_dilations_0 = const()[name = string("x_467_dilations_0"), val = tensor([1])]; + int32 x_467_groups_0 = const()[name = string("x_467_groups_0"), val = int32(1)]; + tensor encoder_layers_17_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(355446272))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(356494912))))[name = string("encoder_layers_17_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_467_cast_fp16 = conv(dilations = x_467_dilations_0, groups = x_467_groups_0, pad = x_467_pad_0, pad_type = x_467_pad_type_0, strides = x_467_strides_0, weight = encoder_layers_17_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_943_cast_fp16)[name = string("x_467_cast_fp16")]; + tensor input_945_perm_0 = const()[name = string("input_945_perm_0"), val = tensor([0, 2, 1])]; + tensor input_945_cast_fp16 = transpose(perm = input_945_perm_0, x = x_467_cast_fp16)[name = string("transpose_201")]; + tensor input_947_cast_fp16 = add(x = input_931_cast_fp16, y = input_945_cast_fp16)[name = string("input_947_cast_fp16")]; + tensor input_949_axes_0 = const()[name = string("input_949_axes_0"), val = tensor([-1])]; + tensor encoder_layers_17_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_17_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(356497024)))]; + tensor encoder_layers_17_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_17_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(356499136)))]; + tensor input_949_cast_fp16 = layer_norm(axes = input_949_axes_0, beta = encoder_layers_17_norm_feed_forward2_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_17_norm_feed_forward2_weight_to_fp16, x = input_947_cast_fp16)[name = string("input_949_cast_fp16")]; + tensor encoder_layers_17_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(356501248))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(359647040))))[name = string("encoder_layers_17_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_17_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_17_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(359647232)))]; + tensor linear_161_cast_fp16 = linear(bias = encoder_layers_17_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_17_feed_forward2_linear1_weight_to_fp16_palettized, x = input_949_cast_fp16)[name = string("linear_161_cast_fp16")]; + tensor input_953_cast_fp16 = silu(x = linear_161_cast_fp16)[name = string("input_953_cast_fp16")]; + tensor encoder_layers_17_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(359655488))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(362801280))))[name = string("encoder_layers_17_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_17_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_17_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(362801472)))]; + tensor linear_162_cast_fp16 = linear(bias = encoder_layers_17_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_17_feed_forward2_linear2_weight_to_fp16_palettized, x = input_953_cast_fp16)[name = string("linear_162_cast_fp16")]; + fp16 var_4251_to_fp16 = const()[name = string("op_4251_to_fp16"), val = fp16(0x1p-1)]; + tensor var_4252_cast_fp16 = mul(x = linear_162_cast_fp16, y = var_4251_to_fp16)[name = string("op_4252_cast_fp16")]; + tensor input_959_cast_fp16 = add(x = input_947_cast_fp16, y = var_4252_cast_fp16)[name = string("input_959_cast_fp16")]; + tensor input_961_axes_0 = const()[name = string("input_961_axes_0"), val = tensor([-1])]; + tensor encoder_layers_17_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_17_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(362803584)))]; + tensor encoder_layers_17_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_17_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(362805696)))]; + tensor input_961_cast_fp16 = layer_norm(axes = input_961_axes_0, beta = encoder_layers_17_norm_out_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_17_norm_out_weight_to_fp16, x = input_959_cast_fp16)[name = string("input_961_cast_fp16")]; + tensor cache_73_begin_0 = const()[name = string("cache_73_begin_0"), val = tensor([18, 0, 0, 0])]; + tensor cache_73_end_0 = const()[name = string("cache_73_end_0"), val = tensor([19, 1, 42, 1024])]; + tensor cache_73_end_mask_0 = const()[name = string("cache_73_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_73_squeeze_mask_0 = const()[name = string("cache_73_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_73_cast_fp16 = slice_by_index(begin = cache_73_begin_0, end = cache_73_end_0, end_mask = cache_73_end_mask_0, squeeze_mask = cache_73_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_73_cast_fp16")]; + tensor cache_75_begin_0 = const()[name = string("cache_75_begin_0"), val = tensor([18, 0, 0, 0])]; + tensor cache_75_end_0 = const()[name = string("cache_75_end_0"), val = tensor([19, 1, 1024, 8])]; + tensor cache_75_end_mask_0 = const()[name = string("cache_75_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_75_squeeze_mask_0 = const()[name = string("cache_75_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_75_cast_fp16 = slice_by_index(begin = cache_75_begin_0, end = cache_75_end_0, end_mask = cache_75_end_mask_0, squeeze_mask = cache_75_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_75_cast_fp16")]; + tensor input_963_axes_0 = const()[name = string("input_963_axes_0"), val = tensor([-1])]; + tensor encoder_layers_18_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_18_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(362807808)))]; + tensor encoder_layers_18_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_18_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(362809920)))]; + tensor input_963_cast_fp16 = layer_norm(axes = input_963_axes_0, beta = encoder_layers_18_norm_feed_forward1_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_18_norm_feed_forward1_weight_to_fp16, x = input_961_cast_fp16)[name = string("input_963_cast_fp16")]; + tensor encoder_layers_18_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(362812032))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(365957824))))[name = string("encoder_layers_18_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_18_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_18_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(365958016)))]; + tensor linear_163_cast_fp16 = linear(bias = encoder_layers_18_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_18_feed_forward1_linear1_weight_to_fp16_palettized, x = input_963_cast_fp16)[name = string("linear_163_cast_fp16")]; + tensor input_967_cast_fp16 = silu(x = linear_163_cast_fp16)[name = string("input_967_cast_fp16")]; + tensor encoder_layers_18_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(365966272))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369112064))))[name = string("encoder_layers_18_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_18_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_18_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369112256)))]; + tensor linear_164_cast_fp16 = linear(bias = encoder_layers_18_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_18_feed_forward1_linear2_weight_to_fp16_palettized, x = input_967_cast_fp16)[name = string("linear_164_cast_fp16")]; + fp16 var_4288_to_fp16 = const()[name = string("op_4288_to_fp16"), val = fp16(0x1p-1)]; + tensor var_4289_cast_fp16 = mul(x = linear_164_cast_fp16, y = var_4288_to_fp16)[name = string("op_4289_cast_fp16")]; + tensor input_973_cast_fp16 = add(x = input_961_cast_fp16, y = var_4289_cast_fp16)[name = string("input_973_cast_fp16")]; + tensor key_37_axes_0 = const()[name = string("key_37_axes_0"), val = tensor([-1])]; + tensor encoder_layers_18_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_18_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369114368)))]; + tensor encoder_layers_18_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_18_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369116480)))]; + tensor key_37_cast_fp16 = layer_norm(axes = key_37_axes_0, beta = encoder_layers_18_norm_self_att_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_18_norm_self_att_weight_to_fp16, x = input_973_cast_fp16)[name = string("key_37_cast_fp16")]; + bool input_975_interleave_0 = const()[name = string("input_975_interleave_0"), val = bool(false)]; + tensor input_975_cast_fp16 = concat(axis = var_67, interleave = input_975_interleave_0, values = (cache_73_cast_fp16, key_37_cast_fp16))[name = string("input_975_cast_fp16")]; + bool var_4317_interleave_0 = const()[name = string("op_4317_interleave_0"), val = bool(false)]; + tensor var_4317_cast_fp16 = concat(axis = var_67, interleave = var_4317_interleave_0, values = key_37_cast_fp16)[name = string("op_4317_cast_fp16")]; + tensor encoder_layers_18_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369118592))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369905088))))[name = string("encoder_layers_18_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_18_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_18_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369905280)))]; + tensor linear_165_cast_fp16 = linear(bias = encoder_layers_18_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_18_self_attn_linear_q_weight_to_fp16_palettized, x = key_37_cast_fp16)[name = string("linear_165_cast_fp16")]; + tensor var_4322 = const()[name = string("op_4322"), val = tensor([1, -1, 8, 128])]; + tensor q_109_cast_fp16 = reshape(shape = var_4322, x = linear_165_cast_fp16)[name = string("q_109_cast_fp16")]; + tensor encoder_layers_18_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369907392))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(370693888))))[name = string("encoder_layers_18_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_18_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_18_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(370694080)))]; + tensor linear_166_cast_fp16 = linear(bias = encoder_layers_18_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_18_self_attn_linear_k_weight_to_fp16_palettized, x = input_975_cast_fp16)[name = string("linear_166_cast_fp16")]; + tensor var_4327 = const()[name = string("op_4327"), val = tensor([1, -1, 8, 128])]; + tensor k_73_cast_fp16 = reshape(shape = var_4327, x = linear_166_cast_fp16)[name = string("k_73_cast_fp16")]; + tensor encoder_layers_18_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(370696192))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371482688))))[name = string("encoder_layers_18_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_18_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_18_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371482880)))]; + tensor linear_167_cast_fp16 = linear(bias = encoder_layers_18_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_18_self_attn_linear_v_weight_to_fp16_palettized, x = input_975_cast_fp16)[name = string("linear_167_cast_fp16")]; + tensor var_4332 = const()[name = string("op_4332"), val = tensor([1, -1, 8, 128])]; + tensor v_37_cast_fp16 = reshape(shape = var_4332, x = linear_167_cast_fp16)[name = string("v_37_cast_fp16")]; + tensor value_45_perm_0 = const()[name = string("value_45_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_18_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_18_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371484992)))]; + tensor var_4345_cast_fp16 = add(x = q_109_cast_fp16, y = encoder_layers_18_self_attn_pos_bias_u_to_fp16)[name = string("op_4345_cast_fp16")]; + tensor encoder_layers_18_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_18_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371487104)))]; + tensor var_4347_cast_fp16 = add(x = q_109_cast_fp16, y = encoder_layers_18_self_attn_pos_bias_v_to_fp16)[name = string("op_4347_cast_fp16")]; + tensor q_with_bias_v_37_perm_0 = const()[name = string("q_with_bias_v_37_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_475_transpose_x_0 = const()[name = string("x_475_transpose_x_0"), val = bool(false)]; + bool x_475_transpose_y_0 = const()[name = string("x_475_transpose_y_0"), val = bool(false)]; + tensor op_4349_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371489216))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371688960))))[name = string("op_4349_to_fp16_quantized")]; + tensor q_with_bias_v_37_cast_fp16 = transpose(perm = q_with_bias_v_37_perm_0, x = var_4347_cast_fp16)[name = string("transpose_200")]; + tensor x_475_cast_fp16 = matmul(transpose_x = x_475_transpose_x_0, transpose_y = x_475_transpose_y_0, x = q_with_bias_v_37_cast_fp16, y = op_4349_to_fp16_quantized)[name = string("x_475_cast_fp16")]; + tensor x_477_pad_0 = const()[name = string("x_477_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_477_mode_0 = const()[name = string("x_477_mode_0"), val = string("constant")]; + fp16 const_313_to_fp16 = const()[name = string("const_313_to_fp16"), val = fp16(0x0p+0)]; + tensor x_477_cast_fp16 = pad(constant_val = const_313_to_fp16, mode = x_477_mode_0, pad = x_477_pad_0, x = x_475_cast_fp16)[name = string("x_477_cast_fp16")]; + tensor var_4357 = const()[name = string("op_4357"), val = tensor([1, 8, -1, 56])]; + tensor x_479_cast_fp16 = reshape(shape = var_4357, x = x_477_cast_fp16)[name = string("x_479_cast_fp16")]; + tensor var_4361_begin_0 = const()[name = string("op_4361_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_4361_end_0 = const()[name = string("op_4361_end_0"), val = tensor([1, 8, 196, 56])]; + tensor var_4361_end_mask_0 = const()[name = string("op_4361_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_4361_cast_fp16 = slice_by_index(begin = var_4361_begin_0, end = var_4361_end_0, end_mask = var_4361_end_mask_0, x = x_479_cast_fp16)[name = string("op_4361_cast_fp16")]; + tensor var_4362 = const()[name = string("op_4362"), val = tensor([1, 8, 56, 195])]; + tensor matrix_bd_73_cast_fp16 = reshape(shape = var_4362, x = var_4361_cast_fp16)[name = string("matrix_bd_73_cast_fp16")]; + bool matrix_ac_37_transpose_x_0 = const()[name = string("matrix_ac_37_transpose_x_0"), val = bool(false)]; + bool matrix_ac_37_transpose_y_0 = const()[name = string("matrix_ac_37_transpose_y_0"), val = bool(false)]; + tensor transpose_132_perm_0 = const()[name = string("transpose_132_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_133_perm_0 = const()[name = string("transpose_133_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_133 = transpose(perm = transpose_133_perm_0, x = k_73_cast_fp16)[name = string("transpose_198")]; + tensor transpose_132 = transpose(perm = transpose_132_perm_0, x = var_4345_cast_fp16)[name = string("transpose_199")]; + tensor matrix_ac_37_cast_fp16 = matmul(transpose_x = matrix_ac_37_transpose_x_0, transpose_y = matrix_ac_37_transpose_y_0, x = transpose_132, y = transpose_133)[name = string("matrix_ac_37_cast_fp16")]; + tensor matrix_bd_75_begin_0 = const()[name = string("matrix_bd_75_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_75_end_0 = const()[name = string("matrix_bd_75_end_0"), val = tensor([1, 8, 56, 98])]; + tensor matrix_bd_75_end_mask_0 = const()[name = string("matrix_bd_75_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_75_cast_fp16 = slice_by_index(begin = matrix_bd_75_begin_0, end = matrix_bd_75_end_0, end_mask = matrix_bd_75_end_mask_0, x = matrix_bd_73_cast_fp16)[name = string("matrix_bd_75_cast_fp16")]; + tensor var_4371_cast_fp16 = add(x = matrix_ac_37_cast_fp16, y = matrix_bd_75_cast_fp16)[name = string("op_4371_cast_fp16")]; + fp16 _inversed_scores_73_y_0_to_fp16 = const()[name = string("_inversed_scores_73_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_73_cast_fp16 = mul(x = var_4371_cast_fp16, y = _inversed_scores_73_y_0_to_fp16)[name = string("_inversed_scores_73_cast_fp16")]; + tensor scores_75_cast_fp16 = select(a = var_44_to_fp16, b = _inversed_scores_73_cast_fp16, cond = mask_11)[name = string("scores_75_cast_fp16")]; + tensor var_4377_cast_fp16 = softmax(axis = var_58, x = scores_75_cast_fp16)[name = string("op_4377_cast_fp16")]; + tensor input_977_cast_fp16 = select(a = var_43_to_fp16, b = var_4377_cast_fp16, cond = mask_11)[name = string("input_977_cast_fp16")]; + bool x_481_transpose_x_0 = const()[name = string("x_481_transpose_x_0"), val = bool(false)]; + bool x_481_transpose_y_0 = const()[name = string("x_481_transpose_y_0"), val = bool(false)]; + tensor value_45_cast_fp16 = transpose(perm = value_45_perm_0, x = v_37_cast_fp16)[name = string("transpose_197")]; + tensor x_481_cast_fp16 = matmul(transpose_x = x_481_transpose_x_0, transpose_y = x_481_transpose_y_0, x = input_977_cast_fp16, y = value_45_cast_fp16)[name = string("x_481_cast_fp16")]; + tensor var_4381_perm_0 = const()[name = string("op_4381_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_4382 = const()[name = string("op_4382"), val = tensor([1, -1, 1024])]; + tensor var_4381_cast_fp16 = transpose(perm = var_4381_perm_0, x = x_481_cast_fp16)[name = string("transpose_196")]; + tensor input_979_cast_fp16 = reshape(shape = var_4382, x = var_4381_cast_fp16)[name = string("input_979_cast_fp16")]; + tensor encoder_layers_18_self_attn_linear_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371689472))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(372738112))))[name = string("encoder_layers_18_self_attn_linear_out_weight_to_fp16_quantized")]; + tensor encoder_layers_18_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_18_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(372740224)))]; + tensor linear_169_cast_fp16 = linear(bias = encoder_layers_18_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_18_self_attn_linear_out_weight_to_fp16_quantized, x = input_979_cast_fp16)[name = string("linear_169_cast_fp16")]; + tensor input_983_cast_fp16 = add(x = input_973_cast_fp16, y = linear_169_cast_fp16)[name = string("input_983_cast_fp16")]; + tensor x_485_axes_0 = const()[name = string("x_485_axes_0"), val = tensor([-1])]; + tensor encoder_layers_18_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_18_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(372742336)))]; + tensor encoder_layers_18_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_18_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(372744448)))]; + tensor x_485_cast_fp16 = layer_norm(axes = x_485_axes_0, beta = encoder_layers_18_norm_conv_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_18_norm_conv_weight_to_fp16, x = input_983_cast_fp16)[name = string("x_485_cast_fp16")]; + tensor input_985_perm_0 = const()[name = string("input_985_perm_0"), val = tensor([0, 2, 1])]; + string input_987_pad_type_0 = const()[name = string("input_987_pad_type_0"), val = string("valid")]; + tensor input_987_strides_0 = const()[name = string("input_987_strides_0"), val = tensor([1])]; + tensor input_987_pad_0 = const()[name = string("input_987_pad_0"), val = tensor([0, 0])]; + tensor input_987_dilations_0 = const()[name = string("input_987_dilations_0"), val = tensor([1])]; + int32 input_987_groups_0 = const()[name = string("input_987_groups_0"), val = int32(1)]; + tensor encoder_layers_18_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(372746560))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(374843776))))[name = string("encoder_layers_18_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_985_cast_fp16 = transpose(perm = input_985_perm_0, x = x_485_cast_fp16)[name = string("transpose_195")]; + tensor input_987_cast_fp16 = conv(dilations = input_987_dilations_0, groups = input_987_groups_0, pad = input_987_pad_0, pad_type = input_987_pad_type_0, strides = input_987_strides_0, weight = encoder_layers_18_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_985_cast_fp16)[name = string("input_987_cast_fp16")]; + int32 x_487_split_num_splits_0 = const()[name = string("x_487_split_num_splits_0"), val = int32(2)]; + int32 x_487_split_axis_0 = const()[name = string("x_487_split_axis_0"), val = int32(1)]; + tensor x_487_split_cast_fp16_0, tensor x_487_split_cast_fp16_1 = split(axis = x_487_split_axis_0, num_splits = x_487_split_num_splits_0, x = input_987_cast_fp16)[name = string("x_487_split_cast_fp16")]; + tensor x_487_split_1_sigmoid_cast_fp16 = sigmoid(x = x_487_split_cast_fp16_1)[name = string("x_487_split_1_sigmoid_cast_fp16")]; + tensor x_487_cast_fp16 = mul(x = x_487_split_cast_fp16_0, y = x_487_split_1_sigmoid_cast_fp16)[name = string("x_487_cast_fp16")]; + tensor input_989_cast_fp16 = select(a = var_43_to_fp16, b = x_487_cast_fp16, cond = var_574)[name = string("input_989_cast_fp16")]; + bool new_x_75_interleave_0 = const()[name = string("new_x_75_interleave_0"), val = bool(false)]; + tensor new_x_75_cast_fp16 = concat(axis = var_58, interleave = new_x_75_interleave_0, values = (cache_75_cast_fp16, input_989_cast_fp16))[name = string("new_x_75_cast_fp16")]; + tensor var_4421_begin_0 = const()[name = string("op_4421_begin_0"), val = tensor([0, 0, 56])]; + tensor var_4421_end_0 = const()[name = string("op_4421_end_0"), val = tensor([1, 1024, 64])]; + tensor var_4421_end_mask_0 = const()[name = string("op_4421_end_mask_0"), val = tensor([true, true, true])]; + tensor var_4421_cast_fp16 = slice_by_index(begin = var_4421_begin_0, end = var_4421_end_0, end_mask = var_4421_end_mask_0, x = new_x_75_cast_fp16)[name = string("op_4421_cast_fp16")]; + string x_489_pad_type_0 = const()[name = string("x_489_pad_type_0"), val = string("valid")]; + int32 x_489_groups_0 = const()[name = string("x_489_groups_0"), val = int32(1024)]; + tensor x_489_strides_0 = const()[name = string("x_489_strides_0"), val = tensor([1])]; + tensor x_489_pad_0 = const()[name = string("x_489_pad_0"), val = tensor([0, 0])]; + tensor x_489_dilations_0 = const()[name = string("x_489_dilations_0"), val = tensor([1])]; + tensor encoder_layers_18_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(374847936))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(374857216))))[name = string("encoder_layers_18_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_489_cast_fp16 = conv(dilations = x_489_dilations_0, groups = x_489_groups_0, pad = x_489_pad_0, pad_type = x_489_pad_type_0, strides = x_489_strides_0, weight = encoder_layers_18_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_75_cast_fp16)[name = string("x_489_cast_fp16")]; + tensor input_991_perm_0 = const()[name = string("input_991_perm_0"), val = tensor([0, 2, 1])]; + tensor x_491_axes_0 = const()[name = string("x_491_axes_0"), val = tensor([-1])]; + tensor encoder_layers_18_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_18_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(374859328)))]; + tensor encoder_layers_18_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_18_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(374861440)))]; + tensor input_991_cast_fp16 = transpose(perm = input_991_perm_0, x = x_489_cast_fp16)[name = string("transpose_194")]; + tensor x_491_cast_fp16 = layer_norm(axes = x_491_axes_0, beta = encoder_layers_18_conv_batch_norm_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_18_conv_batch_norm_weight_to_fp16, x = input_991_cast_fp16)[name = string("x_491_cast_fp16")]; + tensor input_993_perm_0 = const()[name = string("input_993_perm_0"), val = tensor([0, 2, 1])]; + tensor input_993_cast_fp16 = transpose(perm = input_993_perm_0, x = x_491_cast_fp16)[name = string("transpose_193")]; + tensor input_995_cast_fp16 = silu(x = input_993_cast_fp16)[name = string("input_995_cast_fp16")]; + string x_493_pad_type_0 = const()[name = string("x_493_pad_type_0"), val = string("valid")]; + tensor x_493_strides_0 = const()[name = string("x_493_strides_0"), val = tensor([1])]; + tensor x_493_pad_0 = const()[name = string("x_493_pad_0"), val = tensor([0, 0])]; + tensor x_493_dilations_0 = const()[name = string("x_493_dilations_0"), val = tensor([1])]; + int32 x_493_groups_0 = const()[name = string("x_493_groups_0"), val = int32(1)]; + tensor encoder_layers_18_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(374863552))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(375912192))))[name = string("encoder_layers_18_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_493_cast_fp16 = conv(dilations = x_493_dilations_0, groups = x_493_groups_0, pad = x_493_pad_0, pad_type = x_493_pad_type_0, strides = x_493_strides_0, weight = encoder_layers_18_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_995_cast_fp16)[name = string("x_493_cast_fp16")]; + tensor input_997_perm_0 = const()[name = string("input_997_perm_0"), val = tensor([0, 2, 1])]; + tensor input_997_cast_fp16 = transpose(perm = input_997_perm_0, x = x_493_cast_fp16)[name = string("transpose_192")]; + tensor input_999_cast_fp16 = add(x = input_983_cast_fp16, y = input_997_cast_fp16)[name = string("input_999_cast_fp16")]; + tensor input_1001_axes_0 = const()[name = string("input_1001_axes_0"), val = tensor([-1])]; + tensor encoder_layers_18_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_18_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(375914304)))]; + tensor encoder_layers_18_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_18_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(375916416)))]; + tensor input_1001_cast_fp16 = layer_norm(axes = input_1001_axes_0, beta = encoder_layers_18_norm_feed_forward2_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_18_norm_feed_forward2_weight_to_fp16, x = input_999_cast_fp16)[name = string("input_1001_cast_fp16")]; + tensor encoder_layers_18_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(375918528))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(380112896))))[name = string("encoder_layers_18_feed_forward2_linear1_weight_to_fp16_quantized")]; + tensor encoder_layers_18_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_18_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(380121152)))]; + tensor linear_170_cast_fp16 = linear(bias = encoder_layers_18_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_18_feed_forward2_linear1_weight_to_fp16_quantized, x = input_1001_cast_fp16)[name = string("linear_170_cast_fp16")]; + tensor input_1005_cast_fp16 = silu(x = linear_170_cast_fp16)[name = string("input_1005_cast_fp16")]; + tensor encoder_layers_18_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(380129408))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(384323776))))[name = string("encoder_layers_18_feed_forward2_linear2_weight_to_fp16_quantized")]; + tensor encoder_layers_18_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_18_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(384325888)))]; + tensor linear_171_cast_fp16 = linear(bias = encoder_layers_18_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_18_feed_forward2_linear2_weight_to_fp16_quantized, x = input_1005_cast_fp16)[name = string("linear_171_cast_fp16")]; + fp16 var_4464_to_fp16 = const()[name = string("op_4464_to_fp16"), val = fp16(0x1p-1)]; + tensor var_4465_cast_fp16 = mul(x = linear_171_cast_fp16, y = var_4464_to_fp16)[name = string("op_4465_cast_fp16")]; + tensor input_1011_cast_fp16 = add(x = input_999_cast_fp16, y = var_4465_cast_fp16)[name = string("input_1011_cast_fp16")]; + tensor input_1013_axes_0 = const()[name = string("input_1013_axes_0"), val = tensor([-1])]; + tensor encoder_layers_18_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_18_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(384328000)))]; + tensor encoder_layers_18_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_18_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(384330112)))]; + tensor input_1013_cast_fp16 = layer_norm(axes = input_1013_axes_0, beta = encoder_layers_18_norm_out_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_18_norm_out_weight_to_fp16, x = input_1011_cast_fp16)[name = string("input_1013_cast_fp16")]; + tensor cache_77_begin_0 = const()[name = string("cache_77_begin_0"), val = tensor([19, 0, 0, 0])]; + tensor cache_77_end_0 = const()[name = string("cache_77_end_0"), val = tensor([20, 1, 42, 1024])]; + tensor cache_77_end_mask_0 = const()[name = string("cache_77_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_77_squeeze_mask_0 = const()[name = string("cache_77_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_77_cast_fp16 = slice_by_index(begin = cache_77_begin_0, end = cache_77_end_0, end_mask = cache_77_end_mask_0, squeeze_mask = cache_77_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_77_cast_fp16")]; + tensor cache_79_begin_0 = const()[name = string("cache_79_begin_0"), val = tensor([19, 0, 0, 0])]; + tensor cache_79_end_0 = const()[name = string("cache_79_end_0"), val = tensor([20, 1, 1024, 8])]; + tensor cache_79_end_mask_0 = const()[name = string("cache_79_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_79_squeeze_mask_0 = const()[name = string("cache_79_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_79_cast_fp16 = slice_by_index(begin = cache_79_begin_0, end = cache_79_end_0, end_mask = cache_79_end_mask_0, squeeze_mask = cache_79_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_79_cast_fp16")]; + tensor input_1015_axes_0 = const()[name = string("input_1015_axes_0"), val = tensor([-1])]; + tensor encoder_layers_19_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_19_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(384332224)))]; + tensor encoder_layers_19_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_19_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(384334336)))]; + tensor input_1015_cast_fp16 = layer_norm(axes = input_1015_axes_0, beta = encoder_layers_19_norm_feed_forward1_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_19_norm_feed_forward1_weight_to_fp16, x = input_1013_cast_fp16)[name = string("input_1015_cast_fp16")]; + tensor encoder_layers_19_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(384336448))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(388530816))))[name = string("encoder_layers_19_feed_forward1_linear1_weight_to_fp16_quantized")]; + tensor encoder_layers_19_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_19_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(388539072)))]; + tensor linear_172_cast_fp16 = linear(bias = encoder_layers_19_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_19_feed_forward1_linear1_weight_to_fp16_quantized, x = input_1015_cast_fp16)[name = string("linear_172_cast_fp16")]; + tensor input_1019_cast_fp16 = silu(x = linear_172_cast_fp16)[name = string("input_1019_cast_fp16")]; + tensor encoder_layers_19_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(388547328))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(392741696))))[name = string("encoder_layers_19_feed_forward1_linear2_weight_to_fp16_quantized")]; + tensor encoder_layers_19_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_19_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(392743808)))]; + tensor linear_173_cast_fp16 = linear(bias = encoder_layers_19_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_19_feed_forward1_linear2_weight_to_fp16_quantized, x = input_1019_cast_fp16)[name = string("linear_173_cast_fp16")]; + fp16 var_4501_to_fp16 = const()[name = string("op_4501_to_fp16"), val = fp16(0x1p-1)]; + tensor var_4502_cast_fp16 = mul(x = linear_173_cast_fp16, y = var_4501_to_fp16)[name = string("op_4502_cast_fp16")]; + tensor input_1025_cast_fp16 = add(x = input_1013_cast_fp16, y = var_4502_cast_fp16)[name = string("input_1025_cast_fp16")]; + tensor key_39_axes_0 = const()[name = string("key_39_axes_0"), val = tensor([-1])]; + tensor encoder_layers_19_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_19_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(392745920)))]; + tensor encoder_layers_19_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_19_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(392748032)))]; + tensor key_39_cast_fp16 = layer_norm(axes = key_39_axes_0, beta = encoder_layers_19_norm_self_att_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_19_norm_self_att_weight_to_fp16, x = input_1025_cast_fp16)[name = string("key_39_cast_fp16")]; + bool input_1027_interleave_0 = const()[name = string("input_1027_interleave_0"), val = bool(false)]; + tensor input_1027_cast_fp16 = concat(axis = var_67, interleave = input_1027_interleave_0, values = (cache_77_cast_fp16, key_39_cast_fp16))[name = string("input_1027_cast_fp16")]; + bool var_4530_interleave_0 = const()[name = string("op_4530_interleave_0"), val = bool(false)]; + tensor var_4530_cast_fp16 = concat(axis = var_67, interleave = var_4530_interleave_0, values = key_39_cast_fp16)[name = string("op_4530_cast_fp16")]; + tensor encoder_layers_19_self_attn_linear_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(392750144))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(393798784))))[name = string("encoder_layers_19_self_attn_linear_q_weight_to_fp16_quantized")]; + tensor encoder_layers_19_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_19_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(393800896)))]; + tensor linear_174_cast_fp16 = linear(bias = encoder_layers_19_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_19_self_attn_linear_q_weight_to_fp16_quantized, x = key_39_cast_fp16)[name = string("linear_174_cast_fp16")]; + tensor var_4535 = const()[name = string("op_4535"), val = tensor([1, -1, 8, 128])]; + tensor q_115_cast_fp16 = reshape(shape = var_4535, x = linear_174_cast_fp16)[name = string("q_115_cast_fp16")]; + tensor encoder_layers_19_self_attn_linear_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(393803008))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(394851648))))[name = string("encoder_layers_19_self_attn_linear_k_weight_to_fp16_quantized")]; + tensor encoder_layers_19_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_19_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(394853760)))]; + tensor linear_175_cast_fp16 = linear(bias = encoder_layers_19_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_19_self_attn_linear_k_weight_to_fp16_quantized, x = input_1027_cast_fp16)[name = string("linear_175_cast_fp16")]; + tensor var_4540 = const()[name = string("op_4540"), val = tensor([1, -1, 8, 128])]; + tensor k_77_cast_fp16 = reshape(shape = var_4540, x = linear_175_cast_fp16)[name = string("k_77_cast_fp16")]; + tensor encoder_layers_19_self_attn_linear_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(394855872))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(395904512))))[name = string("encoder_layers_19_self_attn_linear_v_weight_to_fp16_quantized")]; + tensor encoder_layers_19_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_19_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(395906624)))]; + tensor linear_176_cast_fp16 = linear(bias = encoder_layers_19_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_19_self_attn_linear_v_weight_to_fp16_quantized, x = input_1027_cast_fp16)[name = string("linear_176_cast_fp16")]; + tensor var_4545 = const()[name = string("op_4545"), val = tensor([1, -1, 8, 128])]; + tensor v_39_cast_fp16 = reshape(shape = var_4545, x = linear_176_cast_fp16)[name = string("v_39_cast_fp16")]; + tensor value_47_perm_0 = const()[name = string("value_47_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_19_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_19_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(395908736)))]; + tensor var_4558_cast_fp16 = add(x = q_115_cast_fp16, y = encoder_layers_19_self_attn_pos_bias_u_to_fp16)[name = string("op_4558_cast_fp16")]; + tensor encoder_layers_19_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_19_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(395910848)))]; + tensor var_4560_cast_fp16 = add(x = q_115_cast_fp16, y = encoder_layers_19_self_attn_pos_bias_v_to_fp16)[name = string("op_4560_cast_fp16")]; + tensor q_with_bias_v_39_perm_0 = const()[name = string("q_with_bias_v_39_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_501_transpose_x_0 = const()[name = string("x_501_transpose_x_0"), val = bool(false)]; + bool x_501_transpose_y_0 = const()[name = string("x_501_transpose_y_0"), val = bool(false)]; + tensor op_4562_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(395912960))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(396112704))))[name = string("op_4562_to_fp16_quantized")]; + tensor q_with_bias_v_39_cast_fp16 = transpose(perm = q_with_bias_v_39_perm_0, x = var_4560_cast_fp16)[name = string("transpose_191")]; + tensor x_501_cast_fp16 = matmul(transpose_x = x_501_transpose_x_0, transpose_y = x_501_transpose_y_0, x = q_with_bias_v_39_cast_fp16, y = op_4562_to_fp16_quantized)[name = string("x_501_cast_fp16")]; + tensor x_503_pad_0 = const()[name = string("x_503_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_503_mode_0 = const()[name = string("x_503_mode_0"), val = string("constant")]; + fp16 const_326_to_fp16 = const()[name = string("const_326_to_fp16"), val = fp16(0x0p+0)]; + tensor x_503_cast_fp16 = pad(constant_val = const_326_to_fp16, mode = x_503_mode_0, pad = x_503_pad_0, x = x_501_cast_fp16)[name = string("x_503_cast_fp16")]; + tensor var_4570 = const()[name = string("op_4570"), val = tensor([1, 8, -1, 56])]; + tensor x_505_cast_fp16 = reshape(shape = var_4570, x = x_503_cast_fp16)[name = string("x_505_cast_fp16")]; + tensor var_4574_begin_0 = const()[name = string("op_4574_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_4574_end_0 = const()[name = string("op_4574_end_0"), val = tensor([1, 8, 196, 56])]; + tensor var_4574_end_mask_0 = const()[name = string("op_4574_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_4574_cast_fp16 = slice_by_index(begin = var_4574_begin_0, end = var_4574_end_0, end_mask = var_4574_end_mask_0, x = x_505_cast_fp16)[name = string("op_4574_cast_fp16")]; + tensor var_4575 = const()[name = string("op_4575"), val = tensor([1, 8, 56, 195])]; + tensor matrix_bd_77_cast_fp16 = reshape(shape = var_4575, x = var_4574_cast_fp16)[name = string("matrix_bd_77_cast_fp16")]; + bool matrix_ac_39_transpose_x_0 = const()[name = string("matrix_ac_39_transpose_x_0"), val = bool(false)]; + bool matrix_ac_39_transpose_y_0 = const()[name = string("matrix_ac_39_transpose_y_0"), val = bool(false)]; + tensor transpose_134_perm_0 = const()[name = string("transpose_134_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_135_perm_0 = const()[name = string("transpose_135_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_135 = transpose(perm = transpose_135_perm_0, x = k_77_cast_fp16)[name = string("transpose_189")]; + tensor transpose_134 = transpose(perm = transpose_134_perm_0, x = var_4558_cast_fp16)[name = string("transpose_190")]; + tensor matrix_ac_39_cast_fp16 = matmul(transpose_x = matrix_ac_39_transpose_x_0, transpose_y = matrix_ac_39_transpose_y_0, x = transpose_134, y = transpose_135)[name = string("matrix_ac_39_cast_fp16")]; + tensor matrix_bd_79_begin_0 = const()[name = string("matrix_bd_79_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_79_end_0 = const()[name = string("matrix_bd_79_end_0"), val = tensor([1, 8, 56, 98])]; + tensor matrix_bd_79_end_mask_0 = const()[name = string("matrix_bd_79_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_79_cast_fp16 = slice_by_index(begin = matrix_bd_79_begin_0, end = matrix_bd_79_end_0, end_mask = matrix_bd_79_end_mask_0, x = matrix_bd_77_cast_fp16)[name = string("matrix_bd_79_cast_fp16")]; + tensor var_4584_cast_fp16 = add(x = matrix_ac_39_cast_fp16, y = matrix_bd_79_cast_fp16)[name = string("op_4584_cast_fp16")]; + fp16 _inversed_scores_77_y_0_to_fp16 = const()[name = string("_inversed_scores_77_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_77_cast_fp16 = mul(x = var_4584_cast_fp16, y = _inversed_scores_77_y_0_to_fp16)[name = string("_inversed_scores_77_cast_fp16")]; + tensor scores_79_cast_fp16 = select(a = var_44_to_fp16, b = _inversed_scores_77_cast_fp16, cond = mask_11)[name = string("scores_79_cast_fp16")]; + tensor var_4590_cast_fp16 = softmax(axis = var_58, x = scores_79_cast_fp16)[name = string("op_4590_cast_fp16")]; + tensor input_1029_cast_fp16 = select(a = var_43_to_fp16, b = var_4590_cast_fp16, cond = mask_11)[name = string("input_1029_cast_fp16")]; + bool x_507_transpose_x_0 = const()[name = string("x_507_transpose_x_0"), val = bool(false)]; + bool x_507_transpose_y_0 = const()[name = string("x_507_transpose_y_0"), val = bool(false)]; + tensor value_47_cast_fp16 = transpose(perm = value_47_perm_0, x = v_39_cast_fp16)[name = string("transpose_188")]; + tensor x_507_cast_fp16 = matmul(transpose_x = x_507_transpose_x_0, transpose_y = x_507_transpose_y_0, x = input_1029_cast_fp16, y = value_47_cast_fp16)[name = string("x_507_cast_fp16")]; + tensor var_4594_perm_0 = const()[name = string("op_4594_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_4595 = const()[name = string("op_4595"), val = tensor([1, -1, 1024])]; + tensor var_4594_cast_fp16 = transpose(perm = var_4594_perm_0, x = x_507_cast_fp16)[name = string("transpose_187")]; + tensor input_1031_cast_fp16 = reshape(shape = var_4595, x = var_4594_cast_fp16)[name = string("input_1031_cast_fp16")]; + tensor encoder_layers_19_self_attn_linear_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(396113216))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397161856))))[name = string("encoder_layers_19_self_attn_linear_out_weight_to_fp16_quantized")]; + tensor encoder_layers_19_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_19_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397163968)))]; + tensor linear_178_cast_fp16 = linear(bias = encoder_layers_19_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_19_self_attn_linear_out_weight_to_fp16_quantized, x = input_1031_cast_fp16)[name = string("linear_178_cast_fp16")]; + tensor input_1035_cast_fp16 = add(x = input_1025_cast_fp16, y = linear_178_cast_fp16)[name = string("input_1035_cast_fp16")]; + tensor x_511_axes_0 = const()[name = string("x_511_axes_0"), val = tensor([-1])]; + tensor encoder_layers_19_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_19_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397166080)))]; + tensor encoder_layers_19_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_19_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397168192)))]; + tensor x_511_cast_fp16 = layer_norm(axes = x_511_axes_0, beta = encoder_layers_19_norm_conv_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_19_norm_conv_weight_to_fp16, x = input_1035_cast_fp16)[name = string("x_511_cast_fp16")]; + tensor input_1037_perm_0 = const()[name = string("input_1037_perm_0"), val = tensor([0, 2, 1])]; + string input_1039_pad_type_0 = const()[name = string("input_1039_pad_type_0"), val = string("valid")]; + tensor input_1039_strides_0 = const()[name = string("input_1039_strides_0"), val = tensor([1])]; + tensor input_1039_pad_0 = const()[name = string("input_1039_pad_0"), val = tensor([0, 0])]; + tensor input_1039_dilations_0 = const()[name = string("input_1039_dilations_0"), val = tensor([1])]; + int32 input_1039_groups_0 = const()[name = string("input_1039_groups_0"), val = int32(1)]; + tensor encoder_layers_19_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397170304))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(399267520))))[name = string("encoder_layers_19_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_1037_cast_fp16 = transpose(perm = input_1037_perm_0, x = x_511_cast_fp16)[name = string("transpose_186")]; + tensor input_1039_cast_fp16 = conv(dilations = input_1039_dilations_0, groups = input_1039_groups_0, pad = input_1039_pad_0, pad_type = input_1039_pad_type_0, strides = input_1039_strides_0, weight = encoder_layers_19_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_1037_cast_fp16)[name = string("input_1039_cast_fp16")]; + int32 x_513_split_num_splits_0 = const()[name = string("x_513_split_num_splits_0"), val = int32(2)]; + int32 x_513_split_axis_0 = const()[name = string("x_513_split_axis_0"), val = int32(1)]; + tensor x_513_split_cast_fp16_0, tensor x_513_split_cast_fp16_1 = split(axis = x_513_split_axis_0, num_splits = x_513_split_num_splits_0, x = input_1039_cast_fp16)[name = string("x_513_split_cast_fp16")]; + tensor x_513_split_1_sigmoid_cast_fp16 = sigmoid(x = x_513_split_cast_fp16_1)[name = string("x_513_split_1_sigmoid_cast_fp16")]; + tensor x_513_cast_fp16 = mul(x = x_513_split_cast_fp16_0, y = x_513_split_1_sigmoid_cast_fp16)[name = string("x_513_cast_fp16")]; + tensor input_1041_cast_fp16 = select(a = var_43_to_fp16, b = x_513_cast_fp16, cond = var_574)[name = string("input_1041_cast_fp16")]; + bool new_x_79_interleave_0 = const()[name = string("new_x_79_interleave_0"), val = bool(false)]; + tensor new_x_79_cast_fp16 = concat(axis = var_58, interleave = new_x_79_interleave_0, values = (cache_79_cast_fp16, input_1041_cast_fp16))[name = string("new_x_79_cast_fp16")]; + tensor var_4634_begin_0 = const()[name = string("op_4634_begin_0"), val = tensor([0, 0, 56])]; + tensor var_4634_end_0 = const()[name = string("op_4634_end_0"), val = tensor([1, 1024, 64])]; + tensor var_4634_end_mask_0 = const()[name = string("op_4634_end_mask_0"), val = tensor([true, true, true])]; + tensor var_4634_cast_fp16 = slice_by_index(begin = var_4634_begin_0, end = var_4634_end_0, end_mask = var_4634_end_mask_0, x = new_x_79_cast_fp16)[name = string("op_4634_cast_fp16")]; + string x_515_pad_type_0 = const()[name = string("x_515_pad_type_0"), val = string("valid")]; + int32 x_515_groups_0 = const()[name = string("x_515_groups_0"), val = int32(1024)]; + tensor x_515_strides_0 = const()[name = string("x_515_strides_0"), val = tensor([1])]; + tensor x_515_pad_0 = const()[name = string("x_515_pad_0"), val = tensor([0, 0])]; + tensor x_515_dilations_0 = const()[name = string("x_515_dilations_0"), val = tensor([1])]; + tensor encoder_layers_19_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(399271680))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(399280960))))[name = string("encoder_layers_19_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_515_cast_fp16 = conv(dilations = x_515_dilations_0, groups = x_515_groups_0, pad = x_515_pad_0, pad_type = x_515_pad_type_0, strides = x_515_strides_0, weight = encoder_layers_19_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_79_cast_fp16)[name = string("x_515_cast_fp16")]; + tensor input_1043_perm_0 = const()[name = string("input_1043_perm_0"), val = tensor([0, 2, 1])]; + tensor x_517_axes_0 = const()[name = string("x_517_axes_0"), val = tensor([-1])]; + tensor encoder_layers_19_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_19_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(399283072)))]; + tensor encoder_layers_19_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_19_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(399285184)))]; + tensor input_1043_cast_fp16 = transpose(perm = input_1043_perm_0, x = x_515_cast_fp16)[name = string("transpose_185")]; + tensor x_517_cast_fp16 = layer_norm(axes = x_517_axes_0, beta = encoder_layers_19_conv_batch_norm_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_19_conv_batch_norm_weight_to_fp16, x = input_1043_cast_fp16)[name = string("x_517_cast_fp16")]; + tensor input_1045_perm_0 = const()[name = string("input_1045_perm_0"), val = tensor([0, 2, 1])]; + tensor input_1045_cast_fp16 = transpose(perm = input_1045_perm_0, x = x_517_cast_fp16)[name = string("transpose_184")]; + tensor input_1047_cast_fp16 = silu(x = input_1045_cast_fp16)[name = string("input_1047_cast_fp16")]; + string x_519_pad_type_0 = const()[name = string("x_519_pad_type_0"), val = string("valid")]; + tensor x_519_strides_0 = const()[name = string("x_519_strides_0"), val = tensor([1])]; + tensor x_519_pad_0 = const()[name = string("x_519_pad_0"), val = tensor([0, 0])]; + tensor x_519_dilations_0 = const()[name = string("x_519_dilations_0"), val = tensor([1])]; + int32 x_519_groups_0 = const()[name = string("x_519_groups_0"), val = int32(1)]; + tensor encoder_layers_19_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(399287296))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(400335936))))[name = string("encoder_layers_19_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_519_cast_fp16 = conv(dilations = x_519_dilations_0, groups = x_519_groups_0, pad = x_519_pad_0, pad_type = x_519_pad_type_0, strides = x_519_strides_0, weight = encoder_layers_19_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_1047_cast_fp16)[name = string("x_519_cast_fp16")]; + tensor input_1049_perm_0 = const()[name = string("input_1049_perm_0"), val = tensor([0, 2, 1])]; + tensor input_1049_cast_fp16 = transpose(perm = input_1049_perm_0, x = x_519_cast_fp16)[name = string("transpose_183")]; + tensor input_1051_cast_fp16 = add(x = input_1035_cast_fp16, y = input_1049_cast_fp16)[name = string("input_1051_cast_fp16")]; + tensor input_1053_axes_0 = const()[name = string("input_1053_axes_0"), val = tensor([-1])]; + tensor encoder_layers_19_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_19_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(400338048)))]; + tensor encoder_layers_19_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_19_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(400340160)))]; + tensor input_1053_cast_fp16 = layer_norm(axes = input_1053_axes_0, beta = encoder_layers_19_norm_feed_forward2_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_19_norm_feed_forward2_weight_to_fp16, x = input_1051_cast_fp16)[name = string("input_1053_cast_fp16")]; + tensor encoder_layers_19_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(400342272))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(404536640))))[name = string("encoder_layers_19_feed_forward2_linear1_weight_to_fp16_quantized")]; + tensor encoder_layers_19_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_19_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(404544896)))]; + tensor linear_179_cast_fp16 = linear(bias = encoder_layers_19_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_19_feed_forward2_linear1_weight_to_fp16_quantized, x = input_1053_cast_fp16)[name = string("linear_179_cast_fp16")]; + tensor input_1057_cast_fp16 = silu(x = linear_179_cast_fp16)[name = string("input_1057_cast_fp16")]; + tensor encoder_layers_19_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(404553152))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(408747520))))[name = string("encoder_layers_19_feed_forward2_linear2_weight_to_fp16_quantized")]; + tensor encoder_layers_19_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_19_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(408749632)))]; + tensor linear_180_cast_fp16 = linear(bias = encoder_layers_19_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_19_feed_forward2_linear2_weight_to_fp16_quantized, x = input_1057_cast_fp16)[name = string("linear_180_cast_fp16")]; + fp16 var_4677_to_fp16 = const()[name = string("op_4677_to_fp16"), val = fp16(0x1p-1)]; + tensor var_4678_cast_fp16 = mul(x = linear_180_cast_fp16, y = var_4677_to_fp16)[name = string("op_4678_cast_fp16")]; + tensor input_1063_cast_fp16 = add(x = input_1051_cast_fp16, y = var_4678_cast_fp16)[name = string("input_1063_cast_fp16")]; + tensor input_1065_axes_0 = const()[name = string("input_1065_axes_0"), val = tensor([-1])]; + tensor encoder_layers_19_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_19_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(408751744)))]; + tensor encoder_layers_19_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_19_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(408753856)))]; + tensor input_1065_cast_fp16 = layer_norm(axes = input_1065_axes_0, beta = encoder_layers_19_norm_out_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_19_norm_out_weight_to_fp16, x = input_1063_cast_fp16)[name = string("input_1065_cast_fp16")]; + tensor cache_81_begin_0 = const()[name = string("cache_81_begin_0"), val = tensor([20, 0, 0, 0])]; + tensor cache_81_end_0 = const()[name = string("cache_81_end_0"), val = tensor([21, 1, 42, 1024])]; + tensor cache_81_end_mask_0 = const()[name = string("cache_81_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_81_squeeze_mask_0 = const()[name = string("cache_81_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_81_cast_fp16 = slice_by_index(begin = cache_81_begin_0, end = cache_81_end_0, end_mask = cache_81_end_mask_0, squeeze_mask = cache_81_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_81_cast_fp16")]; + tensor cache_83_begin_0 = const()[name = string("cache_83_begin_0"), val = tensor([20, 0, 0, 0])]; + tensor cache_83_end_0 = const()[name = string("cache_83_end_0"), val = tensor([21, 1, 1024, 8])]; + tensor cache_83_end_mask_0 = const()[name = string("cache_83_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_83_squeeze_mask_0 = const()[name = string("cache_83_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_83_cast_fp16 = slice_by_index(begin = cache_83_begin_0, end = cache_83_end_0, end_mask = cache_83_end_mask_0, squeeze_mask = cache_83_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_83_cast_fp16")]; + tensor input_1067_axes_0 = const()[name = string("input_1067_axes_0"), val = tensor([-1])]; + tensor encoder_layers_20_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_20_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(408755968)))]; + tensor encoder_layers_20_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_20_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(408758080)))]; + tensor input_1067_cast_fp16 = layer_norm(axes = input_1067_axes_0, beta = encoder_layers_20_norm_feed_forward1_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_20_norm_feed_forward1_weight_to_fp16, x = input_1065_cast_fp16)[name = string("input_1067_cast_fp16")]; + tensor encoder_layers_20_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(408760192))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(412954560))))[name = string("encoder_layers_20_feed_forward1_linear1_weight_to_fp16_quantized")]; + tensor encoder_layers_20_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_20_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(412962816)))]; + tensor linear_181_cast_fp16 = linear(bias = encoder_layers_20_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_20_feed_forward1_linear1_weight_to_fp16_quantized, x = input_1067_cast_fp16)[name = string("linear_181_cast_fp16")]; + tensor input_1071_cast_fp16 = silu(x = linear_181_cast_fp16)[name = string("input_1071_cast_fp16")]; + tensor encoder_layers_20_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(412971072))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(417165440))))[name = string("encoder_layers_20_feed_forward1_linear2_weight_to_fp16_quantized")]; + tensor encoder_layers_20_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_20_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(417167552)))]; + tensor linear_182_cast_fp16 = linear(bias = encoder_layers_20_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_20_feed_forward1_linear2_weight_to_fp16_quantized, x = input_1071_cast_fp16)[name = string("linear_182_cast_fp16")]; + fp16 var_4714_to_fp16 = const()[name = string("op_4714_to_fp16"), val = fp16(0x1p-1)]; + tensor var_4715_cast_fp16 = mul(x = linear_182_cast_fp16, y = var_4714_to_fp16)[name = string("op_4715_cast_fp16")]; + tensor input_1077_cast_fp16 = add(x = input_1065_cast_fp16, y = var_4715_cast_fp16)[name = string("input_1077_cast_fp16")]; + tensor key_41_axes_0 = const()[name = string("key_41_axes_0"), val = tensor([-1])]; + tensor encoder_layers_20_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_20_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(417169664)))]; + tensor encoder_layers_20_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_20_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(417171776)))]; + tensor key_41_cast_fp16 = layer_norm(axes = key_41_axes_0, beta = encoder_layers_20_norm_self_att_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_20_norm_self_att_weight_to_fp16, x = input_1077_cast_fp16)[name = string("key_41_cast_fp16")]; + bool input_1079_interleave_0 = const()[name = string("input_1079_interleave_0"), val = bool(false)]; + tensor input_1079_cast_fp16 = concat(axis = var_67, interleave = input_1079_interleave_0, values = (cache_81_cast_fp16, key_41_cast_fp16))[name = string("input_1079_cast_fp16")]; + bool var_4743_interleave_0 = const()[name = string("op_4743_interleave_0"), val = bool(false)]; + tensor var_4743_cast_fp16 = concat(axis = var_67, interleave = var_4743_interleave_0, values = key_41_cast_fp16)[name = string("op_4743_cast_fp16")]; + tensor encoder_layers_20_self_attn_linear_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(417173888))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(418222528))))[name = string("encoder_layers_20_self_attn_linear_q_weight_to_fp16_quantized")]; + tensor encoder_layers_20_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_20_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(418224640)))]; + tensor linear_183_cast_fp16 = linear(bias = encoder_layers_20_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_20_self_attn_linear_q_weight_to_fp16_quantized, x = key_41_cast_fp16)[name = string("linear_183_cast_fp16")]; + tensor var_4748 = const()[name = string("op_4748"), val = tensor([1, -1, 8, 128])]; + tensor q_121_cast_fp16 = reshape(shape = var_4748, x = linear_183_cast_fp16)[name = string("q_121_cast_fp16")]; + tensor encoder_layers_20_self_attn_linear_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(418226752))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(419275392))))[name = string("encoder_layers_20_self_attn_linear_k_weight_to_fp16_quantized")]; + tensor encoder_layers_20_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_20_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(419277504)))]; + tensor linear_184_cast_fp16 = linear(bias = encoder_layers_20_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_20_self_attn_linear_k_weight_to_fp16_quantized, x = input_1079_cast_fp16)[name = string("linear_184_cast_fp16")]; + tensor var_4753 = const()[name = string("op_4753"), val = tensor([1, -1, 8, 128])]; + tensor k_81_cast_fp16 = reshape(shape = var_4753, x = linear_184_cast_fp16)[name = string("k_81_cast_fp16")]; + tensor encoder_layers_20_self_attn_linear_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(419279616))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(420328256))))[name = string("encoder_layers_20_self_attn_linear_v_weight_to_fp16_quantized")]; + tensor encoder_layers_20_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_20_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(420330368)))]; + tensor linear_185_cast_fp16 = linear(bias = encoder_layers_20_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_20_self_attn_linear_v_weight_to_fp16_quantized, x = input_1079_cast_fp16)[name = string("linear_185_cast_fp16")]; + tensor var_4758 = const()[name = string("op_4758"), val = tensor([1, -1, 8, 128])]; + tensor v_41_cast_fp16 = reshape(shape = var_4758, x = linear_185_cast_fp16)[name = string("v_41_cast_fp16")]; + tensor value_49_perm_0 = const()[name = string("value_49_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_20_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_20_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(420332480)))]; + tensor var_4771_cast_fp16 = add(x = q_121_cast_fp16, y = encoder_layers_20_self_attn_pos_bias_u_to_fp16)[name = string("op_4771_cast_fp16")]; + tensor encoder_layers_20_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_20_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(420334592)))]; + tensor var_4773_cast_fp16 = add(x = q_121_cast_fp16, y = encoder_layers_20_self_attn_pos_bias_v_to_fp16)[name = string("op_4773_cast_fp16")]; + tensor q_with_bias_v_41_perm_0 = const()[name = string("q_with_bias_v_41_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_527_transpose_x_0 = const()[name = string("x_527_transpose_x_0"), val = bool(false)]; + bool x_527_transpose_y_0 = const()[name = string("x_527_transpose_y_0"), val = bool(false)]; + tensor op_4775_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(420336704))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(420536448))))[name = string("op_4775_to_fp16_quantized")]; + tensor q_with_bias_v_41_cast_fp16 = transpose(perm = q_with_bias_v_41_perm_0, x = var_4773_cast_fp16)[name = string("transpose_182")]; + tensor x_527_cast_fp16 = matmul(transpose_x = x_527_transpose_x_0, transpose_y = x_527_transpose_y_0, x = q_with_bias_v_41_cast_fp16, y = op_4775_to_fp16_quantized)[name = string("x_527_cast_fp16")]; + tensor x_529_pad_0 = const()[name = string("x_529_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_529_mode_0 = const()[name = string("x_529_mode_0"), val = string("constant")]; + fp16 const_339_to_fp16 = const()[name = string("const_339_to_fp16"), val = fp16(0x0p+0)]; + tensor x_529_cast_fp16 = pad(constant_val = const_339_to_fp16, mode = x_529_mode_0, pad = x_529_pad_0, x = x_527_cast_fp16)[name = string("x_529_cast_fp16")]; + tensor var_4783 = const()[name = string("op_4783"), val = tensor([1, 8, -1, 56])]; + tensor x_531_cast_fp16 = reshape(shape = var_4783, x = x_529_cast_fp16)[name = string("x_531_cast_fp16")]; + tensor var_4787_begin_0 = const()[name = string("op_4787_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_4787_end_0 = const()[name = string("op_4787_end_0"), val = tensor([1, 8, 196, 56])]; + tensor var_4787_end_mask_0 = const()[name = string("op_4787_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_4787_cast_fp16 = slice_by_index(begin = var_4787_begin_0, end = var_4787_end_0, end_mask = var_4787_end_mask_0, x = x_531_cast_fp16)[name = string("op_4787_cast_fp16")]; + tensor var_4788 = const()[name = string("op_4788"), val = tensor([1, 8, 56, 195])]; + tensor matrix_bd_81_cast_fp16 = reshape(shape = var_4788, x = var_4787_cast_fp16)[name = string("matrix_bd_81_cast_fp16")]; + bool matrix_ac_41_transpose_x_0 = const()[name = string("matrix_ac_41_transpose_x_0"), val = bool(false)]; + bool matrix_ac_41_transpose_y_0 = const()[name = string("matrix_ac_41_transpose_y_0"), val = bool(false)]; + tensor transpose_136_perm_0 = const()[name = string("transpose_136_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_137_perm_0 = const()[name = string("transpose_137_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_137 = transpose(perm = transpose_137_perm_0, x = k_81_cast_fp16)[name = string("transpose_180")]; + tensor transpose_136 = transpose(perm = transpose_136_perm_0, x = var_4771_cast_fp16)[name = string("transpose_181")]; + tensor matrix_ac_41_cast_fp16 = matmul(transpose_x = matrix_ac_41_transpose_x_0, transpose_y = matrix_ac_41_transpose_y_0, x = transpose_136, y = transpose_137)[name = string("matrix_ac_41_cast_fp16")]; + tensor matrix_bd_83_begin_0 = const()[name = string("matrix_bd_83_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_83_end_0 = const()[name = string("matrix_bd_83_end_0"), val = tensor([1, 8, 56, 98])]; + tensor matrix_bd_83_end_mask_0 = const()[name = string("matrix_bd_83_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_83_cast_fp16 = slice_by_index(begin = matrix_bd_83_begin_0, end = matrix_bd_83_end_0, end_mask = matrix_bd_83_end_mask_0, x = matrix_bd_81_cast_fp16)[name = string("matrix_bd_83_cast_fp16")]; + tensor var_4797_cast_fp16 = add(x = matrix_ac_41_cast_fp16, y = matrix_bd_83_cast_fp16)[name = string("op_4797_cast_fp16")]; + fp16 _inversed_scores_81_y_0_to_fp16 = const()[name = string("_inversed_scores_81_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_81_cast_fp16 = mul(x = var_4797_cast_fp16, y = _inversed_scores_81_y_0_to_fp16)[name = string("_inversed_scores_81_cast_fp16")]; + tensor scores_83_cast_fp16 = select(a = var_44_to_fp16, b = _inversed_scores_81_cast_fp16, cond = mask_11)[name = string("scores_83_cast_fp16")]; + tensor var_4803_cast_fp16 = softmax(axis = var_58, x = scores_83_cast_fp16)[name = string("op_4803_cast_fp16")]; + tensor input_1081_cast_fp16 = select(a = var_43_to_fp16, b = var_4803_cast_fp16, cond = mask_11)[name = string("input_1081_cast_fp16")]; + bool x_533_transpose_x_0 = const()[name = string("x_533_transpose_x_0"), val = bool(false)]; + bool x_533_transpose_y_0 = const()[name = string("x_533_transpose_y_0"), val = bool(false)]; + tensor value_49_cast_fp16 = transpose(perm = value_49_perm_0, x = v_41_cast_fp16)[name = string("transpose_179")]; + tensor x_533_cast_fp16 = matmul(transpose_x = x_533_transpose_x_0, transpose_y = x_533_transpose_y_0, x = input_1081_cast_fp16, y = value_49_cast_fp16)[name = string("x_533_cast_fp16")]; + tensor var_4807_perm_0 = const()[name = string("op_4807_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_4808 = const()[name = string("op_4808"), val = tensor([1, -1, 1024])]; + tensor var_4807_cast_fp16 = transpose(perm = var_4807_perm_0, x = x_533_cast_fp16)[name = string("transpose_178")]; + tensor input_1083_cast_fp16 = reshape(shape = var_4808, x = var_4807_cast_fp16)[name = string("input_1083_cast_fp16")]; + tensor encoder_layers_20_self_attn_linear_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(420536960))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421585600))))[name = string("encoder_layers_20_self_attn_linear_out_weight_to_fp16_quantized")]; + tensor encoder_layers_20_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_20_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421587712)))]; + tensor linear_187_cast_fp16 = linear(bias = encoder_layers_20_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_20_self_attn_linear_out_weight_to_fp16_quantized, x = input_1083_cast_fp16)[name = string("linear_187_cast_fp16")]; + tensor input_1087_cast_fp16 = add(x = input_1077_cast_fp16, y = linear_187_cast_fp16)[name = string("input_1087_cast_fp16")]; + tensor x_537_axes_0 = const()[name = string("x_537_axes_0"), val = tensor([-1])]; + tensor encoder_layers_20_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_20_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421589824)))]; + tensor encoder_layers_20_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_20_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421591936)))]; + tensor x_537_cast_fp16 = layer_norm(axes = x_537_axes_0, beta = encoder_layers_20_norm_conv_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_20_norm_conv_weight_to_fp16, x = input_1087_cast_fp16)[name = string("x_537_cast_fp16")]; + tensor input_1089_perm_0 = const()[name = string("input_1089_perm_0"), val = tensor([0, 2, 1])]; + string input_1091_pad_type_0 = const()[name = string("input_1091_pad_type_0"), val = string("valid")]; + tensor input_1091_strides_0 = const()[name = string("input_1091_strides_0"), val = tensor([1])]; + tensor input_1091_pad_0 = const()[name = string("input_1091_pad_0"), val = tensor([0, 0])]; + tensor input_1091_dilations_0 = const()[name = string("input_1091_dilations_0"), val = tensor([1])]; + int32 input_1091_groups_0 = const()[name = string("input_1091_groups_0"), val = int32(1)]; + tensor encoder_layers_20_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421594048))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(423691264))))[name = string("encoder_layers_20_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_1089_cast_fp16 = transpose(perm = input_1089_perm_0, x = x_537_cast_fp16)[name = string("transpose_177")]; + tensor input_1091_cast_fp16 = conv(dilations = input_1091_dilations_0, groups = input_1091_groups_0, pad = input_1091_pad_0, pad_type = input_1091_pad_type_0, strides = input_1091_strides_0, weight = encoder_layers_20_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_1089_cast_fp16)[name = string("input_1091_cast_fp16")]; + int32 x_539_split_num_splits_0 = const()[name = string("x_539_split_num_splits_0"), val = int32(2)]; + int32 x_539_split_axis_0 = const()[name = string("x_539_split_axis_0"), val = int32(1)]; + tensor x_539_split_cast_fp16_0, tensor x_539_split_cast_fp16_1 = split(axis = x_539_split_axis_0, num_splits = x_539_split_num_splits_0, x = input_1091_cast_fp16)[name = string("x_539_split_cast_fp16")]; + tensor x_539_split_1_sigmoid_cast_fp16 = sigmoid(x = x_539_split_cast_fp16_1)[name = string("x_539_split_1_sigmoid_cast_fp16")]; + tensor x_539_cast_fp16 = mul(x = x_539_split_cast_fp16_0, y = x_539_split_1_sigmoid_cast_fp16)[name = string("x_539_cast_fp16")]; + tensor input_1093_cast_fp16 = select(a = var_43_to_fp16, b = x_539_cast_fp16, cond = var_574)[name = string("input_1093_cast_fp16")]; + bool new_x_83_interleave_0 = const()[name = string("new_x_83_interleave_0"), val = bool(false)]; + tensor new_x_83_cast_fp16 = concat(axis = var_58, interleave = new_x_83_interleave_0, values = (cache_83_cast_fp16, input_1093_cast_fp16))[name = string("new_x_83_cast_fp16")]; + tensor var_4847_begin_0 = const()[name = string("op_4847_begin_0"), val = tensor([0, 0, 56])]; + tensor var_4847_end_0 = const()[name = string("op_4847_end_0"), val = tensor([1, 1024, 64])]; + tensor var_4847_end_mask_0 = const()[name = string("op_4847_end_mask_0"), val = tensor([true, true, true])]; + tensor var_4847_cast_fp16 = slice_by_index(begin = var_4847_begin_0, end = var_4847_end_0, end_mask = var_4847_end_mask_0, x = new_x_83_cast_fp16)[name = string("op_4847_cast_fp16")]; + string x_541_pad_type_0 = const()[name = string("x_541_pad_type_0"), val = string("valid")]; + int32 x_541_groups_0 = const()[name = string("x_541_groups_0"), val = int32(1024)]; + tensor x_541_strides_0 = const()[name = string("x_541_strides_0"), val = tensor([1])]; + tensor x_541_pad_0 = const()[name = string("x_541_pad_0"), val = tensor([0, 0])]; + tensor x_541_dilations_0 = const()[name = string("x_541_dilations_0"), val = tensor([1])]; + tensor encoder_layers_20_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(423695424))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(423704704))))[name = string("encoder_layers_20_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_541_cast_fp16 = conv(dilations = x_541_dilations_0, groups = x_541_groups_0, pad = x_541_pad_0, pad_type = x_541_pad_type_0, strides = x_541_strides_0, weight = encoder_layers_20_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_83_cast_fp16)[name = string("x_541_cast_fp16")]; + tensor input_1095_perm_0 = const()[name = string("input_1095_perm_0"), val = tensor([0, 2, 1])]; + tensor x_543_axes_0 = const()[name = string("x_543_axes_0"), val = tensor([-1])]; + tensor encoder_layers_20_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_20_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(423706816)))]; + tensor encoder_layers_20_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_20_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(423708928)))]; + tensor input_1095_cast_fp16 = transpose(perm = input_1095_perm_0, x = x_541_cast_fp16)[name = string("transpose_176")]; + tensor x_543_cast_fp16 = layer_norm(axes = x_543_axes_0, beta = encoder_layers_20_conv_batch_norm_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_20_conv_batch_norm_weight_to_fp16, x = input_1095_cast_fp16)[name = string("x_543_cast_fp16")]; + tensor input_1097_perm_0 = const()[name = string("input_1097_perm_0"), val = tensor([0, 2, 1])]; + tensor input_1097_cast_fp16 = transpose(perm = input_1097_perm_0, x = x_543_cast_fp16)[name = string("transpose_175")]; + tensor input_1099_cast_fp16 = silu(x = input_1097_cast_fp16)[name = string("input_1099_cast_fp16")]; + string x_545_pad_type_0 = const()[name = string("x_545_pad_type_0"), val = string("valid")]; + tensor x_545_strides_0 = const()[name = string("x_545_strides_0"), val = tensor([1])]; + tensor x_545_pad_0 = const()[name = string("x_545_pad_0"), val = tensor([0, 0])]; + tensor x_545_dilations_0 = const()[name = string("x_545_dilations_0"), val = tensor([1])]; + int32 x_545_groups_0 = const()[name = string("x_545_groups_0"), val = int32(1)]; + tensor encoder_layers_20_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(423711040))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(424759680))))[name = string("encoder_layers_20_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_545_cast_fp16 = conv(dilations = x_545_dilations_0, groups = x_545_groups_0, pad = x_545_pad_0, pad_type = x_545_pad_type_0, strides = x_545_strides_0, weight = encoder_layers_20_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_1099_cast_fp16)[name = string("x_545_cast_fp16")]; + tensor input_1101_perm_0 = const()[name = string("input_1101_perm_0"), val = tensor([0, 2, 1])]; + tensor input_1101_cast_fp16 = transpose(perm = input_1101_perm_0, x = x_545_cast_fp16)[name = string("transpose_174")]; + tensor input_1103_cast_fp16 = add(x = input_1087_cast_fp16, y = input_1101_cast_fp16)[name = string("input_1103_cast_fp16")]; + tensor input_1105_axes_0 = const()[name = string("input_1105_axes_0"), val = tensor([-1])]; + tensor encoder_layers_20_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_20_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(424761792)))]; + tensor encoder_layers_20_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_20_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(424763904)))]; + tensor input_1105_cast_fp16 = layer_norm(axes = input_1105_axes_0, beta = encoder_layers_20_norm_feed_forward2_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_20_norm_feed_forward2_weight_to_fp16, x = input_1103_cast_fp16)[name = string("input_1105_cast_fp16")]; + tensor encoder_layers_20_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(424766016))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(428960384))))[name = string("encoder_layers_20_feed_forward2_linear1_weight_to_fp16_quantized")]; + tensor encoder_layers_20_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_20_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(428968640)))]; + tensor linear_188_cast_fp16 = linear(bias = encoder_layers_20_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_20_feed_forward2_linear1_weight_to_fp16_quantized, x = input_1105_cast_fp16)[name = string("linear_188_cast_fp16")]; + tensor input_1109_cast_fp16 = silu(x = linear_188_cast_fp16)[name = string("input_1109_cast_fp16")]; + tensor encoder_layers_20_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(428976896))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(433171264))))[name = string("encoder_layers_20_feed_forward2_linear2_weight_to_fp16_quantized")]; + tensor encoder_layers_20_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_20_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(433173376)))]; + tensor linear_189_cast_fp16 = linear(bias = encoder_layers_20_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_20_feed_forward2_linear2_weight_to_fp16_quantized, x = input_1109_cast_fp16)[name = string("linear_189_cast_fp16")]; + fp16 var_4890_to_fp16 = const()[name = string("op_4890_to_fp16"), val = fp16(0x1p-1)]; + tensor var_4891_cast_fp16 = mul(x = linear_189_cast_fp16, y = var_4890_to_fp16)[name = string("op_4891_cast_fp16")]; + tensor input_1115_cast_fp16 = add(x = input_1103_cast_fp16, y = var_4891_cast_fp16)[name = string("input_1115_cast_fp16")]; + tensor input_1117_axes_0 = const()[name = string("input_1117_axes_0"), val = tensor([-1])]; + tensor encoder_layers_20_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_20_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(433175488)))]; + tensor encoder_layers_20_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_20_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(433177600)))]; + tensor input_1117_cast_fp16 = layer_norm(axes = input_1117_axes_0, beta = encoder_layers_20_norm_out_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_20_norm_out_weight_to_fp16, x = input_1115_cast_fp16)[name = string("input_1117_cast_fp16")]; + tensor cache_85_begin_0 = const()[name = string("cache_85_begin_0"), val = tensor([21, 0, 0, 0])]; + tensor cache_85_end_0 = const()[name = string("cache_85_end_0"), val = tensor([22, 1, 42, 1024])]; + tensor cache_85_end_mask_0 = const()[name = string("cache_85_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_85_squeeze_mask_0 = const()[name = string("cache_85_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_85_cast_fp16 = slice_by_index(begin = cache_85_begin_0, end = cache_85_end_0, end_mask = cache_85_end_mask_0, squeeze_mask = cache_85_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_85_cast_fp16")]; + tensor cache_87_begin_0 = const()[name = string("cache_87_begin_0"), val = tensor([21, 0, 0, 0])]; + tensor cache_87_end_0 = const()[name = string("cache_87_end_0"), val = tensor([22, 1, 1024, 8])]; + tensor cache_87_end_mask_0 = const()[name = string("cache_87_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_87_squeeze_mask_0 = const()[name = string("cache_87_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_87_cast_fp16 = slice_by_index(begin = cache_87_begin_0, end = cache_87_end_0, end_mask = cache_87_end_mask_0, squeeze_mask = cache_87_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_87_cast_fp16")]; + tensor input_1119_axes_0 = const()[name = string("input_1119_axes_0"), val = tensor([-1])]; + tensor encoder_layers_21_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_21_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(433179712)))]; + tensor encoder_layers_21_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_21_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(433181824)))]; + tensor input_1119_cast_fp16 = layer_norm(axes = input_1119_axes_0, beta = encoder_layers_21_norm_feed_forward1_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_21_norm_feed_forward1_weight_to_fp16, x = input_1117_cast_fp16)[name = string("input_1119_cast_fp16")]; + tensor encoder_layers_21_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(433183936))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(437378304))))[name = string("encoder_layers_21_feed_forward1_linear1_weight_to_fp16_quantized")]; + tensor encoder_layers_21_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_21_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(437386560)))]; + tensor linear_190_cast_fp16 = linear(bias = encoder_layers_21_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_21_feed_forward1_linear1_weight_to_fp16_quantized, x = input_1119_cast_fp16)[name = string("linear_190_cast_fp16")]; + tensor input_1123_cast_fp16 = silu(x = linear_190_cast_fp16)[name = string("input_1123_cast_fp16")]; + tensor encoder_layers_21_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(437394816))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(441589184))))[name = string("encoder_layers_21_feed_forward1_linear2_weight_to_fp16_quantized")]; + tensor encoder_layers_21_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_21_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(441591296)))]; + tensor linear_191_cast_fp16 = linear(bias = encoder_layers_21_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_21_feed_forward1_linear2_weight_to_fp16_quantized, x = input_1123_cast_fp16)[name = string("linear_191_cast_fp16")]; + fp16 var_4927_to_fp16 = const()[name = string("op_4927_to_fp16"), val = fp16(0x1p-1)]; + tensor var_4928_cast_fp16 = mul(x = linear_191_cast_fp16, y = var_4927_to_fp16)[name = string("op_4928_cast_fp16")]; + tensor input_1129_cast_fp16 = add(x = input_1117_cast_fp16, y = var_4928_cast_fp16)[name = string("input_1129_cast_fp16")]; + tensor key_43_axes_0 = const()[name = string("key_43_axes_0"), val = tensor([-1])]; + tensor encoder_layers_21_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_21_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(441593408)))]; + tensor encoder_layers_21_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_21_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(441595520)))]; + tensor key_43_cast_fp16 = layer_norm(axes = key_43_axes_0, beta = encoder_layers_21_norm_self_att_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_21_norm_self_att_weight_to_fp16, x = input_1129_cast_fp16)[name = string("key_43_cast_fp16")]; + bool input_1131_interleave_0 = const()[name = string("input_1131_interleave_0"), val = bool(false)]; + tensor input_1131_cast_fp16 = concat(axis = var_67, interleave = input_1131_interleave_0, values = (cache_85_cast_fp16, key_43_cast_fp16))[name = string("input_1131_cast_fp16")]; + bool var_4956_interleave_0 = const()[name = string("op_4956_interleave_0"), val = bool(false)]; + tensor var_4956_cast_fp16 = concat(axis = var_67, interleave = var_4956_interleave_0, values = key_43_cast_fp16)[name = string("op_4956_cast_fp16")]; + tensor encoder_layers_21_self_attn_linear_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(441597632))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(442646272))))[name = string("encoder_layers_21_self_attn_linear_q_weight_to_fp16_quantized")]; + tensor encoder_layers_21_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_21_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(442648384)))]; + tensor linear_192_cast_fp16 = linear(bias = encoder_layers_21_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_21_self_attn_linear_q_weight_to_fp16_quantized, x = key_43_cast_fp16)[name = string("linear_192_cast_fp16")]; + tensor var_4961 = const()[name = string("op_4961"), val = tensor([1, -1, 8, 128])]; + tensor q_127_cast_fp16 = reshape(shape = var_4961, x = linear_192_cast_fp16)[name = string("q_127_cast_fp16")]; + tensor encoder_layers_21_self_attn_linear_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(442650496))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(443699136))))[name = string("encoder_layers_21_self_attn_linear_k_weight_to_fp16_quantized")]; + tensor encoder_layers_21_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_21_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(443701248)))]; + tensor linear_193_cast_fp16 = linear(bias = encoder_layers_21_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_21_self_attn_linear_k_weight_to_fp16_quantized, x = input_1131_cast_fp16)[name = string("linear_193_cast_fp16")]; + tensor var_4966 = const()[name = string("op_4966"), val = tensor([1, -1, 8, 128])]; + tensor k_85_cast_fp16 = reshape(shape = var_4966, x = linear_193_cast_fp16)[name = string("k_85_cast_fp16")]; + tensor encoder_layers_21_self_attn_linear_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(443703360))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(444752000))))[name = string("encoder_layers_21_self_attn_linear_v_weight_to_fp16_quantized")]; + tensor encoder_layers_21_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_21_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(444754112)))]; + tensor linear_194_cast_fp16 = linear(bias = encoder_layers_21_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_21_self_attn_linear_v_weight_to_fp16_quantized, x = input_1131_cast_fp16)[name = string("linear_194_cast_fp16")]; + tensor var_4971 = const()[name = string("op_4971"), val = tensor([1, -1, 8, 128])]; + tensor v_43_cast_fp16 = reshape(shape = var_4971, x = linear_194_cast_fp16)[name = string("v_43_cast_fp16")]; + tensor value_51_perm_0 = const()[name = string("value_51_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_21_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_21_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(444756224)))]; + tensor var_4984_cast_fp16 = add(x = q_127_cast_fp16, y = encoder_layers_21_self_attn_pos_bias_u_to_fp16)[name = string("op_4984_cast_fp16")]; + tensor encoder_layers_21_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_21_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(444758336)))]; + tensor var_4986_cast_fp16 = add(x = q_127_cast_fp16, y = encoder_layers_21_self_attn_pos_bias_v_to_fp16)[name = string("op_4986_cast_fp16")]; + tensor q_with_bias_v_43_perm_0 = const()[name = string("q_with_bias_v_43_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_553_transpose_x_0 = const()[name = string("x_553_transpose_x_0"), val = bool(false)]; + bool x_553_transpose_y_0 = const()[name = string("x_553_transpose_y_0"), val = bool(false)]; + tensor op_4988_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(444760448))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(444960192))))[name = string("op_4988_to_fp16_quantized")]; + tensor q_with_bias_v_43_cast_fp16 = transpose(perm = q_with_bias_v_43_perm_0, x = var_4986_cast_fp16)[name = string("transpose_173")]; + tensor x_553_cast_fp16 = matmul(transpose_x = x_553_transpose_x_0, transpose_y = x_553_transpose_y_0, x = q_with_bias_v_43_cast_fp16, y = op_4988_to_fp16_quantized)[name = string("x_553_cast_fp16")]; + tensor x_555_pad_0 = const()[name = string("x_555_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_555_mode_0 = const()[name = string("x_555_mode_0"), val = string("constant")]; + fp16 const_352_to_fp16 = const()[name = string("const_352_to_fp16"), val = fp16(0x0p+0)]; + tensor x_555_cast_fp16 = pad(constant_val = const_352_to_fp16, mode = x_555_mode_0, pad = x_555_pad_0, x = x_553_cast_fp16)[name = string("x_555_cast_fp16")]; + tensor var_4996 = const()[name = string("op_4996"), val = tensor([1, 8, -1, 56])]; + tensor x_557_cast_fp16 = reshape(shape = var_4996, x = x_555_cast_fp16)[name = string("x_557_cast_fp16")]; + tensor var_5000_begin_0 = const()[name = string("op_5000_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_5000_end_0 = const()[name = string("op_5000_end_0"), val = tensor([1, 8, 196, 56])]; + tensor var_5000_end_mask_0 = const()[name = string("op_5000_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_5000_cast_fp16 = slice_by_index(begin = var_5000_begin_0, end = var_5000_end_0, end_mask = var_5000_end_mask_0, x = x_557_cast_fp16)[name = string("op_5000_cast_fp16")]; + tensor var_5001 = const()[name = string("op_5001"), val = tensor([1, 8, 56, 195])]; + tensor matrix_bd_85_cast_fp16 = reshape(shape = var_5001, x = var_5000_cast_fp16)[name = string("matrix_bd_85_cast_fp16")]; + bool matrix_ac_43_transpose_x_0 = const()[name = string("matrix_ac_43_transpose_x_0"), val = bool(false)]; + bool matrix_ac_43_transpose_y_0 = const()[name = string("matrix_ac_43_transpose_y_0"), val = bool(false)]; + tensor transpose_138_perm_0 = const()[name = string("transpose_138_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_139_perm_0 = const()[name = string("transpose_139_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_139 = transpose(perm = transpose_139_perm_0, x = k_85_cast_fp16)[name = string("transpose_171")]; + tensor transpose_138 = transpose(perm = transpose_138_perm_0, x = var_4984_cast_fp16)[name = string("transpose_172")]; + tensor matrix_ac_43_cast_fp16 = matmul(transpose_x = matrix_ac_43_transpose_x_0, transpose_y = matrix_ac_43_transpose_y_0, x = transpose_138, y = transpose_139)[name = string("matrix_ac_43_cast_fp16")]; + tensor matrix_bd_87_begin_0 = const()[name = string("matrix_bd_87_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_87_end_0 = const()[name = string("matrix_bd_87_end_0"), val = tensor([1, 8, 56, 98])]; + tensor matrix_bd_87_end_mask_0 = const()[name = string("matrix_bd_87_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_87_cast_fp16 = slice_by_index(begin = matrix_bd_87_begin_0, end = matrix_bd_87_end_0, end_mask = matrix_bd_87_end_mask_0, x = matrix_bd_85_cast_fp16)[name = string("matrix_bd_87_cast_fp16")]; + tensor var_5010_cast_fp16 = add(x = matrix_ac_43_cast_fp16, y = matrix_bd_87_cast_fp16)[name = string("op_5010_cast_fp16")]; + fp16 _inversed_scores_85_y_0_to_fp16 = const()[name = string("_inversed_scores_85_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_85_cast_fp16 = mul(x = var_5010_cast_fp16, y = _inversed_scores_85_y_0_to_fp16)[name = string("_inversed_scores_85_cast_fp16")]; + tensor scores_87_cast_fp16 = select(a = var_44_to_fp16, b = _inversed_scores_85_cast_fp16, cond = mask_11)[name = string("scores_87_cast_fp16")]; + tensor var_5016_cast_fp16 = softmax(axis = var_58, x = scores_87_cast_fp16)[name = string("op_5016_cast_fp16")]; + tensor input_1133_cast_fp16 = select(a = var_43_to_fp16, b = var_5016_cast_fp16, cond = mask_11)[name = string("input_1133_cast_fp16")]; + bool x_559_transpose_x_0 = const()[name = string("x_559_transpose_x_0"), val = bool(false)]; + bool x_559_transpose_y_0 = const()[name = string("x_559_transpose_y_0"), val = bool(false)]; + tensor value_51_cast_fp16 = transpose(perm = value_51_perm_0, x = v_43_cast_fp16)[name = string("transpose_170")]; + tensor x_559_cast_fp16 = matmul(transpose_x = x_559_transpose_x_0, transpose_y = x_559_transpose_y_0, x = input_1133_cast_fp16, y = value_51_cast_fp16)[name = string("x_559_cast_fp16")]; + tensor var_5020_perm_0 = const()[name = string("op_5020_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_5021 = const()[name = string("op_5021"), val = tensor([1, -1, 1024])]; + tensor var_5020_cast_fp16 = transpose(perm = var_5020_perm_0, x = x_559_cast_fp16)[name = string("transpose_169")]; + tensor input_1135_cast_fp16 = reshape(shape = var_5021, x = var_5020_cast_fp16)[name = string("input_1135_cast_fp16")]; + tensor encoder_layers_21_self_attn_linear_out_weight_to_fp16 = const()[name = string("encoder_layers_21_self_attn_linear_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(444960704)))]; + tensor encoder_layers_21_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_21_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(447057920)))]; + tensor linear_196_cast_fp16 = linear(bias = encoder_layers_21_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_21_self_attn_linear_out_weight_to_fp16, x = input_1135_cast_fp16)[name = string("linear_196_cast_fp16")]; + tensor input_1139_cast_fp16 = add(x = input_1129_cast_fp16, y = linear_196_cast_fp16)[name = string("input_1139_cast_fp16")]; + tensor x_563_axes_0 = const()[name = string("x_563_axes_0"), val = tensor([-1])]; + tensor encoder_layers_21_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_21_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(447060032)))]; + tensor encoder_layers_21_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_21_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(447062144)))]; + tensor x_563_cast_fp16 = layer_norm(axes = x_563_axes_0, beta = encoder_layers_21_norm_conv_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_21_norm_conv_weight_to_fp16, x = input_1139_cast_fp16)[name = string("x_563_cast_fp16")]; + tensor input_1141_perm_0 = const()[name = string("input_1141_perm_0"), val = tensor([0, 2, 1])]; + string input_1143_pad_type_0 = const()[name = string("input_1143_pad_type_0"), val = string("valid")]; + tensor input_1143_strides_0 = const()[name = string("input_1143_strides_0"), val = tensor([1])]; + tensor input_1143_pad_0 = const()[name = string("input_1143_pad_0"), val = tensor([0, 0])]; + tensor input_1143_dilations_0 = const()[name = string("input_1143_dilations_0"), val = tensor([1])]; + int32 input_1143_groups_0 = const()[name = string("input_1143_groups_0"), val = int32(1)]; + tensor encoder_layers_21_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(447064256))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(449161472))))[name = string("encoder_layers_21_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_1141_cast_fp16 = transpose(perm = input_1141_perm_0, x = x_563_cast_fp16)[name = string("transpose_168")]; + tensor input_1143_cast_fp16 = conv(dilations = input_1143_dilations_0, groups = input_1143_groups_0, pad = input_1143_pad_0, pad_type = input_1143_pad_type_0, strides = input_1143_strides_0, weight = encoder_layers_21_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_1141_cast_fp16)[name = string("input_1143_cast_fp16")]; + int32 x_565_split_num_splits_0 = const()[name = string("x_565_split_num_splits_0"), val = int32(2)]; + int32 x_565_split_axis_0 = const()[name = string("x_565_split_axis_0"), val = int32(1)]; + tensor x_565_split_cast_fp16_0, tensor x_565_split_cast_fp16_1 = split(axis = x_565_split_axis_0, num_splits = x_565_split_num_splits_0, x = input_1143_cast_fp16)[name = string("x_565_split_cast_fp16")]; + tensor x_565_split_1_sigmoid_cast_fp16 = sigmoid(x = x_565_split_cast_fp16_1)[name = string("x_565_split_1_sigmoid_cast_fp16")]; + tensor x_565_cast_fp16 = mul(x = x_565_split_cast_fp16_0, y = x_565_split_1_sigmoid_cast_fp16)[name = string("x_565_cast_fp16")]; + tensor input_1145_cast_fp16 = select(a = var_43_to_fp16, b = x_565_cast_fp16, cond = var_574)[name = string("input_1145_cast_fp16")]; + bool new_x_87_interleave_0 = const()[name = string("new_x_87_interleave_0"), val = bool(false)]; + tensor new_x_87_cast_fp16 = concat(axis = var_58, interleave = new_x_87_interleave_0, values = (cache_87_cast_fp16, input_1145_cast_fp16))[name = string("new_x_87_cast_fp16")]; + tensor var_5060_begin_0 = const()[name = string("op_5060_begin_0"), val = tensor([0, 0, 56])]; + tensor var_5060_end_0 = const()[name = string("op_5060_end_0"), val = tensor([1, 1024, 64])]; + tensor var_5060_end_mask_0 = const()[name = string("op_5060_end_mask_0"), val = tensor([true, true, true])]; + tensor var_5060_cast_fp16 = slice_by_index(begin = var_5060_begin_0, end = var_5060_end_0, end_mask = var_5060_end_mask_0, x = new_x_87_cast_fp16)[name = string("op_5060_cast_fp16")]; + string x_567_pad_type_0 = const()[name = string("x_567_pad_type_0"), val = string("valid")]; + int32 x_567_groups_0 = const()[name = string("x_567_groups_0"), val = int32(1024)]; + tensor x_567_strides_0 = const()[name = string("x_567_strides_0"), val = tensor([1])]; + tensor x_567_pad_0 = const()[name = string("x_567_pad_0"), val = tensor([0, 0])]; + tensor x_567_dilations_0 = const()[name = string("x_567_dilations_0"), val = tensor([1])]; + tensor encoder_layers_21_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(449165632))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(449174912))))[name = string("encoder_layers_21_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_567_cast_fp16 = conv(dilations = x_567_dilations_0, groups = x_567_groups_0, pad = x_567_pad_0, pad_type = x_567_pad_type_0, strides = x_567_strides_0, weight = encoder_layers_21_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_87_cast_fp16)[name = string("x_567_cast_fp16")]; + tensor input_1147_perm_0 = const()[name = string("input_1147_perm_0"), val = tensor([0, 2, 1])]; + tensor x_569_axes_0 = const()[name = string("x_569_axes_0"), val = tensor([-1])]; + tensor encoder_layers_21_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_21_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(449177024)))]; + tensor encoder_layers_21_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_21_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(449179136)))]; + tensor input_1147_cast_fp16 = transpose(perm = input_1147_perm_0, x = x_567_cast_fp16)[name = string("transpose_167")]; + tensor x_569_cast_fp16 = layer_norm(axes = x_569_axes_0, beta = encoder_layers_21_conv_batch_norm_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_21_conv_batch_norm_weight_to_fp16, x = input_1147_cast_fp16)[name = string("x_569_cast_fp16")]; + tensor input_1149_perm_0 = const()[name = string("input_1149_perm_0"), val = tensor([0, 2, 1])]; + tensor input_1149_cast_fp16 = transpose(perm = input_1149_perm_0, x = x_569_cast_fp16)[name = string("transpose_166")]; + tensor input_1151_cast_fp16 = silu(x = input_1149_cast_fp16)[name = string("input_1151_cast_fp16")]; + string x_571_pad_type_0 = const()[name = string("x_571_pad_type_0"), val = string("valid")]; + tensor x_571_strides_0 = const()[name = string("x_571_strides_0"), val = tensor([1])]; + tensor x_571_pad_0 = const()[name = string("x_571_pad_0"), val = tensor([0, 0])]; + tensor x_571_dilations_0 = const()[name = string("x_571_dilations_0"), val = tensor([1])]; + int32 x_571_groups_0 = const()[name = string("x_571_groups_0"), val = int32(1)]; + tensor encoder_layers_21_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(449181248))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(450229888))))[name = string("encoder_layers_21_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_571_cast_fp16 = conv(dilations = x_571_dilations_0, groups = x_571_groups_0, pad = x_571_pad_0, pad_type = x_571_pad_type_0, strides = x_571_strides_0, weight = encoder_layers_21_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_1151_cast_fp16)[name = string("x_571_cast_fp16")]; + tensor input_1153_perm_0 = const()[name = string("input_1153_perm_0"), val = tensor([0, 2, 1])]; + tensor input_1153_cast_fp16 = transpose(perm = input_1153_perm_0, x = x_571_cast_fp16)[name = string("transpose_165")]; + tensor input_1155_cast_fp16 = add(x = input_1139_cast_fp16, y = input_1153_cast_fp16)[name = string("input_1155_cast_fp16")]; + tensor input_1157_axes_0 = const()[name = string("input_1157_axes_0"), val = tensor([-1])]; + tensor encoder_layers_21_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_21_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(450232000)))]; + tensor encoder_layers_21_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_21_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(450234112)))]; + tensor input_1157_cast_fp16 = layer_norm(axes = input_1157_axes_0, beta = encoder_layers_21_norm_feed_forward2_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_21_norm_feed_forward2_weight_to_fp16, x = input_1155_cast_fp16)[name = string("input_1157_cast_fp16")]; + tensor encoder_layers_21_feed_forward2_linear1_weight_to_fp16 = const()[name = string("encoder_layers_21_feed_forward2_linear1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(450236224)))]; + tensor encoder_layers_21_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_21_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(458624896)))]; + tensor linear_197_cast_fp16 = linear(bias = encoder_layers_21_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_21_feed_forward2_linear1_weight_to_fp16, x = input_1157_cast_fp16)[name = string("linear_197_cast_fp16")]; + tensor input_1161_cast_fp16 = silu(x = linear_197_cast_fp16)[name = string("input_1161_cast_fp16")]; + tensor encoder_layers_21_feed_forward2_linear2_weight_to_fp16 = const()[name = string("encoder_layers_21_feed_forward2_linear2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(458633152)))]; + tensor encoder_layers_21_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_21_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(467021824)))]; + tensor linear_198_cast_fp16 = linear(bias = encoder_layers_21_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_21_feed_forward2_linear2_weight_to_fp16, x = input_1161_cast_fp16)[name = string("linear_198_cast_fp16")]; + fp16 var_5103_to_fp16 = const()[name = string("op_5103_to_fp16"), val = fp16(0x1p-1)]; + tensor var_5104_cast_fp16 = mul(x = linear_198_cast_fp16, y = var_5103_to_fp16)[name = string("op_5104_cast_fp16")]; + tensor input_1167_cast_fp16 = add(x = input_1155_cast_fp16, y = var_5104_cast_fp16)[name = string("input_1167_cast_fp16")]; + tensor input_1169_axes_0 = const()[name = string("input_1169_axes_0"), val = tensor([-1])]; + tensor encoder_layers_21_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_21_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(467023936)))]; + tensor encoder_layers_21_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_21_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(467026048)))]; + tensor input_1169_cast_fp16 = layer_norm(axes = input_1169_axes_0, beta = encoder_layers_21_norm_out_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_21_norm_out_weight_to_fp16, x = input_1167_cast_fp16)[name = string("input_1169_cast_fp16")]; + tensor cache_89_begin_0 = const()[name = string("cache_89_begin_0"), val = tensor([22, 0, 0, 0])]; + tensor cache_89_end_0 = const()[name = string("cache_89_end_0"), val = tensor([23, 1, 42, 1024])]; + tensor cache_89_end_mask_0 = const()[name = string("cache_89_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_89_squeeze_mask_0 = const()[name = string("cache_89_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_89_cast_fp16 = slice_by_index(begin = cache_89_begin_0, end = cache_89_end_0, end_mask = cache_89_end_mask_0, squeeze_mask = cache_89_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_89_cast_fp16")]; + tensor cache_91_begin_0 = const()[name = string("cache_91_begin_0"), val = tensor([22, 0, 0, 0])]; + tensor cache_91_end_0 = const()[name = string("cache_91_end_0"), val = tensor([23, 1, 1024, 8])]; + tensor cache_91_end_mask_0 = const()[name = string("cache_91_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_91_squeeze_mask_0 = const()[name = string("cache_91_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_91_cast_fp16 = slice_by_index(begin = cache_91_begin_0, end = cache_91_end_0, end_mask = cache_91_end_mask_0, squeeze_mask = cache_91_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_91_cast_fp16")]; + tensor input_1171_axes_0 = const()[name = string("input_1171_axes_0"), val = tensor([-1])]; + tensor encoder_layers_22_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_22_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(467028160)))]; + tensor encoder_layers_22_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_22_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(467030272)))]; + tensor input_1171_cast_fp16 = layer_norm(axes = input_1171_axes_0, beta = encoder_layers_22_norm_feed_forward1_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_22_norm_feed_forward1_weight_to_fp16, x = input_1169_cast_fp16)[name = string("input_1171_cast_fp16")]; + tensor encoder_layers_22_feed_forward1_linear1_weight_to_fp16 = const()[name = string("encoder_layers_22_feed_forward1_linear1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(467032384)))]; + tensor encoder_layers_22_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_22_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(475421056)))]; + tensor linear_199_cast_fp16 = linear(bias = encoder_layers_22_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_22_feed_forward1_linear1_weight_to_fp16, x = input_1171_cast_fp16)[name = string("linear_199_cast_fp16")]; + tensor input_1175_cast_fp16 = silu(x = linear_199_cast_fp16)[name = string("input_1175_cast_fp16")]; + tensor encoder_layers_22_feed_forward1_linear2_weight_to_fp16 = const()[name = string("encoder_layers_22_feed_forward1_linear2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(475429312)))]; + tensor encoder_layers_22_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_22_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(483817984)))]; + tensor linear_200_cast_fp16 = linear(bias = encoder_layers_22_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_22_feed_forward1_linear2_weight_to_fp16, x = input_1175_cast_fp16)[name = string("linear_200_cast_fp16")]; + fp16 var_5140_to_fp16 = const()[name = string("op_5140_to_fp16"), val = fp16(0x1p-1)]; + tensor var_5141_cast_fp16 = mul(x = linear_200_cast_fp16, y = var_5140_to_fp16)[name = string("op_5141_cast_fp16")]; + tensor input_1181_cast_fp16 = add(x = input_1169_cast_fp16, y = var_5141_cast_fp16)[name = string("input_1181_cast_fp16")]; + tensor key_45_axes_0 = const()[name = string("key_45_axes_0"), val = tensor([-1])]; + tensor encoder_layers_22_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_22_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(483820096)))]; + tensor encoder_layers_22_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_22_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(483822208)))]; + tensor key_45_cast_fp16 = layer_norm(axes = key_45_axes_0, beta = encoder_layers_22_norm_self_att_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_22_norm_self_att_weight_to_fp16, x = input_1181_cast_fp16)[name = string("key_45_cast_fp16")]; + bool input_1183_interleave_0 = const()[name = string("input_1183_interleave_0"), val = bool(false)]; + tensor input_1183_cast_fp16 = concat(axis = var_67, interleave = input_1183_interleave_0, values = (cache_89_cast_fp16, key_45_cast_fp16))[name = string("input_1183_cast_fp16")]; + bool var_5169_interleave_0 = const()[name = string("op_5169_interleave_0"), val = bool(false)]; + tensor var_5169_cast_fp16 = concat(axis = var_67, interleave = var_5169_interleave_0, values = key_45_cast_fp16)[name = string("op_5169_cast_fp16")]; + tensor encoder_layers_22_self_attn_linear_q_weight_to_fp16 = const()[name = string("encoder_layers_22_self_attn_linear_q_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(483824320)))]; + tensor encoder_layers_22_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_22_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(485921536)))]; + tensor linear_201_cast_fp16 = linear(bias = encoder_layers_22_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_22_self_attn_linear_q_weight_to_fp16, x = key_45_cast_fp16)[name = string("linear_201_cast_fp16")]; + tensor var_5174 = const()[name = string("op_5174"), val = tensor([1, -1, 8, 128])]; + tensor q_133_cast_fp16 = reshape(shape = var_5174, x = linear_201_cast_fp16)[name = string("q_133_cast_fp16")]; + tensor encoder_layers_22_self_attn_linear_k_weight_to_fp16 = const()[name = string("encoder_layers_22_self_attn_linear_k_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(485923648)))]; + tensor encoder_layers_22_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_22_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(488020864)))]; + tensor linear_202_cast_fp16 = linear(bias = encoder_layers_22_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_22_self_attn_linear_k_weight_to_fp16, x = input_1183_cast_fp16)[name = string("linear_202_cast_fp16")]; + tensor var_5179 = const()[name = string("op_5179"), val = tensor([1, -1, 8, 128])]; + tensor k_89_cast_fp16 = reshape(shape = var_5179, x = linear_202_cast_fp16)[name = string("k_89_cast_fp16")]; + tensor encoder_layers_22_self_attn_linear_v_weight_to_fp16 = const()[name = string("encoder_layers_22_self_attn_linear_v_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(488022976)))]; + tensor encoder_layers_22_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_22_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(490120192)))]; + tensor linear_203_cast_fp16 = linear(bias = encoder_layers_22_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_22_self_attn_linear_v_weight_to_fp16, x = input_1183_cast_fp16)[name = string("linear_203_cast_fp16")]; + tensor var_5184 = const()[name = string("op_5184"), val = tensor([1, -1, 8, 128])]; + tensor v_45_cast_fp16 = reshape(shape = var_5184, x = linear_203_cast_fp16)[name = string("v_45_cast_fp16")]; + tensor value_53_perm_0 = const()[name = string("value_53_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_22_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_22_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(490122304)))]; + tensor var_5197_cast_fp16 = add(x = q_133_cast_fp16, y = encoder_layers_22_self_attn_pos_bias_u_to_fp16)[name = string("op_5197_cast_fp16")]; + tensor encoder_layers_22_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_22_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(490124416)))]; + tensor var_5199_cast_fp16 = add(x = q_133_cast_fp16, y = encoder_layers_22_self_attn_pos_bias_v_to_fp16)[name = string("op_5199_cast_fp16")]; + tensor q_with_bias_v_45_perm_0 = const()[name = string("q_with_bias_v_45_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_579_transpose_x_0 = const()[name = string("x_579_transpose_x_0"), val = bool(false)]; + bool x_579_transpose_y_0 = const()[name = string("x_579_transpose_y_0"), val = bool(false)]; + tensor op_5201_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(490126528))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(490326272))))[name = string("op_5201_to_fp16_quantized")]; + tensor q_with_bias_v_45_cast_fp16 = transpose(perm = q_with_bias_v_45_perm_0, x = var_5199_cast_fp16)[name = string("transpose_164")]; + tensor x_579_cast_fp16 = matmul(transpose_x = x_579_transpose_x_0, transpose_y = x_579_transpose_y_0, x = q_with_bias_v_45_cast_fp16, y = op_5201_to_fp16_quantized)[name = string("x_579_cast_fp16")]; + tensor x_581_pad_0 = const()[name = string("x_581_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_581_mode_0 = const()[name = string("x_581_mode_0"), val = string("constant")]; + fp16 const_365_to_fp16 = const()[name = string("const_365_to_fp16"), val = fp16(0x0p+0)]; + tensor x_581_cast_fp16 = pad(constant_val = const_365_to_fp16, mode = x_581_mode_0, pad = x_581_pad_0, x = x_579_cast_fp16)[name = string("x_581_cast_fp16")]; + tensor var_5209 = const()[name = string("op_5209"), val = tensor([1, 8, -1, 56])]; + tensor x_583_cast_fp16 = reshape(shape = var_5209, x = x_581_cast_fp16)[name = string("x_583_cast_fp16")]; + tensor var_5213_begin_0 = const()[name = string("op_5213_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_5213_end_0 = const()[name = string("op_5213_end_0"), val = tensor([1, 8, 196, 56])]; + tensor var_5213_end_mask_0 = const()[name = string("op_5213_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_5213_cast_fp16 = slice_by_index(begin = var_5213_begin_0, end = var_5213_end_0, end_mask = var_5213_end_mask_0, x = x_583_cast_fp16)[name = string("op_5213_cast_fp16")]; + tensor var_5214 = const()[name = string("op_5214"), val = tensor([1, 8, 56, 195])]; + tensor matrix_bd_89_cast_fp16 = reshape(shape = var_5214, x = var_5213_cast_fp16)[name = string("matrix_bd_89_cast_fp16")]; + bool matrix_ac_45_transpose_x_0 = const()[name = string("matrix_ac_45_transpose_x_0"), val = bool(false)]; + bool matrix_ac_45_transpose_y_0 = const()[name = string("matrix_ac_45_transpose_y_0"), val = bool(false)]; + tensor transpose_140_perm_0 = const()[name = string("transpose_140_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_141_perm_0 = const()[name = string("transpose_141_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_141 = transpose(perm = transpose_141_perm_0, x = k_89_cast_fp16)[name = string("transpose_162")]; + tensor transpose_140 = transpose(perm = transpose_140_perm_0, x = var_5197_cast_fp16)[name = string("transpose_163")]; + tensor matrix_ac_45_cast_fp16 = matmul(transpose_x = matrix_ac_45_transpose_x_0, transpose_y = matrix_ac_45_transpose_y_0, x = transpose_140, y = transpose_141)[name = string("matrix_ac_45_cast_fp16")]; + tensor matrix_bd_91_begin_0 = const()[name = string("matrix_bd_91_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_91_end_0 = const()[name = string("matrix_bd_91_end_0"), val = tensor([1, 8, 56, 98])]; + tensor matrix_bd_91_end_mask_0 = const()[name = string("matrix_bd_91_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_91_cast_fp16 = slice_by_index(begin = matrix_bd_91_begin_0, end = matrix_bd_91_end_0, end_mask = matrix_bd_91_end_mask_0, x = matrix_bd_89_cast_fp16)[name = string("matrix_bd_91_cast_fp16")]; + tensor var_5223_cast_fp16 = add(x = matrix_ac_45_cast_fp16, y = matrix_bd_91_cast_fp16)[name = string("op_5223_cast_fp16")]; + fp16 _inversed_scores_89_y_0_to_fp16 = const()[name = string("_inversed_scores_89_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_89_cast_fp16 = mul(x = var_5223_cast_fp16, y = _inversed_scores_89_y_0_to_fp16)[name = string("_inversed_scores_89_cast_fp16")]; + tensor scores_91_cast_fp16 = select(a = var_44_to_fp16, b = _inversed_scores_89_cast_fp16, cond = mask_11)[name = string("scores_91_cast_fp16")]; + tensor var_5229_cast_fp16 = softmax(axis = var_58, x = scores_91_cast_fp16)[name = string("op_5229_cast_fp16")]; + tensor input_1185_cast_fp16 = select(a = var_43_to_fp16, b = var_5229_cast_fp16, cond = mask_11)[name = string("input_1185_cast_fp16")]; + bool x_585_transpose_x_0 = const()[name = string("x_585_transpose_x_0"), val = bool(false)]; + bool x_585_transpose_y_0 = const()[name = string("x_585_transpose_y_0"), val = bool(false)]; + tensor value_53_cast_fp16 = transpose(perm = value_53_perm_0, x = v_45_cast_fp16)[name = string("transpose_161")]; + tensor x_585_cast_fp16 = matmul(transpose_x = x_585_transpose_x_0, transpose_y = x_585_transpose_y_0, x = input_1185_cast_fp16, y = value_53_cast_fp16)[name = string("x_585_cast_fp16")]; + tensor var_5233_perm_0 = const()[name = string("op_5233_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_5234 = const()[name = string("op_5234"), val = tensor([1, -1, 1024])]; + tensor var_5233_cast_fp16 = transpose(perm = var_5233_perm_0, x = x_585_cast_fp16)[name = string("transpose_160")]; + tensor input_1187_cast_fp16 = reshape(shape = var_5234, x = var_5233_cast_fp16)[name = string("input_1187_cast_fp16")]; + tensor encoder_layers_22_self_attn_linear_out_weight_to_fp16 = const()[name = string("encoder_layers_22_self_attn_linear_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(490326784)))]; + tensor encoder_layers_22_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_22_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(492424000)))]; + tensor linear_205_cast_fp16 = linear(bias = encoder_layers_22_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_22_self_attn_linear_out_weight_to_fp16, x = input_1187_cast_fp16)[name = string("linear_205_cast_fp16")]; + tensor input_1191_cast_fp16 = add(x = input_1181_cast_fp16, y = linear_205_cast_fp16)[name = string("input_1191_cast_fp16")]; + tensor x_589_axes_0 = const()[name = string("x_589_axes_0"), val = tensor([-1])]; + tensor encoder_layers_22_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_22_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(492426112)))]; + tensor encoder_layers_22_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_22_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(492428224)))]; + tensor x_589_cast_fp16 = layer_norm(axes = x_589_axes_0, beta = encoder_layers_22_norm_conv_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_22_norm_conv_weight_to_fp16, x = input_1191_cast_fp16)[name = string("x_589_cast_fp16")]; + tensor input_1193_perm_0 = const()[name = string("input_1193_perm_0"), val = tensor([0, 2, 1])]; + string input_1195_pad_type_0 = const()[name = string("input_1195_pad_type_0"), val = string("valid")]; + tensor input_1195_strides_0 = const()[name = string("input_1195_strides_0"), val = tensor([1])]; + tensor input_1195_pad_0 = const()[name = string("input_1195_pad_0"), val = tensor([0, 0])]; + tensor input_1195_dilations_0 = const()[name = string("input_1195_dilations_0"), val = tensor([1])]; + int32 input_1195_groups_0 = const()[name = string("input_1195_groups_0"), val = int32(1)]; + tensor encoder_layers_22_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(492430336))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(494527552))))[name = string("encoder_layers_22_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_1193_cast_fp16 = transpose(perm = input_1193_perm_0, x = x_589_cast_fp16)[name = string("transpose_159")]; + tensor input_1195_cast_fp16 = conv(dilations = input_1195_dilations_0, groups = input_1195_groups_0, pad = input_1195_pad_0, pad_type = input_1195_pad_type_0, strides = input_1195_strides_0, weight = encoder_layers_22_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_1193_cast_fp16)[name = string("input_1195_cast_fp16")]; + int32 x_591_split_num_splits_0 = const()[name = string("x_591_split_num_splits_0"), val = int32(2)]; + int32 x_591_split_axis_0 = const()[name = string("x_591_split_axis_0"), val = int32(1)]; + tensor x_591_split_cast_fp16_0, tensor x_591_split_cast_fp16_1 = split(axis = x_591_split_axis_0, num_splits = x_591_split_num_splits_0, x = input_1195_cast_fp16)[name = string("x_591_split_cast_fp16")]; + tensor x_591_split_1_sigmoid_cast_fp16 = sigmoid(x = x_591_split_cast_fp16_1)[name = string("x_591_split_1_sigmoid_cast_fp16")]; + tensor x_591_cast_fp16 = mul(x = x_591_split_cast_fp16_0, y = x_591_split_1_sigmoid_cast_fp16)[name = string("x_591_cast_fp16")]; + tensor input_1197_cast_fp16 = select(a = var_43_to_fp16, b = x_591_cast_fp16, cond = var_574)[name = string("input_1197_cast_fp16")]; + bool new_x_91_interleave_0 = const()[name = string("new_x_91_interleave_0"), val = bool(false)]; + tensor new_x_91_cast_fp16 = concat(axis = var_58, interleave = new_x_91_interleave_0, values = (cache_91_cast_fp16, input_1197_cast_fp16))[name = string("new_x_91_cast_fp16")]; + tensor var_5273_begin_0 = const()[name = string("op_5273_begin_0"), val = tensor([0, 0, 56])]; + tensor var_5273_end_0 = const()[name = string("op_5273_end_0"), val = tensor([1, 1024, 64])]; + tensor var_5273_end_mask_0 = const()[name = string("op_5273_end_mask_0"), val = tensor([true, true, true])]; + tensor var_5273_cast_fp16 = slice_by_index(begin = var_5273_begin_0, end = var_5273_end_0, end_mask = var_5273_end_mask_0, x = new_x_91_cast_fp16)[name = string("op_5273_cast_fp16")]; + string x_593_pad_type_0 = const()[name = string("x_593_pad_type_0"), val = string("valid")]; + int32 x_593_groups_0 = const()[name = string("x_593_groups_0"), val = int32(1024)]; + tensor x_593_strides_0 = const()[name = string("x_593_strides_0"), val = tensor([1])]; + tensor x_593_pad_0 = const()[name = string("x_593_pad_0"), val = tensor([0, 0])]; + tensor x_593_dilations_0 = const()[name = string("x_593_dilations_0"), val = tensor([1])]; + tensor encoder_layers_22_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(494531712))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(494540992))))[name = string("encoder_layers_22_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_593_cast_fp16 = conv(dilations = x_593_dilations_0, groups = x_593_groups_0, pad = x_593_pad_0, pad_type = x_593_pad_type_0, strides = x_593_strides_0, weight = encoder_layers_22_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_91_cast_fp16)[name = string("x_593_cast_fp16")]; + tensor input_1199_perm_0 = const()[name = string("input_1199_perm_0"), val = tensor([0, 2, 1])]; + tensor x_595_axes_0 = const()[name = string("x_595_axes_0"), val = tensor([-1])]; + tensor encoder_layers_22_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_22_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(494543104)))]; + tensor encoder_layers_22_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_22_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(494545216)))]; + tensor input_1199_cast_fp16 = transpose(perm = input_1199_perm_0, x = x_593_cast_fp16)[name = string("transpose_158")]; + tensor x_595_cast_fp16 = layer_norm(axes = x_595_axes_0, beta = encoder_layers_22_conv_batch_norm_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_22_conv_batch_norm_weight_to_fp16, x = input_1199_cast_fp16)[name = string("x_595_cast_fp16")]; + tensor input_1201_perm_0 = const()[name = string("input_1201_perm_0"), val = tensor([0, 2, 1])]; + tensor input_1201_cast_fp16 = transpose(perm = input_1201_perm_0, x = x_595_cast_fp16)[name = string("transpose_157")]; + tensor input_1203_cast_fp16 = silu(x = input_1201_cast_fp16)[name = string("input_1203_cast_fp16")]; + string x_597_pad_type_0 = const()[name = string("x_597_pad_type_0"), val = string("valid")]; + tensor x_597_strides_0 = const()[name = string("x_597_strides_0"), val = tensor([1])]; + tensor x_597_pad_0 = const()[name = string("x_597_pad_0"), val = tensor([0, 0])]; + tensor x_597_dilations_0 = const()[name = string("x_597_dilations_0"), val = tensor([1])]; + int32 x_597_groups_0 = const()[name = string("x_597_groups_0"), val = int32(1)]; + tensor encoder_layers_22_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(494547328))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(495595968))))[name = string("encoder_layers_22_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_597_cast_fp16 = conv(dilations = x_597_dilations_0, groups = x_597_groups_0, pad = x_597_pad_0, pad_type = x_597_pad_type_0, strides = x_597_strides_0, weight = encoder_layers_22_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_1203_cast_fp16)[name = string("x_597_cast_fp16")]; + tensor input_1205_perm_0 = const()[name = string("input_1205_perm_0"), val = tensor([0, 2, 1])]; + tensor input_1205_cast_fp16 = transpose(perm = input_1205_perm_0, x = x_597_cast_fp16)[name = string("transpose_156")]; + tensor input_1207_cast_fp16 = add(x = input_1191_cast_fp16, y = input_1205_cast_fp16)[name = string("input_1207_cast_fp16")]; + tensor input_1209_axes_0 = const()[name = string("input_1209_axes_0"), val = tensor([-1])]; + tensor encoder_layers_22_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_22_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(495598080)))]; + tensor encoder_layers_22_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_22_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(495600192)))]; + tensor input_1209_cast_fp16 = layer_norm(axes = input_1209_axes_0, beta = encoder_layers_22_norm_feed_forward2_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_22_norm_feed_forward2_weight_to_fp16, x = input_1207_cast_fp16)[name = string("input_1209_cast_fp16")]; + tensor encoder_layers_22_feed_forward2_linear1_weight_to_fp16 = const()[name = string("encoder_layers_22_feed_forward2_linear1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(495602304)))]; + tensor encoder_layers_22_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_22_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(503990976)))]; + tensor linear_206_cast_fp16 = linear(bias = encoder_layers_22_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_22_feed_forward2_linear1_weight_to_fp16, x = input_1209_cast_fp16)[name = string("linear_206_cast_fp16")]; + tensor input_1213_cast_fp16 = silu(x = linear_206_cast_fp16)[name = string("input_1213_cast_fp16")]; + tensor encoder_layers_22_feed_forward2_linear2_weight_to_fp16 = const()[name = string("encoder_layers_22_feed_forward2_linear2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(503999232)))]; + tensor encoder_layers_22_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_22_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(512387904)))]; + tensor linear_207_cast_fp16 = linear(bias = encoder_layers_22_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_22_feed_forward2_linear2_weight_to_fp16, x = input_1213_cast_fp16)[name = string("linear_207_cast_fp16")]; + fp16 var_5316_to_fp16 = const()[name = string("op_5316_to_fp16"), val = fp16(0x1p-1)]; + tensor var_5317_cast_fp16 = mul(x = linear_207_cast_fp16, y = var_5316_to_fp16)[name = string("op_5317_cast_fp16")]; + tensor input_1219_cast_fp16 = add(x = input_1207_cast_fp16, y = var_5317_cast_fp16)[name = string("input_1219_cast_fp16")]; + tensor input_1221_axes_0 = const()[name = string("input_1221_axes_0"), val = tensor([-1])]; + tensor encoder_layers_22_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_22_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(512390016)))]; + tensor encoder_layers_22_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_22_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(512392128)))]; + tensor input_1221_cast_fp16 = layer_norm(axes = input_1221_axes_0, beta = encoder_layers_22_norm_out_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_22_norm_out_weight_to_fp16, x = input_1219_cast_fp16)[name = string("input_1221_cast_fp16")]; + tensor cache_93_begin_0 = const()[name = string("cache_93_begin_0"), val = tensor([23, 0, 0, 0])]; + tensor cache_93_end_0 = const()[name = string("cache_93_end_0"), val = tensor([24, 1, 42, 1024])]; + tensor cache_93_end_mask_0 = const()[name = string("cache_93_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_93_squeeze_mask_0 = const()[name = string("cache_93_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_93_cast_fp16 = slice_by_index(begin = cache_93_begin_0, end = cache_93_end_0, end_mask = cache_93_end_mask_0, squeeze_mask = cache_93_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_93_cast_fp16")]; + tensor cache_begin_0 = const()[name = string("cache_begin_0"), val = tensor([23, 0, 0, 0])]; + tensor cache_end_0 = const()[name = string("cache_end_0"), val = tensor([24, 1, 1024, 8])]; + tensor cache_end_mask_0 = const()[name = string("cache_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_squeeze_mask_0 = const()[name = string("cache_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_cast_fp16 = slice_by_index(begin = cache_begin_0, end = cache_end_0, end_mask = cache_end_mask_0, squeeze_mask = cache_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_cast_fp16")]; + tensor input_1223_axes_0 = const()[name = string("input_1223_axes_0"), val = tensor([-1])]; + tensor encoder_layers_23_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_23_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(512394240)))]; + tensor encoder_layers_23_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_23_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(512396352)))]; + tensor input_1223_cast_fp16 = layer_norm(axes = input_1223_axes_0, beta = encoder_layers_23_norm_feed_forward1_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_23_norm_feed_forward1_weight_to_fp16, x = input_1221_cast_fp16)[name = string("input_1223_cast_fp16")]; + tensor encoder_layers_23_feed_forward1_linear1_weight_to_fp16 = const()[name = string("encoder_layers_23_feed_forward1_linear1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(512398464)))]; + tensor encoder_layers_23_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_23_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(520787136)))]; + tensor linear_208_cast_fp16 = linear(bias = encoder_layers_23_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_23_feed_forward1_linear1_weight_to_fp16, x = input_1223_cast_fp16)[name = string("linear_208_cast_fp16")]; + tensor input_1227_cast_fp16 = silu(x = linear_208_cast_fp16)[name = string("input_1227_cast_fp16")]; + tensor encoder_layers_23_feed_forward1_linear2_weight_to_fp16 = const()[name = string("encoder_layers_23_feed_forward1_linear2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(520795392)))]; + tensor encoder_layers_23_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_23_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(529184064)))]; + tensor linear_209_cast_fp16 = linear(bias = encoder_layers_23_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_23_feed_forward1_linear2_weight_to_fp16, x = input_1227_cast_fp16)[name = string("linear_209_cast_fp16")]; + fp16 var_5353_to_fp16 = const()[name = string("op_5353_to_fp16"), val = fp16(0x1p-1)]; + tensor var_5354_cast_fp16 = mul(x = linear_209_cast_fp16, y = var_5353_to_fp16)[name = string("op_5354_cast_fp16")]; + tensor input_1233_cast_fp16 = add(x = input_1221_cast_fp16, y = var_5354_cast_fp16)[name = string("input_1233_cast_fp16")]; + tensor key_axes_0 = const()[name = string("key_axes_0"), val = tensor([-1])]; + tensor encoder_layers_23_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_23_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(529186176)))]; + tensor encoder_layers_23_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_23_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(529188288)))]; + tensor key_cast_fp16 = layer_norm(axes = key_axes_0, beta = encoder_layers_23_norm_self_att_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_23_norm_self_att_weight_to_fp16, x = input_1233_cast_fp16)[name = string("key_cast_fp16")]; + bool input_1235_interleave_0 = const()[name = string("input_1235_interleave_0"), val = bool(false)]; + tensor input_1235_cast_fp16 = concat(axis = var_67, interleave = input_1235_interleave_0, values = (cache_93_cast_fp16, key_cast_fp16))[name = string("input_1235_cast_fp16")]; + bool cache_last_channel_cur_interleave_0 = const()[name = string("cache_last_channel_cur_interleave_0"), val = bool(false)]; + tensor cache_last_channel_cur_cast_fp16 = concat(axis = var_67, interleave = cache_last_channel_cur_interleave_0, values = key_cast_fp16)[name = string("cache_last_channel_cur_cast_fp16")]; + tensor encoder_layers_23_self_attn_linear_q_weight_to_fp16 = const()[name = string("encoder_layers_23_self_attn_linear_q_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(529190400)))]; + tensor encoder_layers_23_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_23_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(531287616)))]; + tensor linear_210_cast_fp16 = linear(bias = encoder_layers_23_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_23_self_attn_linear_q_weight_to_fp16, x = key_cast_fp16)[name = string("linear_210_cast_fp16")]; + tensor var_5387 = const()[name = string("op_5387"), val = tensor([1, -1, 8, 128])]; + tensor q_139_cast_fp16 = reshape(shape = var_5387, x = linear_210_cast_fp16)[name = string("q_139_cast_fp16")]; + tensor encoder_layers_23_self_attn_linear_k_weight_to_fp16 = const()[name = string("encoder_layers_23_self_attn_linear_k_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(531289728)))]; + tensor encoder_layers_23_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_23_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(533386944)))]; + tensor linear_211_cast_fp16 = linear(bias = encoder_layers_23_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_23_self_attn_linear_k_weight_to_fp16, x = input_1235_cast_fp16)[name = string("linear_211_cast_fp16")]; + tensor var_5392 = const()[name = string("op_5392"), val = tensor([1, -1, 8, 128])]; + tensor k_93_cast_fp16 = reshape(shape = var_5392, x = linear_211_cast_fp16)[name = string("k_93_cast_fp16")]; + tensor encoder_layers_23_self_attn_linear_v_weight_to_fp16 = const()[name = string("encoder_layers_23_self_attn_linear_v_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(533389056)))]; + tensor encoder_layers_23_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_23_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(535486272)))]; + tensor linear_212_cast_fp16 = linear(bias = encoder_layers_23_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_23_self_attn_linear_v_weight_to_fp16, x = input_1235_cast_fp16)[name = string("linear_212_cast_fp16")]; + tensor var_5397 = const()[name = string("op_5397"), val = tensor([1, -1, 8, 128])]; + tensor v_cast_fp16 = reshape(shape = var_5397, x = linear_212_cast_fp16)[name = string("v_cast_fp16")]; + tensor value_perm_0 = const()[name = string("value_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_23_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_23_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(535488384)))]; + tensor var_5410_cast_fp16 = add(x = q_139_cast_fp16, y = encoder_layers_23_self_attn_pos_bias_u_to_fp16)[name = string("op_5410_cast_fp16")]; + tensor encoder_layers_23_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_23_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(535490496)))]; + tensor var_5412_cast_fp16 = add(x = q_139_cast_fp16, y = encoder_layers_23_self_attn_pos_bias_v_to_fp16)[name = string("op_5412_cast_fp16")]; + tensor q_with_bias_v_perm_0 = const()[name = string("q_with_bias_v_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_605_transpose_x_0 = const()[name = string("x_605_transpose_x_0"), val = bool(false)]; + bool x_605_transpose_y_0 = const()[name = string("x_605_transpose_y_0"), val = bool(false)]; + tensor op_5414_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(535492608))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(535692352))))[name = string("op_5414_to_fp16_quantized")]; + tensor q_with_bias_v_cast_fp16 = transpose(perm = q_with_bias_v_perm_0, x = var_5412_cast_fp16)[name = string("transpose_155")]; + tensor x_605_cast_fp16 = matmul(transpose_x = x_605_transpose_x_0, transpose_y = x_605_transpose_y_0, x = q_with_bias_v_cast_fp16, y = op_5414_to_fp16_quantized)[name = string("x_605_cast_fp16")]; + tensor x_607_pad_0 = const()[name = string("x_607_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_607_mode_0 = const()[name = string("x_607_mode_0"), val = string("constant")]; + fp16 const_378_to_fp16 = const()[name = string("const_378_to_fp16"), val = fp16(0x0p+0)]; + tensor x_607_cast_fp16 = pad(constant_val = const_378_to_fp16, mode = x_607_mode_0, pad = x_607_pad_0, x = x_605_cast_fp16)[name = string("x_607_cast_fp16")]; + tensor var_5422 = const()[name = string("op_5422"), val = tensor([1, 8, -1, 56])]; + tensor x_609_cast_fp16 = reshape(shape = var_5422, x = x_607_cast_fp16)[name = string("x_609_cast_fp16")]; + tensor var_5426_begin_0 = const()[name = string("op_5426_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_5426_end_0 = const()[name = string("op_5426_end_0"), val = tensor([1, 8, 196, 56])]; + tensor var_5426_end_mask_0 = const()[name = string("op_5426_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_5426_cast_fp16 = slice_by_index(begin = var_5426_begin_0, end = var_5426_end_0, end_mask = var_5426_end_mask_0, x = x_609_cast_fp16)[name = string("op_5426_cast_fp16")]; + tensor var_5427 = const()[name = string("op_5427"), val = tensor([1, 8, 56, 195])]; + tensor matrix_bd_93_cast_fp16 = reshape(shape = var_5427, x = var_5426_cast_fp16)[name = string("matrix_bd_93_cast_fp16")]; + bool matrix_ac_transpose_x_0 = const()[name = string("matrix_ac_transpose_x_0"), val = bool(false)]; + bool matrix_ac_transpose_y_0 = const()[name = string("matrix_ac_transpose_y_0"), val = bool(false)]; + tensor transpose_142_perm_0 = const()[name = string("transpose_142_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_143_perm_0 = const()[name = string("transpose_143_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_143 = transpose(perm = transpose_143_perm_0, x = k_93_cast_fp16)[name = string("transpose_153")]; + tensor transpose_142 = transpose(perm = transpose_142_perm_0, x = var_5410_cast_fp16)[name = string("transpose_154")]; + tensor matrix_ac_cast_fp16 = matmul(transpose_x = matrix_ac_transpose_x_0, transpose_y = matrix_ac_transpose_y_0, x = transpose_142, y = transpose_143)[name = string("matrix_ac_cast_fp16")]; + tensor matrix_bd_begin_0 = const()[name = string("matrix_bd_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_end_0 = const()[name = string("matrix_bd_end_0"), val = tensor([1, 8, 56, 98])]; + tensor matrix_bd_end_mask_0 = const()[name = string("matrix_bd_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_cast_fp16 = slice_by_index(begin = matrix_bd_begin_0, end = matrix_bd_end_0, end_mask = matrix_bd_end_mask_0, x = matrix_bd_93_cast_fp16)[name = string("matrix_bd_cast_fp16")]; + tensor var_5436_cast_fp16 = add(x = matrix_ac_cast_fp16, y = matrix_bd_cast_fp16)[name = string("op_5436_cast_fp16")]; + fp16 _inversed_scores_93_y_0_to_fp16 = const()[name = string("_inversed_scores_93_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_93_cast_fp16 = mul(x = var_5436_cast_fp16, y = _inversed_scores_93_y_0_to_fp16)[name = string("_inversed_scores_93_cast_fp16")]; + tensor scores_cast_fp16 = select(a = var_44_to_fp16, b = _inversed_scores_93_cast_fp16, cond = mask_11)[name = string("scores_cast_fp16")]; + tensor var_5442_cast_fp16 = softmax(axis = var_58, x = scores_cast_fp16)[name = string("op_5442_cast_fp16")]; + tensor input_1237_cast_fp16 = select(a = var_43_to_fp16, b = var_5442_cast_fp16, cond = mask_11)[name = string("input_1237_cast_fp16")]; + bool x_611_transpose_x_0 = const()[name = string("x_611_transpose_x_0"), val = bool(false)]; + bool x_611_transpose_y_0 = const()[name = string("x_611_transpose_y_0"), val = bool(false)]; + tensor value_cast_fp16 = transpose(perm = value_perm_0, x = v_cast_fp16)[name = string("transpose_152")]; + tensor x_611_cast_fp16 = matmul(transpose_x = x_611_transpose_x_0, transpose_y = x_611_transpose_y_0, x = input_1237_cast_fp16, y = value_cast_fp16)[name = string("x_611_cast_fp16")]; + tensor var_5446_perm_0 = const()[name = string("op_5446_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_5447 = const()[name = string("op_5447"), val = tensor([1, -1, 1024])]; + tensor var_5446_cast_fp16 = transpose(perm = var_5446_perm_0, x = x_611_cast_fp16)[name = string("transpose_151")]; + tensor input_1239_cast_fp16 = reshape(shape = var_5447, x = var_5446_cast_fp16)[name = string("input_1239_cast_fp16")]; + tensor encoder_layers_23_self_attn_linear_out_weight_to_fp16 = const()[name = string("encoder_layers_23_self_attn_linear_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(535692864)))]; + tensor encoder_layers_23_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_23_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(537790080)))]; + tensor linear_214_cast_fp16 = linear(bias = encoder_layers_23_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_23_self_attn_linear_out_weight_to_fp16, x = input_1239_cast_fp16)[name = string("linear_214_cast_fp16")]; + tensor input_1243_cast_fp16 = add(x = input_1233_cast_fp16, y = linear_214_cast_fp16)[name = string("input_1243_cast_fp16")]; + tensor x_615_axes_0 = const()[name = string("x_615_axes_0"), val = tensor([-1])]; + tensor encoder_layers_23_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_23_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(537792192)))]; + tensor encoder_layers_23_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_23_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(537794304)))]; + tensor x_615_cast_fp16 = layer_norm(axes = x_615_axes_0, beta = encoder_layers_23_norm_conv_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_23_norm_conv_weight_to_fp16, x = input_1243_cast_fp16)[name = string("x_615_cast_fp16")]; + tensor input_1245_perm_0 = const()[name = string("input_1245_perm_0"), val = tensor([0, 2, 1])]; + string input_1247_pad_type_0 = const()[name = string("input_1247_pad_type_0"), val = string("valid")]; + tensor input_1247_strides_0 = const()[name = string("input_1247_strides_0"), val = tensor([1])]; + tensor input_1247_pad_0 = const()[name = string("input_1247_pad_0"), val = tensor([0, 0])]; + tensor input_1247_dilations_0 = const()[name = string("input_1247_dilations_0"), val = tensor([1])]; + int32 input_1247_groups_0 = const()[name = string("input_1247_groups_0"), val = int32(1)]; + tensor encoder_layers_23_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(537796416))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(539893632))))[name = string("encoder_layers_23_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_1245_cast_fp16 = transpose(perm = input_1245_perm_0, x = x_615_cast_fp16)[name = string("transpose_150")]; + tensor input_1247_cast_fp16 = conv(dilations = input_1247_dilations_0, groups = input_1247_groups_0, pad = input_1247_pad_0, pad_type = input_1247_pad_type_0, strides = input_1247_strides_0, weight = encoder_layers_23_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_1245_cast_fp16)[name = string("input_1247_cast_fp16")]; + int32 x_617_split_num_splits_0 = const()[name = string("x_617_split_num_splits_0"), val = int32(2)]; + int32 x_617_split_axis_0 = const()[name = string("x_617_split_axis_0"), val = int32(1)]; + tensor x_617_split_cast_fp16_0, tensor x_617_split_cast_fp16_1 = split(axis = x_617_split_axis_0, num_splits = x_617_split_num_splits_0, x = input_1247_cast_fp16)[name = string("x_617_split_cast_fp16")]; + tensor x_617_split_1_sigmoid_cast_fp16 = sigmoid(x = x_617_split_cast_fp16_1)[name = string("x_617_split_1_sigmoid_cast_fp16")]; + tensor x_617_cast_fp16 = mul(x = x_617_split_cast_fp16_0, y = x_617_split_1_sigmoid_cast_fp16)[name = string("x_617_cast_fp16")]; + tensor input_1249_cast_fp16 = select(a = var_43_to_fp16, b = x_617_cast_fp16, cond = var_574)[name = string("input_1249_cast_fp16")]; + bool new_x_interleave_0 = const()[name = string("new_x_interleave_0"), val = bool(false)]; + tensor new_x_cast_fp16 = concat(axis = var_58, interleave = new_x_interleave_0, values = (cache_cast_fp16, input_1249_cast_fp16))[name = string("new_x_cast_fp16")]; + tensor cache_last_time_cur_begin_0 = const()[name = string("cache_last_time_cur_begin_0"), val = tensor([0, 0, 56])]; + tensor cache_last_time_cur_end_0 = const()[name = string("cache_last_time_cur_end_0"), val = tensor([1, 1024, 64])]; + tensor cache_last_time_cur_end_mask_0 = const()[name = string("cache_last_time_cur_end_mask_0"), val = tensor([true, true, true])]; + tensor cache_last_time_cur_cast_fp16 = slice_by_index(begin = cache_last_time_cur_begin_0, end = cache_last_time_cur_end_0, end_mask = cache_last_time_cur_end_mask_0, x = new_x_cast_fp16)[name = string("cache_last_time_cur_cast_fp16")]; + string x_619_pad_type_0 = const()[name = string("x_619_pad_type_0"), val = string("valid")]; + int32 x_619_groups_0 = const()[name = string("x_619_groups_0"), val = int32(1024)]; + tensor x_619_strides_0 = const()[name = string("x_619_strides_0"), val = tensor([1])]; + tensor x_619_pad_0 = const()[name = string("x_619_pad_0"), val = tensor([0, 0])]; + tensor x_619_dilations_0 = const()[name = string("x_619_dilations_0"), val = tensor([1])]; + tensor encoder_layers_23_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(539897792))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(539907072))))[name = string("encoder_layers_23_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_619_cast_fp16 = conv(dilations = x_619_dilations_0, groups = x_619_groups_0, pad = x_619_pad_0, pad_type = x_619_pad_type_0, strides = x_619_strides_0, weight = encoder_layers_23_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_cast_fp16)[name = string("x_619_cast_fp16")]; + tensor input_1251_perm_0 = const()[name = string("input_1251_perm_0"), val = tensor([0, 2, 1])]; + tensor x_621_axes_0 = const()[name = string("x_621_axes_0"), val = tensor([-1])]; + tensor encoder_layers_23_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_23_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(539909184)))]; + tensor encoder_layers_23_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_23_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(539911296)))]; + tensor input_1251_cast_fp16 = transpose(perm = input_1251_perm_0, x = x_619_cast_fp16)[name = string("transpose_149")]; + tensor x_621_cast_fp16 = layer_norm(axes = x_621_axes_0, beta = encoder_layers_23_conv_batch_norm_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_23_conv_batch_norm_weight_to_fp16, x = input_1251_cast_fp16)[name = string("x_621_cast_fp16")]; + tensor input_1253_perm_0 = const()[name = string("input_1253_perm_0"), val = tensor([0, 2, 1])]; + tensor input_1253_cast_fp16 = transpose(perm = input_1253_perm_0, x = x_621_cast_fp16)[name = string("transpose_148")]; + tensor input_1255_cast_fp16 = silu(x = input_1253_cast_fp16)[name = string("input_1255_cast_fp16")]; + string x_623_pad_type_0 = const()[name = string("x_623_pad_type_0"), val = string("valid")]; + tensor x_623_strides_0 = const()[name = string("x_623_strides_0"), val = tensor([1])]; + tensor x_623_pad_0 = const()[name = string("x_623_pad_0"), val = tensor([0, 0])]; + tensor x_623_dilations_0 = const()[name = string("x_623_dilations_0"), val = tensor([1])]; + int32 x_623_groups_0 = const()[name = string("x_623_groups_0"), val = int32(1)]; + tensor encoder_layers_23_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(539913408))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(540962048))))[name = string("encoder_layers_23_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_623_cast_fp16 = conv(dilations = x_623_dilations_0, groups = x_623_groups_0, pad = x_623_pad_0, pad_type = x_623_pad_type_0, strides = x_623_strides_0, weight = encoder_layers_23_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_1255_cast_fp16)[name = string("x_623_cast_fp16")]; + tensor input_1257_perm_0 = const()[name = string("input_1257_perm_0"), val = tensor([0, 2, 1])]; + tensor input_1257_cast_fp16 = transpose(perm = input_1257_perm_0, x = x_623_cast_fp16)[name = string("transpose_147")]; + tensor input_1259_cast_fp16 = add(x = input_1243_cast_fp16, y = input_1257_cast_fp16)[name = string("input_1259_cast_fp16")]; + tensor input_1261_axes_0 = const()[name = string("input_1261_axes_0"), val = tensor([-1])]; + tensor encoder_layers_23_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_23_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(540964160)))]; + tensor encoder_layers_23_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_23_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(540966272)))]; + tensor input_1261_cast_fp16 = layer_norm(axes = input_1261_axes_0, beta = encoder_layers_23_norm_feed_forward2_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_23_norm_feed_forward2_weight_to_fp16, x = input_1259_cast_fp16)[name = string("input_1261_cast_fp16")]; + tensor encoder_layers_23_feed_forward2_linear1_weight_to_fp16 = const()[name = string("encoder_layers_23_feed_forward2_linear1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(540968384)))]; + tensor encoder_layers_23_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_23_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(549357056)))]; + tensor linear_215_cast_fp16 = linear(bias = encoder_layers_23_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_23_feed_forward2_linear1_weight_to_fp16, x = input_1261_cast_fp16)[name = string("linear_215_cast_fp16")]; + tensor input_1265_cast_fp16 = silu(x = linear_215_cast_fp16)[name = string("input_1265_cast_fp16")]; + tensor encoder_layers_23_feed_forward2_linear2_weight_to_fp16 = const()[name = string("encoder_layers_23_feed_forward2_linear2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(549365312)))]; + tensor encoder_layers_23_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_23_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(557753984)))]; + tensor linear_216_cast_fp16 = linear(bias = encoder_layers_23_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_23_feed_forward2_linear2_weight_to_fp16, x = input_1265_cast_fp16)[name = string("linear_216_cast_fp16")]; + fp16 var_5529_to_fp16 = const()[name = string("op_5529_to_fp16"), val = fp16(0x1p-1)]; + tensor var_5530_cast_fp16 = mul(x = linear_216_cast_fp16, y = var_5529_to_fp16)[name = string("op_5530_cast_fp16")]; + tensor input_1271_cast_fp16 = add(x = input_1259_cast_fp16, y = var_5530_cast_fp16)[name = string("input_1271_cast_fp16")]; + tensor audio_signal_axes_0 = const()[name = string("audio_signal_axes_0"), val = tensor([-1])]; + tensor encoder_layers_23_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_23_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(557756096)))]; + tensor encoder_layers_23_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_23_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(557758208)))]; + tensor audio_signal_cast_fp16 = layer_norm(axes = audio_signal_axes_0, beta = encoder_layers_23_norm_out_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_23_norm_out_weight_to_fp16, x = input_1271_cast_fp16)[name = string("audio_signal_cast_fp16")]; + int32 obj_5_axis_0 = const()[name = string("obj_5_axis_0"), val = int32(0)]; + tensor obj_5_cast_fp16 = stack(axis = obj_5_axis_0, values = (var_483_cast_fp16, var_696_cast_fp16, var_909_cast_fp16, var_1122_cast_fp16, var_1335_cast_fp16, var_1548_cast_fp16, var_1761_cast_fp16, var_1974_cast_fp16, var_2187_cast_fp16, var_2400_cast_fp16, var_2613_cast_fp16, var_2826_cast_fp16, var_3039_cast_fp16, var_3252_cast_fp16, var_3465_cast_fp16, var_3678_cast_fp16, var_3891_cast_fp16, var_4104_cast_fp16, var_4317_cast_fp16, var_4530_cast_fp16, var_4743_cast_fp16, var_4956_cast_fp16, var_5169_cast_fp16, cache_last_channel_cur_cast_fp16))[name = string("obj_5_cast_fp16")]; + int32 obj_7_axis_0 = const()[name = string("obj_7_axis_0"), val = int32(0)]; + tensor obj_7_cast_fp16 = stack(axis = obj_7_axis_0, values = (var_587_cast_fp16, var_800_cast_fp16, var_1013_cast_fp16, var_1226_cast_fp16, var_1439_cast_fp16, var_1652_cast_fp16, var_1865_cast_fp16, var_2078_cast_fp16, var_2291_cast_fp16, var_2504_cast_fp16, var_2717_cast_fp16, var_2930_cast_fp16, var_3143_cast_fp16, var_3356_cast_fp16, var_3569_cast_fp16, var_3782_cast_fp16, var_3995_cast_fp16, var_4208_cast_fp16, var_4421_cast_fp16, var_4634_cast_fp16, var_4847_cast_fp16, var_5060_cast_fp16, var_5273_cast_fp16, cache_last_time_cur_cast_fp16))[name = string("obj_7_cast_fp16")]; + tensor var_5546 = add(x = cache_len, y = max_audio_length_1)[name = string("op_5546")]; + string var_5546_promoted_to_fp16_dtype_0 = const()[name = string("op_5546_promoted_to_fp16_dtype_0"), val = string("fp16")]; + fp16 const_384_to_fp16 = const()[name = string("const_384_to_fp16"), val = fp16(-inf)]; + fp16 var_48_promoted_to_fp16 = const()[name = string("op_48_promoted_to_fp16"), val = fp16(0x1.5p+5)]; + tensor var_5546_to_fp16 = cast(dtype = var_5546_promoted_to_fp16_dtype_0, x = var_5546)[name = string("cast_6")]; + tensor clip_1_cast_fp16 = clip(alpha = const_384_to_fp16, beta = var_48_promoted_to_fp16, x = var_5546_to_fp16)[name = string("clip_1_cast_fp16")]; + tensor var_5568_begin_0 = const()[name = string("op_5568_begin_0"), val = tensor([0, 0, 14, 0])]; + tensor var_5568_end_0 = const()[name = string("op_5568_end_0"), val = tensor([24, 1, 56, 1024])]; + tensor var_5568_end_mask_0 = const()[name = string("op_5568_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_5568_cast_fp16 = slice_by_index(begin = var_5568_begin_0, end = var_5568_end_0, end_mask = var_5568_end_mask_0, x = obj_5_cast_fp16)[name = string("op_5568_cast_fp16")]; + int32 var_5588_one_hot_vector_size_0 = const()[name = string("op_5588_one_hot_vector_size_0"), val = int32(128)]; + int32 var_5588_axis_0 = const()[name = string("op_5588_axis_0"), val = int32(-1)]; + int32 var_5588_on_value_0 = const()[name = string("op_5588_on_value_0"), val = int32(1)]; + int32 var_5588_off_value_0 = const()[name = string("op_5588_off_value_0"), val = int32(0)]; + tensor var_5588 = one_hot(axis = var_5588_axis_0, indices = prompt_id, off_value = var_5588_off_value_0, on_value = var_5588_on_value_0, one_hot_vector_size = var_5588_one_hot_vector_size_0)[name = string("op_5588")]; + tensor var_5591_axes_0 = const()[name = string("op_5591_axes_0"), val = tensor([1])]; + string cast_245_to_fp16_dtype_0 = const()[name = string("cast_245_to_fp16_dtype_0"), val = string("fp16")]; + tensor var_5588_to_fp16 = cast(dtype = cast_245_to_fp16_dtype_0, x = var_5588)[name = string("cast_5")]; + tensor var_5591_cast_fp16 = expand_dims(axes = var_5591_axes_0, x = var_5588_to_fp16)[name = string("op_5591_cast_fp16")]; + tensor one_hot_reps_0 = const()[name = string("one_hot_reps_0"), val = tensor([1, 56, 1])]; + tensor one_hot_cast_fp16 = tile(reps = one_hot_reps_0, x = var_5591_cast_fp16)[name = string("one_hot_cast_fp16")]; + int32 var_5600 = const()[name = string("op_5600"), val = int32(-1)]; + bool input_1273_interleave_0 = const()[name = string("input_1273_interleave_0"), val = bool(false)]; + tensor input_1273_cast_fp16 = concat(axis = var_5600, interleave = input_1273_interleave_0, values = (audio_signal_cast_fp16, one_hot_cast_fp16))[name = string("input_1273_cast_fp16")]; + tensor prompt_kernel_0_weight_to_fp16 = const()[name = string("prompt_kernel_0_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(557760320)))]; + tensor prompt_kernel_0_bias_to_fp16 = const()[name = string("prompt_kernel_0_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(562478976)))]; + tensor linear_217_cast_fp16 = linear(bias = prompt_kernel_0_bias_to_fp16, weight = prompt_kernel_0_weight_to_fp16, x = input_1273_cast_fp16)[name = string("linear_217_cast_fp16")]; + tensor input_cast_fp16 = relu(x = linear_217_cast_fp16)[name = string("input_cast_fp16")]; + tensor prompt_kernel_2_weight_to_fp16 = const()[name = string("prompt_kernel_2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(562483136)))]; + tensor prompt_kernel_2_bias_to_fp16 = const()[name = string("prompt_kernel_2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(566677504)))]; + tensor linear_218_cast_fp16 = linear(bias = prompt_kernel_2_bias_to_fp16, weight = prompt_kernel_2_weight_to_fp16, x = input_cast_fp16)[name = string("linear_218_cast_fp16")]; + tensor var_5613_perm_0 = const()[name = string("op_5613_perm_0"), val = tensor([0, 2, 1])]; + string var_5613_cast_fp16_to_fp32_dtype_0 = const()[name = string("op_5613_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + string cast_246_dtype_0 = const()[name = string("cast_246_dtype_0"), val = string("int32")]; + tensor var_5621_perm_0 = const()[name = string("op_5621_perm_0"), val = tensor([1, 0, 2, 3])]; + string var_5621_cast_fp16_to_fp32_dtype_0 = const()[name = string("op_5621_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + tensor var_5624_perm_0 = const()[name = string("op_5624_perm_0"), val = tensor([1, 0, 2, 3])]; + string var_5624_cast_fp16_to_fp32_dtype_0 = const()[name = string("op_5624_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + string cast_247_dtype_0 = const()[name = string("cast_247_dtype_0"), val = string("int32")]; + tensor cache_len_out = cast(dtype = cast_247_dtype_0, x = clip_1_cast_fp16)[name = string("cast_0")]; + tensor var_5624_cast_fp16 = transpose(perm = var_5624_perm_0, x = obj_7_cast_fp16)[name = string("transpose_144")]; + tensor cache_time_out = cast(dtype = var_5624_cast_fp16_to_fp32_dtype_0, x = var_5624_cast_fp16)[name = string("cast_1")]; + tensor var_5621_cast_fp16 = transpose(perm = var_5621_perm_0, x = var_5568_cast_fp16)[name = string("transpose_145")]; + tensor cache_channel_out = cast(dtype = var_5621_cast_fp16_to_fp32_dtype_0, x = var_5621_cast_fp16)[name = string("cast_2")]; + tensor encoded_length = cast(dtype = cast_246_dtype_0, x = clip_0_cast_fp16)[name = string("cast_3")]; + tensor var_5613_cast_fp16 = transpose(perm = var_5613_perm_0, x = linear_218_cast_fp16)[name = string("transpose_146")]; + tensor encoded = cast(dtype = var_5613_cast_fp16_to_fp32_dtype_0, x = var_5613_cast_fp16)[name = string("cast_4")]; + } -> (encoded, encoded_length, cache_channel_out, cache_time_out, cache_len_out); +} \ No newline at end of file diff --git a/it/4480ms/encoder.mlmodelc/weights/weight.bin b/it/4480ms/encoder.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..e770743e8f69de769852fc42baf0304b71ade176 --- /dev/null +++ b/it/4480ms/encoder.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6a88ef3a70bab53d01290a8c9d3f29c171843a29356f93213384700ab3525ab0 +size 566679616 diff --git a/it/4480ms/encoder.mlpackage/Data/com.apple.CoreML/model.mlmodel b/it/4480ms/encoder.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..919b21f24dae459f97bb5fc50c99482ee289fbc1 --- /dev/null +++ b/it/4480ms/encoder.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b9f318539d59a79e6726cf6511a8bf6df036ffc5e1efccfcac77e99cc94591ba +size 793485 diff --git a/it/4480ms/encoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/it/4480ms/encoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..e770743e8f69de769852fc42baf0304b71ade176 --- /dev/null +++ b/it/4480ms/encoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6a88ef3a70bab53d01290a8c9d3f29c171843a29356f93213384700ab3525ab0 +size 566679616 diff --git a/it/4480ms/encoder.mlpackage/Manifest.json b/it/4480ms/encoder.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..335e50da2f499f257db872690cc2898e328624b5 --- /dev/null +++ b/it/4480ms/encoder.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "20F6F84B-098E-4410-8D1D-975A130376A0": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + }, + "BF510505-EFFA-4C71-924F-46A668751948": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + } + }, + "rootModelIdentifier": "20F6F84B-098E-4410-8D1D-975A130376A0" +} diff --git a/it/4480ms/joint.mlmodelc/analytics/coremldata.bin b/it/4480ms/joint.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..ee72f6da06176879d781b1a46f441fa01048d662 --- /dev/null +++ b/it/4480ms/joint.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cae47ef3b8cb372625305f4cdd31159b5ce56be1470c94613ff946be342f6d7a +size 243 diff --git a/it/4480ms/joint.mlmodelc/coremldata.bin b/it/4480ms/joint.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..3768aff44e7aaca9bce1d363a00f6a10df573c08 --- /dev/null +++ b/it/4480ms/joint.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6830391eb54a5590fa90d0de9dddd023f149a1d6dc2f13217287be524243de43 +size 341 diff --git a/it/4480ms/joint.mlmodelc/model.mil b/it/4480ms/joint.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..bfde40ec94bf61746424d2d3e196a4fba198de2d --- /dev/null +++ b/it/4480ms/joint.mlmodelc/model.mil @@ -0,0 +1,31 @@ +program(1.3) +[buildInfo = dict({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.5.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.3.0"}})] +{ + func main(tensor decoder, tensor encoder) { + tensor input_1_perm_0 = const()[name = string("input_1_perm_0"), val = tensor([0, 2, 1])]; + string encoder_to_fp16_dtype_0 = const()[name = string("encoder_to_fp16_dtype_0"), val = string("fp16")]; + tensor input_3_perm_0 = const()[name = string("input_3_perm_0"), val = tensor([0, 2, 1])]; + string decoder_to_fp16_dtype_0 = const()[name = string("decoder_to_fp16_dtype_0"), val = string("fp16")]; + tensor module_enc_weight_to_fp16 = const()[name = string("module_enc_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))]; + tensor module_enc_bias_to_fp16 = const()[name = string("module_enc_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1310848)))]; + tensor encoder_to_fp16 = cast(dtype = encoder_to_fp16_dtype_0, x = encoder)[name = string("cast_2")]; + tensor input_1_cast_fp16 = transpose(perm = input_1_perm_0, x = encoder_to_fp16)[name = string("transpose_1")]; + tensor linear_0_cast_fp16 = linear(bias = module_enc_bias_to_fp16, weight = module_enc_weight_to_fp16, x = input_1_cast_fp16)[name = string("linear_0_cast_fp16")]; + tensor module_pred_weight_to_fp16 = const()[name = string("module_pred_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1312192)))]; + tensor module_pred_bias_to_fp16 = const()[name = string("module_pred_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2131456)))]; + tensor decoder_to_fp16 = cast(dtype = decoder_to_fp16_dtype_0, x = decoder)[name = string("cast_1")]; + tensor input_3_cast_fp16 = transpose(perm = input_3_perm_0, x = decoder_to_fp16)[name = string("transpose_0")]; + tensor linear_1_cast_fp16 = linear(bias = module_pred_bias_to_fp16, weight = module_pred_weight_to_fp16, x = input_3_cast_fp16)[name = string("linear_1_cast_fp16")]; + tensor var_23_axes_0 = const()[name = string("op_23_axes_0"), val = tensor([2])]; + tensor var_23_cast_fp16 = expand_dims(axes = var_23_axes_0, x = linear_0_cast_fp16)[name = string("op_23_cast_fp16")]; + tensor var_25_axes_0 = const()[name = string("op_25_axes_0"), val = tensor([1])]; + tensor var_25_cast_fp16 = expand_dims(axes = var_25_axes_0, x = linear_1_cast_fp16)[name = string("op_25_cast_fp16")]; + tensor input_5_cast_fp16 = add(x = var_23_cast_fp16, y = var_25_cast_fp16)[name = string("input_5_cast_fp16")]; + tensor input_7_cast_fp16 = relu(x = input_5_cast_fp16)[name = string("input_7_cast_fp16")]; + tensor module_joint_net_2_weight_to_fp16 = const()[name = string("module_joint_net_2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2132800)))]; + tensor module_joint_net_2_bias_to_fp16 = const()[name = string("module_joint_net_2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3164544)))]; + tensor linear_2_cast_fp16 = linear(bias = module_joint_net_2_bias_to_fp16, weight = module_joint_net_2_weight_to_fp16, x = input_7_cast_fp16)[name = string("linear_2_cast_fp16")]; + string linear_2_cast_fp16_to_fp32_dtype_0 = const()[name = string("linear_2_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + tensor logits = cast(dtype = linear_2_cast_fp16_to_fp32_dtype_0, x = linear_2_cast_fp16)[name = string("cast_0")]; + } -> (logits); +} \ No newline at end of file diff --git a/it/4480ms/joint.mlmodelc/weights/weight.bin b/it/4480ms/joint.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..3673a1d2a1060ecda8626c76ffacdfd89c5f00c1 --- /dev/null +++ b/it/4480ms/joint.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1740c99cbe34ebeaa0163c8421135b4586df09960ef07fe02abb2a94b5693411 +size 3166220 diff --git a/it/4480ms/joint.mlpackage/Data/com.apple.CoreML/model.mlmodel b/it/4480ms/joint.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..f7f468ddd814131e36b8af9ed7a3358576bffcf0 --- /dev/null +++ b/it/4480ms/joint.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d902ae83b8a93f0f4240a4a6939466dbd1a6b2291f1615d81d7ac26d9115bc23 +size 4486 diff --git a/it/4480ms/joint.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/it/4480ms/joint.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..3673a1d2a1060ecda8626c76ffacdfd89c5f00c1 --- /dev/null +++ b/it/4480ms/joint.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1740c99cbe34ebeaa0163c8421135b4586df09960ef07fe02abb2a94b5693411 +size 3166220 diff --git a/it/4480ms/joint.mlpackage/Manifest.json b/it/4480ms/joint.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..fea9f1ee7eee62ace96d28134fe38a74b32b40c9 --- /dev/null +++ b/it/4480ms/joint.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "6D4EAD3B-A17D-4807-8572-3E0A513C0C7E": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + }, + "8242AF76-3606-4A98-8B3D-E0BB730201A4": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + } + }, + "rootModelIdentifier": "6D4EAD3B-A17D-4807-8572-3E0A513C0C7E" +} diff --git a/it/4480ms/joint_noencproj_batched.mlmodelc/analytics/coremldata.bin b/it/4480ms/joint_noencproj_batched.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..d2ffa5c9579dd51ce490b9c8fe47e3c27a05c723 --- /dev/null +++ b/it/4480ms/joint_noencproj_batched.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:24510a0ca90eea355a0e446ba54d63004a3f2c51d30eb3e415e4af938ac1acba +size 243 diff --git a/it/4480ms/joint_noencproj_batched.mlmodelc/coremldata.bin b/it/4480ms/joint_noencproj_batched.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..ea11e14b407bfea4bd50e9ff32006fcc1f4efa3f --- /dev/null +++ b/it/4480ms/joint_noencproj_batched.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ae7fee916e3edf31504829d2c651dbf0173b4c31ecca79c37587d988436a3faf +size 406 diff --git a/it/4480ms/joint_noencproj_batched.mlmodelc/model.mil b/it/4480ms/joint_noencproj_batched.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..58942374262618031d39a52f7a009b81c7f24c24 --- /dev/null +++ b/it/4480ms/joint_noencproj_batched.mlmodelc/model.mil @@ -0,0 +1,26 @@ +program(1.3) +[buildInfo = dict({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.10.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor decoder, tensor encoder_proj) { + tensor input_1_perm_0 = const()[name = string("input_1_perm_0"), val = tensor([0, 2, 1])]; + string decoder_to_fp16_dtype_0 = const()[name = string("decoder_to_fp16_dtype_0"), val = string("fp16")]; + tensor joint_module_pred_weight_to_fp16 = const()[name = string("joint_module_pred_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))]; + tensor joint_module_pred_bias_to_fp16 = const()[name = string("joint_module_pred_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(819328)))]; + tensor decoder_to_fp16 = cast(dtype = decoder_to_fp16_dtype_0, x = decoder)[name = string("cast_2")]; + tensor input_1_cast_fp16 = transpose(perm = input_1_perm_0, x = decoder_to_fp16)[name = string("transpose_0")]; + tensor linear_0_cast_fp16 = linear(bias = joint_module_pred_bias_to_fp16, weight = joint_module_pred_weight_to_fp16, x = input_1_cast_fp16)[name = string("linear_0_cast_fp16")]; + tensor var_15_axes_0 = const()[name = string("op_15_axes_0"), val = tensor([2])]; + string encoder_proj_to_fp16_dtype_0 = const()[name = string("encoder_proj_to_fp16_dtype_0"), val = string("fp16")]; + tensor encoder_proj_to_fp16 = cast(dtype = encoder_proj_to_fp16_dtype_0, x = encoder_proj)[name = string("cast_1")]; + tensor var_15_cast_fp16 = expand_dims(axes = var_15_axes_0, x = encoder_proj_to_fp16)[name = string("op_15_cast_fp16")]; + tensor var_17_axes_0 = const()[name = string("op_17_axes_0"), val = tensor([1])]; + tensor var_17_cast_fp16 = expand_dims(axes = var_17_axes_0, x = linear_0_cast_fp16)[name = string("op_17_cast_fp16")]; + tensor input_3_cast_fp16 = add(x = var_15_cast_fp16, y = var_17_cast_fp16)[name = string("input_3_cast_fp16")]; + tensor input_5_cast_fp16 = relu(x = input_3_cast_fp16)[name = string("input_5_cast_fp16")]; + tensor joint_module_joint_net_2_weight_to_fp16 = const()[name = string("joint_module_joint_net_2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(820672)))]; + tensor joint_module_joint_net_2_bias_to_fp16 = const()[name = string("joint_module_joint_net_2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1852416)))]; + tensor linear_1_cast_fp16 = linear(bias = joint_module_joint_net_2_bias_to_fp16, weight = joint_module_joint_net_2_weight_to_fp16, x = input_5_cast_fp16)[name = string("linear_1_cast_fp16")]; + string linear_1_cast_fp16_to_fp32_dtype_0 = const()[name = string("linear_1_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + tensor logits = cast(dtype = linear_1_cast_fp16_to_fp32_dtype_0, x = linear_1_cast_fp16)[name = string("cast_0")]; + } -> (logits); +} \ No newline at end of file diff --git a/it/4480ms/joint_noencproj_batched.mlmodelc/weights/weight.bin b/it/4480ms/joint_noencproj_batched.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..40ceadd152241059aa378e2ddb6cc9f649e0b59c --- /dev/null +++ b/it/4480ms/joint_noencproj_batched.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cd83e82dcfec315f28c8a8872b0d7f22e668a2c485821de86a0379ae2b3864ad +size 1854092 diff --git a/it/4480ms/joint_noencproj_batched.mlpackage/Data/com.apple.CoreML/model.mlmodel b/it/4480ms/joint_noencproj_batched.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..49f1fd816a2725ab32d473dfe600f911656f675b --- /dev/null +++ b/it/4480ms/joint_noencproj_batched.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:26a19f2e269811070e57eda978e8226ca9dcca6177b46b95f916538b5f815716 +size 3839 diff --git a/it/4480ms/joint_noencproj_batched.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/it/4480ms/joint_noencproj_batched.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..40ceadd152241059aa378e2ddb6cc9f649e0b59c --- /dev/null +++ b/it/4480ms/joint_noencproj_batched.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cd83e82dcfec315f28c8a8872b0d7f22e668a2c485821de86a0379ae2b3864ad +size 1854092 diff --git a/it/4480ms/joint_noencproj_batched.mlpackage/Manifest.json b/it/4480ms/joint_noencproj_batched.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..e3857a03022f3cec16823ed1d8a741fee56cfce4 --- /dev/null +++ b/it/4480ms/joint_noencproj_batched.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "4233CE8E-FB95-4FF9-BCD8-2A834D55C580": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + }, + "96E0F26C-90DC-49EE-B510-D0FB3FC812CC": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + } + }, + "rootModelIdentifier": "96E0F26C-90DC-49EE-B510-D0FB3FC812CC" +} diff --git a/it/4480ms/metadata.json b/it/4480ms/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..76e8991aa36387cfce7557cdc45b7d746cf53ddf --- /dev/null +++ b/it/4480ms/metadata.json @@ -0,0 +1,198 @@ +{ + "model": "nvidia/nemotron-asr-streaming-multilingual-0.6b", + "model_class": "nemo.collections.asr.models.rnnt_bpe_models_prompt.EncDecRNNTBPEModelWithPrompt", + "sample_rate": 16000, + "mel_features": 128, + "chunk_mel_frames": 448, + "pre_encode_cache": 9, + "total_mel_frames": 457, + "att_context_size": [ + 42, + 13 + ], + "vocab_size": 805, + "blank_idx": 805, + "vocab_pruned": true, + "vocab_pruned_original_size": 13087, + "cache_channel_shape": [ + 1, + 24, + 42, + 1024 + ], + "cache_time_shape": [ + 1, + 24, + 1024, + 8 + ], + "decoder_hidden": 640, + "decoder_layers": 2, + "encoder_dim": 1024, + "num_prompts": 128, + "prompt_dictionary": { + "af-ZA": 54, + "am-ET": 49, + "ar": 7, + "ar-AR": 7, + "auto": 101, + "ay-BO": 81, + "az-AZ": 66, + "bg": 30, + "bg-BG": 30, + "bn-IN": 36, + "cs": 22, + "cs-CZ": 22, + "da": 25, + "da-DK": 25, + "de": 9, + "de-DE": 9, + "el": 21, + "el-GR": 21, + "en": 0, + "en-GB": 1, + "en-US": 0, + "enGB": 1, + "es": 3, + "es-ES": 2, + "es-US": 3, + "esES": 2, + "et": 60, + "et-EE": 60, + "fa-IR": 38, + "fi": 26, + "fi-FI": 26, + "fr": 8, + "fr-CA": 100, + "fr-FR": 8, + "gn-PY": 82, + "gu-IN": 42, + "ha-NG": 50, + "haw-US": 97, + "he-IL": 64, + "hi": 6, + "hi-HI": 6, + "hi-IN": 6, + "hr": 29, + "hr-HR": 29, + "hu": 23, + "hu-HU": 23, + "hy-AM": 68, + "id-ID": 34, + "ig-NG": 53, + "it": 15, + "it-IT": 15, + "ja-JA": 10, + "ja-JP": 10, + "ka-GE": 67, + "km-KH": 47, + "kn-IN": 43, + "ko": 14, + "ko-KO": 14, + "ko-KR": 14, + "ku-TR": 65, + "ky-KG": 71, + "ln-CD": 58, + "lt": 31, + "lt-LT": 31, + "lv": 61, + "lv-LV": 61, + "mi-NZ": 96, + "ml-IN": 44, + "mr-IN": 41, + "ms-MY": 35, + "mt-MT": 102, + "nah-MX": 83, + "nb": 103, + "nb-NO": 103, + "ne-NP": 46, + "nl": 16, + "nl-NL": 16, + "nn": 104, + "nn-NO": 104, + "no": 27, + "no-NO": 27, + "ny-MW": 57, + "or-KE": 59, + "pl": 17, + "pl-PL": 17, + "pt": 13, + "pt-BR": 12, + "pt-PT": 13, + "qu-PE": 80, + "ro": 20, + "ro-RO": 20, + "ru": 11, + "ru-RU": 11, + "rw-RW": 55, + "si-LK": 45, + "sk": 28, + "sk-SK": 28, + "sl": 62, + "sl-SI": 62, + "sm-WS": 98, + "so-SO": 56, + "sv": 24, + "sv-SE": 24, + "sw-KE": 48, + "ta-IN": 39, + "te-IN": 40, + "tg-TJ": 70, + "th-TH": 32, + "to-TO": 99, + "tr": 18, + "tr-TR": 18, + "uk": 19, + "uk-UA": 19, + "ur-PK": 37, + "uz-UZ": 69, + "vi-VN": 33, + "yo-NG": 52, + "zh-CN": 4, + "zh-TW": 5, + "zh-ZH": 4, + "zu-ZA": 51 + }, + "default_prompt_id": 101, + "lang_tag_token_ids": [ + 1, + 63, + 115, + 167, + 226, + 227, + 259, + 276, + 328, + 353, + 368, + 462, + 481, + 499, + 518, + 542, + 571, + 602, + 603, + 612, + 624, + 646, + 647, + 667, + 689, + 699, + 720, + 727, + 747, + 748, + 750, + 751, + 752, + 756, + 774, + 787, + 788, + 801, + 802 + ] +} \ No newline at end of file diff --git a/it/4480ms/preprocessor.mlmodelc/analytics/coremldata.bin b/it/4480ms/preprocessor.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..caf6367f51e931ec1ff5cc630cadd6c50bd7d4ab --- /dev/null +++ b/it/4480ms/preprocessor.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:97be07beee7a6a28b7db85d47087d5b018ebcd1fc0b1565707141d574244bdc9 +size 243 diff --git a/it/4480ms/preprocessor.mlmodelc/coremldata.bin b/it/4480ms/preprocessor.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..ec3eb8270e4d6bb997c705e7e4e55167406c124f --- /dev/null +++ b/it/4480ms/preprocessor.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f0722dcca24c40139a73b9133f318ca41b9204768d3f14b999b6ef68cab6d8ad +size 371 diff --git a/it/4480ms/preprocessor.mlmodelc/model.mil b/it/4480ms/preprocessor.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..0b8261362f9cbf465b530a0d2d0ee9a2b2f462cd --- /dev/null +++ b/it/4480ms/preprocessor.mlmodelc/model.mil @@ -0,0 +1,122 @@ +program(1.3) +[buildInfo = dict({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.5.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.3.0"}})] +{ + func main(tensor audio, tensor audio_length) [FlexibleShapeInformation = tuple>>, tuple, ?>>>>((("DefaultShapes", {{"audio", [1, 1]}}), ("RangeDims", {{"audio", [[1, 1], [1, 480000]]}})))] { + int32 var_9 = const()[name = string("op_9"), val = int32(1)]; + int32 var_10 = const()[name = string("op_10"), val = int32(160)]; + int32 var_12 = const()[name = string("op_12"), val = int32(0)]; + int32 var_33 = const()[name = string("op_33"), val = int32(512)]; + tensor var_34 = add(x = audio_length, y = var_33)[name = string("op_34")]; + int32 var_35 = const()[name = string("op_35"), val = int32(512)]; + tensor var_36 = sub(x = var_34, y = var_35)[name = string("op_36")]; + tensor floor_div_0 = floor_div(x = var_36, y = var_10)[name = string("floor_div_0")]; + tensor var_39 = equal(x = audio_length, y = var_12)[name = string("op_39")]; + tensor var_40 = const()[name = string("op_40"), val = tensor([0])]; + tensor mel_length = select(a = var_40, b = floor_div_0, cond = var_39)[name = string("seq_len")]; + string audio_to_fp16_dtype_0 = const()[name = string("audio_to_fp16_dtype_0"), val = string("fp16")]; + tensor audio_to_fp16 = cast(dtype = audio_to_fp16_dtype_0, x = audio)[name = string("cast_14")]; + tensor var_42_shape_cast_fp16 = shape(x = audio_to_fp16)[name = string("op_42_shape_cast_fp16")]; + int32 gather_0_axis_0 = const()[name = string("gather_0_axis_0"), val = int32(0)]; + int32 gather_0_batch_dims_0 = const()[name = string("gather_0_batch_dims_0"), val = int32(0)]; + bool gather_0_validate_indices_0 = const()[name = string("gather_0_validate_indices_0"), val = bool(false)]; + string var_42_shape_cast_fp16_to_int16_dtype_0 = const()[name = string("op_42_shape_cast_fp16_to_int16_dtype_0"), val = string("int16")]; + uint16 select_0_to_uint16 = const()[name = string("select_0_to_uint16"), val = uint16(1)]; + tensor var_42_shape_cast_fp16_to_int16 = cast(dtype = var_42_shape_cast_fp16_to_int16_dtype_0, x = var_42_shape_cast_fp16)[name = string("cast_13")]; + int16 gather_0_cast_uint16 = gather(axis = gather_0_axis_0, batch_dims = gather_0_batch_dims_0, indices = select_0_to_uint16, validate_indices = gather_0_validate_indices_0, x = var_42_shape_cast_fp16_to_int16)[name = string("gather_0_cast_uint16")]; + string gather_0_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_0_cast_uint16_to_int32_dtype_0"), val = string("int32")]; + int32 const_0 = const()[name = string("const_0"), val = int32(0)]; + int32 const_1 = const()[name = string("const_1"), val = int32(1)]; + int32 gather_0_cast_uint16_to_int32 = cast(dtype = gather_0_cast_uint16_to_int32_dtype_0, x = gather_0_cast_uint16)[name = string("cast_12")]; + tensor var_43 = range_1d(end = gather_0_cast_uint16_to_int32, start = const_0, step = const_1)[name = string("op_43")]; + tensor var_44_axes_0 = const()[name = string("op_44_axes_0"), val = tensor([0])]; + tensor var_44 = expand_dims(axes = var_44_axes_0, x = var_43)[name = string("op_44")]; + tensor var_45_axes_0 = const()[name = string("op_45_axes_0"), val = tensor([1])]; + tensor var_45 = expand_dims(axes = var_45_axes_0, x = audio_length)[name = string("op_45")]; + tensor timemask = less(x = var_44, y = var_45)[name = string("timemask")]; + tensor var_48_begin_0 = const()[name = string("op_48_begin_0"), val = tensor([0, 0])]; + tensor var_48_end_0 = const()[name = string("op_48_end_0"), val = tensor([1, 1])]; + tensor var_48_end_mask_0 = const()[name = string("op_48_end_mask_0"), val = tensor([true, false])]; + tensor var_48_squeeze_mask_0 = const()[name = string("op_48_squeeze_mask_0"), val = tensor([false, true])]; + tensor var_48_cast_fp16 = slice_by_index(begin = var_48_begin_0, end = var_48_end_0, end_mask = var_48_end_mask_0, squeeze_mask = var_48_squeeze_mask_0, x = audio_to_fp16)[name = string("op_48_cast_fp16")]; + tensor var_49_axes_0 = const()[name = string("op_49_axes_0"), val = tensor([1])]; + tensor var_49_cast_fp16 = expand_dims(axes = var_49_axes_0, x = var_48_cast_fp16)[name = string("op_49_cast_fp16")]; + tensor var_51_begin_0 = const()[name = string("op_51_begin_0"), val = tensor([0, 1])]; + tensor var_51_end_0 = const()[name = string("op_51_end_0"), val = tensor([1, 0])]; + tensor var_51_end_mask_0 = const()[name = string("op_51_end_mask_0"), val = tensor([true, true])]; + tensor var_51_cast_fp16 = slice_by_index(begin = var_51_begin_0, end = var_51_end_0, end_mask = var_51_end_mask_0, x = audio_to_fp16)[name = string("op_51_cast_fp16")]; + tensor var_53_begin_0 = const()[name = string("op_53_begin_0"), val = tensor([0, 0])]; + tensor var_53_end_0 = const()[name = string("op_53_end_0"), val = tensor([1, -1])]; + tensor var_53_end_mask_0 = const()[name = string("op_53_end_mask_0"), val = tensor([true, false])]; + tensor var_53_cast_fp16 = slice_by_index(begin = var_53_begin_0, end = var_53_end_0, end_mask = var_53_end_mask_0, x = audio_to_fp16)[name = string("op_53_cast_fp16")]; + fp16 var_54_to_fp16 = const()[name = string("op_54_to_fp16"), val = fp16(0x1.f0cp-1)]; + tensor var_55_cast_fp16 = mul(x = var_53_cast_fp16, y = var_54_to_fp16)[name = string("op_55_cast_fp16")]; + tensor var_56_cast_fp16 = sub(x = var_51_cast_fp16, y = var_55_cast_fp16)[name = string("op_56_cast_fp16")]; + bool x_3_interleave_0 = const()[name = string("x_3_interleave_0"), val = bool(false)]; + tensor x_3_cast_fp16 = concat(axis = var_9, interleave = x_3_interleave_0, values = (var_49_cast_fp16, var_56_cast_fp16))[name = string("x_3_cast_fp16")]; + tensor var_59 = logical_not(x = timemask)[name = string("op_59")]; + fp16 var_16_to_fp16 = const()[name = string("op_16_to_fp16"), val = fp16(0x0p+0)]; + tensor input_1_cast_fp16 = select(a = var_16_to_fp16, b = x_3_cast_fp16, cond = var_59)[name = string("input_1_cast_fp16")]; + tensor concat_1x = const()[name = string("concat_1x"), val = tensor([1, 1, -1])]; + tensor input_3_cast_fp16 = reshape(shape = concat_1x, x = input_1_cast_fp16)[name = string("input_3_cast_fp16")]; + tensor input_5_pad_0 = const()[name = string("input_5_pad_0"), val = tensor([0, 0, 0, 0, 256, 256])]; + string input_5_mode_0 = const()[name = string("input_5_mode_0"), val = string("constant")]; + fp16 const_3_to_fp16 = const()[name = string("const_3_to_fp16"), val = fp16(0x0p+0)]; + tensor input_5_cast_fp16 = pad(constant_val = const_3_to_fp16, mode = input_5_mode_0, pad = input_5_pad_0, x = input_3_cast_fp16)[name = string("input_5_cast_fp16")]; + tensor concat_2x = const()[name = string("concat_2x"), val = tensor([1, -1])]; + tensor input_cast_fp16 = reshape(shape = concat_2x, x = input_5_cast_fp16)[name = string("input_cast_fp16")]; + tensor expand_dims_3 = const()[name = string("expand_dims_3"), val = tensor([160])]; + tensor expand_dims_4_axes_0 = const()[name = string("expand_dims_4_axes_0"), val = tensor([1])]; + tensor expand_dims_4_cast_fp16 = expand_dims(axes = expand_dims_4_axes_0, x = input_cast_fp16)[name = string("expand_dims_4_cast_fp16")]; + string conv_0_pad_type_0 = const()[name = string("conv_0_pad_type_0"), val = string("valid")]; + tensor conv_0_pad_0 = const()[name = string("conv_0_pad_0"), val = tensor([0, 0])]; + tensor conv_0_dilations_0 = const()[name = string("conv_0_dilations_0"), val = tensor([1])]; + int32 conv_0_groups_0 = const()[name = string("conv_0_groups_0"), val = int32(1)]; + tensor expand_dims_1_to_fp16 = const()[name = string("expand_dims_1_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))]; + tensor conv_0_cast_fp16 = conv(dilations = conv_0_dilations_0, groups = conv_0_groups_0, pad = conv_0_pad_0, pad_type = conv_0_pad_type_0, strides = expand_dims_3, weight = expand_dims_1_to_fp16, x = expand_dims_4_cast_fp16)[name = string("conv_0_cast_fp16")]; + string conv_1_pad_type_0 = const()[name = string("conv_1_pad_type_0"), val = string("valid")]; + tensor conv_1_pad_0 = const()[name = string("conv_1_pad_0"), val = tensor([0, 0])]; + tensor conv_1_dilations_0 = const()[name = string("conv_1_dilations_0"), val = tensor([1])]; + int32 conv_1_groups_0 = const()[name = string("conv_1_groups_0"), val = int32(1)]; + tensor expand_dims_2_to_fp16 = const()[name = string("expand_dims_2_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(263296)))]; + tensor conv_1_cast_fp16 = conv(dilations = conv_1_dilations_0, groups = conv_1_groups_0, pad = conv_1_pad_0, pad_type = conv_1_pad_type_0, strides = expand_dims_3, weight = expand_dims_2_to_fp16, x = expand_dims_4_cast_fp16)[name = string("conv_1_cast_fp16")]; + int32 stack_0_axis_0 = const()[name = string("stack_0_axis_0"), val = int32(-1)]; + tensor stack_0_cast_fp16 = stack(axis = stack_0_axis_0, values = (conv_0_cast_fp16, conv_1_cast_fp16))[name = string("stack_0_cast_fp16")]; + fp16 var_19_promoted_to_fp16 = const()[name = string("op_19_promoted_to_fp16"), val = fp16(0x1p+1)]; + tensor var_74_cast_fp16 = pow(x = stack_0_cast_fp16, y = var_19_promoted_to_fp16)[name = string("op_74_cast_fp16")]; + tensor var_76_axes_0 = const()[name = string("op_76_axes_0"), val = tensor([-1])]; + bool var_76_keep_dims_0 = const()[name = string("op_76_keep_dims_0"), val = bool(false)]; + tensor var_76_cast_fp16 = reduce_sum(axes = var_76_axes_0, keep_dims = var_76_keep_dims_0, x = var_74_cast_fp16)[name = string("op_76_cast_fp16")]; + tensor x_11_cast_fp16 = identity(x = var_76_cast_fp16)[name = string("x_11_cast_fp16")]; + bool x_13_transpose_x_0 = const()[name = string("x_13_transpose_x_0"), val = bool(false)]; + bool x_13_transpose_y_0 = const()[name = string("x_13_transpose_y_0"), val = bool(false)]; + tensor const_4_to_fp16 = const()[name = string("const_4_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(526528)))]; + tensor x_13_cast_fp16 = matmul(transpose_x = x_13_transpose_x_0, transpose_y = x_13_transpose_y_0, x = const_4_to_fp16, y = x_11_cast_fp16)[name = string("x_13_cast_fp16")]; + fp16 var_83_to_fp16 = const()[name = string("op_83_to_fp16"), val = fp16(0x1p-24)]; + tensor var_84_cast_fp16 = add(x = x_13_cast_fp16, y = var_83_to_fp16)[name = string("op_84_cast_fp16")]; + fp32 x_epsilon_0 = const()[name = string("x_epsilon_0"), val = fp32(0x1p-149)]; + tensor x_cast_fp16 = log(epsilon = x_epsilon_0, x = var_84_cast_fp16)[name = string("x_cast_fp16")]; + tensor var_86_shape_cast_fp16 = shape(x = x_cast_fp16)[name = string("op_86_shape_cast_fp16")]; + int32 gather_5_axis_0 = const()[name = string("gather_5_axis_0"), val = int32(0)]; + int32 gather_5_batch_dims_0 = const()[name = string("gather_5_batch_dims_0"), val = int32(0)]; + bool gather_5_validate_indices_0 = const()[name = string("gather_5_validate_indices_0"), val = bool(false)]; + string var_86_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_86_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")]; + uint16 select_5_to_uint16 = const()[name = string("select_5_to_uint16"), val = uint16(2)]; + tensor var_86_shape_cast_fp16_to_uint16 = cast(dtype = var_86_shape_cast_fp16_to_uint16_dtype_0, x = var_86_shape_cast_fp16)[name = string("cast_11")]; + uint16 gather_5_cast_uint16 = gather(axis = gather_5_axis_0, batch_dims = gather_5_batch_dims_0, indices = select_5_to_uint16, validate_indices = gather_5_validate_indices_0, x = var_86_shape_cast_fp16_to_uint16)[name = string("gather_5_cast_uint16")]; + string gather_5_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_5_cast_uint16_to_int32_dtype_0"), val = string("int32")]; + int32 const_5 = const()[name = string("const_5"), val = int32(0)]; + int32 const_6 = const()[name = string("const_6"), val = int32(1)]; + int32 gather_5_cast_uint16_to_int32 = cast(dtype = gather_5_cast_uint16_to_int32_dtype_0, x = gather_5_cast_uint16)[name = string("cast_10")]; + tensor mask_1 = range_1d(end = gather_5_cast_uint16_to_int32, start = const_5, step = const_6)[name = string("mask_1")]; + tensor expand_dims_0_axes_0 = const()[name = string("expand_dims_0_axes_0"), val = tensor([0])]; + tensor expand_dims_0 = expand_dims(axes = expand_dims_0_axes_0, x = mask_1)[name = string("expand_dims_0")]; + tensor var_91_axes_0 = const()[name = string("op_91_axes_0"), val = tensor([1])]; + tensor var_91 = expand_dims(axes = var_91_axes_0, x = mel_length)[name = string("op_91")]; + tensor mask = greater_equal(x = expand_dims_0, y = var_91)[name = string("mask")]; + tensor var_93_axes_0 = const()[name = string("op_93_axes_0"), val = tensor([1])]; + tensor var_93 = expand_dims(axes = var_93_axes_0, x = mask)[name = string("op_93")]; + tensor processed_signal_cast_fp16 = select(a = var_16_to_fp16, b = x_cast_fp16, cond = var_93)[name = string("processed_signal_cast_fp16")]; + string processed_signal_cast_fp16_to_fp32_dtype_0 = const()[name = string("processed_signal_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + tensor mel = cast(dtype = processed_signal_cast_fp16_to_fp32_dtype_0, x = processed_signal_cast_fp16)[name = string("cast_9")]; + } -> (mel, mel_length); +} \ No newline at end of file diff --git a/it/4480ms/preprocessor.mlmodelc/weights/weight.bin b/it/4480ms/preprocessor.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..86dd375f6649d262d58c9cf8b89006ceab216411 --- /dev/null +++ b/it/4480ms/preprocessor.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:297514e2b211d14b0e53cb97193d679bb89ead98d28e578f3f1d049ddbcc36b3 +size 592384 diff --git a/it/4480ms/preprocessor.mlpackage/Data/com.apple.CoreML/model.mlmodel b/it/4480ms/preprocessor.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..a760b77aa823c74ebff0487b3aea510a88447fd7 --- /dev/null +++ b/it/4480ms/preprocessor.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f1df1cad6f43ce4fa090e0f9a33bb5ddf25a0aaeca2be136339878b5b9de45c0 +size 15878 diff --git a/it/4480ms/preprocessor.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/it/4480ms/preprocessor.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..86dd375f6649d262d58c9cf8b89006ceab216411 --- /dev/null +++ b/it/4480ms/preprocessor.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:297514e2b211d14b0e53cb97193d679bb89ead98d28e578f3f1d049ddbcc36b3 +size 592384 diff --git a/it/4480ms/preprocessor.mlpackage/Manifest.json b/it/4480ms/preprocessor.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..d289893c92607e581c8cdaebf26772223814cb0d --- /dev/null +++ b/it/4480ms/preprocessor.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "162F4F34-AB45-4D09-8F9D-2F76AAE79E0F": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + }, + "B5F24360-A5E1-4F41-9872-3D77094622F0": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + } + }, + "rootModelIdentifier": "B5F24360-A5E1-4F41-9872-3D77094622F0" +} diff --git a/it/4480ms/tokenizer.json b/it/4480ms/tokenizer.json new file mode 100644 index 0000000000000000000000000000000000000000..033007249c80c55ca310e0c90339004b867968f6 --- /dev/null +++ b/it/4480ms/tokenizer.json @@ -0,0 +1,808 @@ +{ + "0": "", + "1": "", + "2": "▁", + "3": ".", + "4": ",", + "5": "e", + "6": "t", + "7": "a", + "8": "s", + "9": "o", + "10": "i", + "11": "r", + "12": "l", + "13": "u", + "14": "d", + "15": "c", + "16": "h", + "17": "m", + "18": "p", + "19": "n", + "20": "g", + "21": "f", + "22": "en", + "23": "in", + "24": "on", + "25": "y", + "26": "er", + "27": "an", + "28": "w", + "29": "0", + "30": "b", + "31": "v", + "32": "2", + "33": "1", + "34": "k", + "35": "3", + "36": "5", + "37": "q", + "38": "I", + "39": "é", + "40": "4", + "41": "z", + "42": "A", + "43": "6", + "44": "j", + "45": "E", + "46": "7", + "47": "8", + "48": "9", + "49": "S", + "50": "x", + "51": "C", + "52": "P", + "53": "U", + "54": "N", + "55": "O", + "56": "à", + "57": "L", + "58": "è", + "59": "V", + "60": "R", + "61": "ó", + "62": "J", + "63": "", + "64": "st", + "65": "ch", + "66": "le", + "67": "li", + "68": "▁po", + "69": "no", + "70": "to", + "71": "me", + "72": "te", + "73": "ho", + "74": "▁pro", + "75": "ro", + "76": "▁na", + "77": "ce", + "78": "la", + "79": "ni", + "80": "ra", + "81": "ti", + "82": "lo", + "83": "ko", + "84": "po", + "85": "je", + "86": "de", + "87": "na", + "88": "mi", + "89": "ci", + "90": "▁by", + "91": "ve", + "92": "▁za", + "93": "▁A", + "94": "re", + "95": "ou", + "96": "vo", + "97": "né", + "98": "va", + "99": "mo", + "100": "ze", + "101": "ne", + "102": "ka", + "103": "ky", + "104": "ovat", + "105": "▁jak", + "106": "ny", + "107": "vi", + "108": "ent", + "109": "prav", + "110": "Z", + "111": "▁Je", + "112": "oval", + "113": "ă", + "114": "X", + "115": "", + "116": "▁for", + "117": "▁det", + "118": "▁at", + "119": "et", + "120": "▁og", + "121": "▁vi", + "122": "al", + "123": "▁de", + "124": "▁der", + "125": "or", + "126": "om", + "127": "and", + "128": "▁har", + "129": "at", + "130": "▁af", + "131": "ge", + "132": "ar", + "133": "is", + "134": "▁med", + "135": "▁be", + "136": "un", + "137": "lig", + "138": "▁man", + "139": "ig", + "140": "▁som", + "141": "▁Og", + "142": "el", + "143": "ag", + "144": "erne", + "145": "▁den", + "146": "ste", + "147": "id", + "148": "▁kan", + "149": "ske", + "150": "iv", + "151": "ion", + "152": "am", + "153": "ur", + "154": "for", + "155": "else", + "156": "▁sig", + "157": "▁ind", + "158": "ende", + "159": "▁Vi", + "160": "ation", + "161": "mme", + "162": "▁op", + "163": "▁fra", + "164": "▁alle", + "165": "▁Men", + "166": "▁var", + "167": "", + "168": "▁die", + "169": "▁und", + "170": "sch", + "171": "it", + "172": "gen", + "173": "▁W", + "174": "▁B", + "175": "▁E", + "176": "▁F", + "177": "ll", + "178": "▁es", + "179": "▁K", + "180": "ie", + "181": "au", + "182": "ich", + "183": "ck", + "184": "ten", + "185": "mal", + "186": "ein", + "187": "▁T", + "188": "▁dann", + "189": "▁mit", + "190": "ter", + "191": "tz", + "192": "▁G", + "193": "ben", + "194": "um", + "195": "us", + "196": "il", + "197": "ut", + "198": "▁ver", + "199": "ri", + "200": "ach", + "201": "ol", + "202": "▁Da", + "203": "sp", + "204": "ell", + "205": "▁was", + "206": "▁ja", + "207": "wi", + "208": "rei", + "209": "▁Ge", + "210": "und", + "211": "▁St", + "212": "▁sie", + "213": "▁Ja", + "214": "▁du", + "215": "▁Sch", + "216": "▁Ma", + "217": "▁De", + "218": "▁Sie", + "219": "▁vor", + "220": "▁Le", + "221": "▁In", + "222": "▁Ver", + "223": "▁Re", + "224": "▁Mi", + "225": "▁Ha", + "226": "", + "227": "", + "228": "ma", + "229": "ta", + "230": "se", + "231": "da", + "232": "si", + "233": "ks", + "234": "ga", + "235": "he", + "236": "mu", + "237": "tu", + "238": "ha", + "239": "ja", + "240": "gi", + "241": "gu", + "242": "ju", + "243": "est", + "244": "▁pa", + "245": "tud", + "246": "nda", + "247": "ke", + "248": "sta", + "249": "sed", + "250": "di", + "251": "▁su", + "252": "ide", + "253": "pool", + "254": "val", + "255": "▁Me", + "256": "ment", + "257": "ndus", + "258": "Q", + "259": "", + "260": "ssa", + "261": "lla", + "262": "ki", + "263": "pa", + "264": "lle", + "265": "lu", + "266": "tta", + "267": "isi", + "268": "tte", + "269": "ista", + "270": "llis", + "271": "vu", + "272": "▁voi", + "273": "utta", + "274": "iden", + "275": "parlament", + "276": "", + "277": "▁est", + "278": "▁c", + "279": "▁que", + "280": "es", + "281": "▁un", + "282": "▁pas", + "283": "▁qui", + "284": "▁il", + "285": "▁des", + "286": "▁qu", + "287": "▁par", + "288": "ant", + "289": "▁C", + "290": "tre", + "291": "ir", + "292": "elle", + "293": "eur", + "294": "▁sur", + "295": "▁con", + "296": "ement", + "297": "tion", + "298": "mp", + "299": "▁comme", + "300": "ac", + "301": "▁là", + "302": "che", + "303": "que", + "304": "ul", + "305": "▁bien", + "306": "age", + "307": "▁mon", + "308": "end", + "309": "ver", + "310": "tra", + "311": "ille", + "312": "ff", + "313": "▁ex", + "314": "▁Il", + "315": "im", + "316": "▁dire", + "317": "ance", + "318": "aire", + "319": "▁app", + "320": "onne", + "321": "mb", + "322": "man", + "323": "▁quand", + "324": "port", + "325": "form", + "326": "ture", + "327": "ù", + "328": "", + "329": "ok", + "330": "gy", + "331": "ek", + "332": "em", + "333": "▁is", + "334": "os", + "335": "ak", + "336": "ban", + "337": "ik", + "338": "▁nem", + "339": "oz", + "340": "cs", + "341": "nek", + "342": "bb", + "343": "▁ha", + "344": "ott", + "345": "▁van", + "346": "▁fel", + "347": "leg", + "348": "▁ami", + "349": "tart", + "350": "rend", + "351": "▁fog", + "352": "▁volt", + "353": "", + "354": "▁bi", + "355": "▁sa", + "356": "ru", + "357": "go", + "358": "sti", + "359": "▁pri", + "360": "ima", + "361": "nu", + "362": "▁pre", + "363": "zi", + "364": "vr", + "365": "ca", + "366": "ba", + "367": "▁raz", + "368": "", + "369": "▁di", + "370": "▁che", + "371": "▁è", + "372": "co", + "373": "▁per", + "374": "▁non", + "375": "do", + "376": "gli", + "377": "so", + "378": "amo", + "379": "sa", + "380": "ndo", + "381": "▁una", + "382": "fi", + "383": "pi", + "384": "nti", + "385": "tto", + "386": "tro", + "387": "▁fa", + "388": "chi", + "389": "▁qua", + "390": "zione", + "391": "bi", + "392": "▁del", + "393": "mente", + "394": "pe", + "395": "ssi", + "396": "▁sono", + "397": "▁questo", + "398": "nte", + "399": "tti", + "400": "tà", + "401": "▁nel", + "402": "▁anche", + "403": "sso", + "404": "▁perché", + "405": "▁più", + "406": "nta", + "407": "▁come", + "408": "cu", + "409": "▁quindi", + "410": "ggi", + "411": "nza", + "412": "sto", + "413": "ò", + "414": "▁della", + "415": "gra", + "416": "▁fare", + "417": "spe", + "418": "cco", + "419": "nde", + "420": "mento", + "421": "fe", + "422": "gio", + "423": "pu", + "424": "▁questa", + "425": "zza", + "426": "sci", + "427": "▁dei", + "428": "▁poi", + "429": "sco", + "430": "stra", + "431": "▁quel", + "432": "qui", + "433": "▁delle", + "434": "▁cosa", + "435": "▁molto", + "436": "sse", + "437": "zioni", + "438": "▁inter", + "439": "sce", + "440": "▁fatto", + "441": "▁com", + "442": "▁quello", + "443": "▁essere", + "444": "▁due", + "445": "▁abbiamo", + "446": "▁comp", + "447": "▁tutti", + "448": "ì", + "449": "▁prima", + "450": "▁parte", + "451": "▁così", + "452": "▁sempre", + "453": "▁tutto", + "454": "▁video", + "455": "▁imp", + "456": "▁cui", + "457": "▁dove", + "458": "▁Quindi", + "459": "sione", + "460": "rebbe", + "461": "scri", + "462": "", + "463": "ai", + "464": "▁ir", + "465": "as", + "466": "▁tai", + "467": "uo", + "468": "tin", + "469": "▁vis", + "470": "ly", + "471": "gal", + "472": "tar", + "473": "▁Ir", + "474": "▁turi", + "475": "▁Tai", + "476": "▁nu", + "477": "▁mes", + "478": "išk", + "479": "imas", + "480": "▁pra", + "481": "", + "482": "iem", + "483": "▁pie", + "484": "ies", + "485": "ot", + "486": "▁vien", + "487": "▁Un", + "488": "iet", + "489": "▁Ta", + "490": "dar", + "491": "sim", + "492": "gan", + "493": "▁ap", + "494": "▁nav", + "495": "▁Nu", + "496": "▁cit", + "497": "▁Ne", + "498": "▁20", + "499": "", + "500": "▁dat", + "501": "▁we", + "502": "▁En", + "503": "▁dan", + "504": "▁zo", + "505": "▁met", + "506": "▁wat", + "507": "der", + "508": "ui", + "509": "den", + "510": "op", + "511": "oor", + "512": "▁of", + "513": "ven", + "514": "acht", + "515": "▁even", + "516": "▁wil", + "517": "vol", + "518": "", + "519": "nie", + "520": "wa", + "521": "cie", + "522": "nia", + "523": "wo", + "524": "rze", + "525": "by", + "526": "za", + "527": "dy", + "528": "ry", + "529": "ego", + "530": "mie", + "531": "rz", + "532": "pie", + "533": "▁bo", + "534": "bie", + "535": "▁Po", + "536": "ski", + "537": "nego", + "538": "▁No", + "539": "▁Na", + "540": "▁Was", + "541": "▁musi", + "542": "", + "543": "-", + "544": "▁para", + "545": "▁uma", + "546": "▁pe", + "547": "▁tem", + "548": "▁gente", + "549": "▁O", + "550": "▁ele", + "551": "pre", + "552": "ria", + "553": "▁fo", + "554": "mos", + "555": "bo", + "556": "▁por", + "557": "nto", + "558": "be", + "559": "▁esse", + "560": "ente", + "561": "▁essa", + "562": "▁mas", + "563": "qua", + "564": "fica", + "565": "▁Se", + "566": "▁Por", + "567": "▁Co", + "568": "iza", + "569": "▁sua", + "570": "▁quando", + "571": "", + "572": "▁mai", + "573": "sc", + "574": "are", + "575": "▁din", + "576": "▁este", + "577": "rea", + "578": "ele", + "579": "du", + "580": "▁M", + "581": "▁fac", + "582": "lor", + "583": "▁mult", + "584": "per", + "585": "cur", + "586": "tor", + "587": "inte", + "588": "▁sau", + "589": "tat", + "590": "ori", + "591": "▁prim", + "592": "▁spun", + "593": "▁lui", + "594": "▁sub", + "595": "itate", + "596": "▁prin", + "597": "▁alt", + "598": "stru", + "599": "▁vede", + "600": "fer", + "601": "▁chiar", + "602": "", + "603": "", + "604": "ov", + "605": "ob", + "606": "▁bol", + "607": "ali", + "608": "rov", + "609": "rob", + "610": "▁spo", + "611": "osti", + "612": "", + "613": "sl", + "614": "udi", + "615": "del", + "616": "▁sem", + "617": "▁samo", + "618": "▁pred", + "619": "nost", + "620": "▁Pre", + "621": "▁prot", + "622": "▁internet", + "623": "▁film", + "624": "", + "625": "▁att", + "626": "▁inte", + "627": "▁av", + "628": "all", + "629": "era", + "630": "pp", + "631": "▁upp", + "632": "isk", + "633": "het", + "634": "▁vill", + "635": "erna", + "636": "ande", + "637": "ade", + "638": "bil", + "639": "▁min", + "640": "▁alla", + "641": "lev", + "642": "▁oss", + "643": "land", + "644": "▁Vad", + "645": "person", + "646": "", + "647": "", + "648": "vy", + "649": "ft", + "650": "lige", + "651": "ved", + "652": "'", + "653": "▁H", + "654": "▁D", + "655": "aus", + "656": "▁N", + "657": "▁Be", + "658": "mm", + "659": "ab", + "660": "▁Er", + "661": "ssen", + "662": "rie", + "663": "lei", + "664": "▁An", + "665": "rau", + "666": "▁So", + "667": "", + "668": "▁and", + "669": "▁can", + "670": "ed", + "671": "ay", + "672": "th", + "673": "ic", + "674": "hi", + "675": "▁Oh", + "676": "▁not", + "677": "ight", + "678": "ex", + "679": "▁great", + "680": "ill", + "681": "▁don", + "682": "▁problem", + "683": "▁fine", + "684": "▁month", + "685": "▁check", + "686": "▁zero", + "687": "▁first", + "688": "▁question", + "689": "", + "690": "ive", + "691": "ate", + "692": "ad", + "693": "ng", + "694": "ity", + "695": "ther", + "696": "act", + "697": "side", + "698": "\"", + "699": "", + "700": "ción", + "701": "▁Es", + "702": "res", + "703": "▁La", + "704": "dos", + "705": "▁El", + "706": "▁las", + "707": "men", + "708": "par", + "709": "rio", + "710": "enta", + "711": "▁Ca", + "712": "▁Su", + "713": "▁son", + "714": "ncia", + "715": "▁Con", + "716": "ones", + "717": "▁San", + "718": "▁persona", + "719": "▁Com", + "720": "", + "721": "cia", + "722": "▁Y", + "723": "ron", + "724": "les", + "725": "cio", + "726": "bu", + "727": "", + "728": "ré", + "729": "▁Les", + "730": "our", + "731": "▁Ce", + "732": "com", + "733": "ale", + "734": "if", + "735": "iste", + "736": "▁parti", + "737": "avec", + "738": "app", + "739": "gue", + "740": "▁grand", + "741": "Une", + "742": "È", + "743": "av", + "744": "pri", + "745": "sion", + "746": "ard", + "747": "", + "748": "", + "749": "!", + "750": "", + "751": "", + "752": "", + "753": "ene", + "754": "opp", + "755": "▁han", + "756": "", + "757": "eg", + "758": "kk", + "759": "▁god", + "760": "dde", + "761": "inn", + "762": "dig", + "763": "ord", + "764": "▁tru", + "765": "▁sei", + "766": "ller", + "767": "car", + "768": "ito", + "769": "ram", + "770": "fa", + "771": "▁mil", + "772": "▁passa", + "773": "▁casa", + "774": "", + "775": "▁Pa", + "776": "tura", + "777": "forma", + "778": "tua", + "779": "mar", + "780": "este", + "781": "fun", + "782": "gua", + "783": "▁grande", + "784": "▁nome", + "785": "▁Sua", + "786": "var", + "787": "", + "788": "", + "789": "ş", + "790": "ğ", + "791": "ya", + "792": "▁ve", + "793": "lar", + "794": "ler", + "795": "leri", + "796": "▁bu", + "797": "lan", + "798": "ara", + "799": "▁Bu", + "800": "yo", + "801": "", + "802": "", + "803": "▁t", + "804": "nh", + "805": "" +} \ No newline at end of file diff --git a/it/560ms/decoder.mlmodelc/analytics/coremldata.bin b/it/560ms/decoder.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..85afd8d84c262c9e1ba71c6b460a5beb4d6b94c3 --- /dev/null +++ b/it/560ms/decoder.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4cdca6bf678463f31354072f526088e5bdf5115ae94c04e387bb35b2c7a607d6 +size 243 diff --git a/it/560ms/decoder.mlmodelc/coremldata.bin b/it/560ms/decoder.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..8bb44f2c4c669bd785344007b33e6273bd87aa8c --- /dev/null +++ b/it/560ms/decoder.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3c993f8b96ce22027cd2ed42d99b7e61f93a01197bb17cadada8eb989e946dec +size 433 diff --git a/it/560ms/decoder.mlmodelc/model.mil b/it/560ms/decoder.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..e034376bbf9a1dff11539e03ae80e7a65ea4f393 --- /dev/null +++ b/it/560ms/decoder.mlmodelc/model.mil @@ -0,0 +1,64 @@ +program(1.3) +[buildInfo = dict({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.5.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.3.0"}})] +{ + func main(tensor c_in, tensor h_in, tensor token, tensor token_length) { + int32 y_axis_0 = const()[name = string("y_axis_0"), val = int32(0)]; + int32 y_batch_dims_0 = const()[name = string("y_batch_dims_0"), val = int32(0)]; + bool y_validate_indices_0 = const()[name = string("y_validate_indices_0"), val = bool(false)]; + tensor module_prediction_embed_weight_to_fp16 = const()[name = string("module_prediction_embed_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))]; + string token_to_int16_dtype_0 = const()[name = string("token_to_int16_dtype_0"), val = string("int16")]; + tensor token_to_int16 = cast(dtype = token_to_int16_dtype_0, x = token)[name = string("cast_8")]; + tensor y_cast_fp16_cast_uint16 = gather(axis = y_axis_0, batch_dims = y_batch_dims_0, indices = token_to_int16, validate_indices = y_validate_indices_0, x = module_prediction_embed_weight_to_fp16)[name = string("y_cast_fp16_cast_uint16")]; + tensor input_3_perm_0 = const()[name = string("input_3_perm_0"), val = tensor([1, 0, 2])]; + int32 split_0_num_splits_0 = const()[name = string("split_0_num_splits_0"), val = int32(2)]; + int32 split_0_axis_0 = const()[name = string("split_0_axis_0"), val = int32(0)]; + string h_in_to_fp16_dtype_0 = const()[name = string("h_in_to_fp16_dtype_0"), val = string("fp16")]; + tensor h_in_to_fp16 = cast(dtype = h_in_to_fp16_dtype_0, x = h_in)[name = string("cast_7")]; + tensor split_0_cast_fp16_0, tensor split_0_cast_fp16_1 = split(axis = split_0_axis_0, num_splits = split_0_num_splits_0, x = h_in_to_fp16)[name = string("split_0_cast_fp16")]; + int32 split_1_num_splits_0 = const()[name = string("split_1_num_splits_0"), val = int32(2)]; + int32 split_1_axis_0 = const()[name = string("split_1_axis_0"), val = int32(0)]; + string c_in_to_fp16_dtype_0 = const()[name = string("c_in_to_fp16_dtype_0"), val = string("fp16")]; + tensor c_in_to_fp16 = cast(dtype = c_in_to_fp16_dtype_0, x = c_in)[name = string("cast_6")]; + tensor split_1_cast_fp16_0, tensor split_1_cast_fp16_1 = split(axis = split_1_axis_0, num_splits = split_1_num_splits_0, x = c_in_to_fp16)[name = string("split_1_cast_fp16")]; + tensor input_lstm_layer_0_lstm_h0_squeeze_axes_0 = const()[name = string("input_lstm_layer_0_lstm_h0_squeeze_axes_0"), val = tensor([0])]; + tensor input_lstm_layer_0_lstm_h0_squeeze_cast_fp16 = squeeze(axes = input_lstm_layer_0_lstm_h0_squeeze_axes_0, x = split_0_cast_fp16_0)[name = string("input_lstm_layer_0_lstm_h0_squeeze_cast_fp16")]; + tensor input_lstm_layer_0_lstm_c0_squeeze_axes_0 = const()[name = string("input_lstm_layer_0_lstm_c0_squeeze_axes_0"), val = tensor([0])]; + tensor input_lstm_layer_0_lstm_c0_squeeze_cast_fp16 = squeeze(axes = input_lstm_layer_0_lstm_c0_squeeze_axes_0, x = split_1_cast_fp16_0)[name = string("input_lstm_layer_0_lstm_c0_squeeze_cast_fp16")]; + string input_lstm_layer_0_direction_0 = const()[name = string("input_lstm_layer_0_direction_0"), val = string("forward")]; + bool input_lstm_layer_0_output_sequence_0 = const()[name = string("input_lstm_layer_0_output_sequence_0"), val = bool(true)]; + string input_lstm_layer_0_recurrent_activation_0 = const()[name = string("input_lstm_layer_0_recurrent_activation_0"), val = string("sigmoid")]; + string input_lstm_layer_0_cell_activation_0 = const()[name = string("input_lstm_layer_0_cell_activation_0"), val = string("tanh")]; + string input_lstm_layer_0_activation_0 = const()[name = string("input_lstm_layer_0_activation_0"), val = string("tanh")]; + tensor concat_1_to_fp16 = const()[name = string("concat_1_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1031808)))]; + tensor concat_2_to_fp16 = const()[name = string("concat_2_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4308672)))]; + tensor concat_0_to_fp16 = const()[name = string("concat_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7585536)))]; + tensor input_3_cast_fp16 = transpose(perm = input_3_perm_0, x = y_cast_fp16_cast_uint16)[name = string("transpose_2")]; + tensor input_lstm_layer_0_cast_fp16_0, tensor input_lstm_layer_0_cast_fp16_1, tensor input_lstm_layer_0_cast_fp16_2 = lstm(activation = input_lstm_layer_0_activation_0, bias = concat_0_to_fp16, cell_activation = input_lstm_layer_0_cell_activation_0, direction = input_lstm_layer_0_direction_0, initial_c = input_lstm_layer_0_lstm_c0_squeeze_cast_fp16, initial_h = input_lstm_layer_0_lstm_h0_squeeze_cast_fp16, output_sequence = input_lstm_layer_0_output_sequence_0, recurrent_activation = input_lstm_layer_0_recurrent_activation_0, weight_hh = concat_2_to_fp16, weight_ih = concat_1_to_fp16, x = input_3_cast_fp16)[name = string("input_lstm_layer_0_cast_fp16")]; + tensor input_lstm_h0_squeeze_axes_0 = const()[name = string("input_lstm_h0_squeeze_axes_0"), val = tensor([0])]; + tensor input_lstm_h0_squeeze_cast_fp16 = squeeze(axes = input_lstm_h0_squeeze_axes_0, x = split_0_cast_fp16_1)[name = string("input_lstm_h0_squeeze_cast_fp16")]; + tensor input_lstm_c0_squeeze_axes_0 = const()[name = string("input_lstm_c0_squeeze_axes_0"), val = tensor([0])]; + tensor input_lstm_c0_squeeze_cast_fp16 = squeeze(axes = input_lstm_c0_squeeze_axes_0, x = split_1_cast_fp16_1)[name = string("input_lstm_c0_squeeze_cast_fp16")]; + string input_direction_0 = const()[name = string("input_direction_0"), val = string("forward")]; + bool input_output_sequence_0 = const()[name = string("input_output_sequence_0"), val = bool(true)]; + string input_recurrent_activation_0 = const()[name = string("input_recurrent_activation_0"), val = string("sigmoid")]; + string input_cell_activation_0 = const()[name = string("input_cell_activation_0"), val = string("tanh")]; + string input_activation_0 = const()[name = string("input_activation_0"), val = string("tanh")]; + tensor concat_4_to_fp16 = const()[name = string("concat_4_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7590720)))]; + tensor concat_5_to_fp16 = const()[name = string("concat_5_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10867584)))]; + tensor concat_3_to_fp16 = const()[name = string("concat_3_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14144448)))]; + tensor input_cast_fp16_0, tensor input_cast_fp16_1, tensor input_cast_fp16_2 = lstm(activation = input_activation_0, bias = concat_3_to_fp16, cell_activation = input_cell_activation_0, direction = input_direction_0, initial_c = input_lstm_c0_squeeze_cast_fp16, initial_h = input_lstm_h0_squeeze_cast_fp16, output_sequence = input_output_sequence_0, recurrent_activation = input_recurrent_activation_0, weight_hh = concat_5_to_fp16, weight_ih = concat_4_to_fp16, x = input_lstm_layer_0_cast_fp16_0)[name = string("input_cast_fp16")]; + int32 obj_3_axis_0 = const()[name = string("obj_3_axis_0"), val = int32(0)]; + tensor obj_3_cast_fp16 = stack(axis = obj_3_axis_0, values = (input_lstm_layer_0_cast_fp16_1, input_cast_fp16_1))[name = string("obj_3_cast_fp16")]; + string obj_3_cast_fp16_to_fp32_dtype_0 = const()[name = string("obj_3_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + int32 obj_axis_0 = const()[name = string("obj_axis_0"), val = int32(0)]; + tensor obj_cast_fp16 = stack(axis = obj_axis_0, values = (input_lstm_layer_0_cast_fp16_2, input_cast_fp16_2))[name = string("obj_cast_fp16")]; + string obj_cast_fp16_to_fp32_dtype_0 = const()[name = string("obj_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + tensor transpose_0_perm_0 = const()[name = string("transpose_0_perm_0"), val = tensor([1, 2, 0])]; + string transpose_0_cast_fp16_to_fp32_dtype_0 = const()[name = string("transpose_0_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + tensor transpose_0_cast_fp16 = transpose(perm = transpose_0_perm_0, x = input_cast_fp16_0)[name = string("transpose_1")]; + tensor decoder_out = cast(dtype = transpose_0_cast_fp16_to_fp32_dtype_0, x = transpose_0_cast_fp16)[name = string("cast_3")]; + tensor c_out = cast(dtype = obj_cast_fp16_to_fp32_dtype_0, x = obj_cast_fp16)[name = string("cast_4")]; + tensor h_out = cast(dtype = obj_3_cast_fp16_to_fp32_dtype_0, x = obj_3_cast_fp16)[name = string("cast_5")]; + tensor token_length_tmp = identity(x = token_length)[name = string("token_length_tmp")]; + } -> (decoder_out, h_out, c_out); +} \ No newline at end of file diff --git a/it/560ms/decoder.mlmodelc/weights/weight.bin b/it/560ms/decoder.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..2fac0f3a92f8a80de4d92a62819f6dba98aa4983 --- /dev/null +++ b/it/560ms/decoder.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1360d93c68c3e9c54bda4adaec860753949f3b0dc93bc98f4edc9d6f8dd5595c +size 14149632 diff --git a/it/560ms/decoder.mlpackage/Data/com.apple.CoreML/model.mlmodel b/it/560ms/decoder.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..567c038e1e42f382639a9ececec8bb38c22cbde0 --- /dev/null +++ b/it/560ms/decoder.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:73bb3afa62698bc822b6d32b3731d0bc40521e03737e3139e10a768542fca1fe +size 10359 diff --git a/it/560ms/decoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/it/560ms/decoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..2fac0f3a92f8a80de4d92a62819f6dba98aa4983 --- /dev/null +++ b/it/560ms/decoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1360d93c68c3e9c54bda4adaec860753949f3b0dc93bc98f4edc9d6f8dd5595c +size 14149632 diff --git a/it/560ms/decoder.mlpackage/Manifest.json b/it/560ms/decoder.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..8fc74b00a3e9885d54546160ecae1f6da7736d01 --- /dev/null +++ b/it/560ms/decoder.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "7CBCED8D-FA6A-45B0-BF60-30DB0A653074": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + }, + "AFD197FC-BECC-451A-961C-C0CA05D58065": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + } + }, + "rootModelIdentifier": "AFD197FC-BECC-451A-961C-C0CA05D58065" +} diff --git a/it/560ms/decoder_joint.mlmodelc/analytics/coremldata.bin b/it/560ms/decoder_joint.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..6063f90a9756de97c8450a89ef53ef04317ef653 --- /dev/null +++ b/it/560ms/decoder_joint.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:12f4bcf5114baa2b3a37b8ebeab6c519109bd857e50ec345c458b7a6c4deb20e +size 243 diff --git a/it/560ms/decoder_joint.mlmodelc/coremldata.bin b/it/560ms/decoder_joint.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..6f49a6f6923a6d68c50bdf11730215b1db8a2d62 --- /dev/null +++ b/it/560ms/decoder_joint.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:53754e2eaa7e0f7435220b47b621c5f3d8c5f2da83edd46efa5950fa723ef1d9 +size 454 diff --git a/it/560ms/decoder_joint.mlmodelc/model.mil b/it/560ms/decoder_joint.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..9e96b62349b7d1c4bd97fe8db2d7755704041510 --- /dev/null +++ b/it/560ms/decoder_joint.mlmodelc/model.mil @@ -0,0 +1,83 @@ +program(1.3) +[buildInfo = dict({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.5.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.3.0"}})] +{ + func main(tensor c_in, tensor encoder, tensor h_in, tensor token, tensor token_length) { + int32 y_axis_0 = const()[name = string("y_axis_0"), val = int32(0)]; + int32 y_batch_dims_0 = const()[name = string("y_batch_dims_0"), val = int32(0)]; + bool y_validate_indices_0 = const()[name = string("y_validate_indices_0"), val = bool(false)]; + tensor decoder_module_prediction_embed_weight_to_fp16 = const()[name = string("decoder_module_prediction_embed_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))]; + string token_to_int16_dtype_0 = const()[name = string("token_to_int16_dtype_0"), val = string("int16")]; + tensor token_to_int16 = cast(dtype = token_to_int16_dtype_0, x = token)[name = string("cast_9")]; + tensor y_cast_fp16_cast_uint16 = gather(axis = y_axis_0, batch_dims = y_batch_dims_0, indices = token_to_int16, validate_indices = y_validate_indices_0, x = decoder_module_prediction_embed_weight_to_fp16)[name = string("y_cast_fp16_cast_uint16")]; + tensor input_3_perm_0 = const()[name = string("input_3_perm_0"), val = tensor([1, 0, 2])]; + int32 split_0_num_splits_0 = const()[name = string("split_0_num_splits_0"), val = int32(2)]; + int32 split_0_axis_0 = const()[name = string("split_0_axis_0"), val = int32(0)]; + string h_in_to_fp16_dtype_0 = const()[name = string("h_in_to_fp16_dtype_0"), val = string("fp16")]; + tensor h_in_to_fp16 = cast(dtype = h_in_to_fp16_dtype_0, x = h_in)[name = string("cast_8")]; + tensor split_0_cast_fp16_0, tensor split_0_cast_fp16_1 = split(axis = split_0_axis_0, num_splits = split_0_num_splits_0, x = h_in_to_fp16)[name = string("split_0_cast_fp16")]; + int32 split_1_num_splits_0 = const()[name = string("split_1_num_splits_0"), val = int32(2)]; + int32 split_1_axis_0 = const()[name = string("split_1_axis_0"), val = int32(0)]; + string c_in_to_fp16_dtype_0 = const()[name = string("c_in_to_fp16_dtype_0"), val = string("fp16")]; + tensor c_in_to_fp16 = cast(dtype = c_in_to_fp16_dtype_0, x = c_in)[name = string("cast_7")]; + tensor split_1_cast_fp16_0, tensor split_1_cast_fp16_1 = split(axis = split_1_axis_0, num_splits = split_1_num_splits_0, x = c_in_to_fp16)[name = string("split_1_cast_fp16")]; + tensor input_5_lstm_layer_0_lstm_h0_squeeze_axes_0 = const()[name = string("input_5_lstm_layer_0_lstm_h0_squeeze_axes_0"), val = tensor([0])]; + tensor input_5_lstm_layer_0_lstm_h0_squeeze_cast_fp16 = squeeze(axes = input_5_lstm_layer_0_lstm_h0_squeeze_axes_0, x = split_0_cast_fp16_0)[name = string("input_5_lstm_layer_0_lstm_h0_squeeze_cast_fp16")]; + tensor input_5_lstm_layer_0_lstm_c0_squeeze_axes_0 = const()[name = string("input_5_lstm_layer_0_lstm_c0_squeeze_axes_0"), val = tensor([0])]; + tensor input_5_lstm_layer_0_lstm_c0_squeeze_cast_fp16 = squeeze(axes = input_5_lstm_layer_0_lstm_c0_squeeze_axes_0, x = split_1_cast_fp16_0)[name = string("input_5_lstm_layer_0_lstm_c0_squeeze_cast_fp16")]; + string input_5_lstm_layer_0_direction_0 = const()[name = string("input_5_lstm_layer_0_direction_0"), val = string("forward")]; + bool input_5_lstm_layer_0_output_sequence_0 = const()[name = string("input_5_lstm_layer_0_output_sequence_0"), val = bool(true)]; + string input_5_lstm_layer_0_recurrent_activation_0 = const()[name = string("input_5_lstm_layer_0_recurrent_activation_0"), val = string("sigmoid")]; + string input_5_lstm_layer_0_cell_activation_0 = const()[name = string("input_5_lstm_layer_0_cell_activation_0"), val = string("tanh")]; + string input_5_lstm_layer_0_activation_0 = const()[name = string("input_5_lstm_layer_0_activation_0"), val = string("tanh")]; + tensor concat_1_to_fp16 = const()[name = string("concat_1_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1031808)))]; + tensor concat_2_to_fp16 = const()[name = string("concat_2_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4308672)))]; + tensor concat_0_to_fp16 = const()[name = string("concat_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7585536)))]; + tensor input_3_cast_fp16 = transpose(perm = input_3_perm_0, x = y_cast_fp16_cast_uint16)[name = string("transpose_4")]; + tensor input_5_lstm_layer_0_cast_fp16_0, tensor input_5_lstm_layer_0_cast_fp16_1, tensor input_5_lstm_layer_0_cast_fp16_2 = lstm(activation = input_5_lstm_layer_0_activation_0, bias = concat_0_to_fp16, cell_activation = input_5_lstm_layer_0_cell_activation_0, direction = input_5_lstm_layer_0_direction_0, initial_c = input_5_lstm_layer_0_lstm_c0_squeeze_cast_fp16, initial_h = input_5_lstm_layer_0_lstm_h0_squeeze_cast_fp16, output_sequence = input_5_lstm_layer_0_output_sequence_0, recurrent_activation = input_5_lstm_layer_0_recurrent_activation_0, weight_hh = concat_2_to_fp16, weight_ih = concat_1_to_fp16, x = input_3_cast_fp16)[name = string("input_5_lstm_layer_0_cast_fp16")]; + tensor input_5_lstm_h0_squeeze_axes_0 = const()[name = string("input_5_lstm_h0_squeeze_axes_0"), val = tensor([0])]; + tensor input_5_lstm_h0_squeeze_cast_fp16 = squeeze(axes = input_5_lstm_h0_squeeze_axes_0, x = split_0_cast_fp16_1)[name = string("input_5_lstm_h0_squeeze_cast_fp16")]; + tensor input_5_lstm_c0_squeeze_axes_0 = const()[name = string("input_5_lstm_c0_squeeze_axes_0"), val = tensor([0])]; + tensor input_5_lstm_c0_squeeze_cast_fp16 = squeeze(axes = input_5_lstm_c0_squeeze_axes_0, x = split_1_cast_fp16_1)[name = string("input_5_lstm_c0_squeeze_cast_fp16")]; + string input_5_direction_0 = const()[name = string("input_5_direction_0"), val = string("forward")]; + bool input_5_output_sequence_0 = const()[name = string("input_5_output_sequence_0"), val = bool(true)]; + string input_5_recurrent_activation_0 = const()[name = string("input_5_recurrent_activation_0"), val = string("sigmoid")]; + string input_5_cell_activation_0 = const()[name = string("input_5_cell_activation_0"), val = string("tanh")]; + string input_5_activation_0 = const()[name = string("input_5_activation_0"), val = string("tanh")]; + tensor concat_4_to_fp16 = const()[name = string("concat_4_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7590720)))]; + tensor concat_5_to_fp16 = const()[name = string("concat_5_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10867584)))]; + tensor concat_3_to_fp16 = const()[name = string("concat_3_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14144448)))]; + tensor input_5_cast_fp16_0, tensor input_5_cast_fp16_1, tensor input_5_cast_fp16_2 = lstm(activation = input_5_activation_0, bias = concat_3_to_fp16, cell_activation = input_5_cell_activation_0, direction = input_5_direction_0, initial_c = input_5_lstm_c0_squeeze_cast_fp16, initial_h = input_5_lstm_h0_squeeze_cast_fp16, output_sequence = input_5_output_sequence_0, recurrent_activation = input_5_recurrent_activation_0, weight_hh = concat_5_to_fp16, weight_ih = concat_4_to_fp16, x = input_5_lstm_layer_0_cast_fp16_0)[name = string("input_5_cast_fp16")]; + int32 obj_3_axis_0 = const()[name = string("obj_3_axis_0"), val = int32(0)]; + tensor obj_3_cast_fp16 = stack(axis = obj_3_axis_0, values = (input_5_lstm_layer_0_cast_fp16_1, input_5_cast_fp16_1))[name = string("obj_3_cast_fp16")]; + string obj_3_cast_fp16_to_fp32_dtype_0 = const()[name = string("obj_3_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + int32 obj_axis_0 = const()[name = string("obj_axis_0"), val = int32(0)]; + tensor obj_cast_fp16 = stack(axis = obj_axis_0, values = (input_5_lstm_layer_0_cast_fp16_2, input_5_cast_fp16_2))[name = string("obj_cast_fp16")]; + string obj_cast_fp16_to_fp32_dtype_0 = const()[name = string("obj_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + tensor transpose_1_perm_0 = const()[name = string("transpose_1_perm_0"), val = tensor([1, 0, 2])]; + tensor input_7_perm_0 = const()[name = string("input_7_perm_0"), val = tensor([0, 2, 1])]; + string encoder_to_fp16_dtype_0 = const()[name = string("encoder_to_fp16_dtype_0"), val = string("fp16")]; + tensor joint_module_enc_weight_to_fp16 = const()[name = string("joint_module_enc_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14149632)))]; + tensor joint_module_enc_bias_to_fp16 = const()[name = string("joint_module_enc_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15460416)))]; + tensor encoder_to_fp16 = cast(dtype = encoder_to_fp16_dtype_0, x = encoder)[name = string("cast_4")]; + tensor input_7_cast_fp16 = transpose(perm = input_7_perm_0, x = encoder_to_fp16)[name = string("transpose_2")]; + tensor linear_0_cast_fp16 = linear(bias = joint_module_enc_bias_to_fp16, weight = joint_module_enc_weight_to_fp16, x = input_7_cast_fp16)[name = string("linear_0_cast_fp16")]; + tensor joint_module_pred_weight_to_fp16 = const()[name = string("joint_module_pred_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15461760)))]; + tensor joint_module_pred_bias_to_fp16 = const()[name = string("joint_module_pred_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16281024)))]; + tensor transpose_1_cast_fp16 = transpose(perm = transpose_1_perm_0, x = input_5_cast_fp16_0)[name = string("transpose_3")]; + tensor linear_1_cast_fp16 = linear(bias = joint_module_pred_bias_to_fp16, weight = joint_module_pred_weight_to_fp16, x = transpose_1_cast_fp16)[name = string("linear_1_cast_fp16")]; + tensor var_79_axes_0 = const()[name = string("op_79_axes_0"), val = tensor([2])]; + tensor var_79_cast_fp16 = expand_dims(axes = var_79_axes_0, x = linear_0_cast_fp16)[name = string("op_79_cast_fp16")]; + tensor var_80_axes_0 = const()[name = string("op_80_axes_0"), val = tensor([1])]; + tensor var_80_cast_fp16 = expand_dims(axes = var_80_axes_0, x = linear_1_cast_fp16)[name = string("op_80_cast_fp16")]; + tensor input_11_cast_fp16 = add(x = var_79_cast_fp16, y = var_80_cast_fp16)[name = string("input_11_cast_fp16")]; + tensor input_13_cast_fp16 = relu(x = input_11_cast_fp16)[name = string("input_13_cast_fp16")]; + tensor joint_module_joint_net_2_weight_to_fp16 = const()[name = string("joint_module_joint_net_2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16282368)))]; + tensor joint_module_joint_net_2_bias_to_fp16 = const()[name = string("joint_module_joint_net_2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17314112)))]; + tensor linear_2_cast_fp16 = linear(bias = joint_module_joint_net_2_bias_to_fp16, weight = joint_module_joint_net_2_weight_to_fp16, x = input_13_cast_fp16)[name = string("linear_2_cast_fp16")]; + string linear_2_cast_fp16_to_fp32_dtype_0 = const()[name = string("linear_2_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + tensor logits = cast(dtype = linear_2_cast_fp16_to_fp32_dtype_0, x = linear_2_cast_fp16)[name = string("cast_3")]; + tensor c_out = cast(dtype = obj_cast_fp16_to_fp32_dtype_0, x = obj_cast_fp16)[name = string("cast_5")]; + tensor h_out = cast(dtype = obj_3_cast_fp16_to_fp32_dtype_0, x = obj_3_cast_fp16)[name = string("cast_6")]; + tensor token_length_tmp = identity(x = token_length)[name = string("token_length_tmp")]; + } -> (logits, h_out, c_out); +} \ No newline at end of file diff --git a/it/560ms/decoder_joint.mlmodelc/weights/weight.bin b/it/560ms/decoder_joint.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..33ea2ec2f210db3873bae9b152a8fa5b13171f2e --- /dev/null +++ b/it/560ms/decoder_joint.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e507a69196a04e30adfafc302b6a5f5f527e45c1965c65dd81d63a621cae2064 +size 17315788 diff --git a/it/560ms/decoder_joint.mlpackage/Data/com.apple.CoreML/model.mlmodel b/it/560ms/decoder_joint.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..cc3525ebd701acd827b72a5d6a05caf5ddff80e9 --- /dev/null +++ b/it/560ms/decoder_joint.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5fa6a6d89c1c07ae16f162f5a3b6809b12aafe57a663f5cdb270be3dec7b1427 +size 13745 diff --git a/it/560ms/decoder_joint.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/it/560ms/decoder_joint.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..33ea2ec2f210db3873bae9b152a8fa5b13171f2e --- /dev/null +++ b/it/560ms/decoder_joint.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e507a69196a04e30adfafc302b6a5f5f527e45c1965c65dd81d63a621cae2064 +size 17315788 diff --git a/it/560ms/decoder_joint.mlpackage/Manifest.json b/it/560ms/decoder_joint.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..48ab93415f34542ace6e74a66d563506a10f114a --- /dev/null +++ b/it/560ms/decoder_joint.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "2B19A50C-1D16-4D97-BE3C-D9BCF35884CF": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + }, + "9CA734BC-CFD2-4F39-B068-BE69ABCAAD1F": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + } + }, + "rootModelIdentifier": "2B19A50C-1D16-4D97-BE3C-D9BCF35884CF" +} diff --git a/it/560ms/decoder_joint_noencproj.mlmodelc/analytics/coremldata.bin b/it/560ms/decoder_joint_noencproj.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..3573ed8dea8350501693449f8d9e59b9543d1e3b --- /dev/null +++ b/it/560ms/decoder_joint_noencproj.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:40ec657603479e7dbf8cdb3d6368349eb8b766a52439a26a735d1fadf1b4281d +size 243 diff --git a/it/560ms/decoder_joint_noencproj.mlmodelc/coremldata.bin b/it/560ms/decoder_joint_noencproj.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..94d83d6a74b8a602fbbc8c932d43ab754ba51b88 --- /dev/null +++ b/it/560ms/decoder_joint_noencproj.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c682ba0e028fb7ab6557f8ac1006febc8ec8dd81e4ef8d3a2c05d876e2dbcc8e +size 519 diff --git a/it/560ms/decoder_joint_noencproj.mlmodelc/model.mil b/it/560ms/decoder_joint_noencproj.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..5d5cbd528590956dead59657945f5dab997a7da9 --- /dev/null +++ b/it/560ms/decoder_joint_noencproj.mlmodelc/model.mil @@ -0,0 +1,91 @@ +program(1.3) +[buildInfo = dict({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.10.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor c_in, tensor encoder_proj, tensor h_in, tensor token, tensor token_length) { + int32 y_batch_dims_0 = const()[name = string("y_batch_dims_0"), val = int32(0)]; + bool y_validate_indices_0 = const()[name = string("y_validate_indices_0"), val = bool(false)]; + tensor decoder_module_prediction_embed_weight_to_fp16 = const()[name = string("decoder_module_prediction_embed_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))]; + string token_to_int16_dtype_0 = const()[name = string("token_to_int16_dtype_0"), val = string("int16")]; + string cast_1_dtype_0 = const()[name = string("cast_1_dtype_0"), val = string("int32")]; + int32 greater_equal_0_y_0 = const()[name = string("greater_equal_0_y_0"), val = int32(0)]; + tensor token_to_int16 = cast(dtype = token_to_int16_dtype_0, x = token)[name = string("cast_10")]; + tensor cast_1 = cast(dtype = cast_1_dtype_0, x = token_to_int16)[name = string("cast_9")]; + tensor greater_equal_0 = greater_equal(x = cast_1, y = greater_equal_0_y_0)[name = string("greater_equal_0")]; + int32 slice_by_index_0 = const()[name = string("slice_by_index_0"), val = int32(806)]; + tensor add_2 = add(x = cast_1, y = slice_by_index_0)[name = string("add_2")]; + tensor select_0 = select(a = cast_1, b = add_2, cond = greater_equal_0)[name = string("select_0")]; + int32 y_cast_fp16_cast_uint16_axis_0 = const()[name = string("y_cast_fp16_cast_uint16_axis_0"), val = int32(0)]; + string select_0_to_int16_dtype_0 = const()[name = string("select_0_to_int16_dtype_0"), val = string("int16")]; + tensor select_0_to_int16 = cast(dtype = select_0_to_int16_dtype_0, x = select_0)[name = string("cast_8")]; + tensor y_cast_fp16_cast_uint16_cast_uint16 = gather(axis = y_cast_fp16_cast_uint16_axis_0, batch_dims = y_batch_dims_0, indices = select_0_to_int16, validate_indices = y_validate_indices_0, x = decoder_module_prediction_embed_weight_to_fp16)[name = string("y_cast_fp16_cast_uint16_cast_uint16")]; + tensor input_3_perm_0 = const()[name = string("input_3_perm_0"), val = tensor([1, 0, 2])]; + int32 split_0_num_splits_0 = const()[name = string("split_0_num_splits_0"), val = int32(2)]; + int32 split_0_axis_0 = const()[name = string("split_0_axis_0"), val = int32(0)]; + string h_in_to_fp16_dtype_0 = const()[name = string("h_in_to_fp16_dtype_0"), val = string("fp16")]; + tensor h_in_to_fp16 = cast(dtype = h_in_to_fp16_dtype_0, x = h_in)[name = string("cast_7")]; + tensor split_0_cast_fp16_0, tensor split_0_cast_fp16_1 = split(axis = split_0_axis_0, num_splits = split_0_num_splits_0, x = h_in_to_fp16)[name = string("split_0_cast_fp16")]; + int32 split_1_num_splits_0 = const()[name = string("split_1_num_splits_0"), val = int32(2)]; + int32 split_1_axis_0 = const()[name = string("split_1_axis_0"), val = int32(0)]; + string c_in_to_fp16_dtype_0 = const()[name = string("c_in_to_fp16_dtype_0"), val = string("fp16")]; + tensor c_in_to_fp16 = cast(dtype = c_in_to_fp16_dtype_0, x = c_in)[name = string("cast_6")]; + tensor split_1_cast_fp16_0, tensor split_1_cast_fp16_1 = split(axis = split_1_axis_0, num_splits = split_1_num_splits_0, x = c_in_to_fp16)[name = string("split_1_cast_fp16")]; + tensor input_5_lstm_layer_0_lstm_h0_squeeze_axes_0 = const()[name = string("input_5_lstm_layer_0_lstm_h0_squeeze_axes_0"), val = tensor([0])]; + tensor input_5_lstm_layer_0_lstm_h0_squeeze_cast_fp16 = squeeze(axes = input_5_lstm_layer_0_lstm_h0_squeeze_axes_0, x = split_0_cast_fp16_0)[name = string("input_5_lstm_layer_0_lstm_h0_squeeze_cast_fp16")]; + tensor input_5_lstm_layer_0_lstm_c0_squeeze_axes_0 = const()[name = string("input_5_lstm_layer_0_lstm_c0_squeeze_axes_0"), val = tensor([0])]; + tensor input_5_lstm_layer_0_lstm_c0_squeeze_cast_fp16 = squeeze(axes = input_5_lstm_layer_0_lstm_c0_squeeze_axes_0, x = split_1_cast_fp16_0)[name = string("input_5_lstm_layer_0_lstm_c0_squeeze_cast_fp16")]; + string input_5_lstm_layer_0_direction_0 = const()[name = string("input_5_lstm_layer_0_direction_0"), val = string("forward")]; + bool input_5_lstm_layer_0_output_sequence_0 = const()[name = string("input_5_lstm_layer_0_output_sequence_0"), val = bool(true)]; + string input_5_lstm_layer_0_recurrent_activation_0 = const()[name = string("input_5_lstm_layer_0_recurrent_activation_0"), val = string("sigmoid")]; + string input_5_lstm_layer_0_cell_activation_0 = const()[name = string("input_5_lstm_layer_0_cell_activation_0"), val = string("tanh")]; + string input_5_lstm_layer_0_activation_0 = const()[name = string("input_5_lstm_layer_0_activation_0"), val = string("tanh")]; + tensor concat_1_to_fp16 = const()[name = string("concat_1_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1031808)))]; + tensor concat_2_to_fp16 = const()[name = string("concat_2_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4308672)))]; + tensor concat_0_to_fp16 = const()[name = string("concat_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7585536)))]; + tensor input_3_cast_fp16 = transpose(perm = input_3_perm_0, x = y_cast_fp16_cast_uint16_cast_uint16)[name = string("transpose_3")]; + tensor input_5_lstm_layer_0_cast_fp16_0, tensor input_5_lstm_layer_0_cast_fp16_1, tensor input_5_lstm_layer_0_cast_fp16_2 = lstm(activation = input_5_lstm_layer_0_activation_0, bias = concat_0_to_fp16, cell_activation = input_5_lstm_layer_0_cell_activation_0, direction = input_5_lstm_layer_0_direction_0, initial_c = input_5_lstm_layer_0_lstm_c0_squeeze_cast_fp16, initial_h = input_5_lstm_layer_0_lstm_h0_squeeze_cast_fp16, output_sequence = input_5_lstm_layer_0_output_sequence_0, recurrent_activation = input_5_lstm_layer_0_recurrent_activation_0, weight_hh = concat_2_to_fp16, weight_ih = concat_1_to_fp16, x = input_3_cast_fp16)[name = string("input_5_lstm_layer_0_cast_fp16")]; + tensor input_5_lstm_h0_squeeze_axes_0 = const()[name = string("input_5_lstm_h0_squeeze_axes_0"), val = tensor([0])]; + tensor input_5_lstm_h0_squeeze_cast_fp16 = squeeze(axes = input_5_lstm_h0_squeeze_axes_0, x = split_0_cast_fp16_1)[name = string("input_5_lstm_h0_squeeze_cast_fp16")]; + tensor input_5_lstm_c0_squeeze_axes_0 = const()[name = string("input_5_lstm_c0_squeeze_axes_0"), val = tensor([0])]; + tensor input_5_lstm_c0_squeeze_cast_fp16 = squeeze(axes = input_5_lstm_c0_squeeze_axes_0, x = split_1_cast_fp16_1)[name = string("input_5_lstm_c0_squeeze_cast_fp16")]; + string input_5_direction_0 = const()[name = string("input_5_direction_0"), val = string("forward")]; + bool input_5_output_sequence_0 = const()[name = string("input_5_output_sequence_0"), val = bool(true)]; + string input_5_recurrent_activation_0 = const()[name = string("input_5_recurrent_activation_0"), val = string("sigmoid")]; + string input_5_cell_activation_0 = const()[name = string("input_5_cell_activation_0"), val = string("tanh")]; + string input_5_activation_0 = const()[name = string("input_5_activation_0"), val = string("tanh")]; + tensor concat_4_to_fp16 = const()[name = string("concat_4_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7590720)))]; + tensor concat_5_to_fp16 = const()[name = string("concat_5_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10867584)))]; + tensor concat_3_to_fp16 = const()[name = string("concat_3_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14144448)))]; + tensor input_5_cast_fp16_0, tensor input_5_cast_fp16_1, tensor input_5_cast_fp16_2 = lstm(activation = input_5_activation_0, bias = concat_3_to_fp16, cell_activation = input_5_cell_activation_0, direction = input_5_direction_0, initial_c = input_5_lstm_c0_squeeze_cast_fp16, initial_h = input_5_lstm_h0_squeeze_cast_fp16, output_sequence = input_5_output_sequence_0, recurrent_activation = input_5_recurrent_activation_0, weight_hh = concat_5_to_fp16, weight_ih = concat_4_to_fp16, x = input_5_lstm_layer_0_cast_fp16_0)[name = string("input_5_cast_fp16")]; + int32 obj_3_axis_0 = const()[name = string("obj_3_axis_0"), val = int32(0)]; + tensor obj_3_cast_fp16 = stack(axis = obj_3_axis_0, values = (input_5_lstm_layer_0_cast_fp16_1, input_5_cast_fp16_1))[name = string("obj_3_cast_fp16")]; + string obj_3_cast_fp16_to_fp32_dtype_0 = const()[name = string("obj_3_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + int32 obj_axis_0 = const()[name = string("obj_axis_0"), val = int32(0)]; + tensor obj_cast_fp16 = stack(axis = obj_axis_0, values = (input_5_lstm_layer_0_cast_fp16_2, input_5_cast_fp16_2))[name = string("obj_cast_fp16")]; + string obj_cast_fp16_to_fp32_dtype_0 = const()[name = string("obj_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + tensor transpose_1_perm_0 = const()[name = string("transpose_1_perm_0"), val = tensor([1, 0, 2])]; + tensor joint_module_pred_weight_to_fp16 = const()[name = string("joint_module_pred_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14149632)))]; + tensor joint_module_pred_bias_to_fp16 = const()[name = string("joint_module_pred_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14968896)))]; + tensor transpose_1_cast_fp16 = transpose(perm = transpose_1_perm_0, x = input_5_cast_fp16_0)[name = string("transpose_2")]; + tensor linear_0_cast_fp16 = linear(bias = joint_module_pred_bias_to_fp16, weight = joint_module_pred_weight_to_fp16, x = transpose_1_cast_fp16)[name = string("linear_0_cast_fp16")]; + tensor f_axes_0 = const()[name = string("f_axes_0"), val = tensor([2])]; + string encoder_proj_to_fp16_dtype_0 = const()[name = string("encoder_proj_to_fp16_dtype_0"), val = string("fp16")]; + tensor encoder_proj_to_fp16 = cast(dtype = encoder_proj_to_fp16_dtype_0, x = encoder_proj)[name = string("cast_3")]; + tensor f_cast_fp16 = expand_dims(axes = f_axes_0, x = encoder_proj_to_fp16)[name = string("f_cast_fp16")]; + tensor g_axes_0 = const()[name = string("g_axes_0"), val = tensor([1])]; + tensor g_cast_fp16 = expand_dims(axes = g_axes_0, x = linear_0_cast_fp16)[name = string("g_cast_fp16")]; + tensor input_9_cast_fp16 = add(x = f_cast_fp16, y = g_cast_fp16)[name = string("input_9_cast_fp16")]; + tensor input_11_cast_fp16 = relu(x = input_9_cast_fp16)[name = string("input_11_cast_fp16")]; + tensor joint_module_joint_net_2_weight_to_fp16 = const()[name = string("joint_module_joint_net_2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14970240)))]; + tensor joint_module_joint_net_2_bias_to_fp16 = const()[name = string("joint_module_joint_net_2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16001984)))]; + tensor linear_1_cast_fp16 = linear(bias = joint_module_joint_net_2_bias_to_fp16, weight = joint_module_joint_net_2_weight_to_fp16, x = input_11_cast_fp16)[name = string("linear_1_cast_fp16")]; + int32 var_83 = const()[name = string("op_83"), val = int32(-1)]; + tensor var_85_softmax_cast_fp16 = softmax(axis = var_83, x = linear_1_cast_fp16)[name = string("op_85_softmax_cast_fp16")]; + fp32 var_85_epsilon_0 = const()[name = string("op_85_epsilon_0"), val = fp32(0x1p-149)]; + tensor var_85_cast_fp16 = log(epsilon = var_85_epsilon_0, x = var_85_softmax_cast_fp16)[name = string("op_85_cast_fp16")]; + string var_85_cast_fp16_to_fp32_dtype_0 = const()[name = string("op_85_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + tensor logits = cast(dtype = var_85_cast_fp16_to_fp32_dtype_0, x = var_85_cast_fp16)[name = string("cast_2")]; + tensor c_out = cast(dtype = obj_cast_fp16_to_fp32_dtype_0, x = obj_cast_fp16)[name = string("cast_4")]; + tensor h_out = cast(dtype = obj_3_cast_fp16_to_fp32_dtype_0, x = obj_3_cast_fp16)[name = string("cast_5")]; + tensor token_length_tmp = identity(x = token_length)[name = string("token_length_tmp")]; + } -> (logits, h_out, c_out); +} \ No newline at end of file diff --git a/it/560ms/decoder_joint_noencproj.mlmodelc/weights/weight.bin b/it/560ms/decoder_joint_noencproj.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..80af5cec724e7e6b117fcd6a7bc8046c27c26e75 --- /dev/null +++ b/it/560ms/decoder_joint_noencproj.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5ca6a467ebf44612a8032c6c4ddf323e35a9ffaa15c01822528fb97144ec9439 +size 16003660 diff --git a/it/560ms/decoder_joint_noencproj.mlpackage/Data/com.apple.CoreML/model.mlmodel b/it/560ms/decoder_joint_noencproj.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..492f62656a116fa907f67e5963520bcaa37729f8 --- /dev/null +++ b/it/560ms/decoder_joint_noencproj.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1a46ef6436e8c1e5af6f379658d871f2a40dff3a13c946d4f8df736b19cdc45e +size 14630 diff --git a/it/560ms/decoder_joint_noencproj.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/it/560ms/decoder_joint_noencproj.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..80af5cec724e7e6b117fcd6a7bc8046c27c26e75 --- /dev/null +++ b/it/560ms/decoder_joint_noencproj.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5ca6a467ebf44612a8032c6c4ddf323e35a9ffaa15c01822528fb97144ec9439 +size 16003660 diff --git a/it/560ms/decoder_joint_noencproj.mlpackage/Manifest.json b/it/560ms/decoder_joint_noencproj.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..2135f9be1a1ec8555bef5683841b910c941cbd7c --- /dev/null +++ b/it/560ms/decoder_joint_noencproj.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "E92AE661-EE59-447C-B9BE-52D50BC4FC37": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + }, + "F7138F3F-8633-4C54-A3F4-D0F6E8220E39": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + } + }, + "rootModelIdentifier": "F7138F3F-8633-4C54-A3F4-D0F6E8220E39" +} diff --git a/it/560ms/encoder.mlmodelc/analytics/coremldata.bin b/it/560ms/encoder.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..f631009aa2b7c2a2f48501a3e27f90a0688e220e --- /dev/null +++ b/it/560ms/encoder.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c7172b10d4d8cda6411e6356b4ba90d9d72d946859055983012fa4a9c9984deb +size 243 diff --git a/it/560ms/encoder.mlmodelc/coremldata.bin b/it/560ms/encoder.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..0704d32b82f55832133510e3cbfd7612958c4682 --- /dev/null +++ b/it/560ms/encoder.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:07e85ef7679c45bc47ce04c6fb7970c46b317dea36d6b2bc4545334637e09aca +size 632 diff --git a/it/560ms/encoder.mlmodelc/model.mil b/it/560ms/encoder.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..d6f0179d3761ce7d81034c775fc7550c16626d14 --- /dev/null +++ b/it/560ms/encoder.mlmodelc/model.mil @@ -0,0 +1,4434 @@ +program(1.3) +[buildInfo = dict({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}})] +{ + func main(tensor cache_channel, tensor cache_len, tensor cache_time, tensor mel, tensor mel_length, tensor prompt_id) { + tensor value_3_perm_0 = const()[name = string("value_3_perm_0"), val = tensor([1, 0, 2, 3])]; + string cache_channel_to_fp16_dtype_0 = const()[name = string("cache_channel_to_fp16_dtype_0"), val = string("fp16")]; + tensor value_5_perm_0 = const()[name = string("value_5_perm_0"), val = tensor([1, 0, 2, 3])]; + string cache_time_to_fp16_dtype_0 = const()[name = string("cache_time_to_fp16_dtype_0"), val = string("fp16")]; + int32 var_59 = const()[name = string("op_59"), val = int32(-1)]; + int32 var_68 = const()[name = string("op_68"), val = int32(1)]; + tensor x_1_perm_0 = const()[name = string("x_1_perm_0"), val = tensor([0, 2, 1])]; + string mel_to_fp16_dtype_0 = const()[name = string("mel_to_fp16_dtype_0"), val = string("fp16")]; + tensor tensor_1_axes_0 = const()[name = string("tensor_1_axes_0"), val = tensor([1])]; + tensor mel_to_fp16 = cast(dtype = mel_to_fp16_dtype_0, x = mel)[name = string("cast_21")]; + tensor x_1_cast_fp16 = transpose(perm = x_1_perm_0, x = mel_to_fp16)[name = string("transpose_367")]; + tensor tensor_1_cast_fp16 = expand_dims(axes = tensor_1_axes_0, x = x_1_cast_fp16)[name = string("tensor_1_cast_fp16")]; + tensor expand_dims_0 = const()[name = string("expand_dims_0"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))]; + tensor var_137_axes_0 = const()[name = string("op_137_axes_0"), val = tensor([1])]; + tensor var_137 = expand_dims(axes = var_137_axes_0, x = mel_length)[name = string("op_137")]; + tensor time_mask_1 = less(x = expand_dims_0, y = var_137)[name = string("time_mask_1")]; + tensor var_139_axes_0 = const()[name = string("op_139_axes_0"), val = tensor([-1])]; + tensor var_139 = expand_dims(axes = var_139_axes_0, x = time_mask_1)[name = string("op_139")]; + tensor var_141_reps_0 = const()[name = string("op_141_reps_0"), val = tensor([1, 1, 128])]; + tensor var_141 = tile(reps = var_141_reps_0, x = var_139)[name = string("op_141")]; + tensor var_147_axes_0 = const()[name = string("op_147_axes_0"), val = tensor([1])]; + string mask_1_to_fp16_dtype_0 = const()[name = string("mask_1_to_fp16_dtype_0"), val = string("fp16")]; + tensor var_141_to_fp16 = cast(dtype = mask_1_to_fp16_dtype_0, x = var_141)[name = string("cast_20")]; + tensor var_147_cast_fp16 = expand_dims(axes = var_147_axes_0, x = var_141_to_fp16)[name = string("op_147_cast_fp16")]; + tensor input_1_cast_fp16 = mul(x = tensor_1_cast_fp16, y = var_147_cast_fp16)[name = string("input_1_cast_fp16")]; + tensor input_3_pad_0 = const()[name = string("input_3_pad_0"), val = tensor([0, 0, 0, 0, 2, 1, 2, 1])]; + string input_3_mode_0 = const()[name = string("input_3_mode_0"), val = string("constant")]; + fp16 const_9_to_fp16 = const()[name = string("const_9_to_fp16"), val = fp16(0x0p+0)]; + tensor input_3_cast_fp16 = pad(constant_val = const_9_to_fp16, mode = input_3_mode_0, pad = input_3_pad_0, x = input_1_cast_fp16)[name = string("input_3_cast_fp16")]; + string tensor_3_pad_type_0 = const()[name = string("tensor_3_pad_type_0"), val = string("valid")]; + tensor tensor_3_strides_0 = const()[name = string("tensor_3_strides_0"), val = tensor([2, 2])]; + tensor tensor_3_pad_0 = const()[name = string("tensor_3_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor tensor_3_dilations_0 = const()[name = string("tensor_3_dilations_0"), val = tensor([1, 1])]; + int32 tensor_3_groups_0 = const()[name = string("tensor_3_groups_0"), val = int32(1)]; + tensor encoder_pre_encode_conv_0_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(448))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2816))))[name = string("encoder_pre_encode_conv_0_weight_to_fp16_quantized")]; + tensor encoder_pre_encode_conv_0_bias_to_fp16 = const()[name = string("encoder_pre_encode_conv_0_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3392)))]; + tensor tensor_3_cast_fp16 = conv(bias = encoder_pre_encode_conv_0_bias_to_fp16, dilations = tensor_3_dilations_0, groups = tensor_3_groups_0, pad = tensor_3_pad_0, pad_type = tensor_3_pad_type_0, strides = tensor_3_strides_0, weight = encoder_pre_encode_conv_0_weight_to_fp16_quantized, x = input_3_cast_fp16)[name = string("tensor_3_cast_fp16")]; + string current_lengths_1_to_fp16_dtype_0 = const()[name = string("current_lengths_1_to_fp16_dtype_0"), val = string("fp16")]; + fp16 var_160_promoted_to_fp16 = const()[name = string("op_160_promoted_to_fp16"), val = fp16(0x1p+1)]; + tensor mel_length_to_fp16 = cast(dtype = current_lengths_1_to_fp16_dtype_0, x = mel_length)[name = string("cast_19")]; + tensor var_161_cast_fp16 = add(x = mel_length_to_fp16, y = var_160_promoted_to_fp16)[name = string("op_161_cast_fp16")]; + fp16 var_162_promoted_to_fp16 = const()[name = string("op_162_promoted_to_fp16"), val = fp16(0x1p+0)]; + tensor var_163_cast_fp16 = add(x = var_161_cast_fp16, y = var_162_promoted_to_fp16)[name = string("op_163_cast_fp16")]; + fp16 var_164_promoted_to_fp16 = const()[name = string("op_164_promoted_to_fp16"), val = fp16(0x1.8p+1)]; + tensor var_165_cast_fp16 = sub(x = var_163_cast_fp16, y = var_164_promoted_to_fp16)[name = string("op_165_cast_fp16")]; + fp16 var_56_promoted_to_fp16 = const()[name = string("op_56_promoted_to_fp16"), val = fp16(0x1p+1)]; + tensor floor_div_0_cast_fp16 = floor_div(x = var_165_cast_fp16, y = var_56_promoted_to_fp16)[name = string("floor_div_0_cast_fp16")]; + fp16 var_167_promoted_to_fp16 = const()[name = string("op_167_promoted_to_fp16"), val = fp16(0x1p+0)]; + tensor current_lengths_3_cast_fp16 = add(x = floor_div_0_cast_fp16, y = var_167_promoted_to_fp16)[name = string("current_lengths_3_cast_fp16")]; + string lengths_19_dtype_0 = const()[name = string("lengths_19_dtype_0"), val = string("int32")]; + tensor expand_dims_1 = const()[name = string("expand_dims_1"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3968)))]; + tensor var_176_axes_0 = const()[name = string("op_176_axes_0"), val = tensor([1])]; + tensor current_lengths_3_cast_fp16_to_int32 = cast(dtype = lengths_19_dtype_0, x = current_lengths_3_cast_fp16)[name = string("cast_18")]; + tensor var_176 = expand_dims(axes = var_176_axes_0, x = current_lengths_3_cast_fp16_to_int32)[name = string("op_176")]; + tensor time_mask_3 = less(x = expand_dims_1, y = var_176)[name = string("time_mask_3")]; + tensor var_178_axes_0 = const()[name = string("op_178_axes_0"), val = tensor([-1])]; + tensor var_178 = expand_dims(axes = var_178_axes_0, x = time_mask_3)[name = string("op_178")]; + tensor var_180_reps_0 = const()[name = string("op_180_reps_0"), val = tensor([1, 1, 65])]; + tensor var_180 = tile(reps = var_180_reps_0, x = var_178)[name = string("op_180")]; + tensor var_186_axes_0 = const()[name = string("op_186_axes_0"), val = tensor([1])]; + string mask_3_to_fp16_dtype_0 = const()[name = string("mask_3_to_fp16_dtype_0"), val = string("fp16")]; + tensor var_180_to_fp16 = cast(dtype = mask_3_to_fp16_dtype_0, x = var_180)[name = string("cast_17")]; + tensor var_186_cast_fp16 = expand_dims(axes = var_186_axes_0, x = var_180_to_fp16)[name = string("op_186_cast_fp16")]; + tensor expanded_mask_3_reps_0 = const()[name = string("expanded_mask_3_reps_0"), val = tensor([1, 256, 1, 1])]; + tensor expanded_mask_3_cast_fp16 = tile(reps = expanded_mask_3_reps_0, x = var_186_cast_fp16)[name = string("expanded_mask_3_cast_fp16")]; + tensor input_5_cast_fp16 = mul(x = tensor_3_cast_fp16, y = expanded_mask_3_cast_fp16)[name = string("input_5_cast_fp16")]; + tensor tensor_5_cast_fp16 = relu(x = input_5_cast_fp16)[name = string("tensor_5_cast_fp16")]; + tensor input_7_cast_fp16 = mul(x = tensor_5_cast_fp16, y = expanded_mask_3_cast_fp16)[name = string("input_7_cast_fp16")]; + tensor input_9_pad_0 = const()[name = string("input_9_pad_0"), val = tensor([0, 0, 0, 0, 2, 1, 2, 1])]; + string input_9_mode_0 = const()[name = string("input_9_mode_0"), val = string("constant")]; + fp16 const_23_to_fp16 = const()[name = string("const_23_to_fp16"), val = fp16(0x0p+0)]; + tensor input_9_cast_fp16 = pad(constant_val = const_23_to_fp16, mode = input_9_mode_0, pad = input_9_pad_0, x = input_7_cast_fp16)[name = string("input_9_cast_fp16")]; + string tensor_7_pad_type_0 = const()[name = string("tensor_7_pad_type_0"), val = string("valid")]; + tensor tensor_7_strides_0 = const()[name = string("tensor_7_strides_0"), val = tensor([2, 2])]; + int32 tensor_7_groups_0 = const()[name = string("tensor_7_groups_0"), val = int32(256)]; + tensor tensor_7_pad_0 = const()[name = string("tensor_7_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor tensor_7_dilations_0 = const()[name = string("tensor_7_dilations_0"), val = tensor([1, 1])]; + tensor encoder_pre_encode_conv_2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4224))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6592))))[name = string("encoder_pre_encode_conv_2_weight_to_fp16_quantized")]; + tensor encoder_pre_encode_conv_2_bias_to_fp16 = const()[name = string("encoder_pre_encode_conv_2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7168)))]; + tensor tensor_7_cast_fp16 = conv(bias = encoder_pre_encode_conv_2_bias_to_fp16, dilations = tensor_7_dilations_0, groups = tensor_7_groups_0, pad = tensor_7_pad_0, pad_type = tensor_7_pad_type_0, strides = tensor_7_strides_0, weight = encoder_pre_encode_conv_2_weight_to_fp16_quantized, x = input_9_cast_fp16)[name = string("tensor_7_cast_fp16")]; + fp16 var_208_promoted_to_fp16 = const()[name = string("op_208_promoted_to_fp16"), val = fp16(0x1p+1)]; + tensor var_209_cast_fp16 = add(x = current_lengths_3_cast_fp16, y = var_208_promoted_to_fp16)[name = string("op_209_cast_fp16")]; + fp16 var_210_promoted_to_fp16 = const()[name = string("op_210_promoted_to_fp16"), val = fp16(0x1p+0)]; + tensor var_211_cast_fp16 = add(x = var_209_cast_fp16, y = var_210_promoted_to_fp16)[name = string("op_211_cast_fp16")]; + fp16 var_212_promoted_to_fp16 = const()[name = string("op_212_promoted_to_fp16"), val = fp16(0x1.8p+1)]; + tensor var_213_cast_fp16 = sub(x = var_211_cast_fp16, y = var_212_promoted_to_fp16)[name = string("op_213_cast_fp16")]; + fp16 var_56_promoted_1_to_fp16 = const()[name = string("op_56_promoted_1_to_fp16"), val = fp16(0x1p+1)]; + tensor floor_div_1_cast_fp16 = floor_div(x = var_213_cast_fp16, y = var_56_promoted_1_to_fp16)[name = string("floor_div_1_cast_fp16")]; + fp16 var_215_promoted_to_fp16 = const()[name = string("op_215_promoted_to_fp16"), val = fp16(0x1p+0)]; + tensor current_lengths_5_cast_fp16 = add(x = floor_div_1_cast_fp16, y = var_215_promoted_to_fp16)[name = string("current_lengths_5_cast_fp16")]; + string lengths_21_dtype_0 = const()[name = string("lengths_21_dtype_0"), val = string("int32")]; + tensor expand_dims_2 = const()[name = string("expand_dims_2"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7744)))]; + tensor var_224_axes_0 = const()[name = string("op_224_axes_0"), val = tensor([1])]; + tensor current_lengths_5_cast_fp16_to_int32 = cast(dtype = lengths_21_dtype_0, x = current_lengths_5_cast_fp16)[name = string("cast_16")]; + tensor var_224 = expand_dims(axes = var_224_axes_0, x = current_lengths_5_cast_fp16_to_int32)[name = string("op_224")]; + tensor time_mask_5 = less(x = expand_dims_2, y = var_224)[name = string("time_mask_5")]; + tensor var_226_axes_0 = const()[name = string("op_226_axes_0"), val = tensor([-1])]; + tensor var_226 = expand_dims(axes = var_226_axes_0, x = time_mask_5)[name = string("op_226")]; + tensor var_228_reps_0 = const()[name = string("op_228_reps_0"), val = tensor([1, 1, 33])]; + tensor var_228 = tile(reps = var_228_reps_0, x = var_226)[name = string("op_228")]; + tensor var_234_axes_0 = const()[name = string("op_234_axes_0"), val = tensor([1])]; + string mask_5_to_fp16_dtype_0 = const()[name = string("mask_5_to_fp16_dtype_0"), val = string("fp16")]; + tensor var_228_to_fp16 = cast(dtype = mask_5_to_fp16_dtype_0, x = var_228)[name = string("cast_15")]; + tensor var_234_cast_fp16 = expand_dims(axes = var_234_axes_0, x = var_228_to_fp16)[name = string("op_234_cast_fp16")]; + tensor expanded_mask_7_reps_0 = const()[name = string("expanded_mask_7_reps_0"), val = tensor([1, 256, 1, 1])]; + tensor expanded_mask_7_cast_fp16 = tile(reps = expanded_mask_7_reps_0, x = var_234_cast_fp16)[name = string("expanded_mask_7_cast_fp16")]; + tensor input_11_cast_fp16 = mul(x = tensor_7_cast_fp16, y = expanded_mask_7_cast_fp16)[name = string("input_11_cast_fp16")]; + string tensor_9_pad_type_0 = const()[name = string("tensor_9_pad_type_0"), val = string("valid")]; + tensor tensor_9_strides_0 = const()[name = string("tensor_9_strides_0"), val = tensor([1, 1])]; + tensor tensor_9_pad_0 = const()[name = string("tensor_9_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor tensor_9_dilations_0 = const()[name = string("tensor_9_dilations_0"), val = tensor([1, 1])]; + int32 tensor_9_groups_0 = const()[name = string("tensor_9_groups_0"), val = int32(1)]; + tensor encoder_pre_encode_conv_3_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7936))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(73536))))[name = string("encoder_pre_encode_conv_3_weight_to_fp16_quantized")]; + tensor encoder_pre_encode_conv_3_bias_to_fp16 = const()[name = string("encoder_pre_encode_conv_3_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(74112)))]; + tensor tensor_9_cast_fp16 = conv(bias = encoder_pre_encode_conv_3_bias_to_fp16, dilations = tensor_9_dilations_0, groups = tensor_9_groups_0, pad = tensor_9_pad_0, pad_type = tensor_9_pad_type_0, strides = tensor_9_strides_0, weight = encoder_pre_encode_conv_3_weight_to_fp16_quantized, x = input_11_cast_fp16)[name = string("tensor_9_cast_fp16")]; + tensor input_13_cast_fp16 = mul(x = tensor_9_cast_fp16, y = expanded_mask_7_cast_fp16)[name = string("input_13_cast_fp16")]; + tensor tensor_11_cast_fp16 = relu(x = input_13_cast_fp16)[name = string("tensor_11_cast_fp16")]; + tensor input_15_cast_fp16 = mul(x = tensor_11_cast_fp16, y = expanded_mask_7_cast_fp16)[name = string("input_15_cast_fp16")]; + tensor input_17_pad_0 = const()[name = string("input_17_pad_0"), val = tensor([0, 0, 0, 0, 2, 1, 2, 1])]; + string input_17_mode_0 = const()[name = string("input_17_mode_0"), val = string("constant")]; + fp16 const_41_to_fp16 = const()[name = string("const_41_to_fp16"), val = fp16(0x0p+0)]; + tensor input_17_cast_fp16 = pad(constant_val = const_41_to_fp16, mode = input_17_mode_0, pad = input_17_pad_0, x = input_15_cast_fp16)[name = string("input_17_cast_fp16")]; + string tensor_13_pad_type_0 = const()[name = string("tensor_13_pad_type_0"), val = string("valid")]; + tensor tensor_13_strides_0 = const()[name = string("tensor_13_strides_0"), val = tensor([2, 2])]; + int32 tensor_13_groups_0 = const()[name = string("tensor_13_groups_0"), val = int32(256)]; + tensor tensor_13_pad_0 = const()[name = string("tensor_13_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor tensor_13_dilations_0 = const()[name = string("tensor_13_dilations_0"), val = tensor([1, 1])]; + tensor encoder_pre_encode_conv_5_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(74688))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77056))))[name = string("encoder_pre_encode_conv_5_weight_to_fp16_quantized")]; + tensor encoder_pre_encode_conv_5_bias_to_fp16 = const()[name = string("encoder_pre_encode_conv_5_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77632)))]; + tensor tensor_13_cast_fp16 = conv(bias = encoder_pre_encode_conv_5_bias_to_fp16, dilations = tensor_13_dilations_0, groups = tensor_13_groups_0, pad = tensor_13_pad_0, pad_type = tensor_13_pad_type_0, strides = tensor_13_strides_0, weight = encoder_pre_encode_conv_5_weight_to_fp16_quantized, x = input_17_cast_fp16)[name = string("tensor_13_cast_fp16")]; + fp16 var_271_promoted_to_fp16 = const()[name = string("op_271_promoted_to_fp16"), val = fp16(0x1p+1)]; + tensor var_272_cast_fp16 = add(x = current_lengths_5_cast_fp16, y = var_271_promoted_to_fp16)[name = string("op_272_cast_fp16")]; + fp16 var_273_promoted_to_fp16 = const()[name = string("op_273_promoted_to_fp16"), val = fp16(0x1p+0)]; + tensor var_274_cast_fp16 = add(x = var_272_cast_fp16, y = var_273_promoted_to_fp16)[name = string("op_274_cast_fp16")]; + fp16 var_275_promoted_to_fp16 = const()[name = string("op_275_promoted_to_fp16"), val = fp16(0x1.8p+1)]; + tensor var_276_cast_fp16 = sub(x = var_274_cast_fp16, y = var_275_promoted_to_fp16)[name = string("op_276_cast_fp16")]; + fp16 var_56_promoted_2_to_fp16 = const()[name = string("op_56_promoted_2_to_fp16"), val = fp16(0x1p+1)]; + tensor floor_div_2_cast_fp16 = floor_div(x = var_276_cast_fp16, y = var_56_promoted_2_to_fp16)[name = string("floor_div_2_cast_fp16")]; + fp16 var_278_promoted_to_fp16 = const()[name = string("op_278_promoted_to_fp16"), val = fp16(0x1p+0)]; + tensor current_lengths_cast_fp16 = add(x = floor_div_2_cast_fp16, y = var_278_promoted_to_fp16)[name = string("current_lengths_cast_fp16")]; + string lengths_dtype_0 = const()[name = string("lengths_dtype_0"), val = string("int32")]; + tensor expand_dims_3 = const()[name = string("expand_dims_3"), val = tensor([[0, 1, 2, 3, 4, 5, 6, 7, 8]])]; + tensor var_287_axes_0 = const()[name = string("op_287_axes_0"), val = tensor([1])]; + tensor current_lengths_cast_fp16_to_int32 = cast(dtype = lengths_dtype_0, x = current_lengths_cast_fp16)[name = string("cast_14")]; + tensor var_287 = expand_dims(axes = var_287_axes_0, x = current_lengths_cast_fp16_to_int32)[name = string("op_287")]; + tensor time_mask = less(x = expand_dims_3, y = var_287)[name = string("time_mask")]; + tensor var_289_axes_0 = const()[name = string("op_289_axes_0"), val = tensor([-1])]; + tensor var_289 = expand_dims(axes = var_289_axes_0, x = time_mask)[name = string("op_289")]; + tensor var_291_reps_0 = const()[name = string("op_291_reps_0"), val = tensor([1, 1, 17])]; + tensor var_291 = tile(reps = var_291_reps_0, x = var_289)[name = string("op_291")]; + tensor var_297_axes_0 = const()[name = string("op_297_axes_0"), val = tensor([1])]; + string mask_7_to_fp16_dtype_0 = const()[name = string("mask_7_to_fp16_dtype_0"), val = string("fp16")]; + tensor var_291_to_fp16 = cast(dtype = mask_7_to_fp16_dtype_0, x = var_291)[name = string("cast_13")]; + tensor var_297_cast_fp16 = expand_dims(axes = var_297_axes_0, x = var_291_to_fp16)[name = string("op_297_cast_fp16")]; + tensor expanded_mask_13_reps_0 = const()[name = string("expanded_mask_13_reps_0"), val = tensor([1, 256, 1, 1])]; + tensor expanded_mask_13_cast_fp16 = tile(reps = expanded_mask_13_reps_0, x = var_297_cast_fp16)[name = string("expanded_mask_13_cast_fp16")]; + tensor input_19_cast_fp16 = mul(x = tensor_13_cast_fp16, y = expanded_mask_13_cast_fp16)[name = string("input_19_cast_fp16")]; + string tensor_15_pad_type_0 = const()[name = string("tensor_15_pad_type_0"), val = string("valid")]; + tensor tensor_15_strides_0 = const()[name = string("tensor_15_strides_0"), val = tensor([1, 1])]; + tensor tensor_15_pad_0 = const()[name = string("tensor_15_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor tensor_15_dilations_0 = const()[name = string("tensor_15_dilations_0"), val = tensor([1, 1])]; + int32 tensor_15_groups_0 = const()[name = string("tensor_15_groups_0"), val = int32(1)]; + tensor encoder_pre_encode_conv_6_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78208))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(143808))))[name = string("encoder_pre_encode_conv_6_weight_to_fp16_quantized")]; + tensor encoder_pre_encode_conv_6_bias_to_fp16 = const()[name = string("encoder_pre_encode_conv_6_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144384)))]; + tensor tensor_15_cast_fp16 = conv(bias = encoder_pre_encode_conv_6_bias_to_fp16, dilations = tensor_15_dilations_0, groups = tensor_15_groups_0, pad = tensor_15_pad_0, pad_type = tensor_15_pad_type_0, strides = tensor_15_strides_0, weight = encoder_pre_encode_conv_6_weight_to_fp16_quantized, x = input_19_cast_fp16)[name = string("tensor_15_cast_fp16")]; + tensor input_21_cast_fp16 = mul(x = tensor_15_cast_fp16, y = expanded_mask_13_cast_fp16)[name = string("input_21_cast_fp16")]; + tensor tensor_cast_fp16 = relu(x = input_21_cast_fp16)[name = string("tensor_cast_fp16")]; + tensor x_3_cast_fp16 = mul(x = tensor_cast_fp16, y = expanded_mask_13_cast_fp16)[name = string("x_3_cast_fp16")]; + tensor var_331_perm_0 = const()[name = string("op_331_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_332 = const()[name = string("op_332"), val = tensor([1, 9, -1])]; + tensor var_331_cast_fp16 = transpose(perm = var_331_perm_0, x = x_3_cast_fp16)[name = string("transpose_366")]; + tensor input_23_cast_fp16 = reshape(shape = var_332, x = var_331_cast_fp16)[name = string("input_23_cast_fp16")]; + tensor encoder_pre_encode_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144960))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4601472))))[name = string("encoder_pre_encode_out_weight_to_fp16_quantized")]; + tensor encoder_pre_encode_out_bias_to_fp16 = const()[name = string("encoder_pre_encode_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4603584)))]; + tensor linear_0_cast_fp16 = linear(bias = encoder_pre_encode_out_bias_to_fp16, weight = encoder_pre_encode_out_weight_to_fp16_quantized, x = input_23_cast_fp16)[name = string("linear_0_cast_fp16")]; + tensor var_342_begin_0 = const()[name = string("op_342_begin_0"), val = tensor([0, 2, 0])]; + tensor var_342_end_0 = const()[name = string("op_342_end_0"), val = tensor([1, 9, 1024])]; + tensor var_342_end_mask_0 = const()[name = string("op_342_end_mask_0"), val = tensor([true, true, true])]; + tensor var_342_cast_fp16 = slice_by_index(begin = var_342_begin_0, end = var_342_end_0, end_mask = var_342_end_mask_0, x = linear_0_cast_fp16)[name = string("op_342_cast_fp16")]; + int32 var_344 = const()[name = string("op_344"), val = int32(2)]; + tensor var_345 = sub(x = current_lengths_cast_fp16_to_int32, y = var_344)[name = string("op_345")]; + string var_345_promoted_to_fp16_dtype_0 = const()[name = string("op_345_promoted_to_fp16_dtype_0"), val = string("fp16")]; + fp16 var_62_promoted_to_fp16 = const()[name = string("op_62_promoted_to_fp16"), val = fp16(0x0p+0)]; + fp16 const_61_to_fp16 = const()[name = string("const_61_to_fp16"), val = fp16(inf)]; + tensor var_345_to_fp16 = cast(dtype = var_345_promoted_to_fp16_dtype_0, x = var_345)[name = string("cast_12")]; + tensor clip_0_cast_fp16 = clip(alpha = var_62_promoted_to_fp16, beta = const_61_to_fp16, x = var_345_to_fp16)[name = string("clip_0_cast_fp16")]; + tensor max_audio_length_1 = const()[name = string("max_audio_length_1"), val = tensor([7])]; + fp16 var_361_promoted_to_fp16 = const()[name = string("op_361_promoted_to_fp16"), val = fp16(0x1.5p+5)]; + tensor padding_length_cast_fp16 = add(x = clip_0_cast_fp16, y = var_361_promoted_to_fp16)[name = string("padding_length_cast_fp16")]; + int32 const_63 = const()[name = string("const_63"), val = int32(-1)]; + tensor var_363 = mul(x = cache_len, y = const_63)[name = string("op_363")]; + int32 var_364 = const()[name = string("op_364"), val = int32(42)]; + tensor offset = add(x = var_363, y = var_364)[name = string("offset")]; + tensor var_404_axes_0 = const()[name = string("op_404_axes_0"), val = tensor([-1])]; + tensor var_404_cast_fp16 = expand_dims(axes = var_404_axes_0, x = padding_length_cast_fp16)[name = string("op_404_cast_fp16")]; + tensor var_403_promoted_to_fp16 = const()[name = string("op_403_promoted_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4605696)))]; + tensor pad_mask_1_cast_fp16 = less(x = var_403_promoted_to_fp16, y = var_404_cast_fp16)[name = string("pad_mask_1_cast_fp16")]; + tensor expand_dims_5 = const()[name = string("expand_dims_5"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4605888)))]; + tensor var_410_axes_0 = const()[name = string("op_410_axes_0"), val = tensor([-1])]; + tensor var_410 = expand_dims(axes = var_410_axes_0, x = offset)[name = string("op_410")]; + tensor pad_mask_off = greater_equal(x = expand_dims_5, y = var_410)[name = string("pad_mask_off")]; + tensor pad_mask_3 = logical_and(x = pad_mask_off, y = pad_mask_1_cast_fp16)[name = string("pad_mask_3")]; + tensor var_413_axes_0 = const()[name = string("op_413_axes_0"), val = tensor([1])]; + tensor var_413 = expand_dims(axes = var_413_axes_0, x = pad_mask_3)[name = string("op_413")]; + tensor var_414 = const()[name = string("op_414"), val = tensor([1, 49, 1])]; + tensor pad_mask_for_att_mask_1 = tile(reps = var_414, x = var_413)[name = string("pad_mask_for_att_mask_1")]; + tensor var_416_perm_0 = const()[name = string("op_416_perm_0"), val = tensor([0, 2, 1])]; + tensor var_416 = transpose(perm = var_416_perm_0, x = pad_mask_for_att_mask_1)[name = string("transpose_365")]; + tensor pad_mask_for_att_mask = logical_and(x = pad_mask_for_att_mask_1, y = var_416)[name = string("pad_mask_for_att_mask")]; + tensor const_71 = const()[name = string("const_71"), val = tensor([[[true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true]]])]; + tensor att_mask_9 = logical_and(x = pad_mask_for_att_mask, y = const_71)[name = string("att_mask_9")]; + tensor att_mask = logical_not(x = att_mask_9)[name = string("att_mask")]; + tensor pad_mask_5 = logical_not(x = pad_mask_3)[name = string("pad_mask_5")]; + tensor pad_mask_begin_0 = const()[name = string("pad_mask_begin_0"), val = tensor([0, 42])]; + tensor pad_mask_end_0 = const()[name = string("pad_mask_end_0"), val = tensor([1, 49])]; + tensor pad_mask_end_mask_0 = const()[name = string("pad_mask_end_mask_0"), val = tensor([true, true])]; + tensor pad_mask = slice_by_index(begin = pad_mask_begin_0, end = pad_mask_end_0, end_mask = pad_mask_end_mask_0, x = pad_mask_5)[name = string("pad_mask")]; + tensor mask_9_begin_0 = const()[name = string("mask_9_begin_0"), val = tensor([0, 42, 0])]; + tensor mask_9_end_0 = const()[name = string("mask_9_end_0"), val = tensor([1, 49, 49])]; + tensor mask_9_end_mask_0 = const()[name = string("mask_9_end_mask_0"), val = tensor([true, true, true])]; + tensor mask_9 = slice_by_index(begin = mask_9_begin_0, end = mask_9_end_0, end_mask = mask_9_end_mask_0, x = att_mask)[name = string("mask_9")]; + tensor cache_1_begin_0 = const()[name = string("cache_1_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor cache_1_end_0 = const()[name = string("cache_1_end_0"), val = tensor([1, 1, 42, 1024])]; + tensor cache_1_end_mask_0 = const()[name = string("cache_1_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_1_squeeze_mask_0 = const()[name = string("cache_1_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_channel_to_fp16 = cast(dtype = cache_channel_to_fp16_dtype_0, x = cache_channel)[name = string("cast_11")]; + tensor value_3_cast_fp16 = transpose(perm = value_3_perm_0, x = cache_channel_to_fp16)[name = string("transpose_364")]; + tensor cache_1_cast_fp16 = slice_by_index(begin = cache_1_begin_0, end = cache_1_end_0, end_mask = cache_1_end_mask_0, squeeze_mask = cache_1_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_1_cast_fp16")]; + tensor cache_3_begin_0 = const()[name = string("cache_3_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor cache_3_end_0 = const()[name = string("cache_3_end_0"), val = tensor([1, 1, 1024, 8])]; + tensor cache_3_end_mask_0 = const()[name = string("cache_3_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_3_squeeze_mask_0 = const()[name = string("cache_3_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_time_to_fp16 = cast(dtype = cache_time_to_fp16_dtype_0, x = cache_time)[name = string("cast_10")]; + tensor value_5_cast_fp16 = transpose(perm = value_5_perm_0, x = cache_time_to_fp16)[name = string("transpose_363")]; + tensor cache_3_cast_fp16 = slice_by_index(begin = cache_3_begin_0, end = cache_3_end_0, end_mask = cache_3_end_mask_0, squeeze_mask = cache_3_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_3_cast_fp16")]; + tensor input_27_axes_0 = const()[name = string("input_27_axes_0"), val = tensor([-1])]; + tensor encoder_layers_0_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_0_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4606208)))]; + tensor encoder_layers_0_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_0_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4608320)))]; + fp16 var_42_to_fp16 = const()[name = string("op_42_to_fp16"), val = fp16(0x1.5p-17)]; + tensor input_27_cast_fp16 = layer_norm(axes = input_27_axes_0, beta = encoder_layers_0_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_0_norm_feed_forward1_weight_to_fp16, x = var_342_cast_fp16)[name = string("input_27_cast_fp16")]; + tensor encoder_layers_0_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4610432))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8804800))))[name = string("encoder_layers_0_feed_forward1_linear1_weight_to_fp16_quantized")]; + tensor encoder_layers_0_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_0_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8813056)))]; + tensor linear_1_cast_fp16 = linear(bias = encoder_layers_0_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_0_feed_forward1_linear1_weight_to_fp16_quantized, x = input_27_cast_fp16)[name = string("linear_1_cast_fp16")]; + tensor input_31_cast_fp16 = silu(x = linear_1_cast_fp16)[name = string("input_31_cast_fp16")]; + tensor encoder_layers_0_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8821312))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13015680))))[name = string("encoder_layers_0_feed_forward1_linear2_weight_to_fp16_quantized")]; + tensor encoder_layers_0_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_0_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13017792)))]; + tensor linear_2_cast_fp16 = linear(bias = encoder_layers_0_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_0_feed_forward1_linear2_weight_to_fp16_quantized, x = input_31_cast_fp16)[name = string("linear_2_cast_fp16")]; + fp16 var_455_to_fp16 = const()[name = string("op_455_to_fp16"), val = fp16(0x1p-1)]; + tensor var_456_cast_fp16 = mul(x = linear_2_cast_fp16, y = var_455_to_fp16)[name = string("op_456_cast_fp16")]; + tensor input_37_cast_fp16 = add(x = var_342_cast_fp16, y = var_456_cast_fp16)[name = string("input_37_cast_fp16")]; + tensor key_1_axes_0 = const()[name = string("key_1_axes_0"), val = tensor([-1])]; + tensor encoder_layers_0_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_0_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13019904)))]; + tensor encoder_layers_0_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_0_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13022016)))]; + tensor key_1_cast_fp16 = layer_norm(axes = key_1_axes_0, beta = encoder_layers_0_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_0_norm_self_att_weight_to_fp16, x = input_37_cast_fp16)[name = string("key_1_cast_fp16")]; + bool input_39_interleave_0 = const()[name = string("input_39_interleave_0"), val = bool(false)]; + tensor input_39_cast_fp16 = concat(axis = var_68, interleave = input_39_interleave_0, values = (cache_1_cast_fp16, key_1_cast_fp16))[name = string("input_39_cast_fp16")]; + tensor var_478_begin_0 = const()[name = string("op_478_begin_0"), val = tensor([0, 7, 0])]; + tensor var_478_end_0 = const()[name = string("op_478_end_0"), val = tensor([1, 42, 1024])]; + tensor var_478_end_mask_0 = const()[name = string("op_478_end_mask_0"), val = tensor([true, true, true])]; + tensor var_478_cast_fp16 = slice_by_index(begin = var_478_begin_0, end = var_478_end_0, end_mask = var_478_end_mask_0, x = cache_1_cast_fp16)[name = string("op_478_cast_fp16")]; + bool var_484_interleave_0 = const()[name = string("op_484_interleave_0"), val = bool(false)]; + tensor var_484_cast_fp16 = concat(axis = var_68, interleave = var_484_interleave_0, values = (var_478_cast_fp16, key_1_cast_fp16))[name = string("op_484_cast_fp16")]; + tensor encoder_layers_0_self_attn_linear_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13024128))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14072768))))[name = string("encoder_layers_0_self_attn_linear_q_weight_to_fp16_quantized")]; + tensor encoder_layers_0_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_0_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14074880)))]; + tensor linear_3_cast_fp16 = linear(bias = encoder_layers_0_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_0_self_attn_linear_q_weight_to_fp16_quantized, x = key_1_cast_fp16)[name = string("linear_3_cast_fp16")]; + tensor var_489 = const()[name = string("op_489"), val = tensor([1, -1, 8, 128])]; + tensor q_1_cast_fp16 = reshape(shape = var_489, x = linear_3_cast_fp16)[name = string("q_1_cast_fp16")]; + tensor encoder_layers_0_self_attn_linear_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14076992))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15125632))))[name = string("encoder_layers_0_self_attn_linear_k_weight_to_fp16_quantized")]; + tensor encoder_layers_0_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_0_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15127744)))]; + tensor linear_4_cast_fp16 = linear(bias = encoder_layers_0_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_0_self_attn_linear_k_weight_to_fp16_quantized, x = input_39_cast_fp16)[name = string("linear_4_cast_fp16")]; + tensor var_494 = const()[name = string("op_494"), val = tensor([1, -1, 8, 128])]; + tensor k_1_cast_fp16 = reshape(shape = var_494, x = linear_4_cast_fp16)[name = string("k_1_cast_fp16")]; + tensor encoder_layers_0_self_attn_linear_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15129856))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16178496))))[name = string("encoder_layers_0_self_attn_linear_v_weight_to_fp16_quantized")]; + tensor encoder_layers_0_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_0_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16180608)))]; + tensor linear_5_cast_fp16 = linear(bias = encoder_layers_0_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_0_self_attn_linear_v_weight_to_fp16_quantized, x = input_39_cast_fp16)[name = string("linear_5_cast_fp16")]; + tensor var_499 = const()[name = string("op_499"), val = tensor([1, -1, 8, 128])]; + tensor v_1_cast_fp16 = reshape(shape = var_499, x = linear_5_cast_fp16)[name = string("v_1_cast_fp16")]; + tensor value_9_perm_0 = const()[name = string("value_9_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_0_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_0_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16182720)))]; + tensor var_512_cast_fp16 = add(x = q_1_cast_fp16, y = encoder_layers_0_self_attn_pos_bias_u_to_fp16)[name = string("op_512_cast_fp16")]; + tensor encoder_layers_0_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_0_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16184832)))]; + tensor var_514_cast_fp16 = add(x = q_1_cast_fp16, y = encoder_layers_0_self_attn_pos_bias_v_to_fp16)[name = string("op_514_cast_fp16")]; + tensor q_with_bias_v_1_perm_0 = const()[name = string("q_with_bias_v_1_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_7_transpose_x_0 = const()[name = string("x_7_transpose_x_0"), val = bool(false)]; + bool x_7_transpose_y_0 = const()[name = string("x_7_transpose_y_0"), val = bool(false)]; + tensor op_516_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16186944))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16286336))))[name = string("op_516_to_fp16_quantized")]; + tensor q_with_bias_v_1_cast_fp16 = transpose(perm = q_with_bias_v_1_perm_0, x = var_514_cast_fp16)[name = string("transpose_362")]; + tensor x_7_cast_fp16 = matmul(transpose_x = x_7_transpose_x_0, transpose_y = x_7_transpose_y_0, x = q_with_bias_v_1_cast_fp16, y = op_516_to_fp16_quantized)[name = string("x_7_cast_fp16")]; + tensor x_9_pad_0 = const()[name = string("x_9_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_9_mode_0 = const()[name = string("x_9_mode_0"), val = string("constant")]; + fp16 const_79_to_fp16 = const()[name = string("const_79_to_fp16"), val = fp16(0x0p+0)]; + tensor x_9_cast_fp16 = pad(constant_val = const_79_to_fp16, mode = x_9_mode_0, pad = x_9_pad_0, x = x_7_cast_fp16)[name = string("x_9_cast_fp16")]; + tensor var_524 = const()[name = string("op_524"), val = tensor([1, 8, -1, 7])]; + tensor x_11_cast_fp16 = reshape(shape = var_524, x = x_9_cast_fp16)[name = string("x_11_cast_fp16")]; + tensor var_528_begin_0 = const()[name = string("op_528_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_528_end_0 = const()[name = string("op_528_end_0"), val = tensor([1, 8, 98, 7])]; + tensor var_528_end_mask_0 = const()[name = string("op_528_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_528_cast_fp16 = slice_by_index(begin = var_528_begin_0, end = var_528_end_0, end_mask = var_528_end_mask_0, x = x_11_cast_fp16)[name = string("op_528_cast_fp16")]; + tensor var_529 = const()[name = string("op_529"), val = tensor([1, 8, 7, 97])]; + tensor matrix_bd_1_cast_fp16 = reshape(shape = var_529, x = var_528_cast_fp16)[name = string("matrix_bd_1_cast_fp16")]; + bool matrix_ac_1_transpose_x_0 = const()[name = string("matrix_ac_1_transpose_x_0"), val = bool(false)]; + bool matrix_ac_1_transpose_y_0 = const()[name = string("matrix_ac_1_transpose_y_0"), val = bool(false)]; + tensor transpose_96_perm_0 = const()[name = string("transpose_96_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_97_perm_0 = const()[name = string("transpose_97_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_97 = transpose(perm = transpose_97_perm_0, x = k_1_cast_fp16)[name = string("transpose_360")]; + tensor transpose_96 = transpose(perm = transpose_96_perm_0, x = var_512_cast_fp16)[name = string("transpose_361")]; + tensor matrix_ac_1_cast_fp16 = matmul(transpose_x = matrix_ac_1_transpose_x_0, transpose_y = matrix_ac_1_transpose_y_0, x = transpose_96, y = transpose_97)[name = string("matrix_ac_1_cast_fp16")]; + tensor matrix_bd_3_begin_0 = const()[name = string("matrix_bd_3_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_3_end_0 = const()[name = string("matrix_bd_3_end_0"), val = tensor([1, 8, 7, 49])]; + tensor matrix_bd_3_end_mask_0 = const()[name = string("matrix_bd_3_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_3_cast_fp16 = slice_by_index(begin = matrix_bd_3_begin_0, end = matrix_bd_3_end_0, end_mask = matrix_bd_3_end_mask_0, x = matrix_bd_1_cast_fp16)[name = string("matrix_bd_3_cast_fp16")]; + tensor var_538_cast_fp16 = add(x = matrix_ac_1_cast_fp16, y = matrix_bd_3_cast_fp16)[name = string("op_538_cast_fp16")]; + fp16 _inversed_scores_1_y_0_to_fp16 = const()[name = string("_inversed_scores_1_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_1_cast_fp16 = mul(x = var_538_cast_fp16, y = _inversed_scores_1_y_0_to_fp16)[name = string("_inversed_scores_1_cast_fp16")]; + tensor mask_11_axes_0 = const()[name = string("mask_11_axes_0"), val = tensor([1])]; + tensor mask_11 = expand_dims(axes = mask_11_axes_0, x = mask_9)[name = string("mask_11")]; + fp16 var_45_to_fp16 = const()[name = string("op_45_to_fp16"), val = fp16(-0x1.388p+13)]; + tensor scores_3_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_1_cast_fp16, cond = mask_11)[name = string("scores_3_cast_fp16")]; + tensor var_544_cast_fp16 = softmax(axis = var_59, x = scores_3_cast_fp16)[name = string("op_544_cast_fp16")]; + fp16 var_44_to_fp16 = const()[name = string("op_44_to_fp16"), val = fp16(0x0p+0)]; + tensor input_41_cast_fp16 = select(a = var_44_to_fp16, b = var_544_cast_fp16, cond = mask_11)[name = string("input_41_cast_fp16")]; + bool x_13_transpose_x_0 = const()[name = string("x_13_transpose_x_0"), val = bool(false)]; + bool x_13_transpose_y_0 = const()[name = string("x_13_transpose_y_0"), val = bool(false)]; + tensor value_9_cast_fp16 = transpose(perm = value_9_perm_0, x = v_1_cast_fp16)[name = string("transpose_359")]; + tensor x_13_cast_fp16 = matmul(transpose_x = x_13_transpose_x_0, transpose_y = x_13_transpose_y_0, x = input_41_cast_fp16, y = value_9_cast_fp16)[name = string("x_13_cast_fp16")]; + tensor var_548_perm_0 = const()[name = string("op_548_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_549 = const()[name = string("op_549"), val = tensor([1, -1, 1024])]; + tensor var_548_cast_fp16 = transpose(perm = var_548_perm_0, x = x_13_cast_fp16)[name = string("transpose_358")]; + tensor input_43_cast_fp16 = reshape(shape = var_549, x = var_548_cast_fp16)[name = string("input_43_cast_fp16")]; + tensor encoder_layers_0_self_attn_linear_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16286656))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17335296))))[name = string("encoder_layers_0_self_attn_linear_out_weight_to_fp16_quantized")]; + tensor encoder_layers_0_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_0_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17337408)))]; + tensor linear_7_cast_fp16 = linear(bias = encoder_layers_0_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_0_self_attn_linear_out_weight_to_fp16_quantized, x = input_43_cast_fp16)[name = string("linear_7_cast_fp16")]; + tensor input_47_cast_fp16 = add(x = input_37_cast_fp16, y = linear_7_cast_fp16)[name = string("input_47_cast_fp16")]; + tensor x_17_axes_0 = const()[name = string("x_17_axes_0"), val = tensor([-1])]; + tensor encoder_layers_0_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_0_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17339520)))]; + tensor encoder_layers_0_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_0_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17341632)))]; + tensor x_17_cast_fp16 = layer_norm(axes = x_17_axes_0, beta = encoder_layers_0_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_0_norm_conv_weight_to_fp16, x = input_47_cast_fp16)[name = string("x_17_cast_fp16")]; + tensor input_49_perm_0 = const()[name = string("input_49_perm_0"), val = tensor([0, 2, 1])]; + string input_51_pad_type_0 = const()[name = string("input_51_pad_type_0"), val = string("valid")]; + tensor input_51_strides_0 = const()[name = string("input_51_strides_0"), val = tensor([1])]; + tensor input_51_pad_0 = const()[name = string("input_51_pad_0"), val = tensor([0, 0])]; + tensor input_51_dilations_0 = const()[name = string("input_51_dilations_0"), val = tensor([1])]; + int32 input_51_groups_0 = const()[name = string("input_51_groups_0"), val = int32(1)]; + tensor encoder_layers_0_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17343744))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19440960))))[name = string("encoder_layers_0_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_49_cast_fp16 = transpose(perm = input_49_perm_0, x = x_17_cast_fp16)[name = string("transpose_357")]; + tensor input_51_cast_fp16 = conv(dilations = input_51_dilations_0, groups = input_51_groups_0, pad = input_51_pad_0, pad_type = input_51_pad_type_0, strides = input_51_strides_0, weight = encoder_layers_0_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_49_cast_fp16)[name = string("input_51_cast_fp16")]; + int32 x_19_split_num_splits_0 = const()[name = string("x_19_split_num_splits_0"), val = int32(2)]; + int32 x_19_split_axis_0 = const()[name = string("x_19_split_axis_0"), val = int32(1)]; + tensor x_19_split_cast_fp16_0, tensor x_19_split_cast_fp16_1 = split(axis = x_19_split_axis_0, num_splits = x_19_split_num_splits_0, x = input_51_cast_fp16)[name = string("x_19_split_cast_fp16")]; + tensor x_19_split_1_sigmoid_cast_fp16 = sigmoid(x = x_19_split_cast_fp16_1)[name = string("x_19_split_1_sigmoid_cast_fp16")]; + tensor x_19_cast_fp16 = mul(x = x_19_split_cast_fp16_0, y = x_19_split_1_sigmoid_cast_fp16)[name = string("x_19_cast_fp16")]; + tensor var_575_axes_0 = const()[name = string("op_575_axes_0"), val = tensor([1])]; + tensor var_575 = expand_dims(axes = var_575_axes_0, x = pad_mask)[name = string("op_575")]; + tensor input_53_cast_fp16 = select(a = var_44_to_fp16, b = x_19_cast_fp16, cond = var_575)[name = string("input_53_cast_fp16")]; + bool new_x_3_interleave_0 = const()[name = string("new_x_3_interleave_0"), val = bool(false)]; + tensor new_x_3_cast_fp16 = concat(axis = var_59, interleave = new_x_3_interleave_0, values = (cache_3_cast_fp16, input_53_cast_fp16))[name = string("new_x_3_cast_fp16")]; + tensor var_588_begin_0 = const()[name = string("op_588_begin_0"), val = tensor([0, 0, 7])]; + tensor var_588_end_0 = const()[name = string("op_588_end_0"), val = tensor([1, 1024, 15])]; + tensor var_588_end_mask_0 = const()[name = string("op_588_end_mask_0"), val = tensor([true, true, true])]; + tensor var_588_cast_fp16 = slice_by_index(begin = var_588_begin_0, end = var_588_end_0, end_mask = var_588_end_mask_0, x = new_x_3_cast_fp16)[name = string("op_588_cast_fp16")]; + string x_21_pad_type_0 = const()[name = string("x_21_pad_type_0"), val = string("valid")]; + int32 x_21_groups_0 = const()[name = string("x_21_groups_0"), val = int32(1024)]; + tensor x_21_strides_0 = const()[name = string("x_21_strides_0"), val = tensor([1])]; + tensor x_21_pad_0 = const()[name = string("x_21_pad_0"), val = tensor([0, 0])]; + tensor x_21_dilations_0 = const()[name = string("x_21_dilations_0"), val = tensor([1])]; + tensor encoder_layers_0_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19445120))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19454400))))[name = string("encoder_layers_0_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_21_cast_fp16 = conv(dilations = x_21_dilations_0, groups = x_21_groups_0, pad = x_21_pad_0, pad_type = x_21_pad_type_0, strides = x_21_strides_0, weight = encoder_layers_0_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_3_cast_fp16)[name = string("x_21_cast_fp16")]; + tensor input_55_perm_0 = const()[name = string("input_55_perm_0"), val = tensor([0, 2, 1])]; + tensor x_23_axes_0 = const()[name = string("x_23_axes_0"), val = tensor([-1])]; + tensor encoder_layers_0_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_0_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19456512)))]; + tensor encoder_layers_0_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_0_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19458624)))]; + tensor input_55_cast_fp16 = transpose(perm = input_55_perm_0, x = x_21_cast_fp16)[name = string("transpose_356")]; + tensor x_23_cast_fp16 = layer_norm(axes = x_23_axes_0, beta = encoder_layers_0_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_0_conv_batch_norm_weight_to_fp16, x = input_55_cast_fp16)[name = string("x_23_cast_fp16")]; + tensor input_57_perm_0 = const()[name = string("input_57_perm_0"), val = tensor([0, 2, 1])]; + tensor input_57_cast_fp16 = transpose(perm = input_57_perm_0, x = x_23_cast_fp16)[name = string("transpose_355")]; + tensor input_59_cast_fp16 = silu(x = input_57_cast_fp16)[name = string("input_59_cast_fp16")]; + string x_25_pad_type_0 = const()[name = string("x_25_pad_type_0"), val = string("valid")]; + tensor x_25_strides_0 = const()[name = string("x_25_strides_0"), val = tensor([1])]; + tensor x_25_pad_0 = const()[name = string("x_25_pad_0"), val = tensor([0, 0])]; + tensor x_25_dilations_0 = const()[name = string("x_25_dilations_0"), val = tensor([1])]; + int32 x_25_groups_0 = const()[name = string("x_25_groups_0"), val = int32(1)]; + tensor encoder_layers_0_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19460736))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20509376))))[name = string("encoder_layers_0_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_25_cast_fp16 = conv(dilations = x_25_dilations_0, groups = x_25_groups_0, pad = x_25_pad_0, pad_type = x_25_pad_type_0, strides = x_25_strides_0, weight = encoder_layers_0_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_59_cast_fp16)[name = string("x_25_cast_fp16")]; + tensor input_61_perm_0 = const()[name = string("input_61_perm_0"), val = tensor([0, 2, 1])]; + tensor input_61_cast_fp16 = transpose(perm = input_61_perm_0, x = x_25_cast_fp16)[name = string("transpose_354")]; + tensor input_63_cast_fp16 = add(x = input_47_cast_fp16, y = input_61_cast_fp16)[name = string("input_63_cast_fp16")]; + tensor input_65_axes_0 = const()[name = string("input_65_axes_0"), val = tensor([-1])]; + tensor encoder_layers_0_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_0_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20511488)))]; + tensor encoder_layers_0_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_0_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20513600)))]; + tensor input_65_cast_fp16 = layer_norm(axes = input_65_axes_0, beta = encoder_layers_0_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_0_norm_feed_forward2_weight_to_fp16, x = input_63_cast_fp16)[name = string("input_65_cast_fp16")]; + tensor encoder_layers_0_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20515712))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24710080))))[name = string("encoder_layers_0_feed_forward2_linear1_weight_to_fp16_quantized")]; + tensor encoder_layers_0_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_0_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24718336)))]; + tensor linear_8_cast_fp16 = linear(bias = encoder_layers_0_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_0_feed_forward2_linear1_weight_to_fp16_quantized, x = input_65_cast_fp16)[name = string("linear_8_cast_fp16")]; + tensor input_69_cast_fp16 = silu(x = linear_8_cast_fp16)[name = string("input_69_cast_fp16")]; + tensor encoder_layers_0_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24726592))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28920960))))[name = string("encoder_layers_0_feed_forward2_linear2_weight_to_fp16_quantized")]; + tensor encoder_layers_0_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_0_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28923072)))]; + tensor linear_9_cast_fp16 = linear(bias = encoder_layers_0_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_0_feed_forward2_linear2_weight_to_fp16_quantized, x = input_69_cast_fp16)[name = string("linear_9_cast_fp16")]; + fp16 var_631_to_fp16 = const()[name = string("op_631_to_fp16"), val = fp16(0x1p-1)]; + tensor var_632_cast_fp16 = mul(x = linear_9_cast_fp16, y = var_631_to_fp16)[name = string("op_632_cast_fp16")]; + tensor input_75_cast_fp16 = add(x = input_63_cast_fp16, y = var_632_cast_fp16)[name = string("input_75_cast_fp16")]; + tensor input_77_axes_0 = const()[name = string("input_77_axes_0"), val = tensor([-1])]; + tensor encoder_layers_0_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_0_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28925184)))]; + tensor encoder_layers_0_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_0_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28927296)))]; + tensor input_77_cast_fp16 = layer_norm(axes = input_77_axes_0, beta = encoder_layers_0_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_0_norm_out_weight_to_fp16, x = input_75_cast_fp16)[name = string("input_77_cast_fp16")]; + tensor cache_5_begin_0 = const()[name = string("cache_5_begin_0"), val = tensor([1, 0, 0, 0])]; + tensor cache_5_end_0 = const()[name = string("cache_5_end_0"), val = tensor([2, 1, 42, 1024])]; + tensor cache_5_end_mask_0 = const()[name = string("cache_5_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_5_squeeze_mask_0 = const()[name = string("cache_5_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_5_cast_fp16 = slice_by_index(begin = cache_5_begin_0, end = cache_5_end_0, end_mask = cache_5_end_mask_0, squeeze_mask = cache_5_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_5_cast_fp16")]; + tensor cache_7_begin_0 = const()[name = string("cache_7_begin_0"), val = tensor([1, 0, 0, 0])]; + tensor cache_7_end_0 = const()[name = string("cache_7_end_0"), val = tensor([2, 1, 1024, 8])]; + tensor cache_7_end_mask_0 = const()[name = string("cache_7_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_7_squeeze_mask_0 = const()[name = string("cache_7_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_7_cast_fp16 = slice_by_index(begin = cache_7_begin_0, end = cache_7_end_0, end_mask = cache_7_end_mask_0, squeeze_mask = cache_7_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_7_cast_fp16")]; + tensor input_79_axes_0 = const()[name = string("input_79_axes_0"), val = tensor([-1])]; + tensor encoder_layers_1_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_1_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28929408)))]; + tensor encoder_layers_1_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_1_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28931520)))]; + tensor input_79_cast_fp16 = layer_norm(axes = input_79_axes_0, beta = encoder_layers_1_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_1_norm_feed_forward1_weight_to_fp16, x = input_77_cast_fp16)[name = string("input_79_cast_fp16")]; + tensor encoder_layers_1_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28933632))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33128000))))[name = string("encoder_layers_1_feed_forward1_linear1_weight_to_fp16_quantized")]; + tensor encoder_layers_1_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_1_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33136256)))]; + tensor linear_10_cast_fp16 = linear(bias = encoder_layers_1_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_1_feed_forward1_linear1_weight_to_fp16_quantized, x = input_79_cast_fp16)[name = string("linear_10_cast_fp16")]; + tensor input_83_cast_fp16 = silu(x = linear_10_cast_fp16)[name = string("input_83_cast_fp16")]; + tensor encoder_layers_1_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33144512))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37338880))))[name = string("encoder_layers_1_feed_forward1_linear2_weight_to_fp16_quantized")]; + tensor encoder_layers_1_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_1_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37340992)))]; + tensor linear_11_cast_fp16 = linear(bias = encoder_layers_1_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_1_feed_forward1_linear2_weight_to_fp16_quantized, x = input_83_cast_fp16)[name = string("linear_11_cast_fp16")]; + fp16 var_668_to_fp16 = const()[name = string("op_668_to_fp16"), val = fp16(0x1p-1)]; + tensor var_669_cast_fp16 = mul(x = linear_11_cast_fp16, y = var_668_to_fp16)[name = string("op_669_cast_fp16")]; + tensor input_89_cast_fp16 = add(x = input_77_cast_fp16, y = var_669_cast_fp16)[name = string("input_89_cast_fp16")]; + tensor key_3_axes_0 = const()[name = string("key_3_axes_0"), val = tensor([-1])]; + tensor encoder_layers_1_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_1_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37343104)))]; + tensor encoder_layers_1_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_1_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37345216)))]; + tensor key_3_cast_fp16 = layer_norm(axes = key_3_axes_0, beta = encoder_layers_1_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_1_norm_self_att_weight_to_fp16, x = input_89_cast_fp16)[name = string("key_3_cast_fp16")]; + bool input_91_interleave_0 = const()[name = string("input_91_interleave_0"), val = bool(false)]; + tensor input_91_cast_fp16 = concat(axis = var_68, interleave = input_91_interleave_0, values = (cache_5_cast_fp16, key_3_cast_fp16))[name = string("input_91_cast_fp16")]; + tensor var_691_begin_0 = const()[name = string("op_691_begin_0"), val = tensor([0, 7, 0])]; + tensor var_691_end_0 = const()[name = string("op_691_end_0"), val = tensor([1, 42, 1024])]; + tensor var_691_end_mask_0 = const()[name = string("op_691_end_mask_0"), val = tensor([true, true, true])]; + tensor var_691_cast_fp16 = slice_by_index(begin = var_691_begin_0, end = var_691_end_0, end_mask = var_691_end_mask_0, x = cache_5_cast_fp16)[name = string("op_691_cast_fp16")]; + bool var_697_interleave_0 = const()[name = string("op_697_interleave_0"), val = bool(false)]; + tensor var_697_cast_fp16 = concat(axis = var_68, interleave = var_697_interleave_0, values = (var_691_cast_fp16, key_3_cast_fp16))[name = string("op_697_cast_fp16")]; + tensor encoder_layers_1_self_attn_linear_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37347328))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38395968))))[name = string("encoder_layers_1_self_attn_linear_q_weight_to_fp16_quantized")]; + tensor encoder_layers_1_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_1_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38398080)))]; + tensor linear_12_cast_fp16 = linear(bias = encoder_layers_1_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_1_self_attn_linear_q_weight_to_fp16_quantized, x = key_3_cast_fp16)[name = string("linear_12_cast_fp16")]; + tensor var_702 = const()[name = string("op_702"), val = tensor([1, -1, 8, 128])]; + tensor q_7_cast_fp16 = reshape(shape = var_702, x = linear_12_cast_fp16)[name = string("q_7_cast_fp16")]; + tensor encoder_layers_1_self_attn_linear_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38400192))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39448832))))[name = string("encoder_layers_1_self_attn_linear_k_weight_to_fp16_quantized")]; + tensor encoder_layers_1_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_1_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39450944)))]; + tensor linear_13_cast_fp16 = linear(bias = encoder_layers_1_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_1_self_attn_linear_k_weight_to_fp16_quantized, x = input_91_cast_fp16)[name = string("linear_13_cast_fp16")]; + tensor var_707 = const()[name = string("op_707"), val = tensor([1, -1, 8, 128])]; + tensor k_5_cast_fp16 = reshape(shape = var_707, x = linear_13_cast_fp16)[name = string("k_5_cast_fp16")]; + tensor encoder_layers_1_self_attn_linear_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39453056))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40501696))))[name = string("encoder_layers_1_self_attn_linear_v_weight_to_fp16_quantized")]; + tensor encoder_layers_1_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_1_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40503808)))]; + tensor linear_14_cast_fp16 = linear(bias = encoder_layers_1_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_1_self_attn_linear_v_weight_to_fp16_quantized, x = input_91_cast_fp16)[name = string("linear_14_cast_fp16")]; + tensor var_712 = const()[name = string("op_712"), val = tensor([1, -1, 8, 128])]; + tensor v_3_cast_fp16 = reshape(shape = var_712, x = linear_14_cast_fp16)[name = string("v_3_cast_fp16")]; + tensor value_11_perm_0 = const()[name = string("value_11_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_1_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_1_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40505920)))]; + tensor var_725_cast_fp16 = add(x = q_7_cast_fp16, y = encoder_layers_1_self_attn_pos_bias_u_to_fp16)[name = string("op_725_cast_fp16")]; + tensor encoder_layers_1_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_1_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40508032)))]; + tensor var_727_cast_fp16 = add(x = q_7_cast_fp16, y = encoder_layers_1_self_attn_pos_bias_v_to_fp16)[name = string("op_727_cast_fp16")]; + tensor q_with_bias_v_3_perm_0 = const()[name = string("q_with_bias_v_3_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_33_transpose_x_0 = const()[name = string("x_33_transpose_x_0"), val = bool(false)]; + bool x_33_transpose_y_0 = const()[name = string("x_33_transpose_y_0"), val = bool(false)]; + tensor op_729_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40510144))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40609536))))[name = string("op_729_to_fp16_quantized")]; + tensor q_with_bias_v_3_cast_fp16 = transpose(perm = q_with_bias_v_3_perm_0, x = var_727_cast_fp16)[name = string("transpose_353")]; + tensor x_33_cast_fp16 = matmul(transpose_x = x_33_transpose_x_0, transpose_y = x_33_transpose_y_0, x = q_with_bias_v_3_cast_fp16, y = op_729_to_fp16_quantized)[name = string("x_33_cast_fp16")]; + tensor x_35_pad_0 = const()[name = string("x_35_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_35_mode_0 = const()[name = string("x_35_mode_0"), val = string("constant")]; + fp16 const_92_to_fp16 = const()[name = string("const_92_to_fp16"), val = fp16(0x0p+0)]; + tensor x_35_cast_fp16 = pad(constant_val = const_92_to_fp16, mode = x_35_mode_0, pad = x_35_pad_0, x = x_33_cast_fp16)[name = string("x_35_cast_fp16")]; + tensor var_737 = const()[name = string("op_737"), val = tensor([1, 8, -1, 7])]; + tensor x_37_cast_fp16 = reshape(shape = var_737, x = x_35_cast_fp16)[name = string("x_37_cast_fp16")]; + tensor var_741_begin_0 = const()[name = string("op_741_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_741_end_0 = const()[name = string("op_741_end_0"), val = tensor([1, 8, 98, 7])]; + tensor var_741_end_mask_0 = const()[name = string("op_741_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_741_cast_fp16 = slice_by_index(begin = var_741_begin_0, end = var_741_end_0, end_mask = var_741_end_mask_0, x = x_37_cast_fp16)[name = string("op_741_cast_fp16")]; + tensor var_742 = const()[name = string("op_742"), val = tensor([1, 8, 7, 97])]; + tensor matrix_bd_5_cast_fp16 = reshape(shape = var_742, x = var_741_cast_fp16)[name = string("matrix_bd_5_cast_fp16")]; + bool matrix_ac_3_transpose_x_0 = const()[name = string("matrix_ac_3_transpose_x_0"), val = bool(false)]; + bool matrix_ac_3_transpose_y_0 = const()[name = string("matrix_ac_3_transpose_y_0"), val = bool(false)]; + tensor transpose_98_perm_0 = const()[name = string("transpose_98_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_99_perm_0 = const()[name = string("transpose_99_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_99 = transpose(perm = transpose_99_perm_0, x = k_5_cast_fp16)[name = string("transpose_351")]; + tensor transpose_98 = transpose(perm = transpose_98_perm_0, x = var_725_cast_fp16)[name = string("transpose_352")]; + tensor matrix_ac_3_cast_fp16 = matmul(transpose_x = matrix_ac_3_transpose_x_0, transpose_y = matrix_ac_3_transpose_y_0, x = transpose_98, y = transpose_99)[name = string("matrix_ac_3_cast_fp16")]; + tensor matrix_bd_7_begin_0 = const()[name = string("matrix_bd_7_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_7_end_0 = const()[name = string("matrix_bd_7_end_0"), val = tensor([1, 8, 7, 49])]; + tensor matrix_bd_7_end_mask_0 = const()[name = string("matrix_bd_7_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_7_cast_fp16 = slice_by_index(begin = matrix_bd_7_begin_0, end = matrix_bd_7_end_0, end_mask = matrix_bd_7_end_mask_0, x = matrix_bd_5_cast_fp16)[name = string("matrix_bd_7_cast_fp16")]; + tensor var_751_cast_fp16 = add(x = matrix_ac_3_cast_fp16, y = matrix_bd_7_cast_fp16)[name = string("op_751_cast_fp16")]; + fp16 _inversed_scores_5_y_0_to_fp16 = const()[name = string("_inversed_scores_5_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_5_cast_fp16 = mul(x = var_751_cast_fp16, y = _inversed_scores_5_y_0_to_fp16)[name = string("_inversed_scores_5_cast_fp16")]; + tensor scores_7_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_5_cast_fp16, cond = mask_11)[name = string("scores_7_cast_fp16")]; + tensor var_757_cast_fp16 = softmax(axis = var_59, x = scores_7_cast_fp16)[name = string("op_757_cast_fp16")]; + tensor input_93_cast_fp16 = select(a = var_44_to_fp16, b = var_757_cast_fp16, cond = mask_11)[name = string("input_93_cast_fp16")]; + bool x_39_transpose_x_0 = const()[name = string("x_39_transpose_x_0"), val = bool(false)]; + bool x_39_transpose_y_0 = const()[name = string("x_39_transpose_y_0"), val = bool(false)]; + tensor value_11_cast_fp16 = transpose(perm = value_11_perm_0, x = v_3_cast_fp16)[name = string("transpose_350")]; + tensor x_39_cast_fp16 = matmul(transpose_x = x_39_transpose_x_0, transpose_y = x_39_transpose_y_0, x = input_93_cast_fp16, y = value_11_cast_fp16)[name = string("x_39_cast_fp16")]; + tensor var_761_perm_0 = const()[name = string("op_761_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_762 = const()[name = string("op_762"), val = tensor([1, -1, 1024])]; + tensor var_761_cast_fp16 = transpose(perm = var_761_perm_0, x = x_39_cast_fp16)[name = string("transpose_349")]; + tensor input_95_cast_fp16 = reshape(shape = var_762, x = var_761_cast_fp16)[name = string("input_95_cast_fp16")]; + tensor encoder_layers_1_self_attn_linear_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40609856))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41658496))))[name = string("encoder_layers_1_self_attn_linear_out_weight_to_fp16_quantized")]; + tensor encoder_layers_1_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_1_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41660608)))]; + tensor linear_16_cast_fp16 = linear(bias = encoder_layers_1_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_1_self_attn_linear_out_weight_to_fp16_quantized, x = input_95_cast_fp16)[name = string("linear_16_cast_fp16")]; + tensor input_99_cast_fp16 = add(x = input_89_cast_fp16, y = linear_16_cast_fp16)[name = string("input_99_cast_fp16")]; + tensor x_43_axes_0 = const()[name = string("x_43_axes_0"), val = tensor([-1])]; + tensor encoder_layers_1_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_1_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41662720)))]; + tensor encoder_layers_1_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_1_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41664832)))]; + tensor x_43_cast_fp16 = layer_norm(axes = x_43_axes_0, beta = encoder_layers_1_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_1_norm_conv_weight_to_fp16, x = input_99_cast_fp16)[name = string("x_43_cast_fp16")]; + tensor input_101_perm_0 = const()[name = string("input_101_perm_0"), val = tensor([0, 2, 1])]; + string input_103_pad_type_0 = const()[name = string("input_103_pad_type_0"), val = string("valid")]; + tensor input_103_strides_0 = const()[name = string("input_103_strides_0"), val = tensor([1])]; + tensor input_103_pad_0 = const()[name = string("input_103_pad_0"), val = tensor([0, 0])]; + tensor input_103_dilations_0 = const()[name = string("input_103_dilations_0"), val = tensor([1])]; + int32 input_103_groups_0 = const()[name = string("input_103_groups_0"), val = int32(1)]; + tensor encoder_layers_1_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41666944))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43764160))))[name = string("encoder_layers_1_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_101_cast_fp16 = transpose(perm = input_101_perm_0, x = x_43_cast_fp16)[name = string("transpose_348")]; + tensor input_103_cast_fp16 = conv(dilations = input_103_dilations_0, groups = input_103_groups_0, pad = input_103_pad_0, pad_type = input_103_pad_type_0, strides = input_103_strides_0, weight = encoder_layers_1_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_101_cast_fp16)[name = string("input_103_cast_fp16")]; + int32 x_45_split_num_splits_0 = const()[name = string("x_45_split_num_splits_0"), val = int32(2)]; + int32 x_45_split_axis_0 = const()[name = string("x_45_split_axis_0"), val = int32(1)]; + tensor x_45_split_cast_fp16_0, tensor x_45_split_cast_fp16_1 = split(axis = x_45_split_axis_0, num_splits = x_45_split_num_splits_0, x = input_103_cast_fp16)[name = string("x_45_split_cast_fp16")]; + tensor x_45_split_1_sigmoid_cast_fp16 = sigmoid(x = x_45_split_cast_fp16_1)[name = string("x_45_split_1_sigmoid_cast_fp16")]; + tensor x_45_cast_fp16 = mul(x = x_45_split_cast_fp16_0, y = x_45_split_1_sigmoid_cast_fp16)[name = string("x_45_cast_fp16")]; + tensor input_105_cast_fp16 = select(a = var_44_to_fp16, b = x_45_cast_fp16, cond = var_575)[name = string("input_105_cast_fp16")]; + bool new_x_7_interleave_0 = const()[name = string("new_x_7_interleave_0"), val = bool(false)]; + tensor new_x_7_cast_fp16 = concat(axis = var_59, interleave = new_x_7_interleave_0, values = (cache_7_cast_fp16, input_105_cast_fp16))[name = string("new_x_7_cast_fp16")]; + tensor var_801_begin_0 = const()[name = string("op_801_begin_0"), val = tensor([0, 0, 7])]; + tensor var_801_end_0 = const()[name = string("op_801_end_0"), val = tensor([1, 1024, 15])]; + tensor var_801_end_mask_0 = const()[name = string("op_801_end_mask_0"), val = tensor([true, true, true])]; + tensor var_801_cast_fp16 = slice_by_index(begin = var_801_begin_0, end = var_801_end_0, end_mask = var_801_end_mask_0, x = new_x_7_cast_fp16)[name = string("op_801_cast_fp16")]; + string x_47_pad_type_0 = const()[name = string("x_47_pad_type_0"), val = string("valid")]; + int32 x_47_groups_0 = const()[name = string("x_47_groups_0"), val = int32(1024)]; + tensor x_47_strides_0 = const()[name = string("x_47_strides_0"), val = tensor([1])]; + tensor x_47_pad_0 = const()[name = string("x_47_pad_0"), val = tensor([0, 0])]; + tensor x_47_dilations_0 = const()[name = string("x_47_dilations_0"), val = tensor([1])]; + tensor encoder_layers_1_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43768320))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43777600))))[name = string("encoder_layers_1_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_47_cast_fp16 = conv(dilations = x_47_dilations_0, groups = x_47_groups_0, pad = x_47_pad_0, pad_type = x_47_pad_type_0, strides = x_47_strides_0, weight = encoder_layers_1_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_7_cast_fp16)[name = string("x_47_cast_fp16")]; + tensor input_107_perm_0 = const()[name = string("input_107_perm_0"), val = tensor([0, 2, 1])]; + tensor x_49_axes_0 = const()[name = string("x_49_axes_0"), val = tensor([-1])]; + tensor encoder_layers_1_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_1_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43779712)))]; + tensor encoder_layers_1_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_1_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43781824)))]; + tensor input_107_cast_fp16 = transpose(perm = input_107_perm_0, x = x_47_cast_fp16)[name = string("transpose_347")]; + tensor x_49_cast_fp16 = layer_norm(axes = x_49_axes_0, beta = encoder_layers_1_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_1_conv_batch_norm_weight_to_fp16, x = input_107_cast_fp16)[name = string("x_49_cast_fp16")]; + tensor input_109_perm_0 = const()[name = string("input_109_perm_0"), val = tensor([0, 2, 1])]; + tensor input_109_cast_fp16 = transpose(perm = input_109_perm_0, x = x_49_cast_fp16)[name = string("transpose_346")]; + tensor input_111_cast_fp16 = silu(x = input_109_cast_fp16)[name = string("input_111_cast_fp16")]; + string x_51_pad_type_0 = const()[name = string("x_51_pad_type_0"), val = string("valid")]; + tensor x_51_strides_0 = const()[name = string("x_51_strides_0"), val = tensor([1])]; + tensor x_51_pad_0 = const()[name = string("x_51_pad_0"), val = tensor([0, 0])]; + tensor x_51_dilations_0 = const()[name = string("x_51_dilations_0"), val = tensor([1])]; + int32 x_51_groups_0 = const()[name = string("x_51_groups_0"), val = int32(1)]; + tensor encoder_layers_1_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43783936))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44832576))))[name = string("encoder_layers_1_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_51_cast_fp16 = conv(dilations = x_51_dilations_0, groups = x_51_groups_0, pad = x_51_pad_0, pad_type = x_51_pad_type_0, strides = x_51_strides_0, weight = encoder_layers_1_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_111_cast_fp16)[name = string("x_51_cast_fp16")]; + tensor input_113_perm_0 = const()[name = string("input_113_perm_0"), val = tensor([0, 2, 1])]; + tensor input_113_cast_fp16 = transpose(perm = input_113_perm_0, x = x_51_cast_fp16)[name = string("transpose_345")]; + tensor input_115_cast_fp16 = add(x = input_99_cast_fp16, y = input_113_cast_fp16)[name = string("input_115_cast_fp16")]; + tensor input_117_axes_0 = const()[name = string("input_117_axes_0"), val = tensor([-1])]; + tensor encoder_layers_1_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_1_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44834688)))]; + tensor encoder_layers_1_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_1_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44836800)))]; + tensor input_117_cast_fp16 = layer_norm(axes = input_117_axes_0, beta = encoder_layers_1_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_1_norm_feed_forward2_weight_to_fp16, x = input_115_cast_fp16)[name = string("input_117_cast_fp16")]; + tensor encoder_layers_1_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44838912))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49033280))))[name = string("encoder_layers_1_feed_forward2_linear1_weight_to_fp16_quantized")]; + tensor encoder_layers_1_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_1_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49041536)))]; + tensor linear_17_cast_fp16 = linear(bias = encoder_layers_1_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_1_feed_forward2_linear1_weight_to_fp16_quantized, x = input_117_cast_fp16)[name = string("linear_17_cast_fp16")]; + tensor input_121_cast_fp16 = silu(x = linear_17_cast_fp16)[name = string("input_121_cast_fp16")]; + tensor encoder_layers_1_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49049792))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53244160))))[name = string("encoder_layers_1_feed_forward2_linear2_weight_to_fp16_quantized")]; + tensor encoder_layers_1_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_1_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53246272)))]; + tensor linear_18_cast_fp16 = linear(bias = encoder_layers_1_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_1_feed_forward2_linear2_weight_to_fp16_quantized, x = input_121_cast_fp16)[name = string("linear_18_cast_fp16")]; + fp16 var_844_to_fp16 = const()[name = string("op_844_to_fp16"), val = fp16(0x1p-1)]; + tensor var_845_cast_fp16 = mul(x = linear_18_cast_fp16, y = var_844_to_fp16)[name = string("op_845_cast_fp16")]; + tensor input_127_cast_fp16 = add(x = input_115_cast_fp16, y = var_845_cast_fp16)[name = string("input_127_cast_fp16")]; + tensor input_129_axes_0 = const()[name = string("input_129_axes_0"), val = tensor([-1])]; + tensor encoder_layers_1_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_1_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53248384)))]; + tensor encoder_layers_1_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_1_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53250496)))]; + tensor input_129_cast_fp16 = layer_norm(axes = input_129_axes_0, beta = encoder_layers_1_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_1_norm_out_weight_to_fp16, x = input_127_cast_fp16)[name = string("input_129_cast_fp16")]; + tensor cache_9_begin_0 = const()[name = string("cache_9_begin_0"), val = tensor([2, 0, 0, 0])]; + tensor cache_9_end_0 = const()[name = string("cache_9_end_0"), val = tensor([3, 1, 42, 1024])]; + tensor cache_9_end_mask_0 = const()[name = string("cache_9_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_9_squeeze_mask_0 = const()[name = string("cache_9_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_9_cast_fp16 = slice_by_index(begin = cache_9_begin_0, end = cache_9_end_0, end_mask = cache_9_end_mask_0, squeeze_mask = cache_9_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_9_cast_fp16")]; + tensor cache_11_begin_0 = const()[name = string("cache_11_begin_0"), val = tensor([2, 0, 0, 0])]; + tensor cache_11_end_0 = const()[name = string("cache_11_end_0"), val = tensor([3, 1, 1024, 8])]; + tensor cache_11_end_mask_0 = const()[name = string("cache_11_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_11_squeeze_mask_0 = const()[name = string("cache_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_11_cast_fp16 = slice_by_index(begin = cache_11_begin_0, end = cache_11_end_0, end_mask = cache_11_end_mask_0, squeeze_mask = cache_11_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_11_cast_fp16")]; + tensor input_131_axes_0 = const()[name = string("input_131_axes_0"), val = tensor([-1])]; + tensor encoder_layers_2_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_2_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53252608)))]; + tensor encoder_layers_2_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_2_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53254720)))]; + tensor input_131_cast_fp16 = layer_norm(axes = input_131_axes_0, beta = encoder_layers_2_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_2_norm_feed_forward1_weight_to_fp16, x = input_129_cast_fp16)[name = string("input_131_cast_fp16")]; + tensor encoder_layers_2_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53256832))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57451200))))[name = string("encoder_layers_2_feed_forward1_linear1_weight_to_fp16_quantized")]; + tensor encoder_layers_2_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_2_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57459456)))]; + tensor linear_19_cast_fp16 = linear(bias = encoder_layers_2_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_2_feed_forward1_linear1_weight_to_fp16_quantized, x = input_131_cast_fp16)[name = string("linear_19_cast_fp16")]; + tensor input_135_cast_fp16 = silu(x = linear_19_cast_fp16)[name = string("input_135_cast_fp16")]; + tensor encoder_layers_2_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57467712))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61662080))))[name = string("encoder_layers_2_feed_forward1_linear2_weight_to_fp16_quantized")]; + tensor encoder_layers_2_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_2_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61664192)))]; + tensor linear_20_cast_fp16 = linear(bias = encoder_layers_2_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_2_feed_forward1_linear2_weight_to_fp16_quantized, x = input_135_cast_fp16)[name = string("linear_20_cast_fp16")]; + fp16 var_881_to_fp16 = const()[name = string("op_881_to_fp16"), val = fp16(0x1p-1)]; + tensor var_882_cast_fp16 = mul(x = linear_20_cast_fp16, y = var_881_to_fp16)[name = string("op_882_cast_fp16")]; + tensor input_141_cast_fp16 = add(x = input_129_cast_fp16, y = var_882_cast_fp16)[name = string("input_141_cast_fp16")]; + tensor key_5_axes_0 = const()[name = string("key_5_axes_0"), val = tensor([-1])]; + tensor encoder_layers_2_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_2_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61666304)))]; + tensor encoder_layers_2_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_2_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61668416)))]; + tensor key_5_cast_fp16 = layer_norm(axes = key_5_axes_0, beta = encoder_layers_2_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_2_norm_self_att_weight_to_fp16, x = input_141_cast_fp16)[name = string("key_5_cast_fp16")]; + bool input_143_interleave_0 = const()[name = string("input_143_interleave_0"), val = bool(false)]; + tensor input_143_cast_fp16 = concat(axis = var_68, interleave = input_143_interleave_0, values = (cache_9_cast_fp16, key_5_cast_fp16))[name = string("input_143_cast_fp16")]; + tensor var_904_begin_0 = const()[name = string("op_904_begin_0"), val = tensor([0, 7, 0])]; + tensor var_904_end_0 = const()[name = string("op_904_end_0"), val = tensor([1, 42, 1024])]; + tensor var_904_end_mask_0 = const()[name = string("op_904_end_mask_0"), val = tensor([true, true, true])]; + tensor var_904_cast_fp16 = slice_by_index(begin = var_904_begin_0, end = var_904_end_0, end_mask = var_904_end_mask_0, x = cache_9_cast_fp16)[name = string("op_904_cast_fp16")]; + bool var_910_interleave_0 = const()[name = string("op_910_interleave_0"), val = bool(false)]; + tensor var_910_cast_fp16 = concat(axis = var_68, interleave = var_910_interleave_0, values = (var_904_cast_fp16, key_5_cast_fp16))[name = string("op_910_cast_fp16")]; + tensor encoder_layers_2_self_attn_linear_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61670528))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62719168))))[name = string("encoder_layers_2_self_attn_linear_q_weight_to_fp16_quantized")]; + tensor encoder_layers_2_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_2_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62721280)))]; + tensor linear_21_cast_fp16 = linear(bias = encoder_layers_2_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_2_self_attn_linear_q_weight_to_fp16_quantized, x = key_5_cast_fp16)[name = string("linear_21_cast_fp16")]; + tensor var_915 = const()[name = string("op_915"), val = tensor([1, -1, 8, 128])]; + tensor q_13_cast_fp16 = reshape(shape = var_915, x = linear_21_cast_fp16)[name = string("q_13_cast_fp16")]; + tensor encoder_layers_2_self_attn_linear_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62723392))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(63772032))))[name = string("encoder_layers_2_self_attn_linear_k_weight_to_fp16_quantized")]; + tensor encoder_layers_2_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_2_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(63774144)))]; + tensor linear_22_cast_fp16 = linear(bias = encoder_layers_2_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_2_self_attn_linear_k_weight_to_fp16_quantized, x = input_143_cast_fp16)[name = string("linear_22_cast_fp16")]; + tensor var_920 = const()[name = string("op_920"), val = tensor([1, -1, 8, 128])]; + tensor k_9_cast_fp16 = reshape(shape = var_920, x = linear_22_cast_fp16)[name = string("k_9_cast_fp16")]; + tensor encoder_layers_2_self_attn_linear_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(63776256))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64824896))))[name = string("encoder_layers_2_self_attn_linear_v_weight_to_fp16_quantized")]; + tensor encoder_layers_2_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_2_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64827008)))]; + tensor linear_23_cast_fp16 = linear(bias = encoder_layers_2_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_2_self_attn_linear_v_weight_to_fp16_quantized, x = input_143_cast_fp16)[name = string("linear_23_cast_fp16")]; + tensor var_925 = const()[name = string("op_925"), val = tensor([1, -1, 8, 128])]; + tensor v_5_cast_fp16 = reshape(shape = var_925, x = linear_23_cast_fp16)[name = string("v_5_cast_fp16")]; + tensor value_13_perm_0 = const()[name = string("value_13_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_2_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_2_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64829120)))]; + tensor var_938_cast_fp16 = add(x = q_13_cast_fp16, y = encoder_layers_2_self_attn_pos_bias_u_to_fp16)[name = string("op_938_cast_fp16")]; + tensor encoder_layers_2_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_2_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64831232)))]; + tensor var_940_cast_fp16 = add(x = q_13_cast_fp16, y = encoder_layers_2_self_attn_pos_bias_v_to_fp16)[name = string("op_940_cast_fp16")]; + tensor q_with_bias_v_5_perm_0 = const()[name = string("q_with_bias_v_5_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_59_transpose_x_0 = const()[name = string("x_59_transpose_x_0"), val = bool(false)]; + bool x_59_transpose_y_0 = const()[name = string("x_59_transpose_y_0"), val = bool(false)]; + tensor op_942_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64833344))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64932736))))[name = string("op_942_to_fp16_quantized")]; + tensor q_with_bias_v_5_cast_fp16 = transpose(perm = q_with_bias_v_5_perm_0, x = var_940_cast_fp16)[name = string("transpose_344")]; + tensor x_59_cast_fp16 = matmul(transpose_x = x_59_transpose_x_0, transpose_y = x_59_transpose_y_0, x = q_with_bias_v_5_cast_fp16, y = op_942_to_fp16_quantized)[name = string("x_59_cast_fp16")]; + tensor x_61_pad_0 = const()[name = string("x_61_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_61_mode_0 = const()[name = string("x_61_mode_0"), val = string("constant")]; + fp16 const_105_to_fp16 = const()[name = string("const_105_to_fp16"), val = fp16(0x0p+0)]; + tensor x_61_cast_fp16 = pad(constant_val = const_105_to_fp16, mode = x_61_mode_0, pad = x_61_pad_0, x = x_59_cast_fp16)[name = string("x_61_cast_fp16")]; + tensor var_950 = const()[name = string("op_950"), val = tensor([1, 8, -1, 7])]; + tensor x_63_cast_fp16 = reshape(shape = var_950, x = x_61_cast_fp16)[name = string("x_63_cast_fp16")]; + tensor var_954_begin_0 = const()[name = string("op_954_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_954_end_0 = const()[name = string("op_954_end_0"), val = tensor([1, 8, 98, 7])]; + tensor var_954_end_mask_0 = const()[name = string("op_954_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_954_cast_fp16 = slice_by_index(begin = var_954_begin_0, end = var_954_end_0, end_mask = var_954_end_mask_0, x = x_63_cast_fp16)[name = string("op_954_cast_fp16")]; + tensor var_955 = const()[name = string("op_955"), val = tensor([1, 8, 7, 97])]; + tensor matrix_bd_9_cast_fp16 = reshape(shape = var_955, x = var_954_cast_fp16)[name = string("matrix_bd_9_cast_fp16")]; + bool matrix_ac_5_transpose_x_0 = const()[name = string("matrix_ac_5_transpose_x_0"), val = bool(false)]; + bool matrix_ac_5_transpose_y_0 = const()[name = string("matrix_ac_5_transpose_y_0"), val = bool(false)]; + tensor transpose_100_perm_0 = const()[name = string("transpose_100_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_101_perm_0 = const()[name = string("transpose_101_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_101 = transpose(perm = transpose_101_perm_0, x = k_9_cast_fp16)[name = string("transpose_342")]; + tensor transpose_100 = transpose(perm = transpose_100_perm_0, x = var_938_cast_fp16)[name = string("transpose_343")]; + tensor matrix_ac_5_cast_fp16 = matmul(transpose_x = matrix_ac_5_transpose_x_0, transpose_y = matrix_ac_5_transpose_y_0, x = transpose_100, y = transpose_101)[name = string("matrix_ac_5_cast_fp16")]; + tensor matrix_bd_11_begin_0 = const()[name = string("matrix_bd_11_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_11_end_0 = const()[name = string("matrix_bd_11_end_0"), val = tensor([1, 8, 7, 49])]; + tensor matrix_bd_11_end_mask_0 = const()[name = string("matrix_bd_11_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_11_cast_fp16 = slice_by_index(begin = matrix_bd_11_begin_0, end = matrix_bd_11_end_0, end_mask = matrix_bd_11_end_mask_0, x = matrix_bd_9_cast_fp16)[name = string("matrix_bd_11_cast_fp16")]; + tensor var_964_cast_fp16 = add(x = matrix_ac_5_cast_fp16, y = matrix_bd_11_cast_fp16)[name = string("op_964_cast_fp16")]; + fp16 _inversed_scores_9_y_0_to_fp16 = const()[name = string("_inversed_scores_9_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_9_cast_fp16 = mul(x = var_964_cast_fp16, y = _inversed_scores_9_y_0_to_fp16)[name = string("_inversed_scores_9_cast_fp16")]; + tensor scores_11_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_9_cast_fp16, cond = mask_11)[name = string("scores_11_cast_fp16")]; + tensor var_970_cast_fp16 = softmax(axis = var_59, x = scores_11_cast_fp16)[name = string("op_970_cast_fp16")]; + tensor input_145_cast_fp16 = select(a = var_44_to_fp16, b = var_970_cast_fp16, cond = mask_11)[name = string("input_145_cast_fp16")]; + bool x_65_transpose_x_0 = const()[name = string("x_65_transpose_x_0"), val = bool(false)]; + bool x_65_transpose_y_0 = const()[name = string("x_65_transpose_y_0"), val = bool(false)]; + tensor value_13_cast_fp16 = transpose(perm = value_13_perm_0, x = v_5_cast_fp16)[name = string("transpose_341")]; + tensor x_65_cast_fp16 = matmul(transpose_x = x_65_transpose_x_0, transpose_y = x_65_transpose_y_0, x = input_145_cast_fp16, y = value_13_cast_fp16)[name = string("x_65_cast_fp16")]; + tensor var_974_perm_0 = const()[name = string("op_974_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_975 = const()[name = string("op_975"), val = tensor([1, -1, 1024])]; + tensor var_974_cast_fp16 = transpose(perm = var_974_perm_0, x = x_65_cast_fp16)[name = string("transpose_340")]; + tensor input_147_cast_fp16 = reshape(shape = var_975, x = var_974_cast_fp16)[name = string("input_147_cast_fp16")]; + tensor encoder_layers_2_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64933056))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65719552))))[name = string("encoder_layers_2_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_2_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_2_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65719744)))]; + tensor linear_25_cast_fp16 = linear(bias = encoder_layers_2_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_2_self_attn_linear_out_weight_to_fp16_palettized, x = input_147_cast_fp16)[name = string("linear_25_cast_fp16")]; + tensor input_151_cast_fp16 = add(x = input_141_cast_fp16, y = linear_25_cast_fp16)[name = string("input_151_cast_fp16")]; + tensor x_69_axes_0 = const()[name = string("x_69_axes_0"), val = tensor([-1])]; + tensor encoder_layers_2_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_2_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65721856)))]; + tensor encoder_layers_2_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_2_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65723968)))]; + tensor x_69_cast_fp16 = layer_norm(axes = x_69_axes_0, beta = encoder_layers_2_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_2_norm_conv_weight_to_fp16, x = input_151_cast_fp16)[name = string("x_69_cast_fp16")]; + tensor input_153_perm_0 = const()[name = string("input_153_perm_0"), val = tensor([0, 2, 1])]; + string input_155_pad_type_0 = const()[name = string("input_155_pad_type_0"), val = string("valid")]; + tensor input_155_strides_0 = const()[name = string("input_155_strides_0"), val = tensor([1])]; + tensor input_155_pad_0 = const()[name = string("input_155_pad_0"), val = tensor([0, 0])]; + tensor input_155_dilations_0 = const()[name = string("input_155_dilations_0"), val = tensor([1])]; + int32 input_155_groups_0 = const()[name = string("input_155_groups_0"), val = int32(1)]; + tensor encoder_layers_2_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65726080))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67823296))))[name = string("encoder_layers_2_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_153_cast_fp16 = transpose(perm = input_153_perm_0, x = x_69_cast_fp16)[name = string("transpose_339")]; + tensor input_155_cast_fp16 = conv(dilations = input_155_dilations_0, groups = input_155_groups_0, pad = input_155_pad_0, pad_type = input_155_pad_type_0, strides = input_155_strides_0, weight = encoder_layers_2_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_153_cast_fp16)[name = string("input_155_cast_fp16")]; + int32 x_71_split_num_splits_0 = const()[name = string("x_71_split_num_splits_0"), val = int32(2)]; + int32 x_71_split_axis_0 = const()[name = string("x_71_split_axis_0"), val = int32(1)]; + tensor x_71_split_cast_fp16_0, tensor x_71_split_cast_fp16_1 = split(axis = x_71_split_axis_0, num_splits = x_71_split_num_splits_0, x = input_155_cast_fp16)[name = string("x_71_split_cast_fp16")]; + tensor x_71_split_1_sigmoid_cast_fp16 = sigmoid(x = x_71_split_cast_fp16_1)[name = string("x_71_split_1_sigmoid_cast_fp16")]; + tensor x_71_cast_fp16 = mul(x = x_71_split_cast_fp16_0, y = x_71_split_1_sigmoid_cast_fp16)[name = string("x_71_cast_fp16")]; + tensor input_157_cast_fp16 = select(a = var_44_to_fp16, b = x_71_cast_fp16, cond = var_575)[name = string("input_157_cast_fp16")]; + bool new_x_11_interleave_0 = const()[name = string("new_x_11_interleave_0"), val = bool(false)]; + tensor new_x_11_cast_fp16 = concat(axis = var_59, interleave = new_x_11_interleave_0, values = (cache_11_cast_fp16, input_157_cast_fp16))[name = string("new_x_11_cast_fp16")]; + tensor var_1014_begin_0 = const()[name = string("op_1014_begin_0"), val = tensor([0, 0, 7])]; + tensor var_1014_end_0 = const()[name = string("op_1014_end_0"), val = tensor([1, 1024, 15])]; + tensor var_1014_end_mask_0 = const()[name = string("op_1014_end_mask_0"), val = tensor([true, true, true])]; + tensor var_1014_cast_fp16 = slice_by_index(begin = var_1014_begin_0, end = var_1014_end_0, end_mask = var_1014_end_mask_0, x = new_x_11_cast_fp16)[name = string("op_1014_cast_fp16")]; + string x_73_pad_type_0 = const()[name = string("x_73_pad_type_0"), val = string("valid")]; + int32 x_73_groups_0 = const()[name = string("x_73_groups_0"), val = int32(1024)]; + tensor x_73_strides_0 = const()[name = string("x_73_strides_0"), val = tensor([1])]; + tensor x_73_pad_0 = const()[name = string("x_73_pad_0"), val = tensor([0, 0])]; + tensor x_73_dilations_0 = const()[name = string("x_73_dilations_0"), val = tensor([1])]; + tensor encoder_layers_2_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67827456))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67836736))))[name = string("encoder_layers_2_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_73_cast_fp16 = conv(dilations = x_73_dilations_0, groups = x_73_groups_0, pad = x_73_pad_0, pad_type = x_73_pad_type_0, strides = x_73_strides_0, weight = encoder_layers_2_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_11_cast_fp16)[name = string("x_73_cast_fp16")]; + tensor input_159_perm_0 = const()[name = string("input_159_perm_0"), val = tensor([0, 2, 1])]; + tensor x_75_axes_0 = const()[name = string("x_75_axes_0"), val = tensor([-1])]; + tensor encoder_layers_2_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_2_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67838848)))]; + tensor encoder_layers_2_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_2_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67840960)))]; + tensor input_159_cast_fp16 = transpose(perm = input_159_perm_0, x = x_73_cast_fp16)[name = string("transpose_338")]; + tensor x_75_cast_fp16 = layer_norm(axes = x_75_axes_0, beta = encoder_layers_2_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_2_conv_batch_norm_weight_to_fp16, x = input_159_cast_fp16)[name = string("x_75_cast_fp16")]; + tensor input_161_perm_0 = const()[name = string("input_161_perm_0"), val = tensor([0, 2, 1])]; + tensor input_161_cast_fp16 = transpose(perm = input_161_perm_0, x = x_75_cast_fp16)[name = string("transpose_337")]; + tensor input_163_cast_fp16 = silu(x = input_161_cast_fp16)[name = string("input_163_cast_fp16")]; + string x_77_pad_type_0 = const()[name = string("x_77_pad_type_0"), val = string("valid")]; + tensor x_77_strides_0 = const()[name = string("x_77_strides_0"), val = tensor([1])]; + tensor x_77_pad_0 = const()[name = string("x_77_pad_0"), val = tensor([0, 0])]; + tensor x_77_dilations_0 = const()[name = string("x_77_dilations_0"), val = tensor([1])]; + int32 x_77_groups_0 = const()[name = string("x_77_groups_0"), val = int32(1)]; + tensor encoder_layers_2_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67843072))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68891712))))[name = string("encoder_layers_2_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_77_cast_fp16 = conv(dilations = x_77_dilations_0, groups = x_77_groups_0, pad = x_77_pad_0, pad_type = x_77_pad_type_0, strides = x_77_strides_0, weight = encoder_layers_2_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_163_cast_fp16)[name = string("x_77_cast_fp16")]; + tensor input_165_perm_0 = const()[name = string("input_165_perm_0"), val = tensor([0, 2, 1])]; + tensor input_165_cast_fp16 = transpose(perm = input_165_perm_0, x = x_77_cast_fp16)[name = string("transpose_336")]; + tensor input_167_cast_fp16 = add(x = input_151_cast_fp16, y = input_165_cast_fp16)[name = string("input_167_cast_fp16")]; + tensor input_169_axes_0 = const()[name = string("input_169_axes_0"), val = tensor([-1])]; + tensor encoder_layers_2_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_2_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68893824)))]; + tensor encoder_layers_2_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_2_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68895936)))]; + tensor input_169_cast_fp16 = layer_norm(axes = input_169_axes_0, beta = encoder_layers_2_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_2_norm_feed_forward2_weight_to_fp16, x = input_167_cast_fp16)[name = string("input_169_cast_fp16")]; + tensor encoder_layers_2_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68898048))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72043840))))[name = string("encoder_layers_2_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_2_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_2_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72044032)))]; + tensor linear_26_cast_fp16 = linear(bias = encoder_layers_2_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_2_feed_forward2_linear1_weight_to_fp16_palettized, x = input_169_cast_fp16)[name = string("linear_26_cast_fp16")]; + tensor input_173_cast_fp16 = silu(x = linear_26_cast_fp16)[name = string("input_173_cast_fp16")]; + tensor encoder_layers_2_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72052288))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75198080))))[name = string("encoder_layers_2_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_2_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_2_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75198272)))]; + tensor linear_27_cast_fp16 = linear(bias = encoder_layers_2_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_2_feed_forward2_linear2_weight_to_fp16_palettized, x = input_173_cast_fp16)[name = string("linear_27_cast_fp16")]; + fp16 var_1057_to_fp16 = const()[name = string("op_1057_to_fp16"), val = fp16(0x1p-1)]; + tensor var_1058_cast_fp16 = mul(x = linear_27_cast_fp16, y = var_1057_to_fp16)[name = string("op_1058_cast_fp16")]; + tensor input_179_cast_fp16 = add(x = input_167_cast_fp16, y = var_1058_cast_fp16)[name = string("input_179_cast_fp16")]; + tensor input_181_axes_0 = const()[name = string("input_181_axes_0"), val = tensor([-1])]; + tensor encoder_layers_2_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_2_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75200384)))]; + tensor encoder_layers_2_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_2_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75202496)))]; + tensor input_181_cast_fp16 = layer_norm(axes = input_181_axes_0, beta = encoder_layers_2_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_2_norm_out_weight_to_fp16, x = input_179_cast_fp16)[name = string("input_181_cast_fp16")]; + tensor cache_13_begin_0 = const()[name = string("cache_13_begin_0"), val = tensor([3, 0, 0, 0])]; + tensor cache_13_end_0 = const()[name = string("cache_13_end_0"), val = tensor([4, 1, 42, 1024])]; + tensor cache_13_end_mask_0 = const()[name = string("cache_13_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_13_squeeze_mask_0 = const()[name = string("cache_13_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_13_cast_fp16 = slice_by_index(begin = cache_13_begin_0, end = cache_13_end_0, end_mask = cache_13_end_mask_0, squeeze_mask = cache_13_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_13_cast_fp16")]; + tensor cache_15_begin_0 = const()[name = string("cache_15_begin_0"), val = tensor([3, 0, 0, 0])]; + tensor cache_15_end_0 = const()[name = string("cache_15_end_0"), val = tensor([4, 1, 1024, 8])]; + tensor cache_15_end_mask_0 = const()[name = string("cache_15_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_15_squeeze_mask_0 = const()[name = string("cache_15_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_15_cast_fp16 = slice_by_index(begin = cache_15_begin_0, end = cache_15_end_0, end_mask = cache_15_end_mask_0, squeeze_mask = cache_15_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_15_cast_fp16")]; + tensor input_183_axes_0 = const()[name = string("input_183_axes_0"), val = tensor([-1])]; + tensor encoder_layers_3_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_3_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75204608)))]; + tensor encoder_layers_3_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_3_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75206720)))]; + tensor input_183_cast_fp16 = layer_norm(axes = input_183_axes_0, beta = encoder_layers_3_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_3_norm_feed_forward1_weight_to_fp16, x = input_181_cast_fp16)[name = string("input_183_cast_fp16")]; + tensor encoder_layers_3_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75208832))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78354624))))[name = string("encoder_layers_3_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_3_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_3_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78354816)))]; + tensor linear_28_cast_fp16 = linear(bias = encoder_layers_3_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_3_feed_forward1_linear1_weight_to_fp16_palettized, x = input_183_cast_fp16)[name = string("linear_28_cast_fp16")]; + tensor input_187_cast_fp16 = silu(x = linear_28_cast_fp16)[name = string("input_187_cast_fp16")]; + tensor encoder_layers_3_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78363072))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81508864))))[name = string("encoder_layers_3_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_3_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_3_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81509056)))]; + tensor linear_29_cast_fp16 = linear(bias = encoder_layers_3_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_3_feed_forward1_linear2_weight_to_fp16_palettized, x = input_187_cast_fp16)[name = string("linear_29_cast_fp16")]; + fp16 var_1094_to_fp16 = const()[name = string("op_1094_to_fp16"), val = fp16(0x1p-1)]; + tensor var_1095_cast_fp16 = mul(x = linear_29_cast_fp16, y = var_1094_to_fp16)[name = string("op_1095_cast_fp16")]; + tensor input_193_cast_fp16 = add(x = input_181_cast_fp16, y = var_1095_cast_fp16)[name = string("input_193_cast_fp16")]; + tensor key_7_axes_0 = const()[name = string("key_7_axes_0"), val = tensor([-1])]; + tensor encoder_layers_3_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_3_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81511168)))]; + tensor encoder_layers_3_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_3_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81513280)))]; + tensor key_7_cast_fp16 = layer_norm(axes = key_7_axes_0, beta = encoder_layers_3_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_3_norm_self_att_weight_to_fp16, x = input_193_cast_fp16)[name = string("key_7_cast_fp16")]; + bool input_195_interleave_0 = const()[name = string("input_195_interleave_0"), val = bool(false)]; + tensor input_195_cast_fp16 = concat(axis = var_68, interleave = input_195_interleave_0, values = (cache_13_cast_fp16, key_7_cast_fp16))[name = string("input_195_cast_fp16")]; + tensor var_1117_begin_0 = const()[name = string("op_1117_begin_0"), val = tensor([0, 7, 0])]; + tensor var_1117_end_0 = const()[name = string("op_1117_end_0"), val = tensor([1, 42, 1024])]; + tensor var_1117_end_mask_0 = const()[name = string("op_1117_end_mask_0"), val = tensor([true, true, true])]; + tensor var_1117_cast_fp16 = slice_by_index(begin = var_1117_begin_0, end = var_1117_end_0, end_mask = var_1117_end_mask_0, x = cache_13_cast_fp16)[name = string("op_1117_cast_fp16")]; + bool var_1123_interleave_0 = const()[name = string("op_1123_interleave_0"), val = bool(false)]; + tensor var_1123_cast_fp16 = concat(axis = var_68, interleave = var_1123_interleave_0, values = (var_1117_cast_fp16, key_7_cast_fp16))[name = string("op_1123_cast_fp16")]; + tensor encoder_layers_3_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81515392))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82301888))))[name = string("encoder_layers_3_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_3_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_3_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82302080)))]; + tensor linear_30_cast_fp16 = linear(bias = encoder_layers_3_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_3_self_attn_linear_q_weight_to_fp16_palettized, x = key_7_cast_fp16)[name = string("linear_30_cast_fp16")]; + tensor var_1128 = const()[name = string("op_1128"), val = tensor([1, -1, 8, 128])]; + tensor q_19_cast_fp16 = reshape(shape = var_1128, x = linear_30_cast_fp16)[name = string("q_19_cast_fp16")]; + tensor encoder_layers_3_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82304192))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83090688))))[name = string("encoder_layers_3_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_3_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_3_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83090880)))]; + tensor linear_31_cast_fp16 = linear(bias = encoder_layers_3_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_3_self_attn_linear_k_weight_to_fp16_palettized, x = input_195_cast_fp16)[name = string("linear_31_cast_fp16")]; + tensor var_1133 = const()[name = string("op_1133"), val = tensor([1, -1, 8, 128])]; + tensor k_13_cast_fp16 = reshape(shape = var_1133, x = linear_31_cast_fp16)[name = string("k_13_cast_fp16")]; + tensor encoder_layers_3_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83092992))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83879488))))[name = string("encoder_layers_3_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_3_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_3_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83879680)))]; + tensor linear_32_cast_fp16 = linear(bias = encoder_layers_3_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_3_self_attn_linear_v_weight_to_fp16_palettized, x = input_195_cast_fp16)[name = string("linear_32_cast_fp16")]; + tensor var_1138 = const()[name = string("op_1138"), val = tensor([1, -1, 8, 128])]; + tensor v_7_cast_fp16 = reshape(shape = var_1138, x = linear_32_cast_fp16)[name = string("v_7_cast_fp16")]; + tensor value_15_perm_0 = const()[name = string("value_15_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_3_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_3_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83881792)))]; + tensor var_1151_cast_fp16 = add(x = q_19_cast_fp16, y = encoder_layers_3_self_attn_pos_bias_u_to_fp16)[name = string("op_1151_cast_fp16")]; + tensor encoder_layers_3_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_3_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83883904)))]; + tensor var_1153_cast_fp16 = add(x = q_19_cast_fp16, y = encoder_layers_3_self_attn_pos_bias_v_to_fp16)[name = string("op_1153_cast_fp16")]; + tensor q_with_bias_v_7_perm_0 = const()[name = string("q_with_bias_v_7_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_85_transpose_x_0 = const()[name = string("x_85_transpose_x_0"), val = bool(false)]; + bool x_85_transpose_y_0 = const()[name = string("x_85_transpose_y_0"), val = bool(false)]; + tensor op_1155_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83886016))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83985408))))[name = string("op_1155_to_fp16_quantized")]; + tensor q_with_bias_v_7_cast_fp16 = transpose(perm = q_with_bias_v_7_perm_0, x = var_1153_cast_fp16)[name = string("transpose_335")]; + tensor x_85_cast_fp16 = matmul(transpose_x = x_85_transpose_x_0, transpose_y = x_85_transpose_y_0, x = q_with_bias_v_7_cast_fp16, y = op_1155_to_fp16_quantized)[name = string("x_85_cast_fp16")]; + tensor x_87_pad_0 = const()[name = string("x_87_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_87_mode_0 = const()[name = string("x_87_mode_0"), val = string("constant")]; + fp16 const_118_to_fp16 = const()[name = string("const_118_to_fp16"), val = fp16(0x0p+0)]; + tensor x_87_cast_fp16 = pad(constant_val = const_118_to_fp16, mode = x_87_mode_0, pad = x_87_pad_0, x = x_85_cast_fp16)[name = string("x_87_cast_fp16")]; + tensor var_1163 = const()[name = string("op_1163"), val = tensor([1, 8, -1, 7])]; + tensor x_89_cast_fp16 = reshape(shape = var_1163, x = x_87_cast_fp16)[name = string("x_89_cast_fp16")]; + tensor var_1167_begin_0 = const()[name = string("op_1167_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_1167_end_0 = const()[name = string("op_1167_end_0"), val = tensor([1, 8, 98, 7])]; + tensor var_1167_end_mask_0 = const()[name = string("op_1167_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_1167_cast_fp16 = slice_by_index(begin = var_1167_begin_0, end = var_1167_end_0, end_mask = var_1167_end_mask_0, x = x_89_cast_fp16)[name = string("op_1167_cast_fp16")]; + tensor var_1168 = const()[name = string("op_1168"), val = tensor([1, 8, 7, 97])]; + tensor matrix_bd_13_cast_fp16 = reshape(shape = var_1168, x = var_1167_cast_fp16)[name = string("matrix_bd_13_cast_fp16")]; + bool matrix_ac_7_transpose_x_0 = const()[name = string("matrix_ac_7_transpose_x_0"), val = bool(false)]; + bool matrix_ac_7_transpose_y_0 = const()[name = string("matrix_ac_7_transpose_y_0"), val = bool(false)]; + tensor transpose_102_perm_0 = const()[name = string("transpose_102_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_103_perm_0 = const()[name = string("transpose_103_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_103 = transpose(perm = transpose_103_perm_0, x = k_13_cast_fp16)[name = string("transpose_333")]; + tensor transpose_102 = transpose(perm = transpose_102_perm_0, x = var_1151_cast_fp16)[name = string("transpose_334")]; + tensor matrix_ac_7_cast_fp16 = matmul(transpose_x = matrix_ac_7_transpose_x_0, transpose_y = matrix_ac_7_transpose_y_0, x = transpose_102, y = transpose_103)[name = string("matrix_ac_7_cast_fp16")]; + tensor matrix_bd_15_begin_0 = const()[name = string("matrix_bd_15_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_15_end_0 = const()[name = string("matrix_bd_15_end_0"), val = tensor([1, 8, 7, 49])]; + tensor matrix_bd_15_end_mask_0 = const()[name = string("matrix_bd_15_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_15_cast_fp16 = slice_by_index(begin = matrix_bd_15_begin_0, end = matrix_bd_15_end_0, end_mask = matrix_bd_15_end_mask_0, x = matrix_bd_13_cast_fp16)[name = string("matrix_bd_15_cast_fp16")]; + tensor var_1177_cast_fp16 = add(x = matrix_ac_7_cast_fp16, y = matrix_bd_15_cast_fp16)[name = string("op_1177_cast_fp16")]; + fp16 _inversed_scores_13_y_0_to_fp16 = const()[name = string("_inversed_scores_13_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_13_cast_fp16 = mul(x = var_1177_cast_fp16, y = _inversed_scores_13_y_0_to_fp16)[name = string("_inversed_scores_13_cast_fp16")]; + tensor scores_15_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_13_cast_fp16, cond = mask_11)[name = string("scores_15_cast_fp16")]; + tensor var_1183_cast_fp16 = softmax(axis = var_59, x = scores_15_cast_fp16)[name = string("op_1183_cast_fp16")]; + tensor input_197_cast_fp16 = select(a = var_44_to_fp16, b = var_1183_cast_fp16, cond = mask_11)[name = string("input_197_cast_fp16")]; + bool x_91_transpose_x_0 = const()[name = string("x_91_transpose_x_0"), val = bool(false)]; + bool x_91_transpose_y_0 = const()[name = string("x_91_transpose_y_0"), val = bool(false)]; + tensor value_15_cast_fp16 = transpose(perm = value_15_perm_0, x = v_7_cast_fp16)[name = string("transpose_332")]; + tensor x_91_cast_fp16 = matmul(transpose_x = x_91_transpose_x_0, transpose_y = x_91_transpose_y_0, x = input_197_cast_fp16, y = value_15_cast_fp16)[name = string("x_91_cast_fp16")]; + tensor var_1187_perm_0 = const()[name = string("op_1187_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1188 = const()[name = string("op_1188"), val = tensor([1, -1, 1024])]; + tensor var_1187_cast_fp16 = transpose(perm = var_1187_perm_0, x = x_91_cast_fp16)[name = string("transpose_331")]; + tensor input_199_cast_fp16 = reshape(shape = var_1188, x = var_1187_cast_fp16)[name = string("input_199_cast_fp16")]; + tensor encoder_layers_3_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83985728))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84772224))))[name = string("encoder_layers_3_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_3_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_3_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84772416)))]; + tensor linear_34_cast_fp16 = linear(bias = encoder_layers_3_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_3_self_attn_linear_out_weight_to_fp16_palettized, x = input_199_cast_fp16)[name = string("linear_34_cast_fp16")]; + tensor input_203_cast_fp16 = add(x = input_193_cast_fp16, y = linear_34_cast_fp16)[name = string("input_203_cast_fp16")]; + tensor x_95_axes_0 = const()[name = string("x_95_axes_0"), val = tensor([-1])]; + tensor encoder_layers_3_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_3_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84774528)))]; + tensor encoder_layers_3_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_3_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84776640)))]; + tensor x_95_cast_fp16 = layer_norm(axes = x_95_axes_0, beta = encoder_layers_3_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_3_norm_conv_weight_to_fp16, x = input_203_cast_fp16)[name = string("x_95_cast_fp16")]; + tensor input_205_perm_0 = const()[name = string("input_205_perm_0"), val = tensor([0, 2, 1])]; + string input_207_pad_type_0 = const()[name = string("input_207_pad_type_0"), val = string("valid")]; + tensor input_207_strides_0 = const()[name = string("input_207_strides_0"), val = tensor([1])]; + tensor input_207_pad_0 = const()[name = string("input_207_pad_0"), val = tensor([0, 0])]; + tensor input_207_dilations_0 = const()[name = string("input_207_dilations_0"), val = tensor([1])]; + int32 input_207_groups_0 = const()[name = string("input_207_groups_0"), val = int32(1)]; + tensor encoder_layers_3_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84778752))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86875968))))[name = string("encoder_layers_3_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_205_cast_fp16 = transpose(perm = input_205_perm_0, x = x_95_cast_fp16)[name = string("transpose_330")]; + tensor input_207_cast_fp16 = conv(dilations = input_207_dilations_0, groups = input_207_groups_0, pad = input_207_pad_0, pad_type = input_207_pad_type_0, strides = input_207_strides_0, weight = encoder_layers_3_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_205_cast_fp16)[name = string("input_207_cast_fp16")]; + int32 x_97_split_num_splits_0 = const()[name = string("x_97_split_num_splits_0"), val = int32(2)]; + int32 x_97_split_axis_0 = const()[name = string("x_97_split_axis_0"), val = int32(1)]; + tensor x_97_split_cast_fp16_0, tensor x_97_split_cast_fp16_1 = split(axis = x_97_split_axis_0, num_splits = x_97_split_num_splits_0, x = input_207_cast_fp16)[name = string("x_97_split_cast_fp16")]; + tensor x_97_split_1_sigmoid_cast_fp16 = sigmoid(x = x_97_split_cast_fp16_1)[name = string("x_97_split_1_sigmoid_cast_fp16")]; + tensor x_97_cast_fp16 = mul(x = x_97_split_cast_fp16_0, y = x_97_split_1_sigmoid_cast_fp16)[name = string("x_97_cast_fp16")]; + tensor input_209_cast_fp16 = select(a = var_44_to_fp16, b = x_97_cast_fp16, cond = var_575)[name = string("input_209_cast_fp16")]; + bool new_x_15_interleave_0 = const()[name = string("new_x_15_interleave_0"), val = bool(false)]; + tensor new_x_15_cast_fp16 = concat(axis = var_59, interleave = new_x_15_interleave_0, values = (cache_15_cast_fp16, input_209_cast_fp16))[name = string("new_x_15_cast_fp16")]; + tensor var_1227_begin_0 = const()[name = string("op_1227_begin_0"), val = tensor([0, 0, 7])]; + tensor var_1227_end_0 = const()[name = string("op_1227_end_0"), val = tensor([1, 1024, 15])]; + tensor var_1227_end_mask_0 = const()[name = string("op_1227_end_mask_0"), val = tensor([true, true, true])]; + tensor var_1227_cast_fp16 = slice_by_index(begin = var_1227_begin_0, end = var_1227_end_0, end_mask = var_1227_end_mask_0, x = new_x_15_cast_fp16)[name = string("op_1227_cast_fp16")]; + string x_99_pad_type_0 = const()[name = string("x_99_pad_type_0"), val = string("valid")]; + int32 x_99_groups_0 = const()[name = string("x_99_groups_0"), val = int32(1024)]; + tensor x_99_strides_0 = const()[name = string("x_99_strides_0"), val = tensor([1])]; + tensor x_99_pad_0 = const()[name = string("x_99_pad_0"), val = tensor([0, 0])]; + tensor x_99_dilations_0 = const()[name = string("x_99_dilations_0"), val = tensor([1])]; + tensor encoder_layers_3_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86880128))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86889408))))[name = string("encoder_layers_3_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_99_cast_fp16 = conv(dilations = x_99_dilations_0, groups = x_99_groups_0, pad = x_99_pad_0, pad_type = x_99_pad_type_0, strides = x_99_strides_0, weight = encoder_layers_3_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_15_cast_fp16)[name = string("x_99_cast_fp16")]; + tensor input_211_perm_0 = const()[name = string("input_211_perm_0"), val = tensor([0, 2, 1])]; + tensor x_101_axes_0 = const()[name = string("x_101_axes_0"), val = tensor([-1])]; + tensor encoder_layers_3_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_3_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86891520)))]; + tensor encoder_layers_3_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_3_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86893632)))]; + tensor input_211_cast_fp16 = transpose(perm = input_211_perm_0, x = x_99_cast_fp16)[name = string("transpose_329")]; + tensor x_101_cast_fp16 = layer_norm(axes = x_101_axes_0, beta = encoder_layers_3_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_3_conv_batch_norm_weight_to_fp16, x = input_211_cast_fp16)[name = string("x_101_cast_fp16")]; + tensor input_213_perm_0 = const()[name = string("input_213_perm_0"), val = tensor([0, 2, 1])]; + tensor input_213_cast_fp16 = transpose(perm = input_213_perm_0, x = x_101_cast_fp16)[name = string("transpose_328")]; + tensor input_215_cast_fp16 = silu(x = input_213_cast_fp16)[name = string("input_215_cast_fp16")]; + string x_103_pad_type_0 = const()[name = string("x_103_pad_type_0"), val = string("valid")]; + tensor x_103_strides_0 = const()[name = string("x_103_strides_0"), val = tensor([1])]; + tensor x_103_pad_0 = const()[name = string("x_103_pad_0"), val = tensor([0, 0])]; + tensor x_103_dilations_0 = const()[name = string("x_103_dilations_0"), val = tensor([1])]; + int32 x_103_groups_0 = const()[name = string("x_103_groups_0"), val = int32(1)]; + tensor encoder_layers_3_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86895744))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87944384))))[name = string("encoder_layers_3_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_103_cast_fp16 = conv(dilations = x_103_dilations_0, groups = x_103_groups_0, pad = x_103_pad_0, pad_type = x_103_pad_type_0, strides = x_103_strides_0, weight = encoder_layers_3_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_215_cast_fp16)[name = string("x_103_cast_fp16")]; + tensor input_217_perm_0 = const()[name = string("input_217_perm_0"), val = tensor([0, 2, 1])]; + tensor input_217_cast_fp16 = transpose(perm = input_217_perm_0, x = x_103_cast_fp16)[name = string("transpose_327")]; + tensor input_219_cast_fp16 = add(x = input_203_cast_fp16, y = input_217_cast_fp16)[name = string("input_219_cast_fp16")]; + tensor input_221_axes_0 = const()[name = string("input_221_axes_0"), val = tensor([-1])]; + tensor encoder_layers_3_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_3_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87946496)))]; + tensor encoder_layers_3_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_3_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87948608)))]; + tensor input_221_cast_fp16 = layer_norm(axes = input_221_axes_0, beta = encoder_layers_3_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_3_norm_feed_forward2_weight_to_fp16, x = input_219_cast_fp16)[name = string("input_221_cast_fp16")]; + tensor encoder_layers_3_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87950720))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(91096512))))[name = string("encoder_layers_3_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_3_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_3_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(91096704)))]; + tensor linear_35_cast_fp16 = linear(bias = encoder_layers_3_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_3_feed_forward2_linear1_weight_to_fp16_palettized, x = input_221_cast_fp16)[name = string("linear_35_cast_fp16")]; + tensor input_225_cast_fp16 = silu(x = linear_35_cast_fp16)[name = string("input_225_cast_fp16")]; + tensor encoder_layers_3_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(91104960))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94250752))))[name = string("encoder_layers_3_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_3_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_3_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94250944)))]; + tensor linear_36_cast_fp16 = linear(bias = encoder_layers_3_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_3_feed_forward2_linear2_weight_to_fp16_palettized, x = input_225_cast_fp16)[name = string("linear_36_cast_fp16")]; + fp16 var_1270_to_fp16 = const()[name = string("op_1270_to_fp16"), val = fp16(0x1p-1)]; + tensor var_1271_cast_fp16 = mul(x = linear_36_cast_fp16, y = var_1270_to_fp16)[name = string("op_1271_cast_fp16")]; + tensor input_231_cast_fp16 = add(x = input_219_cast_fp16, y = var_1271_cast_fp16)[name = string("input_231_cast_fp16")]; + tensor input_233_axes_0 = const()[name = string("input_233_axes_0"), val = tensor([-1])]; + tensor encoder_layers_3_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_3_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94253056)))]; + tensor encoder_layers_3_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_3_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94255168)))]; + tensor input_233_cast_fp16 = layer_norm(axes = input_233_axes_0, beta = encoder_layers_3_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_3_norm_out_weight_to_fp16, x = input_231_cast_fp16)[name = string("input_233_cast_fp16")]; + tensor cache_17_begin_0 = const()[name = string("cache_17_begin_0"), val = tensor([4, 0, 0, 0])]; + tensor cache_17_end_0 = const()[name = string("cache_17_end_0"), val = tensor([5, 1, 42, 1024])]; + tensor cache_17_end_mask_0 = const()[name = string("cache_17_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_17_squeeze_mask_0 = const()[name = string("cache_17_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_17_cast_fp16 = slice_by_index(begin = cache_17_begin_0, end = cache_17_end_0, end_mask = cache_17_end_mask_0, squeeze_mask = cache_17_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_17_cast_fp16")]; + tensor cache_19_begin_0 = const()[name = string("cache_19_begin_0"), val = tensor([4, 0, 0, 0])]; + tensor cache_19_end_0 = const()[name = string("cache_19_end_0"), val = tensor([5, 1, 1024, 8])]; + tensor cache_19_end_mask_0 = const()[name = string("cache_19_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_19_squeeze_mask_0 = const()[name = string("cache_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_19_cast_fp16 = slice_by_index(begin = cache_19_begin_0, end = cache_19_end_0, end_mask = cache_19_end_mask_0, squeeze_mask = cache_19_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_19_cast_fp16")]; + tensor input_235_axes_0 = const()[name = string("input_235_axes_0"), val = tensor([-1])]; + tensor encoder_layers_4_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_4_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94257280)))]; + tensor encoder_layers_4_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_4_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94259392)))]; + tensor input_235_cast_fp16 = layer_norm(axes = input_235_axes_0, beta = encoder_layers_4_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_4_norm_feed_forward1_weight_to_fp16, x = input_233_cast_fp16)[name = string("input_235_cast_fp16")]; + tensor encoder_layers_4_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94261504))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97407296))))[name = string("encoder_layers_4_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_4_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_4_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97407488)))]; + tensor linear_37_cast_fp16 = linear(bias = encoder_layers_4_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_4_feed_forward1_linear1_weight_to_fp16_palettized, x = input_235_cast_fp16)[name = string("linear_37_cast_fp16")]; + tensor input_239_cast_fp16 = silu(x = linear_37_cast_fp16)[name = string("input_239_cast_fp16")]; + tensor encoder_layers_4_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97415744))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100561536))))[name = string("encoder_layers_4_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_4_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_4_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100561728)))]; + tensor linear_38_cast_fp16 = linear(bias = encoder_layers_4_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_4_feed_forward1_linear2_weight_to_fp16_palettized, x = input_239_cast_fp16)[name = string("linear_38_cast_fp16")]; + fp16 var_1307_to_fp16 = const()[name = string("op_1307_to_fp16"), val = fp16(0x1p-1)]; + tensor var_1308_cast_fp16 = mul(x = linear_38_cast_fp16, y = var_1307_to_fp16)[name = string("op_1308_cast_fp16")]; + tensor input_245_cast_fp16 = add(x = input_233_cast_fp16, y = var_1308_cast_fp16)[name = string("input_245_cast_fp16")]; + tensor key_9_axes_0 = const()[name = string("key_9_axes_0"), val = tensor([-1])]; + tensor encoder_layers_4_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_4_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100563840)))]; + tensor encoder_layers_4_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_4_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100565952)))]; + tensor key_9_cast_fp16 = layer_norm(axes = key_9_axes_0, beta = encoder_layers_4_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_4_norm_self_att_weight_to_fp16, x = input_245_cast_fp16)[name = string("key_9_cast_fp16")]; + bool input_247_interleave_0 = const()[name = string("input_247_interleave_0"), val = bool(false)]; + tensor input_247_cast_fp16 = concat(axis = var_68, interleave = input_247_interleave_0, values = (cache_17_cast_fp16, key_9_cast_fp16))[name = string("input_247_cast_fp16")]; + tensor var_1330_begin_0 = const()[name = string("op_1330_begin_0"), val = tensor([0, 7, 0])]; + tensor var_1330_end_0 = const()[name = string("op_1330_end_0"), val = tensor([1, 42, 1024])]; + tensor var_1330_end_mask_0 = const()[name = string("op_1330_end_mask_0"), val = tensor([true, true, true])]; + tensor var_1330_cast_fp16 = slice_by_index(begin = var_1330_begin_0, end = var_1330_end_0, end_mask = var_1330_end_mask_0, x = cache_17_cast_fp16)[name = string("op_1330_cast_fp16")]; + bool var_1336_interleave_0 = const()[name = string("op_1336_interleave_0"), val = bool(false)]; + tensor var_1336_cast_fp16 = concat(axis = var_68, interleave = var_1336_interleave_0, values = (var_1330_cast_fp16, key_9_cast_fp16))[name = string("op_1336_cast_fp16")]; + tensor encoder_layers_4_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100568064))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(101354560))))[name = string("encoder_layers_4_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_4_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_4_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(101354752)))]; + tensor linear_39_cast_fp16 = linear(bias = encoder_layers_4_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_4_self_attn_linear_q_weight_to_fp16_palettized, x = key_9_cast_fp16)[name = string("linear_39_cast_fp16")]; + tensor var_1341 = const()[name = string("op_1341"), val = tensor([1, -1, 8, 128])]; + tensor q_25_cast_fp16 = reshape(shape = var_1341, x = linear_39_cast_fp16)[name = string("q_25_cast_fp16")]; + tensor encoder_layers_4_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(101356864))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102143360))))[name = string("encoder_layers_4_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_4_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_4_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102143552)))]; + tensor linear_40_cast_fp16 = linear(bias = encoder_layers_4_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_4_self_attn_linear_k_weight_to_fp16_palettized, x = input_247_cast_fp16)[name = string("linear_40_cast_fp16")]; + tensor var_1346 = const()[name = string("op_1346"), val = tensor([1, -1, 8, 128])]; + tensor k_17_cast_fp16 = reshape(shape = var_1346, x = linear_40_cast_fp16)[name = string("k_17_cast_fp16")]; + tensor encoder_layers_4_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102145664))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102932160))))[name = string("encoder_layers_4_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_4_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_4_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102932352)))]; + tensor linear_41_cast_fp16 = linear(bias = encoder_layers_4_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_4_self_attn_linear_v_weight_to_fp16_palettized, x = input_247_cast_fp16)[name = string("linear_41_cast_fp16")]; + tensor var_1351 = const()[name = string("op_1351"), val = tensor([1, -1, 8, 128])]; + tensor v_9_cast_fp16 = reshape(shape = var_1351, x = linear_41_cast_fp16)[name = string("v_9_cast_fp16")]; + tensor value_17_perm_0 = const()[name = string("value_17_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_4_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_4_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102934464)))]; + tensor var_1364_cast_fp16 = add(x = q_25_cast_fp16, y = encoder_layers_4_self_attn_pos_bias_u_to_fp16)[name = string("op_1364_cast_fp16")]; + tensor encoder_layers_4_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_4_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102936576)))]; + tensor var_1366_cast_fp16 = add(x = q_25_cast_fp16, y = encoder_layers_4_self_attn_pos_bias_v_to_fp16)[name = string("op_1366_cast_fp16")]; + tensor q_with_bias_v_9_perm_0 = const()[name = string("q_with_bias_v_9_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_111_transpose_x_0 = const()[name = string("x_111_transpose_x_0"), val = bool(false)]; + bool x_111_transpose_y_0 = const()[name = string("x_111_transpose_y_0"), val = bool(false)]; + tensor op_1368_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102938688))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103038080))))[name = string("op_1368_to_fp16_quantized")]; + tensor q_with_bias_v_9_cast_fp16 = transpose(perm = q_with_bias_v_9_perm_0, x = var_1366_cast_fp16)[name = string("transpose_326")]; + tensor x_111_cast_fp16 = matmul(transpose_x = x_111_transpose_x_0, transpose_y = x_111_transpose_y_0, x = q_with_bias_v_9_cast_fp16, y = op_1368_to_fp16_quantized)[name = string("x_111_cast_fp16")]; + tensor x_113_pad_0 = const()[name = string("x_113_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_113_mode_0 = const()[name = string("x_113_mode_0"), val = string("constant")]; + fp16 const_131_to_fp16 = const()[name = string("const_131_to_fp16"), val = fp16(0x0p+0)]; + tensor x_113_cast_fp16 = pad(constant_val = const_131_to_fp16, mode = x_113_mode_0, pad = x_113_pad_0, x = x_111_cast_fp16)[name = string("x_113_cast_fp16")]; + tensor var_1376 = const()[name = string("op_1376"), val = tensor([1, 8, -1, 7])]; + tensor x_115_cast_fp16 = reshape(shape = var_1376, x = x_113_cast_fp16)[name = string("x_115_cast_fp16")]; + tensor var_1380_begin_0 = const()[name = string("op_1380_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_1380_end_0 = const()[name = string("op_1380_end_0"), val = tensor([1, 8, 98, 7])]; + tensor var_1380_end_mask_0 = const()[name = string("op_1380_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_1380_cast_fp16 = slice_by_index(begin = var_1380_begin_0, end = var_1380_end_0, end_mask = var_1380_end_mask_0, x = x_115_cast_fp16)[name = string("op_1380_cast_fp16")]; + tensor var_1381 = const()[name = string("op_1381"), val = tensor([1, 8, 7, 97])]; + tensor matrix_bd_17_cast_fp16 = reshape(shape = var_1381, x = var_1380_cast_fp16)[name = string("matrix_bd_17_cast_fp16")]; + bool matrix_ac_9_transpose_x_0 = const()[name = string("matrix_ac_9_transpose_x_0"), val = bool(false)]; + bool matrix_ac_9_transpose_y_0 = const()[name = string("matrix_ac_9_transpose_y_0"), val = bool(false)]; + tensor transpose_104_perm_0 = const()[name = string("transpose_104_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_105_perm_0 = const()[name = string("transpose_105_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_105 = transpose(perm = transpose_105_perm_0, x = k_17_cast_fp16)[name = string("transpose_324")]; + tensor transpose_104 = transpose(perm = transpose_104_perm_0, x = var_1364_cast_fp16)[name = string("transpose_325")]; + tensor matrix_ac_9_cast_fp16 = matmul(transpose_x = matrix_ac_9_transpose_x_0, transpose_y = matrix_ac_9_transpose_y_0, x = transpose_104, y = transpose_105)[name = string("matrix_ac_9_cast_fp16")]; + tensor matrix_bd_19_begin_0 = const()[name = string("matrix_bd_19_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_19_end_0 = const()[name = string("matrix_bd_19_end_0"), val = tensor([1, 8, 7, 49])]; + tensor matrix_bd_19_end_mask_0 = const()[name = string("matrix_bd_19_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_19_cast_fp16 = slice_by_index(begin = matrix_bd_19_begin_0, end = matrix_bd_19_end_0, end_mask = matrix_bd_19_end_mask_0, x = matrix_bd_17_cast_fp16)[name = string("matrix_bd_19_cast_fp16")]; + tensor var_1390_cast_fp16 = add(x = matrix_ac_9_cast_fp16, y = matrix_bd_19_cast_fp16)[name = string("op_1390_cast_fp16")]; + fp16 _inversed_scores_17_y_0_to_fp16 = const()[name = string("_inversed_scores_17_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_17_cast_fp16 = mul(x = var_1390_cast_fp16, y = _inversed_scores_17_y_0_to_fp16)[name = string("_inversed_scores_17_cast_fp16")]; + tensor scores_19_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_17_cast_fp16, cond = mask_11)[name = string("scores_19_cast_fp16")]; + tensor var_1396_cast_fp16 = softmax(axis = var_59, x = scores_19_cast_fp16)[name = string("op_1396_cast_fp16")]; + tensor input_249_cast_fp16 = select(a = var_44_to_fp16, b = var_1396_cast_fp16, cond = mask_11)[name = string("input_249_cast_fp16")]; + bool x_117_transpose_x_0 = const()[name = string("x_117_transpose_x_0"), val = bool(false)]; + bool x_117_transpose_y_0 = const()[name = string("x_117_transpose_y_0"), val = bool(false)]; + tensor value_17_cast_fp16 = transpose(perm = value_17_perm_0, x = v_9_cast_fp16)[name = string("transpose_323")]; + tensor x_117_cast_fp16 = matmul(transpose_x = x_117_transpose_x_0, transpose_y = x_117_transpose_y_0, x = input_249_cast_fp16, y = value_17_cast_fp16)[name = string("x_117_cast_fp16")]; + tensor var_1400_perm_0 = const()[name = string("op_1400_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1401 = const()[name = string("op_1401"), val = tensor([1, -1, 1024])]; + tensor var_1400_cast_fp16 = transpose(perm = var_1400_perm_0, x = x_117_cast_fp16)[name = string("transpose_322")]; + tensor input_251_cast_fp16 = reshape(shape = var_1401, x = var_1400_cast_fp16)[name = string("input_251_cast_fp16")]; + tensor encoder_layers_4_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103038400))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103824896))))[name = string("encoder_layers_4_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_4_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_4_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103825088)))]; + tensor linear_43_cast_fp16 = linear(bias = encoder_layers_4_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_4_self_attn_linear_out_weight_to_fp16_palettized, x = input_251_cast_fp16)[name = string("linear_43_cast_fp16")]; + tensor input_255_cast_fp16 = add(x = input_245_cast_fp16, y = linear_43_cast_fp16)[name = string("input_255_cast_fp16")]; + tensor x_121_axes_0 = const()[name = string("x_121_axes_0"), val = tensor([-1])]; + tensor encoder_layers_4_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_4_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103827200)))]; + tensor encoder_layers_4_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_4_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103829312)))]; + tensor x_121_cast_fp16 = layer_norm(axes = x_121_axes_0, beta = encoder_layers_4_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_4_norm_conv_weight_to_fp16, x = input_255_cast_fp16)[name = string("x_121_cast_fp16")]; + tensor input_257_perm_0 = const()[name = string("input_257_perm_0"), val = tensor([0, 2, 1])]; + string input_259_pad_type_0 = const()[name = string("input_259_pad_type_0"), val = string("valid")]; + tensor input_259_strides_0 = const()[name = string("input_259_strides_0"), val = tensor([1])]; + tensor input_259_pad_0 = const()[name = string("input_259_pad_0"), val = tensor([0, 0])]; + tensor input_259_dilations_0 = const()[name = string("input_259_dilations_0"), val = tensor([1])]; + int32 input_259_groups_0 = const()[name = string("input_259_groups_0"), val = int32(1)]; + tensor encoder_layers_4_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103831424))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105928640))))[name = string("encoder_layers_4_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_257_cast_fp16 = transpose(perm = input_257_perm_0, x = x_121_cast_fp16)[name = string("transpose_321")]; + tensor input_259_cast_fp16 = conv(dilations = input_259_dilations_0, groups = input_259_groups_0, pad = input_259_pad_0, pad_type = input_259_pad_type_0, strides = input_259_strides_0, weight = encoder_layers_4_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_257_cast_fp16)[name = string("input_259_cast_fp16")]; + int32 x_123_split_num_splits_0 = const()[name = string("x_123_split_num_splits_0"), val = int32(2)]; + int32 x_123_split_axis_0 = const()[name = string("x_123_split_axis_0"), val = int32(1)]; + tensor x_123_split_cast_fp16_0, tensor x_123_split_cast_fp16_1 = split(axis = x_123_split_axis_0, num_splits = x_123_split_num_splits_0, x = input_259_cast_fp16)[name = string("x_123_split_cast_fp16")]; + tensor x_123_split_1_sigmoid_cast_fp16 = sigmoid(x = x_123_split_cast_fp16_1)[name = string("x_123_split_1_sigmoid_cast_fp16")]; + tensor x_123_cast_fp16 = mul(x = x_123_split_cast_fp16_0, y = x_123_split_1_sigmoid_cast_fp16)[name = string("x_123_cast_fp16")]; + tensor input_261_cast_fp16 = select(a = var_44_to_fp16, b = x_123_cast_fp16, cond = var_575)[name = string("input_261_cast_fp16")]; + bool new_x_19_interleave_0 = const()[name = string("new_x_19_interleave_0"), val = bool(false)]; + tensor new_x_19_cast_fp16 = concat(axis = var_59, interleave = new_x_19_interleave_0, values = (cache_19_cast_fp16, input_261_cast_fp16))[name = string("new_x_19_cast_fp16")]; + tensor var_1440_begin_0 = const()[name = string("op_1440_begin_0"), val = tensor([0, 0, 7])]; + tensor var_1440_end_0 = const()[name = string("op_1440_end_0"), val = tensor([1, 1024, 15])]; + tensor var_1440_end_mask_0 = const()[name = string("op_1440_end_mask_0"), val = tensor([true, true, true])]; + tensor var_1440_cast_fp16 = slice_by_index(begin = var_1440_begin_0, end = var_1440_end_0, end_mask = var_1440_end_mask_0, x = new_x_19_cast_fp16)[name = string("op_1440_cast_fp16")]; + string x_125_pad_type_0 = const()[name = string("x_125_pad_type_0"), val = string("valid")]; + int32 x_125_groups_0 = const()[name = string("x_125_groups_0"), val = int32(1024)]; + tensor x_125_strides_0 = const()[name = string("x_125_strides_0"), val = tensor([1])]; + tensor x_125_pad_0 = const()[name = string("x_125_pad_0"), val = tensor([0, 0])]; + tensor x_125_dilations_0 = const()[name = string("x_125_dilations_0"), val = tensor([1])]; + tensor encoder_layers_4_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105932800))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105942080))))[name = string("encoder_layers_4_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_125_cast_fp16 = conv(dilations = x_125_dilations_0, groups = x_125_groups_0, pad = x_125_pad_0, pad_type = x_125_pad_type_0, strides = x_125_strides_0, weight = encoder_layers_4_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_19_cast_fp16)[name = string("x_125_cast_fp16")]; + tensor input_263_perm_0 = const()[name = string("input_263_perm_0"), val = tensor([0, 2, 1])]; + tensor x_127_axes_0 = const()[name = string("x_127_axes_0"), val = tensor([-1])]; + tensor encoder_layers_4_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_4_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105944192)))]; + tensor encoder_layers_4_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_4_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105946304)))]; + tensor input_263_cast_fp16 = transpose(perm = input_263_perm_0, x = x_125_cast_fp16)[name = string("transpose_320")]; + tensor x_127_cast_fp16 = layer_norm(axes = x_127_axes_0, beta = encoder_layers_4_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_4_conv_batch_norm_weight_to_fp16, x = input_263_cast_fp16)[name = string("x_127_cast_fp16")]; + tensor input_265_perm_0 = const()[name = string("input_265_perm_0"), val = tensor([0, 2, 1])]; + tensor input_265_cast_fp16 = transpose(perm = input_265_perm_0, x = x_127_cast_fp16)[name = string("transpose_319")]; + tensor input_267_cast_fp16 = silu(x = input_265_cast_fp16)[name = string("input_267_cast_fp16")]; + string x_129_pad_type_0 = const()[name = string("x_129_pad_type_0"), val = string("valid")]; + tensor x_129_strides_0 = const()[name = string("x_129_strides_0"), val = tensor([1])]; + tensor x_129_pad_0 = const()[name = string("x_129_pad_0"), val = tensor([0, 0])]; + tensor x_129_dilations_0 = const()[name = string("x_129_dilations_0"), val = tensor([1])]; + int32 x_129_groups_0 = const()[name = string("x_129_groups_0"), val = int32(1)]; + tensor encoder_layers_4_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105948416))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106997056))))[name = string("encoder_layers_4_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_129_cast_fp16 = conv(dilations = x_129_dilations_0, groups = x_129_groups_0, pad = x_129_pad_0, pad_type = x_129_pad_type_0, strides = x_129_strides_0, weight = encoder_layers_4_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_267_cast_fp16)[name = string("x_129_cast_fp16")]; + tensor input_269_perm_0 = const()[name = string("input_269_perm_0"), val = tensor([0, 2, 1])]; + tensor input_269_cast_fp16 = transpose(perm = input_269_perm_0, x = x_129_cast_fp16)[name = string("transpose_318")]; + tensor input_271_cast_fp16 = add(x = input_255_cast_fp16, y = input_269_cast_fp16)[name = string("input_271_cast_fp16")]; + tensor input_273_axes_0 = const()[name = string("input_273_axes_0"), val = tensor([-1])]; + tensor encoder_layers_4_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_4_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106999168)))]; + tensor encoder_layers_4_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_4_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107001280)))]; + tensor input_273_cast_fp16 = layer_norm(axes = input_273_axes_0, beta = encoder_layers_4_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_4_norm_feed_forward2_weight_to_fp16, x = input_271_cast_fp16)[name = string("input_273_cast_fp16")]; + tensor encoder_layers_4_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107003392))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110149184))))[name = string("encoder_layers_4_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_4_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_4_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110149376)))]; + tensor linear_44_cast_fp16 = linear(bias = encoder_layers_4_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_4_feed_forward2_linear1_weight_to_fp16_palettized, x = input_273_cast_fp16)[name = string("linear_44_cast_fp16")]; + tensor input_277_cast_fp16 = silu(x = linear_44_cast_fp16)[name = string("input_277_cast_fp16")]; + tensor encoder_layers_4_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110157632))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113303424))))[name = string("encoder_layers_4_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_4_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_4_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113303616)))]; + tensor linear_45_cast_fp16 = linear(bias = encoder_layers_4_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_4_feed_forward2_linear2_weight_to_fp16_palettized, x = input_277_cast_fp16)[name = string("linear_45_cast_fp16")]; + fp16 var_1483_to_fp16 = const()[name = string("op_1483_to_fp16"), val = fp16(0x1p-1)]; + tensor var_1484_cast_fp16 = mul(x = linear_45_cast_fp16, y = var_1483_to_fp16)[name = string("op_1484_cast_fp16")]; + tensor input_283_cast_fp16 = add(x = input_271_cast_fp16, y = var_1484_cast_fp16)[name = string("input_283_cast_fp16")]; + tensor input_285_axes_0 = const()[name = string("input_285_axes_0"), val = tensor([-1])]; + tensor encoder_layers_4_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_4_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113305728)))]; + tensor encoder_layers_4_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_4_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113307840)))]; + tensor input_285_cast_fp16 = layer_norm(axes = input_285_axes_0, beta = encoder_layers_4_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_4_norm_out_weight_to_fp16, x = input_283_cast_fp16)[name = string("input_285_cast_fp16")]; + tensor cache_21_begin_0 = const()[name = string("cache_21_begin_0"), val = tensor([5, 0, 0, 0])]; + tensor cache_21_end_0 = const()[name = string("cache_21_end_0"), val = tensor([6, 1, 42, 1024])]; + tensor cache_21_end_mask_0 = const()[name = string("cache_21_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_21_squeeze_mask_0 = const()[name = string("cache_21_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_21_cast_fp16 = slice_by_index(begin = cache_21_begin_0, end = cache_21_end_0, end_mask = cache_21_end_mask_0, squeeze_mask = cache_21_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_21_cast_fp16")]; + tensor cache_23_begin_0 = const()[name = string("cache_23_begin_0"), val = tensor([5, 0, 0, 0])]; + tensor cache_23_end_0 = const()[name = string("cache_23_end_0"), val = tensor([6, 1, 1024, 8])]; + tensor cache_23_end_mask_0 = const()[name = string("cache_23_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_23_squeeze_mask_0 = const()[name = string("cache_23_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_23_cast_fp16 = slice_by_index(begin = cache_23_begin_0, end = cache_23_end_0, end_mask = cache_23_end_mask_0, squeeze_mask = cache_23_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_23_cast_fp16")]; + tensor input_287_axes_0 = const()[name = string("input_287_axes_0"), val = tensor([-1])]; + tensor encoder_layers_5_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_5_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113309952)))]; + tensor encoder_layers_5_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_5_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113312064)))]; + tensor input_287_cast_fp16 = layer_norm(axes = input_287_axes_0, beta = encoder_layers_5_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_5_norm_feed_forward1_weight_to_fp16, x = input_285_cast_fp16)[name = string("input_287_cast_fp16")]; + tensor encoder_layers_5_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113314176))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116459968))))[name = string("encoder_layers_5_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_5_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_5_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116460160)))]; + tensor linear_46_cast_fp16 = linear(bias = encoder_layers_5_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_5_feed_forward1_linear1_weight_to_fp16_palettized, x = input_287_cast_fp16)[name = string("linear_46_cast_fp16")]; + tensor input_291_cast_fp16 = silu(x = linear_46_cast_fp16)[name = string("input_291_cast_fp16")]; + tensor encoder_layers_5_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116468416))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119614208))))[name = string("encoder_layers_5_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_5_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_5_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119614400)))]; + tensor linear_47_cast_fp16 = linear(bias = encoder_layers_5_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_5_feed_forward1_linear2_weight_to_fp16_palettized, x = input_291_cast_fp16)[name = string("linear_47_cast_fp16")]; + fp16 var_1520_to_fp16 = const()[name = string("op_1520_to_fp16"), val = fp16(0x1p-1)]; + tensor var_1521_cast_fp16 = mul(x = linear_47_cast_fp16, y = var_1520_to_fp16)[name = string("op_1521_cast_fp16")]; + tensor input_297_cast_fp16 = add(x = input_285_cast_fp16, y = var_1521_cast_fp16)[name = string("input_297_cast_fp16")]; + tensor key_11_axes_0 = const()[name = string("key_11_axes_0"), val = tensor([-1])]; + tensor encoder_layers_5_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_5_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119616512)))]; + tensor encoder_layers_5_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_5_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119618624)))]; + tensor key_11_cast_fp16 = layer_norm(axes = key_11_axes_0, beta = encoder_layers_5_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_5_norm_self_att_weight_to_fp16, x = input_297_cast_fp16)[name = string("key_11_cast_fp16")]; + bool input_299_interleave_0 = const()[name = string("input_299_interleave_0"), val = bool(false)]; + tensor input_299_cast_fp16 = concat(axis = var_68, interleave = input_299_interleave_0, values = (cache_21_cast_fp16, key_11_cast_fp16))[name = string("input_299_cast_fp16")]; + tensor var_1543_begin_0 = const()[name = string("op_1543_begin_0"), val = tensor([0, 7, 0])]; + tensor var_1543_end_0 = const()[name = string("op_1543_end_0"), val = tensor([1, 42, 1024])]; + tensor var_1543_end_mask_0 = const()[name = string("op_1543_end_mask_0"), val = tensor([true, true, true])]; + tensor var_1543_cast_fp16 = slice_by_index(begin = var_1543_begin_0, end = var_1543_end_0, end_mask = var_1543_end_mask_0, x = cache_21_cast_fp16)[name = string("op_1543_cast_fp16")]; + bool var_1549_interleave_0 = const()[name = string("op_1549_interleave_0"), val = bool(false)]; + tensor var_1549_cast_fp16 = concat(axis = var_68, interleave = var_1549_interleave_0, values = (var_1543_cast_fp16, key_11_cast_fp16))[name = string("op_1549_cast_fp16")]; + tensor encoder_layers_5_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119620736))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120407232))))[name = string("encoder_layers_5_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_5_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_5_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120407424)))]; + tensor linear_48_cast_fp16 = linear(bias = encoder_layers_5_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_5_self_attn_linear_q_weight_to_fp16_palettized, x = key_11_cast_fp16)[name = string("linear_48_cast_fp16")]; + tensor var_1554 = const()[name = string("op_1554"), val = tensor([1, -1, 8, 128])]; + tensor q_31_cast_fp16 = reshape(shape = var_1554, x = linear_48_cast_fp16)[name = string("q_31_cast_fp16")]; + tensor encoder_layers_5_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120409536))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121196032))))[name = string("encoder_layers_5_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_5_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_5_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121196224)))]; + tensor linear_49_cast_fp16 = linear(bias = encoder_layers_5_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_5_self_attn_linear_k_weight_to_fp16_palettized, x = input_299_cast_fp16)[name = string("linear_49_cast_fp16")]; + tensor var_1559 = const()[name = string("op_1559"), val = tensor([1, -1, 8, 128])]; + tensor k_21_cast_fp16 = reshape(shape = var_1559, x = linear_49_cast_fp16)[name = string("k_21_cast_fp16")]; + tensor encoder_layers_5_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121198336))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121984832))))[name = string("encoder_layers_5_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_5_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_5_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121985024)))]; + tensor linear_50_cast_fp16 = linear(bias = encoder_layers_5_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_5_self_attn_linear_v_weight_to_fp16_palettized, x = input_299_cast_fp16)[name = string("linear_50_cast_fp16")]; + tensor var_1564 = const()[name = string("op_1564"), val = tensor([1, -1, 8, 128])]; + tensor v_11_cast_fp16 = reshape(shape = var_1564, x = linear_50_cast_fp16)[name = string("v_11_cast_fp16")]; + tensor value_19_perm_0 = const()[name = string("value_19_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_5_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_5_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121987136)))]; + tensor var_1577_cast_fp16 = add(x = q_31_cast_fp16, y = encoder_layers_5_self_attn_pos_bias_u_to_fp16)[name = string("op_1577_cast_fp16")]; + tensor encoder_layers_5_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_5_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121989248)))]; + tensor var_1579_cast_fp16 = add(x = q_31_cast_fp16, y = encoder_layers_5_self_attn_pos_bias_v_to_fp16)[name = string("op_1579_cast_fp16")]; + tensor q_with_bias_v_11_perm_0 = const()[name = string("q_with_bias_v_11_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_137_transpose_x_0 = const()[name = string("x_137_transpose_x_0"), val = bool(false)]; + bool x_137_transpose_y_0 = const()[name = string("x_137_transpose_y_0"), val = bool(false)]; + tensor op_1581_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121991360))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122090752))))[name = string("op_1581_to_fp16_quantized")]; + tensor q_with_bias_v_11_cast_fp16 = transpose(perm = q_with_bias_v_11_perm_0, x = var_1579_cast_fp16)[name = string("transpose_317")]; + tensor x_137_cast_fp16 = matmul(transpose_x = x_137_transpose_x_0, transpose_y = x_137_transpose_y_0, x = q_with_bias_v_11_cast_fp16, y = op_1581_to_fp16_quantized)[name = string("x_137_cast_fp16")]; + tensor x_139_pad_0 = const()[name = string("x_139_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_139_mode_0 = const()[name = string("x_139_mode_0"), val = string("constant")]; + fp16 const_144_to_fp16 = const()[name = string("const_144_to_fp16"), val = fp16(0x0p+0)]; + tensor x_139_cast_fp16 = pad(constant_val = const_144_to_fp16, mode = x_139_mode_0, pad = x_139_pad_0, x = x_137_cast_fp16)[name = string("x_139_cast_fp16")]; + tensor var_1589 = const()[name = string("op_1589"), val = tensor([1, 8, -1, 7])]; + tensor x_141_cast_fp16 = reshape(shape = var_1589, x = x_139_cast_fp16)[name = string("x_141_cast_fp16")]; + tensor var_1593_begin_0 = const()[name = string("op_1593_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_1593_end_0 = const()[name = string("op_1593_end_0"), val = tensor([1, 8, 98, 7])]; + tensor var_1593_end_mask_0 = const()[name = string("op_1593_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_1593_cast_fp16 = slice_by_index(begin = var_1593_begin_0, end = var_1593_end_0, end_mask = var_1593_end_mask_0, x = x_141_cast_fp16)[name = string("op_1593_cast_fp16")]; + tensor var_1594 = const()[name = string("op_1594"), val = tensor([1, 8, 7, 97])]; + tensor matrix_bd_21_cast_fp16 = reshape(shape = var_1594, x = var_1593_cast_fp16)[name = string("matrix_bd_21_cast_fp16")]; + bool matrix_ac_11_transpose_x_0 = const()[name = string("matrix_ac_11_transpose_x_0"), val = bool(false)]; + bool matrix_ac_11_transpose_y_0 = const()[name = string("matrix_ac_11_transpose_y_0"), val = bool(false)]; + tensor transpose_106_perm_0 = const()[name = string("transpose_106_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_107_perm_0 = const()[name = string("transpose_107_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_107 = transpose(perm = transpose_107_perm_0, x = k_21_cast_fp16)[name = string("transpose_315")]; + tensor transpose_106 = transpose(perm = transpose_106_perm_0, x = var_1577_cast_fp16)[name = string("transpose_316")]; + tensor matrix_ac_11_cast_fp16 = matmul(transpose_x = matrix_ac_11_transpose_x_0, transpose_y = matrix_ac_11_transpose_y_0, x = transpose_106, y = transpose_107)[name = string("matrix_ac_11_cast_fp16")]; + tensor matrix_bd_23_begin_0 = const()[name = string("matrix_bd_23_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_23_end_0 = const()[name = string("matrix_bd_23_end_0"), val = tensor([1, 8, 7, 49])]; + tensor matrix_bd_23_end_mask_0 = const()[name = string("matrix_bd_23_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_23_cast_fp16 = slice_by_index(begin = matrix_bd_23_begin_0, end = matrix_bd_23_end_0, end_mask = matrix_bd_23_end_mask_0, x = matrix_bd_21_cast_fp16)[name = string("matrix_bd_23_cast_fp16")]; + tensor var_1603_cast_fp16 = add(x = matrix_ac_11_cast_fp16, y = matrix_bd_23_cast_fp16)[name = string("op_1603_cast_fp16")]; + fp16 _inversed_scores_21_y_0_to_fp16 = const()[name = string("_inversed_scores_21_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_21_cast_fp16 = mul(x = var_1603_cast_fp16, y = _inversed_scores_21_y_0_to_fp16)[name = string("_inversed_scores_21_cast_fp16")]; + tensor scores_23_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_21_cast_fp16, cond = mask_11)[name = string("scores_23_cast_fp16")]; + tensor var_1609_cast_fp16 = softmax(axis = var_59, x = scores_23_cast_fp16)[name = string("op_1609_cast_fp16")]; + tensor input_301_cast_fp16 = select(a = var_44_to_fp16, b = var_1609_cast_fp16, cond = mask_11)[name = string("input_301_cast_fp16")]; + bool x_143_transpose_x_0 = const()[name = string("x_143_transpose_x_0"), val = bool(false)]; + bool x_143_transpose_y_0 = const()[name = string("x_143_transpose_y_0"), val = bool(false)]; + tensor value_19_cast_fp16 = transpose(perm = value_19_perm_0, x = v_11_cast_fp16)[name = string("transpose_314")]; + tensor x_143_cast_fp16 = matmul(transpose_x = x_143_transpose_x_0, transpose_y = x_143_transpose_y_0, x = input_301_cast_fp16, y = value_19_cast_fp16)[name = string("x_143_cast_fp16")]; + tensor var_1613_perm_0 = const()[name = string("op_1613_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1614 = const()[name = string("op_1614"), val = tensor([1, -1, 1024])]; + tensor var_1613_cast_fp16 = transpose(perm = var_1613_perm_0, x = x_143_cast_fp16)[name = string("transpose_313")]; + tensor input_303_cast_fp16 = reshape(shape = var_1614, x = var_1613_cast_fp16)[name = string("input_303_cast_fp16")]; + tensor encoder_layers_5_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122091072))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122877568))))[name = string("encoder_layers_5_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_5_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_5_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122877760)))]; + tensor linear_52_cast_fp16 = linear(bias = encoder_layers_5_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_5_self_attn_linear_out_weight_to_fp16_palettized, x = input_303_cast_fp16)[name = string("linear_52_cast_fp16")]; + tensor input_307_cast_fp16 = add(x = input_297_cast_fp16, y = linear_52_cast_fp16)[name = string("input_307_cast_fp16")]; + tensor x_147_axes_0 = const()[name = string("x_147_axes_0"), val = tensor([-1])]; + tensor encoder_layers_5_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_5_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122879872)))]; + tensor encoder_layers_5_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_5_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122881984)))]; + tensor x_147_cast_fp16 = layer_norm(axes = x_147_axes_0, beta = encoder_layers_5_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_5_norm_conv_weight_to_fp16, x = input_307_cast_fp16)[name = string("x_147_cast_fp16")]; + tensor input_309_perm_0 = const()[name = string("input_309_perm_0"), val = tensor([0, 2, 1])]; + string input_311_pad_type_0 = const()[name = string("input_311_pad_type_0"), val = string("valid")]; + tensor input_311_strides_0 = const()[name = string("input_311_strides_0"), val = tensor([1])]; + tensor input_311_pad_0 = const()[name = string("input_311_pad_0"), val = tensor([0, 0])]; + tensor input_311_dilations_0 = const()[name = string("input_311_dilations_0"), val = tensor([1])]; + int32 input_311_groups_0 = const()[name = string("input_311_groups_0"), val = int32(1)]; + tensor encoder_layers_5_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122884096))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(124981312))))[name = string("encoder_layers_5_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_309_cast_fp16 = transpose(perm = input_309_perm_0, x = x_147_cast_fp16)[name = string("transpose_312")]; + tensor input_311_cast_fp16 = conv(dilations = input_311_dilations_0, groups = input_311_groups_0, pad = input_311_pad_0, pad_type = input_311_pad_type_0, strides = input_311_strides_0, weight = encoder_layers_5_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_309_cast_fp16)[name = string("input_311_cast_fp16")]; + int32 x_149_split_num_splits_0 = const()[name = string("x_149_split_num_splits_0"), val = int32(2)]; + int32 x_149_split_axis_0 = const()[name = string("x_149_split_axis_0"), val = int32(1)]; + tensor x_149_split_cast_fp16_0, tensor x_149_split_cast_fp16_1 = split(axis = x_149_split_axis_0, num_splits = x_149_split_num_splits_0, x = input_311_cast_fp16)[name = string("x_149_split_cast_fp16")]; + tensor x_149_split_1_sigmoid_cast_fp16 = sigmoid(x = x_149_split_cast_fp16_1)[name = string("x_149_split_1_sigmoid_cast_fp16")]; + tensor x_149_cast_fp16 = mul(x = x_149_split_cast_fp16_0, y = x_149_split_1_sigmoid_cast_fp16)[name = string("x_149_cast_fp16")]; + tensor input_313_cast_fp16 = select(a = var_44_to_fp16, b = x_149_cast_fp16, cond = var_575)[name = string("input_313_cast_fp16")]; + bool new_x_23_interleave_0 = const()[name = string("new_x_23_interleave_0"), val = bool(false)]; + tensor new_x_23_cast_fp16 = concat(axis = var_59, interleave = new_x_23_interleave_0, values = (cache_23_cast_fp16, input_313_cast_fp16))[name = string("new_x_23_cast_fp16")]; + tensor var_1653_begin_0 = const()[name = string("op_1653_begin_0"), val = tensor([0, 0, 7])]; + tensor var_1653_end_0 = const()[name = string("op_1653_end_0"), val = tensor([1, 1024, 15])]; + tensor var_1653_end_mask_0 = const()[name = string("op_1653_end_mask_0"), val = tensor([true, true, true])]; + tensor var_1653_cast_fp16 = slice_by_index(begin = var_1653_begin_0, end = var_1653_end_0, end_mask = var_1653_end_mask_0, x = new_x_23_cast_fp16)[name = string("op_1653_cast_fp16")]; + string x_151_pad_type_0 = const()[name = string("x_151_pad_type_0"), val = string("valid")]; + int32 x_151_groups_0 = const()[name = string("x_151_groups_0"), val = int32(1024)]; + tensor x_151_strides_0 = const()[name = string("x_151_strides_0"), val = tensor([1])]; + tensor x_151_pad_0 = const()[name = string("x_151_pad_0"), val = tensor([0, 0])]; + tensor x_151_dilations_0 = const()[name = string("x_151_dilations_0"), val = tensor([1])]; + tensor encoder_layers_5_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(124985472))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(124994752))))[name = string("encoder_layers_5_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_151_cast_fp16 = conv(dilations = x_151_dilations_0, groups = x_151_groups_0, pad = x_151_pad_0, pad_type = x_151_pad_type_0, strides = x_151_strides_0, weight = encoder_layers_5_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_23_cast_fp16)[name = string("x_151_cast_fp16")]; + tensor input_315_perm_0 = const()[name = string("input_315_perm_0"), val = tensor([0, 2, 1])]; + tensor x_153_axes_0 = const()[name = string("x_153_axes_0"), val = tensor([-1])]; + tensor encoder_layers_5_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_5_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(124996864)))]; + tensor encoder_layers_5_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_5_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(124998976)))]; + tensor input_315_cast_fp16 = transpose(perm = input_315_perm_0, x = x_151_cast_fp16)[name = string("transpose_311")]; + tensor x_153_cast_fp16 = layer_norm(axes = x_153_axes_0, beta = encoder_layers_5_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_5_conv_batch_norm_weight_to_fp16, x = input_315_cast_fp16)[name = string("x_153_cast_fp16")]; + tensor input_317_perm_0 = const()[name = string("input_317_perm_0"), val = tensor([0, 2, 1])]; + tensor input_317_cast_fp16 = transpose(perm = input_317_perm_0, x = x_153_cast_fp16)[name = string("transpose_310")]; + tensor input_319_cast_fp16 = silu(x = input_317_cast_fp16)[name = string("input_319_cast_fp16")]; + string x_155_pad_type_0 = const()[name = string("x_155_pad_type_0"), val = string("valid")]; + tensor x_155_strides_0 = const()[name = string("x_155_strides_0"), val = tensor([1])]; + tensor x_155_pad_0 = const()[name = string("x_155_pad_0"), val = tensor([0, 0])]; + tensor x_155_dilations_0 = const()[name = string("x_155_dilations_0"), val = tensor([1])]; + int32 x_155_groups_0 = const()[name = string("x_155_groups_0"), val = int32(1)]; + tensor encoder_layers_5_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125001088))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126049728))))[name = string("encoder_layers_5_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_155_cast_fp16 = conv(dilations = x_155_dilations_0, groups = x_155_groups_0, pad = x_155_pad_0, pad_type = x_155_pad_type_0, strides = x_155_strides_0, weight = encoder_layers_5_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_319_cast_fp16)[name = string("x_155_cast_fp16")]; + tensor input_321_perm_0 = const()[name = string("input_321_perm_0"), val = tensor([0, 2, 1])]; + tensor input_321_cast_fp16 = transpose(perm = input_321_perm_0, x = x_155_cast_fp16)[name = string("transpose_309")]; + tensor input_323_cast_fp16 = add(x = input_307_cast_fp16, y = input_321_cast_fp16)[name = string("input_323_cast_fp16")]; + tensor input_325_axes_0 = const()[name = string("input_325_axes_0"), val = tensor([-1])]; + tensor encoder_layers_5_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_5_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126051840)))]; + tensor encoder_layers_5_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_5_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126053952)))]; + tensor input_325_cast_fp16 = layer_norm(axes = input_325_axes_0, beta = encoder_layers_5_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_5_norm_feed_forward2_weight_to_fp16, x = input_323_cast_fp16)[name = string("input_325_cast_fp16")]; + tensor encoder_layers_5_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126056064))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(129201856))))[name = string("encoder_layers_5_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_5_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_5_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(129202048)))]; + tensor linear_53_cast_fp16 = linear(bias = encoder_layers_5_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_5_feed_forward2_linear1_weight_to_fp16_palettized, x = input_325_cast_fp16)[name = string("linear_53_cast_fp16")]; + tensor input_329_cast_fp16 = silu(x = linear_53_cast_fp16)[name = string("input_329_cast_fp16")]; + tensor encoder_layers_5_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(129210304))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132356096))))[name = string("encoder_layers_5_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_5_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_5_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132356288)))]; + tensor linear_54_cast_fp16 = linear(bias = encoder_layers_5_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_5_feed_forward2_linear2_weight_to_fp16_palettized, x = input_329_cast_fp16)[name = string("linear_54_cast_fp16")]; + fp16 var_1696_to_fp16 = const()[name = string("op_1696_to_fp16"), val = fp16(0x1p-1)]; + tensor var_1697_cast_fp16 = mul(x = linear_54_cast_fp16, y = var_1696_to_fp16)[name = string("op_1697_cast_fp16")]; + tensor input_335_cast_fp16 = add(x = input_323_cast_fp16, y = var_1697_cast_fp16)[name = string("input_335_cast_fp16")]; + tensor input_337_axes_0 = const()[name = string("input_337_axes_0"), val = tensor([-1])]; + tensor encoder_layers_5_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_5_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132358400)))]; + tensor encoder_layers_5_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_5_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132360512)))]; + tensor input_337_cast_fp16 = layer_norm(axes = input_337_axes_0, beta = encoder_layers_5_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_5_norm_out_weight_to_fp16, x = input_335_cast_fp16)[name = string("input_337_cast_fp16")]; + tensor cache_25_begin_0 = const()[name = string("cache_25_begin_0"), val = tensor([6, 0, 0, 0])]; + tensor cache_25_end_0 = const()[name = string("cache_25_end_0"), val = tensor([7, 1, 42, 1024])]; + tensor cache_25_end_mask_0 = const()[name = string("cache_25_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_25_squeeze_mask_0 = const()[name = string("cache_25_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_25_cast_fp16 = slice_by_index(begin = cache_25_begin_0, end = cache_25_end_0, end_mask = cache_25_end_mask_0, squeeze_mask = cache_25_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_25_cast_fp16")]; + tensor cache_27_begin_0 = const()[name = string("cache_27_begin_0"), val = tensor([6, 0, 0, 0])]; + tensor cache_27_end_0 = const()[name = string("cache_27_end_0"), val = tensor([7, 1, 1024, 8])]; + tensor cache_27_end_mask_0 = const()[name = string("cache_27_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_27_squeeze_mask_0 = const()[name = string("cache_27_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_27_cast_fp16 = slice_by_index(begin = cache_27_begin_0, end = cache_27_end_0, end_mask = cache_27_end_mask_0, squeeze_mask = cache_27_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_27_cast_fp16")]; + tensor input_339_axes_0 = const()[name = string("input_339_axes_0"), val = tensor([-1])]; + tensor encoder_layers_6_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_6_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132362624)))]; + tensor encoder_layers_6_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_6_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132364736)))]; + tensor input_339_cast_fp16 = layer_norm(axes = input_339_axes_0, beta = encoder_layers_6_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_6_norm_feed_forward1_weight_to_fp16, x = input_337_cast_fp16)[name = string("input_339_cast_fp16")]; + tensor encoder_layers_6_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132366848))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135512640))))[name = string("encoder_layers_6_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_6_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_6_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135512832)))]; + tensor linear_55_cast_fp16 = linear(bias = encoder_layers_6_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_6_feed_forward1_linear1_weight_to_fp16_palettized, x = input_339_cast_fp16)[name = string("linear_55_cast_fp16")]; + tensor input_343_cast_fp16 = silu(x = linear_55_cast_fp16)[name = string("input_343_cast_fp16")]; + tensor encoder_layers_6_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135521088))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138666880))))[name = string("encoder_layers_6_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_6_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_6_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138667072)))]; + tensor linear_56_cast_fp16 = linear(bias = encoder_layers_6_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_6_feed_forward1_linear2_weight_to_fp16_palettized, x = input_343_cast_fp16)[name = string("linear_56_cast_fp16")]; + fp16 var_1733_to_fp16 = const()[name = string("op_1733_to_fp16"), val = fp16(0x1p-1)]; + tensor var_1734_cast_fp16 = mul(x = linear_56_cast_fp16, y = var_1733_to_fp16)[name = string("op_1734_cast_fp16")]; + tensor input_349_cast_fp16 = add(x = input_337_cast_fp16, y = var_1734_cast_fp16)[name = string("input_349_cast_fp16")]; + tensor key_13_axes_0 = const()[name = string("key_13_axes_0"), val = tensor([-1])]; + tensor encoder_layers_6_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_6_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138669184)))]; + tensor encoder_layers_6_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_6_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138671296)))]; + tensor key_13_cast_fp16 = layer_norm(axes = key_13_axes_0, beta = encoder_layers_6_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_6_norm_self_att_weight_to_fp16, x = input_349_cast_fp16)[name = string("key_13_cast_fp16")]; + bool input_351_interleave_0 = const()[name = string("input_351_interleave_0"), val = bool(false)]; + tensor input_351_cast_fp16 = concat(axis = var_68, interleave = input_351_interleave_0, values = (cache_25_cast_fp16, key_13_cast_fp16))[name = string("input_351_cast_fp16")]; + tensor var_1756_begin_0 = const()[name = string("op_1756_begin_0"), val = tensor([0, 7, 0])]; + tensor var_1756_end_0 = const()[name = string("op_1756_end_0"), val = tensor([1, 42, 1024])]; + tensor var_1756_end_mask_0 = const()[name = string("op_1756_end_mask_0"), val = tensor([true, true, true])]; + tensor var_1756_cast_fp16 = slice_by_index(begin = var_1756_begin_0, end = var_1756_end_0, end_mask = var_1756_end_mask_0, x = cache_25_cast_fp16)[name = string("op_1756_cast_fp16")]; + bool var_1762_interleave_0 = const()[name = string("op_1762_interleave_0"), val = bool(false)]; + tensor var_1762_cast_fp16 = concat(axis = var_68, interleave = var_1762_interleave_0, values = (var_1756_cast_fp16, key_13_cast_fp16))[name = string("op_1762_cast_fp16")]; + tensor encoder_layers_6_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138673408))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(139459904))))[name = string("encoder_layers_6_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_6_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_6_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(139460096)))]; + tensor linear_57_cast_fp16 = linear(bias = encoder_layers_6_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_6_self_attn_linear_q_weight_to_fp16_palettized, x = key_13_cast_fp16)[name = string("linear_57_cast_fp16")]; + tensor var_1767 = const()[name = string("op_1767"), val = tensor([1, -1, 8, 128])]; + tensor q_37_cast_fp16 = reshape(shape = var_1767, x = linear_57_cast_fp16)[name = string("q_37_cast_fp16")]; + tensor encoder_layers_6_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(139462208))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140248704))))[name = string("encoder_layers_6_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_6_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_6_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140248896)))]; + tensor linear_58_cast_fp16 = linear(bias = encoder_layers_6_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_6_self_attn_linear_k_weight_to_fp16_palettized, x = input_351_cast_fp16)[name = string("linear_58_cast_fp16")]; + tensor var_1772 = const()[name = string("op_1772"), val = tensor([1, -1, 8, 128])]; + tensor k_25_cast_fp16 = reshape(shape = var_1772, x = linear_58_cast_fp16)[name = string("k_25_cast_fp16")]; + tensor encoder_layers_6_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140251008))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141037504))))[name = string("encoder_layers_6_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_6_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_6_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141037696)))]; + tensor linear_59_cast_fp16 = linear(bias = encoder_layers_6_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_6_self_attn_linear_v_weight_to_fp16_palettized, x = input_351_cast_fp16)[name = string("linear_59_cast_fp16")]; + tensor var_1777 = const()[name = string("op_1777"), val = tensor([1, -1, 8, 128])]; + tensor v_13_cast_fp16 = reshape(shape = var_1777, x = linear_59_cast_fp16)[name = string("v_13_cast_fp16")]; + tensor value_21_perm_0 = const()[name = string("value_21_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_6_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_6_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141039808)))]; + tensor var_1790_cast_fp16 = add(x = q_37_cast_fp16, y = encoder_layers_6_self_attn_pos_bias_u_to_fp16)[name = string("op_1790_cast_fp16")]; + tensor encoder_layers_6_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_6_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141041920)))]; + tensor var_1792_cast_fp16 = add(x = q_37_cast_fp16, y = encoder_layers_6_self_attn_pos_bias_v_to_fp16)[name = string("op_1792_cast_fp16")]; + tensor q_with_bias_v_13_perm_0 = const()[name = string("q_with_bias_v_13_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_163_transpose_x_0 = const()[name = string("x_163_transpose_x_0"), val = bool(false)]; + bool x_163_transpose_y_0 = const()[name = string("x_163_transpose_y_0"), val = bool(false)]; + tensor op_1794_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141044032))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141143424))))[name = string("op_1794_to_fp16_quantized")]; + tensor q_with_bias_v_13_cast_fp16 = transpose(perm = q_with_bias_v_13_perm_0, x = var_1792_cast_fp16)[name = string("transpose_308")]; + tensor x_163_cast_fp16 = matmul(transpose_x = x_163_transpose_x_0, transpose_y = x_163_transpose_y_0, x = q_with_bias_v_13_cast_fp16, y = op_1794_to_fp16_quantized)[name = string("x_163_cast_fp16")]; + tensor x_165_pad_0 = const()[name = string("x_165_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_165_mode_0 = const()[name = string("x_165_mode_0"), val = string("constant")]; + fp16 const_157_to_fp16 = const()[name = string("const_157_to_fp16"), val = fp16(0x0p+0)]; + tensor x_165_cast_fp16 = pad(constant_val = const_157_to_fp16, mode = x_165_mode_0, pad = x_165_pad_0, x = x_163_cast_fp16)[name = string("x_165_cast_fp16")]; + tensor var_1802 = const()[name = string("op_1802"), val = tensor([1, 8, -1, 7])]; + tensor x_167_cast_fp16 = reshape(shape = var_1802, x = x_165_cast_fp16)[name = string("x_167_cast_fp16")]; + tensor var_1806_begin_0 = const()[name = string("op_1806_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_1806_end_0 = const()[name = string("op_1806_end_0"), val = tensor([1, 8, 98, 7])]; + tensor var_1806_end_mask_0 = const()[name = string("op_1806_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_1806_cast_fp16 = slice_by_index(begin = var_1806_begin_0, end = var_1806_end_0, end_mask = var_1806_end_mask_0, x = x_167_cast_fp16)[name = string("op_1806_cast_fp16")]; + tensor var_1807 = const()[name = string("op_1807"), val = tensor([1, 8, 7, 97])]; + tensor matrix_bd_25_cast_fp16 = reshape(shape = var_1807, x = var_1806_cast_fp16)[name = string("matrix_bd_25_cast_fp16")]; + bool matrix_ac_13_transpose_x_0 = const()[name = string("matrix_ac_13_transpose_x_0"), val = bool(false)]; + bool matrix_ac_13_transpose_y_0 = const()[name = string("matrix_ac_13_transpose_y_0"), val = bool(false)]; + tensor transpose_108_perm_0 = const()[name = string("transpose_108_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_109_perm_0 = const()[name = string("transpose_109_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_109 = transpose(perm = transpose_109_perm_0, x = k_25_cast_fp16)[name = string("transpose_306")]; + tensor transpose_108 = transpose(perm = transpose_108_perm_0, x = var_1790_cast_fp16)[name = string("transpose_307")]; + tensor matrix_ac_13_cast_fp16 = matmul(transpose_x = matrix_ac_13_transpose_x_0, transpose_y = matrix_ac_13_transpose_y_0, x = transpose_108, y = transpose_109)[name = string("matrix_ac_13_cast_fp16")]; + tensor matrix_bd_27_begin_0 = const()[name = string("matrix_bd_27_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_27_end_0 = const()[name = string("matrix_bd_27_end_0"), val = tensor([1, 8, 7, 49])]; + tensor matrix_bd_27_end_mask_0 = const()[name = string("matrix_bd_27_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_27_cast_fp16 = slice_by_index(begin = matrix_bd_27_begin_0, end = matrix_bd_27_end_0, end_mask = matrix_bd_27_end_mask_0, x = matrix_bd_25_cast_fp16)[name = string("matrix_bd_27_cast_fp16")]; + tensor var_1816_cast_fp16 = add(x = matrix_ac_13_cast_fp16, y = matrix_bd_27_cast_fp16)[name = string("op_1816_cast_fp16")]; + fp16 _inversed_scores_25_y_0_to_fp16 = const()[name = string("_inversed_scores_25_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_25_cast_fp16 = mul(x = var_1816_cast_fp16, y = _inversed_scores_25_y_0_to_fp16)[name = string("_inversed_scores_25_cast_fp16")]; + tensor scores_27_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_25_cast_fp16, cond = mask_11)[name = string("scores_27_cast_fp16")]; + tensor var_1822_cast_fp16 = softmax(axis = var_59, x = scores_27_cast_fp16)[name = string("op_1822_cast_fp16")]; + tensor input_353_cast_fp16 = select(a = var_44_to_fp16, b = var_1822_cast_fp16, cond = mask_11)[name = string("input_353_cast_fp16")]; + bool x_169_transpose_x_0 = const()[name = string("x_169_transpose_x_0"), val = bool(false)]; + bool x_169_transpose_y_0 = const()[name = string("x_169_transpose_y_0"), val = bool(false)]; + tensor value_21_cast_fp16 = transpose(perm = value_21_perm_0, x = v_13_cast_fp16)[name = string("transpose_305")]; + tensor x_169_cast_fp16 = matmul(transpose_x = x_169_transpose_x_0, transpose_y = x_169_transpose_y_0, x = input_353_cast_fp16, y = value_21_cast_fp16)[name = string("x_169_cast_fp16")]; + tensor var_1826_perm_0 = const()[name = string("op_1826_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1827 = const()[name = string("op_1827"), val = tensor([1, -1, 1024])]; + tensor var_1826_cast_fp16 = transpose(perm = var_1826_perm_0, x = x_169_cast_fp16)[name = string("transpose_304")]; + tensor input_355_cast_fp16 = reshape(shape = var_1827, x = var_1826_cast_fp16)[name = string("input_355_cast_fp16")]; + tensor encoder_layers_6_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141143744))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141930240))))[name = string("encoder_layers_6_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_6_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_6_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141930432)))]; + tensor linear_61_cast_fp16 = linear(bias = encoder_layers_6_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_6_self_attn_linear_out_weight_to_fp16_palettized, x = input_355_cast_fp16)[name = string("linear_61_cast_fp16")]; + tensor input_359_cast_fp16 = add(x = input_349_cast_fp16, y = linear_61_cast_fp16)[name = string("input_359_cast_fp16")]; + tensor x_173_axes_0 = const()[name = string("x_173_axes_0"), val = tensor([-1])]; + tensor encoder_layers_6_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_6_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141932544)))]; + tensor encoder_layers_6_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_6_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141934656)))]; + tensor x_173_cast_fp16 = layer_norm(axes = x_173_axes_0, beta = encoder_layers_6_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_6_norm_conv_weight_to_fp16, x = input_359_cast_fp16)[name = string("x_173_cast_fp16")]; + tensor input_361_perm_0 = const()[name = string("input_361_perm_0"), val = tensor([0, 2, 1])]; + string input_363_pad_type_0 = const()[name = string("input_363_pad_type_0"), val = string("valid")]; + tensor input_363_strides_0 = const()[name = string("input_363_strides_0"), val = tensor([1])]; + tensor input_363_pad_0 = const()[name = string("input_363_pad_0"), val = tensor([0, 0])]; + tensor input_363_dilations_0 = const()[name = string("input_363_dilations_0"), val = tensor([1])]; + int32 input_363_groups_0 = const()[name = string("input_363_groups_0"), val = int32(1)]; + tensor encoder_layers_6_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141936768))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144033984))))[name = string("encoder_layers_6_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_361_cast_fp16 = transpose(perm = input_361_perm_0, x = x_173_cast_fp16)[name = string("transpose_303")]; + tensor input_363_cast_fp16 = conv(dilations = input_363_dilations_0, groups = input_363_groups_0, pad = input_363_pad_0, pad_type = input_363_pad_type_0, strides = input_363_strides_0, weight = encoder_layers_6_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_361_cast_fp16)[name = string("input_363_cast_fp16")]; + int32 x_175_split_num_splits_0 = const()[name = string("x_175_split_num_splits_0"), val = int32(2)]; + int32 x_175_split_axis_0 = const()[name = string("x_175_split_axis_0"), val = int32(1)]; + tensor x_175_split_cast_fp16_0, tensor x_175_split_cast_fp16_1 = split(axis = x_175_split_axis_0, num_splits = x_175_split_num_splits_0, x = input_363_cast_fp16)[name = string("x_175_split_cast_fp16")]; + tensor x_175_split_1_sigmoid_cast_fp16 = sigmoid(x = x_175_split_cast_fp16_1)[name = string("x_175_split_1_sigmoid_cast_fp16")]; + tensor x_175_cast_fp16 = mul(x = x_175_split_cast_fp16_0, y = x_175_split_1_sigmoid_cast_fp16)[name = string("x_175_cast_fp16")]; + tensor input_365_cast_fp16 = select(a = var_44_to_fp16, b = x_175_cast_fp16, cond = var_575)[name = string("input_365_cast_fp16")]; + bool new_x_27_interleave_0 = const()[name = string("new_x_27_interleave_0"), val = bool(false)]; + tensor new_x_27_cast_fp16 = concat(axis = var_59, interleave = new_x_27_interleave_0, values = (cache_27_cast_fp16, input_365_cast_fp16))[name = string("new_x_27_cast_fp16")]; + tensor var_1866_begin_0 = const()[name = string("op_1866_begin_0"), val = tensor([0, 0, 7])]; + tensor var_1866_end_0 = const()[name = string("op_1866_end_0"), val = tensor([1, 1024, 15])]; + tensor var_1866_end_mask_0 = const()[name = string("op_1866_end_mask_0"), val = tensor([true, true, true])]; + tensor var_1866_cast_fp16 = slice_by_index(begin = var_1866_begin_0, end = var_1866_end_0, end_mask = var_1866_end_mask_0, x = new_x_27_cast_fp16)[name = string("op_1866_cast_fp16")]; + string x_177_pad_type_0 = const()[name = string("x_177_pad_type_0"), val = string("valid")]; + int32 x_177_groups_0 = const()[name = string("x_177_groups_0"), val = int32(1024)]; + tensor x_177_strides_0 = const()[name = string("x_177_strides_0"), val = tensor([1])]; + tensor x_177_pad_0 = const()[name = string("x_177_pad_0"), val = tensor([0, 0])]; + tensor x_177_dilations_0 = const()[name = string("x_177_dilations_0"), val = tensor([1])]; + tensor encoder_layers_6_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144038144))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144047424))))[name = string("encoder_layers_6_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_177_cast_fp16 = conv(dilations = x_177_dilations_0, groups = x_177_groups_0, pad = x_177_pad_0, pad_type = x_177_pad_type_0, strides = x_177_strides_0, weight = encoder_layers_6_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_27_cast_fp16)[name = string("x_177_cast_fp16")]; + tensor input_367_perm_0 = const()[name = string("input_367_perm_0"), val = tensor([0, 2, 1])]; + tensor x_179_axes_0 = const()[name = string("x_179_axes_0"), val = tensor([-1])]; + tensor encoder_layers_6_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_6_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144049536)))]; + tensor encoder_layers_6_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_6_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144051648)))]; + tensor input_367_cast_fp16 = transpose(perm = input_367_perm_0, x = x_177_cast_fp16)[name = string("transpose_302")]; + tensor x_179_cast_fp16 = layer_norm(axes = x_179_axes_0, beta = encoder_layers_6_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_6_conv_batch_norm_weight_to_fp16, x = input_367_cast_fp16)[name = string("x_179_cast_fp16")]; + tensor input_369_perm_0 = const()[name = string("input_369_perm_0"), val = tensor([0, 2, 1])]; + tensor input_369_cast_fp16 = transpose(perm = input_369_perm_0, x = x_179_cast_fp16)[name = string("transpose_301")]; + tensor input_371_cast_fp16 = silu(x = input_369_cast_fp16)[name = string("input_371_cast_fp16")]; + string x_181_pad_type_0 = const()[name = string("x_181_pad_type_0"), val = string("valid")]; + tensor x_181_strides_0 = const()[name = string("x_181_strides_0"), val = tensor([1])]; + tensor x_181_pad_0 = const()[name = string("x_181_pad_0"), val = tensor([0, 0])]; + tensor x_181_dilations_0 = const()[name = string("x_181_dilations_0"), val = tensor([1])]; + int32 x_181_groups_0 = const()[name = string("x_181_groups_0"), val = int32(1)]; + tensor encoder_layers_6_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144053760))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(145102400))))[name = string("encoder_layers_6_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_181_cast_fp16 = conv(dilations = x_181_dilations_0, groups = x_181_groups_0, pad = x_181_pad_0, pad_type = x_181_pad_type_0, strides = x_181_strides_0, weight = encoder_layers_6_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_371_cast_fp16)[name = string("x_181_cast_fp16")]; + tensor input_373_perm_0 = const()[name = string("input_373_perm_0"), val = tensor([0, 2, 1])]; + tensor input_373_cast_fp16 = transpose(perm = input_373_perm_0, x = x_181_cast_fp16)[name = string("transpose_300")]; + tensor input_375_cast_fp16 = add(x = input_359_cast_fp16, y = input_373_cast_fp16)[name = string("input_375_cast_fp16")]; + tensor input_377_axes_0 = const()[name = string("input_377_axes_0"), val = tensor([-1])]; + tensor encoder_layers_6_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_6_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(145104512)))]; + tensor encoder_layers_6_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_6_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(145106624)))]; + tensor input_377_cast_fp16 = layer_norm(axes = input_377_axes_0, beta = encoder_layers_6_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_6_norm_feed_forward2_weight_to_fp16, x = input_375_cast_fp16)[name = string("input_377_cast_fp16")]; + tensor encoder_layers_6_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(145108736))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(148254528))))[name = string("encoder_layers_6_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_6_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_6_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(148254720)))]; + tensor linear_62_cast_fp16 = linear(bias = encoder_layers_6_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_6_feed_forward2_linear1_weight_to_fp16_palettized, x = input_377_cast_fp16)[name = string("linear_62_cast_fp16")]; + tensor input_381_cast_fp16 = silu(x = linear_62_cast_fp16)[name = string("input_381_cast_fp16")]; + tensor encoder_layers_6_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(148262976))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151408768))))[name = string("encoder_layers_6_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_6_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_6_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151408960)))]; + tensor linear_63_cast_fp16 = linear(bias = encoder_layers_6_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_6_feed_forward2_linear2_weight_to_fp16_palettized, x = input_381_cast_fp16)[name = string("linear_63_cast_fp16")]; + fp16 var_1909_to_fp16 = const()[name = string("op_1909_to_fp16"), val = fp16(0x1p-1)]; + tensor var_1910_cast_fp16 = mul(x = linear_63_cast_fp16, y = var_1909_to_fp16)[name = string("op_1910_cast_fp16")]; + tensor input_387_cast_fp16 = add(x = input_375_cast_fp16, y = var_1910_cast_fp16)[name = string("input_387_cast_fp16")]; + tensor input_389_axes_0 = const()[name = string("input_389_axes_0"), val = tensor([-1])]; + tensor encoder_layers_6_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_6_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151411072)))]; + tensor encoder_layers_6_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_6_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151413184)))]; + tensor input_389_cast_fp16 = layer_norm(axes = input_389_axes_0, beta = encoder_layers_6_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_6_norm_out_weight_to_fp16, x = input_387_cast_fp16)[name = string("input_389_cast_fp16")]; + tensor cache_29_begin_0 = const()[name = string("cache_29_begin_0"), val = tensor([7, 0, 0, 0])]; + tensor cache_29_end_0 = const()[name = string("cache_29_end_0"), val = tensor([8, 1, 42, 1024])]; + tensor cache_29_end_mask_0 = const()[name = string("cache_29_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_29_squeeze_mask_0 = const()[name = string("cache_29_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_29_cast_fp16 = slice_by_index(begin = cache_29_begin_0, end = cache_29_end_0, end_mask = cache_29_end_mask_0, squeeze_mask = cache_29_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_29_cast_fp16")]; + tensor cache_31_begin_0 = const()[name = string("cache_31_begin_0"), val = tensor([7, 0, 0, 0])]; + tensor cache_31_end_0 = const()[name = string("cache_31_end_0"), val = tensor([8, 1, 1024, 8])]; + tensor cache_31_end_mask_0 = const()[name = string("cache_31_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_31_squeeze_mask_0 = const()[name = string("cache_31_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_31_cast_fp16 = slice_by_index(begin = cache_31_begin_0, end = cache_31_end_0, end_mask = cache_31_end_mask_0, squeeze_mask = cache_31_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_31_cast_fp16")]; + tensor input_391_axes_0 = const()[name = string("input_391_axes_0"), val = tensor([-1])]; + tensor encoder_layers_7_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_7_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151415296)))]; + tensor encoder_layers_7_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_7_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151417408)))]; + tensor input_391_cast_fp16 = layer_norm(axes = input_391_axes_0, beta = encoder_layers_7_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_7_norm_feed_forward1_weight_to_fp16, x = input_389_cast_fp16)[name = string("input_391_cast_fp16")]; + tensor encoder_layers_7_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151419520))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(154565312))))[name = string("encoder_layers_7_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_7_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_7_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(154565504)))]; + tensor linear_64_cast_fp16 = linear(bias = encoder_layers_7_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_7_feed_forward1_linear1_weight_to_fp16_palettized, x = input_391_cast_fp16)[name = string("linear_64_cast_fp16")]; + tensor input_395_cast_fp16 = silu(x = linear_64_cast_fp16)[name = string("input_395_cast_fp16")]; + tensor encoder_layers_7_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(154573760))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(157719552))))[name = string("encoder_layers_7_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_7_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_7_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(157719744)))]; + tensor linear_65_cast_fp16 = linear(bias = encoder_layers_7_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_7_feed_forward1_linear2_weight_to_fp16_palettized, x = input_395_cast_fp16)[name = string("linear_65_cast_fp16")]; + fp16 var_1946_to_fp16 = const()[name = string("op_1946_to_fp16"), val = fp16(0x1p-1)]; + tensor var_1947_cast_fp16 = mul(x = linear_65_cast_fp16, y = var_1946_to_fp16)[name = string("op_1947_cast_fp16")]; + tensor input_401_cast_fp16 = add(x = input_389_cast_fp16, y = var_1947_cast_fp16)[name = string("input_401_cast_fp16")]; + tensor key_15_axes_0 = const()[name = string("key_15_axes_0"), val = tensor([-1])]; + tensor encoder_layers_7_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_7_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(157721856)))]; + tensor encoder_layers_7_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_7_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(157723968)))]; + tensor key_15_cast_fp16 = layer_norm(axes = key_15_axes_0, beta = encoder_layers_7_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_7_norm_self_att_weight_to_fp16, x = input_401_cast_fp16)[name = string("key_15_cast_fp16")]; + bool input_403_interleave_0 = const()[name = string("input_403_interleave_0"), val = bool(false)]; + tensor input_403_cast_fp16 = concat(axis = var_68, interleave = input_403_interleave_0, values = (cache_29_cast_fp16, key_15_cast_fp16))[name = string("input_403_cast_fp16")]; + tensor var_1969_begin_0 = const()[name = string("op_1969_begin_0"), val = tensor([0, 7, 0])]; + tensor var_1969_end_0 = const()[name = string("op_1969_end_0"), val = tensor([1, 42, 1024])]; + tensor var_1969_end_mask_0 = const()[name = string("op_1969_end_mask_0"), val = tensor([true, true, true])]; + tensor var_1969_cast_fp16 = slice_by_index(begin = var_1969_begin_0, end = var_1969_end_0, end_mask = var_1969_end_mask_0, x = cache_29_cast_fp16)[name = string("op_1969_cast_fp16")]; + bool var_1975_interleave_0 = const()[name = string("op_1975_interleave_0"), val = bool(false)]; + tensor var_1975_cast_fp16 = concat(axis = var_68, interleave = var_1975_interleave_0, values = (var_1969_cast_fp16, key_15_cast_fp16))[name = string("op_1975_cast_fp16")]; + tensor encoder_layers_7_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(157726080))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158512576))))[name = string("encoder_layers_7_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_7_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_7_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158512768)))]; + tensor linear_66_cast_fp16 = linear(bias = encoder_layers_7_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_7_self_attn_linear_q_weight_to_fp16_palettized, x = key_15_cast_fp16)[name = string("linear_66_cast_fp16")]; + tensor var_1980 = const()[name = string("op_1980"), val = tensor([1, -1, 8, 128])]; + tensor q_43_cast_fp16 = reshape(shape = var_1980, x = linear_66_cast_fp16)[name = string("q_43_cast_fp16")]; + tensor encoder_layers_7_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158514880))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(159301376))))[name = string("encoder_layers_7_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_7_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_7_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(159301568)))]; + tensor linear_67_cast_fp16 = linear(bias = encoder_layers_7_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_7_self_attn_linear_k_weight_to_fp16_palettized, x = input_403_cast_fp16)[name = string("linear_67_cast_fp16")]; + tensor var_1985 = const()[name = string("op_1985"), val = tensor([1, -1, 8, 128])]; + tensor k_29_cast_fp16 = reshape(shape = var_1985, x = linear_67_cast_fp16)[name = string("k_29_cast_fp16")]; + tensor encoder_layers_7_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(159303680))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160090176))))[name = string("encoder_layers_7_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_7_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_7_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160090368)))]; + tensor linear_68_cast_fp16 = linear(bias = encoder_layers_7_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_7_self_attn_linear_v_weight_to_fp16_palettized, x = input_403_cast_fp16)[name = string("linear_68_cast_fp16")]; + tensor var_1990 = const()[name = string("op_1990"), val = tensor([1, -1, 8, 128])]; + tensor v_15_cast_fp16 = reshape(shape = var_1990, x = linear_68_cast_fp16)[name = string("v_15_cast_fp16")]; + tensor value_23_perm_0 = const()[name = string("value_23_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_7_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_7_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160092480)))]; + tensor var_2003_cast_fp16 = add(x = q_43_cast_fp16, y = encoder_layers_7_self_attn_pos_bias_u_to_fp16)[name = string("op_2003_cast_fp16")]; + tensor encoder_layers_7_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_7_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160094592)))]; + tensor var_2005_cast_fp16 = add(x = q_43_cast_fp16, y = encoder_layers_7_self_attn_pos_bias_v_to_fp16)[name = string("op_2005_cast_fp16")]; + tensor q_with_bias_v_15_perm_0 = const()[name = string("q_with_bias_v_15_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_189_transpose_x_0 = const()[name = string("x_189_transpose_x_0"), val = bool(false)]; + bool x_189_transpose_y_0 = const()[name = string("x_189_transpose_y_0"), val = bool(false)]; + tensor op_2007_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160096704))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160196096))))[name = string("op_2007_to_fp16_quantized")]; + tensor q_with_bias_v_15_cast_fp16 = transpose(perm = q_with_bias_v_15_perm_0, x = var_2005_cast_fp16)[name = string("transpose_299")]; + tensor x_189_cast_fp16 = matmul(transpose_x = x_189_transpose_x_0, transpose_y = x_189_transpose_y_0, x = q_with_bias_v_15_cast_fp16, y = op_2007_to_fp16_quantized)[name = string("x_189_cast_fp16")]; + tensor x_191_pad_0 = const()[name = string("x_191_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_191_mode_0 = const()[name = string("x_191_mode_0"), val = string("constant")]; + fp16 const_170_to_fp16 = const()[name = string("const_170_to_fp16"), val = fp16(0x0p+0)]; + tensor x_191_cast_fp16 = pad(constant_val = const_170_to_fp16, mode = x_191_mode_0, pad = x_191_pad_0, x = x_189_cast_fp16)[name = string("x_191_cast_fp16")]; + tensor var_2015 = const()[name = string("op_2015"), val = tensor([1, 8, -1, 7])]; + tensor x_193_cast_fp16 = reshape(shape = var_2015, x = x_191_cast_fp16)[name = string("x_193_cast_fp16")]; + tensor var_2019_begin_0 = const()[name = string("op_2019_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_2019_end_0 = const()[name = string("op_2019_end_0"), val = tensor([1, 8, 98, 7])]; + tensor var_2019_end_mask_0 = const()[name = string("op_2019_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_2019_cast_fp16 = slice_by_index(begin = var_2019_begin_0, end = var_2019_end_0, end_mask = var_2019_end_mask_0, x = x_193_cast_fp16)[name = string("op_2019_cast_fp16")]; + tensor var_2020 = const()[name = string("op_2020"), val = tensor([1, 8, 7, 97])]; + tensor matrix_bd_29_cast_fp16 = reshape(shape = var_2020, x = var_2019_cast_fp16)[name = string("matrix_bd_29_cast_fp16")]; + bool matrix_ac_15_transpose_x_0 = const()[name = string("matrix_ac_15_transpose_x_0"), val = bool(false)]; + bool matrix_ac_15_transpose_y_0 = const()[name = string("matrix_ac_15_transpose_y_0"), val = bool(false)]; + tensor transpose_110_perm_0 = const()[name = string("transpose_110_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_111_perm_0 = const()[name = string("transpose_111_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_111 = transpose(perm = transpose_111_perm_0, x = k_29_cast_fp16)[name = string("transpose_297")]; + tensor transpose_110 = transpose(perm = transpose_110_perm_0, x = var_2003_cast_fp16)[name = string("transpose_298")]; + tensor matrix_ac_15_cast_fp16 = matmul(transpose_x = matrix_ac_15_transpose_x_0, transpose_y = matrix_ac_15_transpose_y_0, x = transpose_110, y = transpose_111)[name = string("matrix_ac_15_cast_fp16")]; + tensor matrix_bd_31_begin_0 = const()[name = string("matrix_bd_31_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_31_end_0 = const()[name = string("matrix_bd_31_end_0"), val = tensor([1, 8, 7, 49])]; + tensor matrix_bd_31_end_mask_0 = const()[name = string("matrix_bd_31_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_31_cast_fp16 = slice_by_index(begin = matrix_bd_31_begin_0, end = matrix_bd_31_end_0, end_mask = matrix_bd_31_end_mask_0, x = matrix_bd_29_cast_fp16)[name = string("matrix_bd_31_cast_fp16")]; + tensor var_2029_cast_fp16 = add(x = matrix_ac_15_cast_fp16, y = matrix_bd_31_cast_fp16)[name = string("op_2029_cast_fp16")]; + fp16 _inversed_scores_29_y_0_to_fp16 = const()[name = string("_inversed_scores_29_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_29_cast_fp16 = mul(x = var_2029_cast_fp16, y = _inversed_scores_29_y_0_to_fp16)[name = string("_inversed_scores_29_cast_fp16")]; + tensor scores_31_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_29_cast_fp16, cond = mask_11)[name = string("scores_31_cast_fp16")]; + tensor var_2035_cast_fp16 = softmax(axis = var_59, x = scores_31_cast_fp16)[name = string("op_2035_cast_fp16")]; + tensor input_405_cast_fp16 = select(a = var_44_to_fp16, b = var_2035_cast_fp16, cond = mask_11)[name = string("input_405_cast_fp16")]; + bool x_195_transpose_x_0 = const()[name = string("x_195_transpose_x_0"), val = bool(false)]; + bool x_195_transpose_y_0 = const()[name = string("x_195_transpose_y_0"), val = bool(false)]; + tensor value_23_cast_fp16 = transpose(perm = value_23_perm_0, x = v_15_cast_fp16)[name = string("transpose_296")]; + tensor x_195_cast_fp16 = matmul(transpose_x = x_195_transpose_x_0, transpose_y = x_195_transpose_y_0, x = input_405_cast_fp16, y = value_23_cast_fp16)[name = string("x_195_cast_fp16")]; + tensor var_2039_perm_0 = const()[name = string("op_2039_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_2040 = const()[name = string("op_2040"), val = tensor([1, -1, 1024])]; + tensor var_2039_cast_fp16 = transpose(perm = var_2039_perm_0, x = x_195_cast_fp16)[name = string("transpose_295")]; + tensor input_407_cast_fp16 = reshape(shape = var_2040, x = var_2039_cast_fp16)[name = string("input_407_cast_fp16")]; + tensor encoder_layers_7_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160196416))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160982912))))[name = string("encoder_layers_7_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_7_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_7_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160983104)))]; + tensor linear_70_cast_fp16 = linear(bias = encoder_layers_7_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_7_self_attn_linear_out_weight_to_fp16_palettized, x = input_407_cast_fp16)[name = string("linear_70_cast_fp16")]; + tensor input_411_cast_fp16 = add(x = input_401_cast_fp16, y = linear_70_cast_fp16)[name = string("input_411_cast_fp16")]; + tensor x_199_axes_0 = const()[name = string("x_199_axes_0"), val = tensor([-1])]; + tensor encoder_layers_7_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_7_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160985216)))]; + tensor encoder_layers_7_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_7_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160987328)))]; + tensor x_199_cast_fp16 = layer_norm(axes = x_199_axes_0, beta = encoder_layers_7_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_7_norm_conv_weight_to_fp16, x = input_411_cast_fp16)[name = string("x_199_cast_fp16")]; + tensor input_413_perm_0 = const()[name = string("input_413_perm_0"), val = tensor([0, 2, 1])]; + string input_415_pad_type_0 = const()[name = string("input_415_pad_type_0"), val = string("valid")]; + tensor input_415_strides_0 = const()[name = string("input_415_strides_0"), val = tensor([1])]; + tensor input_415_pad_0 = const()[name = string("input_415_pad_0"), val = tensor([0, 0])]; + tensor input_415_dilations_0 = const()[name = string("input_415_dilations_0"), val = tensor([1])]; + int32 input_415_groups_0 = const()[name = string("input_415_groups_0"), val = int32(1)]; + tensor encoder_layers_7_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160989440))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163086656))))[name = string("encoder_layers_7_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_413_cast_fp16 = transpose(perm = input_413_perm_0, x = x_199_cast_fp16)[name = string("transpose_294")]; + tensor input_415_cast_fp16 = conv(dilations = input_415_dilations_0, groups = input_415_groups_0, pad = input_415_pad_0, pad_type = input_415_pad_type_0, strides = input_415_strides_0, weight = encoder_layers_7_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_413_cast_fp16)[name = string("input_415_cast_fp16")]; + int32 x_201_split_num_splits_0 = const()[name = string("x_201_split_num_splits_0"), val = int32(2)]; + int32 x_201_split_axis_0 = const()[name = string("x_201_split_axis_0"), val = int32(1)]; + tensor x_201_split_cast_fp16_0, tensor x_201_split_cast_fp16_1 = split(axis = x_201_split_axis_0, num_splits = x_201_split_num_splits_0, x = input_415_cast_fp16)[name = string("x_201_split_cast_fp16")]; + tensor x_201_split_1_sigmoid_cast_fp16 = sigmoid(x = x_201_split_cast_fp16_1)[name = string("x_201_split_1_sigmoid_cast_fp16")]; + tensor x_201_cast_fp16 = mul(x = x_201_split_cast_fp16_0, y = x_201_split_1_sigmoid_cast_fp16)[name = string("x_201_cast_fp16")]; + tensor input_417_cast_fp16 = select(a = var_44_to_fp16, b = x_201_cast_fp16, cond = var_575)[name = string("input_417_cast_fp16")]; + bool new_x_31_interleave_0 = const()[name = string("new_x_31_interleave_0"), val = bool(false)]; + tensor new_x_31_cast_fp16 = concat(axis = var_59, interleave = new_x_31_interleave_0, values = (cache_31_cast_fp16, input_417_cast_fp16))[name = string("new_x_31_cast_fp16")]; + tensor var_2079_begin_0 = const()[name = string("op_2079_begin_0"), val = tensor([0, 0, 7])]; + tensor var_2079_end_0 = const()[name = string("op_2079_end_0"), val = tensor([1, 1024, 15])]; + tensor var_2079_end_mask_0 = const()[name = string("op_2079_end_mask_0"), val = tensor([true, true, true])]; + tensor var_2079_cast_fp16 = slice_by_index(begin = var_2079_begin_0, end = var_2079_end_0, end_mask = var_2079_end_mask_0, x = new_x_31_cast_fp16)[name = string("op_2079_cast_fp16")]; + string x_203_pad_type_0 = const()[name = string("x_203_pad_type_0"), val = string("valid")]; + int32 x_203_groups_0 = const()[name = string("x_203_groups_0"), val = int32(1024)]; + tensor x_203_strides_0 = const()[name = string("x_203_strides_0"), val = tensor([1])]; + tensor x_203_pad_0 = const()[name = string("x_203_pad_0"), val = tensor([0, 0])]; + tensor x_203_dilations_0 = const()[name = string("x_203_dilations_0"), val = tensor([1])]; + tensor encoder_layers_7_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163090816))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163100096))))[name = string("encoder_layers_7_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_203_cast_fp16 = conv(dilations = x_203_dilations_0, groups = x_203_groups_0, pad = x_203_pad_0, pad_type = x_203_pad_type_0, strides = x_203_strides_0, weight = encoder_layers_7_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_31_cast_fp16)[name = string("x_203_cast_fp16")]; + tensor input_419_perm_0 = const()[name = string("input_419_perm_0"), val = tensor([0, 2, 1])]; + tensor x_205_axes_0 = const()[name = string("x_205_axes_0"), val = tensor([-1])]; + tensor encoder_layers_7_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_7_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163102208)))]; + tensor encoder_layers_7_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_7_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163104320)))]; + tensor input_419_cast_fp16 = transpose(perm = input_419_perm_0, x = x_203_cast_fp16)[name = string("transpose_293")]; + tensor x_205_cast_fp16 = layer_norm(axes = x_205_axes_0, beta = encoder_layers_7_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_7_conv_batch_norm_weight_to_fp16, x = input_419_cast_fp16)[name = string("x_205_cast_fp16")]; + tensor input_421_perm_0 = const()[name = string("input_421_perm_0"), val = tensor([0, 2, 1])]; + tensor input_421_cast_fp16 = transpose(perm = input_421_perm_0, x = x_205_cast_fp16)[name = string("transpose_292")]; + tensor input_423_cast_fp16 = silu(x = input_421_cast_fp16)[name = string("input_423_cast_fp16")]; + string x_207_pad_type_0 = const()[name = string("x_207_pad_type_0"), val = string("valid")]; + tensor x_207_strides_0 = const()[name = string("x_207_strides_0"), val = tensor([1])]; + tensor x_207_pad_0 = const()[name = string("x_207_pad_0"), val = tensor([0, 0])]; + tensor x_207_dilations_0 = const()[name = string("x_207_dilations_0"), val = tensor([1])]; + int32 x_207_groups_0 = const()[name = string("x_207_groups_0"), val = int32(1)]; + tensor encoder_layers_7_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163106432))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(164155072))))[name = string("encoder_layers_7_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_207_cast_fp16 = conv(dilations = x_207_dilations_0, groups = x_207_groups_0, pad = x_207_pad_0, pad_type = x_207_pad_type_0, strides = x_207_strides_0, weight = encoder_layers_7_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_423_cast_fp16)[name = string("x_207_cast_fp16")]; + tensor input_425_perm_0 = const()[name = string("input_425_perm_0"), val = tensor([0, 2, 1])]; + tensor input_425_cast_fp16 = transpose(perm = input_425_perm_0, x = x_207_cast_fp16)[name = string("transpose_291")]; + tensor input_427_cast_fp16 = add(x = input_411_cast_fp16, y = input_425_cast_fp16)[name = string("input_427_cast_fp16")]; + tensor input_429_axes_0 = const()[name = string("input_429_axes_0"), val = tensor([-1])]; + tensor encoder_layers_7_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_7_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(164157184)))]; + tensor encoder_layers_7_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_7_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(164159296)))]; + tensor input_429_cast_fp16 = layer_norm(axes = input_429_axes_0, beta = encoder_layers_7_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_7_norm_feed_forward2_weight_to_fp16, x = input_427_cast_fp16)[name = string("input_429_cast_fp16")]; + tensor encoder_layers_7_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(164161408))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167307200))))[name = string("encoder_layers_7_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_7_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_7_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167307392)))]; + tensor linear_71_cast_fp16 = linear(bias = encoder_layers_7_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_7_feed_forward2_linear1_weight_to_fp16_palettized, x = input_429_cast_fp16)[name = string("linear_71_cast_fp16")]; + tensor input_433_cast_fp16 = silu(x = linear_71_cast_fp16)[name = string("input_433_cast_fp16")]; + tensor encoder_layers_7_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167315648))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170461440))))[name = string("encoder_layers_7_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_7_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_7_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170461632)))]; + tensor linear_72_cast_fp16 = linear(bias = encoder_layers_7_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_7_feed_forward2_linear2_weight_to_fp16_palettized, x = input_433_cast_fp16)[name = string("linear_72_cast_fp16")]; + fp16 var_2122_to_fp16 = const()[name = string("op_2122_to_fp16"), val = fp16(0x1p-1)]; + tensor var_2123_cast_fp16 = mul(x = linear_72_cast_fp16, y = var_2122_to_fp16)[name = string("op_2123_cast_fp16")]; + tensor input_439_cast_fp16 = add(x = input_427_cast_fp16, y = var_2123_cast_fp16)[name = string("input_439_cast_fp16")]; + tensor input_441_axes_0 = const()[name = string("input_441_axes_0"), val = tensor([-1])]; + tensor encoder_layers_7_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_7_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170463744)))]; + tensor encoder_layers_7_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_7_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170465856)))]; + tensor input_441_cast_fp16 = layer_norm(axes = input_441_axes_0, beta = encoder_layers_7_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_7_norm_out_weight_to_fp16, x = input_439_cast_fp16)[name = string("input_441_cast_fp16")]; + tensor cache_33_begin_0 = const()[name = string("cache_33_begin_0"), val = tensor([8, 0, 0, 0])]; + tensor cache_33_end_0 = const()[name = string("cache_33_end_0"), val = tensor([9, 1, 42, 1024])]; + tensor cache_33_end_mask_0 = const()[name = string("cache_33_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_33_squeeze_mask_0 = const()[name = string("cache_33_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_33_cast_fp16 = slice_by_index(begin = cache_33_begin_0, end = cache_33_end_0, end_mask = cache_33_end_mask_0, squeeze_mask = cache_33_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_33_cast_fp16")]; + tensor cache_35_begin_0 = const()[name = string("cache_35_begin_0"), val = tensor([8, 0, 0, 0])]; + tensor cache_35_end_0 = const()[name = string("cache_35_end_0"), val = tensor([9, 1, 1024, 8])]; + tensor cache_35_end_mask_0 = const()[name = string("cache_35_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_35_squeeze_mask_0 = const()[name = string("cache_35_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_35_cast_fp16 = slice_by_index(begin = cache_35_begin_0, end = cache_35_end_0, end_mask = cache_35_end_mask_0, squeeze_mask = cache_35_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_35_cast_fp16")]; + tensor input_443_axes_0 = const()[name = string("input_443_axes_0"), val = tensor([-1])]; + tensor encoder_layers_8_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_8_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170467968)))]; + tensor encoder_layers_8_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_8_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170470080)))]; + tensor input_443_cast_fp16 = layer_norm(axes = input_443_axes_0, beta = encoder_layers_8_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_8_norm_feed_forward1_weight_to_fp16, x = input_441_cast_fp16)[name = string("input_443_cast_fp16")]; + tensor encoder_layers_8_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170472192))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(173617984))))[name = string("encoder_layers_8_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_8_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_8_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(173618176)))]; + tensor linear_73_cast_fp16 = linear(bias = encoder_layers_8_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_8_feed_forward1_linear1_weight_to_fp16_palettized, x = input_443_cast_fp16)[name = string("linear_73_cast_fp16")]; + tensor input_447_cast_fp16 = silu(x = linear_73_cast_fp16)[name = string("input_447_cast_fp16")]; + tensor encoder_layers_8_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(173626432))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(176772224))))[name = string("encoder_layers_8_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_8_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_8_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(176772416)))]; + tensor linear_74_cast_fp16 = linear(bias = encoder_layers_8_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_8_feed_forward1_linear2_weight_to_fp16_palettized, x = input_447_cast_fp16)[name = string("linear_74_cast_fp16")]; + fp16 var_2159_to_fp16 = const()[name = string("op_2159_to_fp16"), val = fp16(0x1p-1)]; + tensor var_2160_cast_fp16 = mul(x = linear_74_cast_fp16, y = var_2159_to_fp16)[name = string("op_2160_cast_fp16")]; + tensor input_453_cast_fp16 = add(x = input_441_cast_fp16, y = var_2160_cast_fp16)[name = string("input_453_cast_fp16")]; + tensor key_17_axes_0 = const()[name = string("key_17_axes_0"), val = tensor([-1])]; + tensor encoder_layers_8_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_8_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(176774528)))]; + tensor encoder_layers_8_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_8_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(176776640)))]; + tensor key_17_cast_fp16 = layer_norm(axes = key_17_axes_0, beta = encoder_layers_8_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_8_norm_self_att_weight_to_fp16, x = input_453_cast_fp16)[name = string("key_17_cast_fp16")]; + bool input_455_interleave_0 = const()[name = string("input_455_interleave_0"), val = bool(false)]; + tensor input_455_cast_fp16 = concat(axis = var_68, interleave = input_455_interleave_0, values = (cache_33_cast_fp16, key_17_cast_fp16))[name = string("input_455_cast_fp16")]; + tensor var_2182_begin_0 = const()[name = string("op_2182_begin_0"), val = tensor([0, 7, 0])]; + tensor var_2182_end_0 = const()[name = string("op_2182_end_0"), val = tensor([1, 42, 1024])]; + tensor var_2182_end_mask_0 = const()[name = string("op_2182_end_mask_0"), val = tensor([true, true, true])]; + tensor var_2182_cast_fp16 = slice_by_index(begin = var_2182_begin_0, end = var_2182_end_0, end_mask = var_2182_end_mask_0, x = cache_33_cast_fp16)[name = string("op_2182_cast_fp16")]; + bool var_2188_interleave_0 = const()[name = string("op_2188_interleave_0"), val = bool(false)]; + tensor var_2188_cast_fp16 = concat(axis = var_68, interleave = var_2188_interleave_0, values = (var_2182_cast_fp16, key_17_cast_fp16))[name = string("op_2188_cast_fp16")]; + tensor encoder_layers_8_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(176778752))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(177565248))))[name = string("encoder_layers_8_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_8_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_8_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(177565440)))]; + tensor linear_75_cast_fp16 = linear(bias = encoder_layers_8_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_8_self_attn_linear_q_weight_to_fp16_palettized, x = key_17_cast_fp16)[name = string("linear_75_cast_fp16")]; + tensor var_2193 = const()[name = string("op_2193"), val = tensor([1, -1, 8, 128])]; + tensor q_49_cast_fp16 = reshape(shape = var_2193, x = linear_75_cast_fp16)[name = string("q_49_cast_fp16")]; + tensor encoder_layers_8_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(177567552))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(178354048))))[name = string("encoder_layers_8_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_8_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_8_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(178354240)))]; + tensor linear_76_cast_fp16 = linear(bias = encoder_layers_8_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_8_self_attn_linear_k_weight_to_fp16_palettized, x = input_455_cast_fp16)[name = string("linear_76_cast_fp16")]; + tensor var_2198 = const()[name = string("op_2198"), val = tensor([1, -1, 8, 128])]; + tensor k_33_cast_fp16 = reshape(shape = var_2198, x = linear_76_cast_fp16)[name = string("k_33_cast_fp16")]; + tensor encoder_layers_8_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(178356352))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179142848))))[name = string("encoder_layers_8_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_8_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_8_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179143040)))]; + tensor linear_77_cast_fp16 = linear(bias = encoder_layers_8_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_8_self_attn_linear_v_weight_to_fp16_palettized, x = input_455_cast_fp16)[name = string("linear_77_cast_fp16")]; + tensor var_2203 = const()[name = string("op_2203"), val = tensor([1, -1, 8, 128])]; + tensor v_17_cast_fp16 = reshape(shape = var_2203, x = linear_77_cast_fp16)[name = string("v_17_cast_fp16")]; + tensor value_25_perm_0 = const()[name = string("value_25_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_8_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_8_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179145152)))]; + tensor var_2216_cast_fp16 = add(x = q_49_cast_fp16, y = encoder_layers_8_self_attn_pos_bias_u_to_fp16)[name = string("op_2216_cast_fp16")]; + tensor encoder_layers_8_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_8_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179147264)))]; + tensor var_2218_cast_fp16 = add(x = q_49_cast_fp16, y = encoder_layers_8_self_attn_pos_bias_v_to_fp16)[name = string("op_2218_cast_fp16")]; + tensor q_with_bias_v_17_perm_0 = const()[name = string("q_with_bias_v_17_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_215_transpose_x_0 = const()[name = string("x_215_transpose_x_0"), val = bool(false)]; + bool x_215_transpose_y_0 = const()[name = string("x_215_transpose_y_0"), val = bool(false)]; + tensor op_2220_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179149376))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179248768))))[name = string("op_2220_to_fp16_quantized")]; + tensor q_with_bias_v_17_cast_fp16 = transpose(perm = q_with_bias_v_17_perm_0, x = var_2218_cast_fp16)[name = string("transpose_290")]; + tensor x_215_cast_fp16 = matmul(transpose_x = x_215_transpose_x_0, transpose_y = x_215_transpose_y_0, x = q_with_bias_v_17_cast_fp16, y = op_2220_to_fp16_quantized)[name = string("x_215_cast_fp16")]; + tensor x_217_pad_0 = const()[name = string("x_217_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_217_mode_0 = const()[name = string("x_217_mode_0"), val = string("constant")]; + fp16 const_183_to_fp16 = const()[name = string("const_183_to_fp16"), val = fp16(0x0p+0)]; + tensor x_217_cast_fp16 = pad(constant_val = const_183_to_fp16, mode = x_217_mode_0, pad = x_217_pad_0, x = x_215_cast_fp16)[name = string("x_217_cast_fp16")]; + tensor var_2228 = const()[name = string("op_2228"), val = tensor([1, 8, -1, 7])]; + tensor x_219_cast_fp16 = reshape(shape = var_2228, x = x_217_cast_fp16)[name = string("x_219_cast_fp16")]; + tensor var_2232_begin_0 = const()[name = string("op_2232_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_2232_end_0 = const()[name = string("op_2232_end_0"), val = tensor([1, 8, 98, 7])]; + tensor var_2232_end_mask_0 = const()[name = string("op_2232_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_2232_cast_fp16 = slice_by_index(begin = var_2232_begin_0, end = var_2232_end_0, end_mask = var_2232_end_mask_0, x = x_219_cast_fp16)[name = string("op_2232_cast_fp16")]; + tensor var_2233 = const()[name = string("op_2233"), val = tensor([1, 8, 7, 97])]; + tensor matrix_bd_33_cast_fp16 = reshape(shape = var_2233, x = var_2232_cast_fp16)[name = string("matrix_bd_33_cast_fp16")]; + bool matrix_ac_17_transpose_x_0 = const()[name = string("matrix_ac_17_transpose_x_0"), val = bool(false)]; + bool matrix_ac_17_transpose_y_0 = const()[name = string("matrix_ac_17_transpose_y_0"), val = bool(false)]; + tensor transpose_112_perm_0 = const()[name = string("transpose_112_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_113_perm_0 = const()[name = string("transpose_113_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_113 = transpose(perm = transpose_113_perm_0, x = k_33_cast_fp16)[name = string("transpose_288")]; + tensor transpose_112 = transpose(perm = transpose_112_perm_0, x = var_2216_cast_fp16)[name = string("transpose_289")]; + tensor matrix_ac_17_cast_fp16 = matmul(transpose_x = matrix_ac_17_transpose_x_0, transpose_y = matrix_ac_17_transpose_y_0, x = transpose_112, y = transpose_113)[name = string("matrix_ac_17_cast_fp16")]; + tensor matrix_bd_35_begin_0 = const()[name = string("matrix_bd_35_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_35_end_0 = const()[name = string("matrix_bd_35_end_0"), val = tensor([1, 8, 7, 49])]; + tensor matrix_bd_35_end_mask_0 = const()[name = string("matrix_bd_35_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_35_cast_fp16 = slice_by_index(begin = matrix_bd_35_begin_0, end = matrix_bd_35_end_0, end_mask = matrix_bd_35_end_mask_0, x = matrix_bd_33_cast_fp16)[name = string("matrix_bd_35_cast_fp16")]; + tensor var_2242_cast_fp16 = add(x = matrix_ac_17_cast_fp16, y = matrix_bd_35_cast_fp16)[name = string("op_2242_cast_fp16")]; + fp16 _inversed_scores_33_y_0_to_fp16 = const()[name = string("_inversed_scores_33_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_33_cast_fp16 = mul(x = var_2242_cast_fp16, y = _inversed_scores_33_y_0_to_fp16)[name = string("_inversed_scores_33_cast_fp16")]; + tensor scores_35_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_33_cast_fp16, cond = mask_11)[name = string("scores_35_cast_fp16")]; + tensor var_2248_cast_fp16 = softmax(axis = var_59, x = scores_35_cast_fp16)[name = string("op_2248_cast_fp16")]; + tensor input_457_cast_fp16 = select(a = var_44_to_fp16, b = var_2248_cast_fp16, cond = mask_11)[name = string("input_457_cast_fp16")]; + bool x_221_transpose_x_0 = const()[name = string("x_221_transpose_x_0"), val = bool(false)]; + bool x_221_transpose_y_0 = const()[name = string("x_221_transpose_y_0"), val = bool(false)]; + tensor value_25_cast_fp16 = transpose(perm = value_25_perm_0, x = v_17_cast_fp16)[name = string("transpose_287")]; + tensor x_221_cast_fp16 = matmul(transpose_x = x_221_transpose_x_0, transpose_y = x_221_transpose_y_0, x = input_457_cast_fp16, y = value_25_cast_fp16)[name = string("x_221_cast_fp16")]; + tensor var_2252_perm_0 = const()[name = string("op_2252_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_2253 = const()[name = string("op_2253"), val = tensor([1, -1, 1024])]; + tensor var_2252_cast_fp16 = transpose(perm = var_2252_perm_0, x = x_221_cast_fp16)[name = string("transpose_286")]; + tensor input_459_cast_fp16 = reshape(shape = var_2253, x = var_2252_cast_fp16)[name = string("input_459_cast_fp16")]; + tensor encoder_layers_8_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179249088))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(180035584))))[name = string("encoder_layers_8_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_8_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_8_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(180035776)))]; + tensor linear_79_cast_fp16 = linear(bias = encoder_layers_8_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_8_self_attn_linear_out_weight_to_fp16_palettized, x = input_459_cast_fp16)[name = string("linear_79_cast_fp16")]; + tensor input_463_cast_fp16 = add(x = input_453_cast_fp16, y = linear_79_cast_fp16)[name = string("input_463_cast_fp16")]; + tensor x_225_axes_0 = const()[name = string("x_225_axes_0"), val = tensor([-1])]; + tensor encoder_layers_8_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_8_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(180037888)))]; + tensor encoder_layers_8_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_8_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(180040000)))]; + tensor x_225_cast_fp16 = layer_norm(axes = x_225_axes_0, beta = encoder_layers_8_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_8_norm_conv_weight_to_fp16, x = input_463_cast_fp16)[name = string("x_225_cast_fp16")]; + tensor input_465_perm_0 = const()[name = string("input_465_perm_0"), val = tensor([0, 2, 1])]; + string input_467_pad_type_0 = const()[name = string("input_467_pad_type_0"), val = string("valid")]; + tensor input_467_strides_0 = const()[name = string("input_467_strides_0"), val = tensor([1])]; + tensor input_467_pad_0 = const()[name = string("input_467_pad_0"), val = tensor([0, 0])]; + tensor input_467_dilations_0 = const()[name = string("input_467_dilations_0"), val = tensor([1])]; + int32 input_467_groups_0 = const()[name = string("input_467_groups_0"), val = int32(1)]; + tensor encoder_layers_8_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(180042112))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182139328))))[name = string("encoder_layers_8_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_465_cast_fp16 = transpose(perm = input_465_perm_0, x = x_225_cast_fp16)[name = string("transpose_285")]; + tensor input_467_cast_fp16 = conv(dilations = input_467_dilations_0, groups = input_467_groups_0, pad = input_467_pad_0, pad_type = input_467_pad_type_0, strides = input_467_strides_0, weight = encoder_layers_8_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_465_cast_fp16)[name = string("input_467_cast_fp16")]; + int32 x_227_split_num_splits_0 = const()[name = string("x_227_split_num_splits_0"), val = int32(2)]; + int32 x_227_split_axis_0 = const()[name = string("x_227_split_axis_0"), val = int32(1)]; + tensor x_227_split_cast_fp16_0, tensor x_227_split_cast_fp16_1 = split(axis = x_227_split_axis_0, num_splits = x_227_split_num_splits_0, x = input_467_cast_fp16)[name = string("x_227_split_cast_fp16")]; + tensor x_227_split_1_sigmoid_cast_fp16 = sigmoid(x = x_227_split_cast_fp16_1)[name = string("x_227_split_1_sigmoid_cast_fp16")]; + tensor x_227_cast_fp16 = mul(x = x_227_split_cast_fp16_0, y = x_227_split_1_sigmoid_cast_fp16)[name = string("x_227_cast_fp16")]; + tensor input_469_cast_fp16 = select(a = var_44_to_fp16, b = x_227_cast_fp16, cond = var_575)[name = string("input_469_cast_fp16")]; + bool new_x_35_interleave_0 = const()[name = string("new_x_35_interleave_0"), val = bool(false)]; + tensor new_x_35_cast_fp16 = concat(axis = var_59, interleave = new_x_35_interleave_0, values = (cache_35_cast_fp16, input_469_cast_fp16))[name = string("new_x_35_cast_fp16")]; + tensor var_2292_begin_0 = const()[name = string("op_2292_begin_0"), val = tensor([0, 0, 7])]; + tensor var_2292_end_0 = const()[name = string("op_2292_end_0"), val = tensor([1, 1024, 15])]; + tensor var_2292_end_mask_0 = const()[name = string("op_2292_end_mask_0"), val = tensor([true, true, true])]; + tensor var_2292_cast_fp16 = slice_by_index(begin = var_2292_begin_0, end = var_2292_end_0, end_mask = var_2292_end_mask_0, x = new_x_35_cast_fp16)[name = string("op_2292_cast_fp16")]; + string x_229_pad_type_0 = const()[name = string("x_229_pad_type_0"), val = string("valid")]; + int32 x_229_groups_0 = const()[name = string("x_229_groups_0"), val = int32(1024)]; + tensor x_229_strides_0 = const()[name = string("x_229_strides_0"), val = tensor([1])]; + tensor x_229_pad_0 = const()[name = string("x_229_pad_0"), val = tensor([0, 0])]; + tensor x_229_dilations_0 = const()[name = string("x_229_dilations_0"), val = tensor([1])]; + tensor encoder_layers_8_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182143488))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182152768))))[name = string("encoder_layers_8_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_229_cast_fp16 = conv(dilations = x_229_dilations_0, groups = x_229_groups_0, pad = x_229_pad_0, pad_type = x_229_pad_type_0, strides = x_229_strides_0, weight = encoder_layers_8_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_35_cast_fp16)[name = string("x_229_cast_fp16")]; + tensor input_471_perm_0 = const()[name = string("input_471_perm_0"), val = tensor([0, 2, 1])]; + tensor x_231_axes_0 = const()[name = string("x_231_axes_0"), val = tensor([-1])]; + tensor encoder_layers_8_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_8_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182154880)))]; + tensor encoder_layers_8_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_8_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182156992)))]; + tensor input_471_cast_fp16 = transpose(perm = input_471_perm_0, x = x_229_cast_fp16)[name = string("transpose_284")]; + tensor x_231_cast_fp16 = layer_norm(axes = x_231_axes_0, beta = encoder_layers_8_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_8_conv_batch_norm_weight_to_fp16, x = input_471_cast_fp16)[name = string("x_231_cast_fp16")]; + tensor input_473_perm_0 = const()[name = string("input_473_perm_0"), val = tensor([0, 2, 1])]; + tensor input_473_cast_fp16 = transpose(perm = input_473_perm_0, x = x_231_cast_fp16)[name = string("transpose_283")]; + tensor input_475_cast_fp16 = silu(x = input_473_cast_fp16)[name = string("input_475_cast_fp16")]; + string x_233_pad_type_0 = const()[name = string("x_233_pad_type_0"), val = string("valid")]; + tensor x_233_strides_0 = const()[name = string("x_233_strides_0"), val = tensor([1])]; + tensor x_233_pad_0 = const()[name = string("x_233_pad_0"), val = tensor([0, 0])]; + tensor x_233_dilations_0 = const()[name = string("x_233_dilations_0"), val = tensor([1])]; + int32 x_233_groups_0 = const()[name = string("x_233_groups_0"), val = int32(1)]; + tensor encoder_layers_8_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182159104))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183207744))))[name = string("encoder_layers_8_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_233_cast_fp16 = conv(dilations = x_233_dilations_0, groups = x_233_groups_0, pad = x_233_pad_0, pad_type = x_233_pad_type_0, strides = x_233_strides_0, weight = encoder_layers_8_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_475_cast_fp16)[name = string("x_233_cast_fp16")]; + tensor input_477_perm_0 = const()[name = string("input_477_perm_0"), val = tensor([0, 2, 1])]; + tensor input_477_cast_fp16 = transpose(perm = input_477_perm_0, x = x_233_cast_fp16)[name = string("transpose_282")]; + tensor input_479_cast_fp16 = add(x = input_463_cast_fp16, y = input_477_cast_fp16)[name = string("input_479_cast_fp16")]; + tensor input_481_axes_0 = const()[name = string("input_481_axes_0"), val = tensor([-1])]; + tensor encoder_layers_8_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_8_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183209856)))]; + tensor encoder_layers_8_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_8_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183211968)))]; + tensor input_481_cast_fp16 = layer_norm(axes = input_481_axes_0, beta = encoder_layers_8_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_8_norm_feed_forward2_weight_to_fp16, x = input_479_cast_fp16)[name = string("input_481_cast_fp16")]; + tensor encoder_layers_8_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183214080))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186359872))))[name = string("encoder_layers_8_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_8_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_8_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186360064)))]; + tensor linear_80_cast_fp16 = linear(bias = encoder_layers_8_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_8_feed_forward2_linear1_weight_to_fp16_palettized, x = input_481_cast_fp16)[name = string("linear_80_cast_fp16")]; + tensor input_485_cast_fp16 = silu(x = linear_80_cast_fp16)[name = string("input_485_cast_fp16")]; + tensor encoder_layers_8_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186368320))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(189514112))))[name = string("encoder_layers_8_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_8_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_8_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(189514304)))]; + tensor linear_81_cast_fp16 = linear(bias = encoder_layers_8_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_8_feed_forward2_linear2_weight_to_fp16_palettized, x = input_485_cast_fp16)[name = string("linear_81_cast_fp16")]; + fp16 var_2335_to_fp16 = const()[name = string("op_2335_to_fp16"), val = fp16(0x1p-1)]; + tensor var_2336_cast_fp16 = mul(x = linear_81_cast_fp16, y = var_2335_to_fp16)[name = string("op_2336_cast_fp16")]; + tensor input_491_cast_fp16 = add(x = input_479_cast_fp16, y = var_2336_cast_fp16)[name = string("input_491_cast_fp16")]; + tensor input_493_axes_0 = const()[name = string("input_493_axes_0"), val = tensor([-1])]; + tensor encoder_layers_8_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_8_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(189516416)))]; + tensor encoder_layers_8_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_8_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(189518528)))]; + tensor input_493_cast_fp16 = layer_norm(axes = input_493_axes_0, beta = encoder_layers_8_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_8_norm_out_weight_to_fp16, x = input_491_cast_fp16)[name = string("input_493_cast_fp16")]; + tensor cache_37_begin_0 = const()[name = string("cache_37_begin_0"), val = tensor([9, 0, 0, 0])]; + tensor cache_37_end_0 = const()[name = string("cache_37_end_0"), val = tensor([10, 1, 42, 1024])]; + tensor cache_37_end_mask_0 = const()[name = string("cache_37_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_37_squeeze_mask_0 = const()[name = string("cache_37_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_37_cast_fp16 = slice_by_index(begin = cache_37_begin_0, end = cache_37_end_0, end_mask = cache_37_end_mask_0, squeeze_mask = cache_37_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_37_cast_fp16")]; + tensor cache_39_begin_0 = const()[name = string("cache_39_begin_0"), val = tensor([9, 0, 0, 0])]; + tensor cache_39_end_0 = const()[name = string("cache_39_end_0"), val = tensor([10, 1, 1024, 8])]; + tensor cache_39_end_mask_0 = const()[name = string("cache_39_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_39_squeeze_mask_0 = const()[name = string("cache_39_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_39_cast_fp16 = slice_by_index(begin = cache_39_begin_0, end = cache_39_end_0, end_mask = cache_39_end_mask_0, squeeze_mask = cache_39_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_39_cast_fp16")]; + tensor input_495_axes_0 = const()[name = string("input_495_axes_0"), val = tensor([-1])]; + tensor encoder_layers_9_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_9_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(189520640)))]; + tensor encoder_layers_9_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_9_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(189522752)))]; + tensor input_495_cast_fp16 = layer_norm(axes = input_495_axes_0, beta = encoder_layers_9_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_9_norm_feed_forward1_weight_to_fp16, x = input_493_cast_fp16)[name = string("input_495_cast_fp16")]; + tensor encoder_layers_9_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(189524864))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(192670656))))[name = string("encoder_layers_9_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_9_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_9_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(192670848)))]; + tensor linear_82_cast_fp16 = linear(bias = encoder_layers_9_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_9_feed_forward1_linear1_weight_to_fp16_palettized, x = input_495_cast_fp16)[name = string("linear_82_cast_fp16")]; + tensor input_499_cast_fp16 = silu(x = linear_82_cast_fp16)[name = string("input_499_cast_fp16")]; + tensor encoder_layers_9_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(192679104))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(195824896))))[name = string("encoder_layers_9_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_9_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_9_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(195825088)))]; + tensor linear_83_cast_fp16 = linear(bias = encoder_layers_9_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_9_feed_forward1_linear2_weight_to_fp16_palettized, x = input_499_cast_fp16)[name = string("linear_83_cast_fp16")]; + fp16 var_2372_to_fp16 = const()[name = string("op_2372_to_fp16"), val = fp16(0x1p-1)]; + tensor var_2373_cast_fp16 = mul(x = linear_83_cast_fp16, y = var_2372_to_fp16)[name = string("op_2373_cast_fp16")]; + tensor input_505_cast_fp16 = add(x = input_493_cast_fp16, y = var_2373_cast_fp16)[name = string("input_505_cast_fp16")]; + tensor key_19_axes_0 = const()[name = string("key_19_axes_0"), val = tensor([-1])]; + tensor encoder_layers_9_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_9_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(195827200)))]; + tensor encoder_layers_9_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_9_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(195829312)))]; + tensor key_19_cast_fp16 = layer_norm(axes = key_19_axes_0, beta = encoder_layers_9_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_9_norm_self_att_weight_to_fp16, x = input_505_cast_fp16)[name = string("key_19_cast_fp16")]; + bool input_507_interleave_0 = const()[name = string("input_507_interleave_0"), val = bool(false)]; + tensor input_507_cast_fp16 = concat(axis = var_68, interleave = input_507_interleave_0, values = (cache_37_cast_fp16, key_19_cast_fp16))[name = string("input_507_cast_fp16")]; + tensor var_2395_begin_0 = const()[name = string("op_2395_begin_0"), val = tensor([0, 7, 0])]; + tensor var_2395_end_0 = const()[name = string("op_2395_end_0"), val = tensor([1, 42, 1024])]; + tensor var_2395_end_mask_0 = const()[name = string("op_2395_end_mask_0"), val = tensor([true, true, true])]; + tensor var_2395_cast_fp16 = slice_by_index(begin = var_2395_begin_0, end = var_2395_end_0, end_mask = var_2395_end_mask_0, x = cache_37_cast_fp16)[name = string("op_2395_cast_fp16")]; + bool var_2401_interleave_0 = const()[name = string("op_2401_interleave_0"), val = bool(false)]; + tensor var_2401_cast_fp16 = concat(axis = var_68, interleave = var_2401_interleave_0, values = (var_2395_cast_fp16, key_19_cast_fp16))[name = string("op_2401_cast_fp16")]; + tensor encoder_layers_9_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(195831424))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196617920))))[name = string("encoder_layers_9_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_9_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_9_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196618112)))]; + tensor linear_84_cast_fp16 = linear(bias = encoder_layers_9_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_9_self_attn_linear_q_weight_to_fp16_palettized, x = key_19_cast_fp16)[name = string("linear_84_cast_fp16")]; + tensor var_2406 = const()[name = string("op_2406"), val = tensor([1, -1, 8, 128])]; + tensor q_55_cast_fp16 = reshape(shape = var_2406, x = linear_84_cast_fp16)[name = string("q_55_cast_fp16")]; + tensor encoder_layers_9_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196620224))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197406720))))[name = string("encoder_layers_9_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_9_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_9_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197406912)))]; + tensor linear_85_cast_fp16 = linear(bias = encoder_layers_9_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_9_self_attn_linear_k_weight_to_fp16_palettized, x = input_507_cast_fp16)[name = string("linear_85_cast_fp16")]; + tensor var_2411 = const()[name = string("op_2411"), val = tensor([1, -1, 8, 128])]; + tensor k_37_cast_fp16 = reshape(shape = var_2411, x = linear_85_cast_fp16)[name = string("k_37_cast_fp16")]; + tensor encoder_layers_9_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197409024))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198195520))))[name = string("encoder_layers_9_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_9_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_9_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198195712)))]; + tensor linear_86_cast_fp16 = linear(bias = encoder_layers_9_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_9_self_attn_linear_v_weight_to_fp16_palettized, x = input_507_cast_fp16)[name = string("linear_86_cast_fp16")]; + tensor var_2416 = const()[name = string("op_2416"), val = tensor([1, -1, 8, 128])]; + tensor v_19_cast_fp16 = reshape(shape = var_2416, x = linear_86_cast_fp16)[name = string("v_19_cast_fp16")]; + tensor value_27_perm_0 = const()[name = string("value_27_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_9_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_9_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198197824)))]; + tensor var_2429_cast_fp16 = add(x = q_55_cast_fp16, y = encoder_layers_9_self_attn_pos_bias_u_to_fp16)[name = string("op_2429_cast_fp16")]; + tensor encoder_layers_9_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_9_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198199936)))]; + tensor var_2431_cast_fp16 = add(x = q_55_cast_fp16, y = encoder_layers_9_self_attn_pos_bias_v_to_fp16)[name = string("op_2431_cast_fp16")]; + tensor q_with_bias_v_19_perm_0 = const()[name = string("q_with_bias_v_19_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_241_transpose_x_0 = const()[name = string("x_241_transpose_x_0"), val = bool(false)]; + bool x_241_transpose_y_0 = const()[name = string("x_241_transpose_y_0"), val = bool(false)]; + tensor op_2433_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198202048))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198301440))))[name = string("op_2433_to_fp16_quantized")]; + tensor q_with_bias_v_19_cast_fp16 = transpose(perm = q_with_bias_v_19_perm_0, x = var_2431_cast_fp16)[name = string("transpose_281")]; + tensor x_241_cast_fp16 = matmul(transpose_x = x_241_transpose_x_0, transpose_y = x_241_transpose_y_0, x = q_with_bias_v_19_cast_fp16, y = op_2433_to_fp16_quantized)[name = string("x_241_cast_fp16")]; + tensor x_243_pad_0 = const()[name = string("x_243_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_243_mode_0 = const()[name = string("x_243_mode_0"), val = string("constant")]; + fp16 const_196_to_fp16 = const()[name = string("const_196_to_fp16"), val = fp16(0x0p+0)]; + tensor x_243_cast_fp16 = pad(constant_val = const_196_to_fp16, mode = x_243_mode_0, pad = x_243_pad_0, x = x_241_cast_fp16)[name = string("x_243_cast_fp16")]; + tensor var_2441 = const()[name = string("op_2441"), val = tensor([1, 8, -1, 7])]; + tensor x_245_cast_fp16 = reshape(shape = var_2441, x = x_243_cast_fp16)[name = string("x_245_cast_fp16")]; + tensor var_2445_begin_0 = const()[name = string("op_2445_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_2445_end_0 = const()[name = string("op_2445_end_0"), val = tensor([1, 8, 98, 7])]; + tensor var_2445_end_mask_0 = const()[name = string("op_2445_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_2445_cast_fp16 = slice_by_index(begin = var_2445_begin_0, end = var_2445_end_0, end_mask = var_2445_end_mask_0, x = x_245_cast_fp16)[name = string("op_2445_cast_fp16")]; + tensor var_2446 = const()[name = string("op_2446"), val = tensor([1, 8, 7, 97])]; + tensor matrix_bd_37_cast_fp16 = reshape(shape = var_2446, x = var_2445_cast_fp16)[name = string("matrix_bd_37_cast_fp16")]; + bool matrix_ac_19_transpose_x_0 = const()[name = string("matrix_ac_19_transpose_x_0"), val = bool(false)]; + bool matrix_ac_19_transpose_y_0 = const()[name = string("matrix_ac_19_transpose_y_0"), val = bool(false)]; + tensor transpose_114_perm_0 = const()[name = string("transpose_114_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_115_perm_0 = const()[name = string("transpose_115_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_115 = transpose(perm = transpose_115_perm_0, x = k_37_cast_fp16)[name = string("transpose_279")]; + tensor transpose_114 = transpose(perm = transpose_114_perm_0, x = var_2429_cast_fp16)[name = string("transpose_280")]; + tensor matrix_ac_19_cast_fp16 = matmul(transpose_x = matrix_ac_19_transpose_x_0, transpose_y = matrix_ac_19_transpose_y_0, x = transpose_114, y = transpose_115)[name = string("matrix_ac_19_cast_fp16")]; + tensor matrix_bd_39_begin_0 = const()[name = string("matrix_bd_39_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_39_end_0 = const()[name = string("matrix_bd_39_end_0"), val = tensor([1, 8, 7, 49])]; + tensor matrix_bd_39_end_mask_0 = const()[name = string("matrix_bd_39_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_39_cast_fp16 = slice_by_index(begin = matrix_bd_39_begin_0, end = matrix_bd_39_end_0, end_mask = matrix_bd_39_end_mask_0, x = matrix_bd_37_cast_fp16)[name = string("matrix_bd_39_cast_fp16")]; + tensor var_2455_cast_fp16 = add(x = matrix_ac_19_cast_fp16, y = matrix_bd_39_cast_fp16)[name = string("op_2455_cast_fp16")]; + fp16 _inversed_scores_37_y_0_to_fp16 = const()[name = string("_inversed_scores_37_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_37_cast_fp16 = mul(x = var_2455_cast_fp16, y = _inversed_scores_37_y_0_to_fp16)[name = string("_inversed_scores_37_cast_fp16")]; + tensor scores_39_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_37_cast_fp16, cond = mask_11)[name = string("scores_39_cast_fp16")]; + tensor var_2461_cast_fp16 = softmax(axis = var_59, x = scores_39_cast_fp16)[name = string("op_2461_cast_fp16")]; + tensor input_509_cast_fp16 = select(a = var_44_to_fp16, b = var_2461_cast_fp16, cond = mask_11)[name = string("input_509_cast_fp16")]; + bool x_247_transpose_x_0 = const()[name = string("x_247_transpose_x_0"), val = bool(false)]; + bool x_247_transpose_y_0 = const()[name = string("x_247_transpose_y_0"), val = bool(false)]; + tensor value_27_cast_fp16 = transpose(perm = value_27_perm_0, x = v_19_cast_fp16)[name = string("transpose_278")]; + tensor x_247_cast_fp16 = matmul(transpose_x = x_247_transpose_x_0, transpose_y = x_247_transpose_y_0, x = input_509_cast_fp16, y = value_27_cast_fp16)[name = string("x_247_cast_fp16")]; + tensor var_2465_perm_0 = const()[name = string("op_2465_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_2466 = const()[name = string("op_2466"), val = tensor([1, -1, 1024])]; + tensor var_2465_cast_fp16 = transpose(perm = var_2465_perm_0, x = x_247_cast_fp16)[name = string("transpose_277")]; + tensor input_511_cast_fp16 = reshape(shape = var_2466, x = var_2465_cast_fp16)[name = string("input_511_cast_fp16")]; + tensor encoder_layers_9_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198301760))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199088256))))[name = string("encoder_layers_9_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_9_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_9_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199088448)))]; + tensor linear_88_cast_fp16 = linear(bias = encoder_layers_9_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_9_self_attn_linear_out_weight_to_fp16_palettized, x = input_511_cast_fp16)[name = string("linear_88_cast_fp16")]; + tensor input_515_cast_fp16 = add(x = input_505_cast_fp16, y = linear_88_cast_fp16)[name = string("input_515_cast_fp16")]; + tensor x_251_axes_0 = const()[name = string("x_251_axes_0"), val = tensor([-1])]; + tensor encoder_layers_9_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_9_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199090560)))]; + tensor encoder_layers_9_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_9_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199092672)))]; + tensor x_251_cast_fp16 = layer_norm(axes = x_251_axes_0, beta = encoder_layers_9_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_9_norm_conv_weight_to_fp16, x = input_515_cast_fp16)[name = string("x_251_cast_fp16")]; + tensor input_517_perm_0 = const()[name = string("input_517_perm_0"), val = tensor([0, 2, 1])]; + string input_519_pad_type_0 = const()[name = string("input_519_pad_type_0"), val = string("valid")]; + tensor input_519_strides_0 = const()[name = string("input_519_strides_0"), val = tensor([1])]; + tensor input_519_pad_0 = const()[name = string("input_519_pad_0"), val = tensor([0, 0])]; + tensor input_519_dilations_0 = const()[name = string("input_519_dilations_0"), val = tensor([1])]; + int32 input_519_groups_0 = const()[name = string("input_519_groups_0"), val = int32(1)]; + tensor encoder_layers_9_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199094784))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(201192000))))[name = string("encoder_layers_9_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_517_cast_fp16 = transpose(perm = input_517_perm_0, x = x_251_cast_fp16)[name = string("transpose_276")]; + tensor input_519_cast_fp16 = conv(dilations = input_519_dilations_0, groups = input_519_groups_0, pad = input_519_pad_0, pad_type = input_519_pad_type_0, strides = input_519_strides_0, weight = encoder_layers_9_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_517_cast_fp16)[name = string("input_519_cast_fp16")]; + int32 x_253_split_num_splits_0 = const()[name = string("x_253_split_num_splits_0"), val = int32(2)]; + int32 x_253_split_axis_0 = const()[name = string("x_253_split_axis_0"), val = int32(1)]; + tensor x_253_split_cast_fp16_0, tensor x_253_split_cast_fp16_1 = split(axis = x_253_split_axis_0, num_splits = x_253_split_num_splits_0, x = input_519_cast_fp16)[name = string("x_253_split_cast_fp16")]; + tensor x_253_split_1_sigmoid_cast_fp16 = sigmoid(x = x_253_split_cast_fp16_1)[name = string("x_253_split_1_sigmoid_cast_fp16")]; + tensor x_253_cast_fp16 = mul(x = x_253_split_cast_fp16_0, y = x_253_split_1_sigmoid_cast_fp16)[name = string("x_253_cast_fp16")]; + tensor input_521_cast_fp16 = select(a = var_44_to_fp16, b = x_253_cast_fp16, cond = var_575)[name = string("input_521_cast_fp16")]; + bool new_x_39_interleave_0 = const()[name = string("new_x_39_interleave_0"), val = bool(false)]; + tensor new_x_39_cast_fp16 = concat(axis = var_59, interleave = new_x_39_interleave_0, values = (cache_39_cast_fp16, input_521_cast_fp16))[name = string("new_x_39_cast_fp16")]; + tensor var_2505_begin_0 = const()[name = string("op_2505_begin_0"), val = tensor([0, 0, 7])]; + tensor var_2505_end_0 = const()[name = string("op_2505_end_0"), val = tensor([1, 1024, 15])]; + tensor var_2505_end_mask_0 = const()[name = string("op_2505_end_mask_0"), val = tensor([true, true, true])]; + tensor var_2505_cast_fp16 = slice_by_index(begin = var_2505_begin_0, end = var_2505_end_0, end_mask = var_2505_end_mask_0, x = new_x_39_cast_fp16)[name = string("op_2505_cast_fp16")]; + string x_255_pad_type_0 = const()[name = string("x_255_pad_type_0"), val = string("valid")]; + int32 x_255_groups_0 = const()[name = string("x_255_groups_0"), val = int32(1024)]; + tensor x_255_strides_0 = const()[name = string("x_255_strides_0"), val = tensor([1])]; + tensor x_255_pad_0 = const()[name = string("x_255_pad_0"), val = tensor([0, 0])]; + tensor x_255_dilations_0 = const()[name = string("x_255_dilations_0"), val = tensor([1])]; + tensor encoder_layers_9_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(201196160))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(201205440))))[name = string("encoder_layers_9_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_255_cast_fp16 = conv(dilations = x_255_dilations_0, groups = x_255_groups_0, pad = x_255_pad_0, pad_type = x_255_pad_type_0, strides = x_255_strides_0, weight = encoder_layers_9_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_39_cast_fp16)[name = string("x_255_cast_fp16")]; + tensor input_523_perm_0 = const()[name = string("input_523_perm_0"), val = tensor([0, 2, 1])]; + tensor x_257_axes_0 = const()[name = string("x_257_axes_0"), val = tensor([-1])]; + tensor encoder_layers_9_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_9_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(201207552)))]; + tensor encoder_layers_9_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_9_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(201209664)))]; + tensor input_523_cast_fp16 = transpose(perm = input_523_perm_0, x = x_255_cast_fp16)[name = string("transpose_275")]; + tensor x_257_cast_fp16 = layer_norm(axes = x_257_axes_0, beta = encoder_layers_9_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_9_conv_batch_norm_weight_to_fp16, x = input_523_cast_fp16)[name = string("x_257_cast_fp16")]; + tensor input_525_perm_0 = const()[name = string("input_525_perm_0"), val = tensor([0, 2, 1])]; + tensor input_525_cast_fp16 = transpose(perm = input_525_perm_0, x = x_257_cast_fp16)[name = string("transpose_274")]; + tensor input_527_cast_fp16 = silu(x = input_525_cast_fp16)[name = string("input_527_cast_fp16")]; + string x_259_pad_type_0 = const()[name = string("x_259_pad_type_0"), val = string("valid")]; + tensor x_259_strides_0 = const()[name = string("x_259_strides_0"), val = tensor([1])]; + tensor x_259_pad_0 = const()[name = string("x_259_pad_0"), val = tensor([0, 0])]; + tensor x_259_dilations_0 = const()[name = string("x_259_dilations_0"), val = tensor([1])]; + int32 x_259_groups_0 = const()[name = string("x_259_groups_0"), val = int32(1)]; + tensor encoder_layers_9_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(201211776))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202260416))))[name = string("encoder_layers_9_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_259_cast_fp16 = conv(dilations = x_259_dilations_0, groups = x_259_groups_0, pad = x_259_pad_0, pad_type = x_259_pad_type_0, strides = x_259_strides_0, weight = encoder_layers_9_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_527_cast_fp16)[name = string("x_259_cast_fp16")]; + tensor input_529_perm_0 = const()[name = string("input_529_perm_0"), val = tensor([0, 2, 1])]; + tensor input_529_cast_fp16 = transpose(perm = input_529_perm_0, x = x_259_cast_fp16)[name = string("transpose_273")]; + tensor input_531_cast_fp16 = add(x = input_515_cast_fp16, y = input_529_cast_fp16)[name = string("input_531_cast_fp16")]; + tensor input_533_axes_0 = const()[name = string("input_533_axes_0"), val = tensor([-1])]; + tensor encoder_layers_9_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_9_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202262528)))]; + tensor encoder_layers_9_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_9_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202264640)))]; + tensor input_533_cast_fp16 = layer_norm(axes = input_533_axes_0, beta = encoder_layers_9_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_9_norm_feed_forward2_weight_to_fp16, x = input_531_cast_fp16)[name = string("input_533_cast_fp16")]; + tensor encoder_layers_9_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202266752))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(205412544))))[name = string("encoder_layers_9_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_9_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_9_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(205412736)))]; + tensor linear_89_cast_fp16 = linear(bias = encoder_layers_9_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_9_feed_forward2_linear1_weight_to_fp16_palettized, x = input_533_cast_fp16)[name = string("linear_89_cast_fp16")]; + tensor input_537_cast_fp16 = silu(x = linear_89_cast_fp16)[name = string("input_537_cast_fp16")]; + tensor encoder_layers_9_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(205420992))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(208566784))))[name = string("encoder_layers_9_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_9_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_9_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(208566976)))]; + tensor linear_90_cast_fp16 = linear(bias = encoder_layers_9_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_9_feed_forward2_linear2_weight_to_fp16_palettized, x = input_537_cast_fp16)[name = string("linear_90_cast_fp16")]; + fp16 var_2548_to_fp16 = const()[name = string("op_2548_to_fp16"), val = fp16(0x1p-1)]; + tensor var_2549_cast_fp16 = mul(x = linear_90_cast_fp16, y = var_2548_to_fp16)[name = string("op_2549_cast_fp16")]; + tensor input_543_cast_fp16 = add(x = input_531_cast_fp16, y = var_2549_cast_fp16)[name = string("input_543_cast_fp16")]; + tensor input_545_axes_0 = const()[name = string("input_545_axes_0"), val = tensor([-1])]; + tensor encoder_layers_9_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_9_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(208569088)))]; + tensor encoder_layers_9_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_9_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(208571200)))]; + tensor input_545_cast_fp16 = layer_norm(axes = input_545_axes_0, beta = encoder_layers_9_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_9_norm_out_weight_to_fp16, x = input_543_cast_fp16)[name = string("input_545_cast_fp16")]; + tensor cache_41_begin_0 = const()[name = string("cache_41_begin_0"), val = tensor([10, 0, 0, 0])]; + tensor cache_41_end_0 = const()[name = string("cache_41_end_0"), val = tensor([11, 1, 42, 1024])]; + tensor cache_41_end_mask_0 = const()[name = string("cache_41_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_41_squeeze_mask_0 = const()[name = string("cache_41_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_41_cast_fp16 = slice_by_index(begin = cache_41_begin_0, end = cache_41_end_0, end_mask = cache_41_end_mask_0, squeeze_mask = cache_41_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_41_cast_fp16")]; + tensor cache_43_begin_0 = const()[name = string("cache_43_begin_0"), val = tensor([10, 0, 0, 0])]; + tensor cache_43_end_0 = const()[name = string("cache_43_end_0"), val = tensor([11, 1, 1024, 8])]; + tensor cache_43_end_mask_0 = const()[name = string("cache_43_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_43_squeeze_mask_0 = const()[name = string("cache_43_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_43_cast_fp16 = slice_by_index(begin = cache_43_begin_0, end = cache_43_end_0, end_mask = cache_43_end_mask_0, squeeze_mask = cache_43_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_43_cast_fp16")]; + tensor input_547_axes_0 = const()[name = string("input_547_axes_0"), val = tensor([-1])]; + tensor encoder_layers_10_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_10_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(208573312)))]; + tensor encoder_layers_10_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_10_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(208575424)))]; + tensor input_547_cast_fp16 = layer_norm(axes = input_547_axes_0, beta = encoder_layers_10_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_10_norm_feed_forward1_weight_to_fp16, x = input_545_cast_fp16)[name = string("input_547_cast_fp16")]; + tensor encoder_layers_10_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(208577536))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211723328))))[name = string("encoder_layers_10_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_10_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_10_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211723520)))]; + tensor linear_91_cast_fp16 = linear(bias = encoder_layers_10_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_10_feed_forward1_linear1_weight_to_fp16_palettized, x = input_547_cast_fp16)[name = string("linear_91_cast_fp16")]; + tensor input_551_cast_fp16 = silu(x = linear_91_cast_fp16)[name = string("input_551_cast_fp16")]; + tensor encoder_layers_10_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211731776))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(214877568))))[name = string("encoder_layers_10_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_10_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_10_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(214877760)))]; + tensor linear_92_cast_fp16 = linear(bias = encoder_layers_10_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_10_feed_forward1_linear2_weight_to_fp16_palettized, x = input_551_cast_fp16)[name = string("linear_92_cast_fp16")]; + fp16 var_2585_to_fp16 = const()[name = string("op_2585_to_fp16"), val = fp16(0x1p-1)]; + tensor var_2586_cast_fp16 = mul(x = linear_92_cast_fp16, y = var_2585_to_fp16)[name = string("op_2586_cast_fp16")]; + tensor input_557_cast_fp16 = add(x = input_545_cast_fp16, y = var_2586_cast_fp16)[name = string("input_557_cast_fp16")]; + tensor key_21_axes_0 = const()[name = string("key_21_axes_0"), val = tensor([-1])]; + tensor encoder_layers_10_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_10_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(214879872)))]; + tensor encoder_layers_10_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_10_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(214881984)))]; + tensor key_21_cast_fp16 = layer_norm(axes = key_21_axes_0, beta = encoder_layers_10_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_10_norm_self_att_weight_to_fp16, x = input_557_cast_fp16)[name = string("key_21_cast_fp16")]; + bool input_559_interleave_0 = const()[name = string("input_559_interleave_0"), val = bool(false)]; + tensor input_559_cast_fp16 = concat(axis = var_68, interleave = input_559_interleave_0, values = (cache_41_cast_fp16, key_21_cast_fp16))[name = string("input_559_cast_fp16")]; + tensor var_2608_begin_0 = const()[name = string("op_2608_begin_0"), val = tensor([0, 7, 0])]; + tensor var_2608_end_0 = const()[name = string("op_2608_end_0"), val = tensor([1, 42, 1024])]; + tensor var_2608_end_mask_0 = const()[name = string("op_2608_end_mask_0"), val = tensor([true, true, true])]; + tensor var_2608_cast_fp16 = slice_by_index(begin = var_2608_begin_0, end = var_2608_end_0, end_mask = var_2608_end_mask_0, x = cache_41_cast_fp16)[name = string("op_2608_cast_fp16")]; + bool var_2614_interleave_0 = const()[name = string("op_2614_interleave_0"), val = bool(false)]; + tensor var_2614_cast_fp16 = concat(axis = var_68, interleave = var_2614_interleave_0, values = (var_2608_cast_fp16, key_21_cast_fp16))[name = string("op_2614_cast_fp16")]; + tensor encoder_layers_10_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(214884096))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215670592))))[name = string("encoder_layers_10_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_10_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_10_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215670784)))]; + tensor linear_93_cast_fp16 = linear(bias = encoder_layers_10_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_10_self_attn_linear_q_weight_to_fp16_palettized, x = key_21_cast_fp16)[name = string("linear_93_cast_fp16")]; + tensor var_2619 = const()[name = string("op_2619"), val = tensor([1, -1, 8, 128])]; + tensor q_61_cast_fp16 = reshape(shape = var_2619, x = linear_93_cast_fp16)[name = string("q_61_cast_fp16")]; + tensor encoder_layers_10_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215672896))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216459392))))[name = string("encoder_layers_10_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_10_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_10_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216459584)))]; + tensor linear_94_cast_fp16 = linear(bias = encoder_layers_10_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_10_self_attn_linear_k_weight_to_fp16_palettized, x = input_559_cast_fp16)[name = string("linear_94_cast_fp16")]; + tensor var_2624 = const()[name = string("op_2624"), val = tensor([1, -1, 8, 128])]; + tensor k_41_cast_fp16 = reshape(shape = var_2624, x = linear_94_cast_fp16)[name = string("k_41_cast_fp16")]; + tensor encoder_layers_10_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216461696))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217248192))))[name = string("encoder_layers_10_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_10_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_10_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217248384)))]; + tensor linear_95_cast_fp16 = linear(bias = encoder_layers_10_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_10_self_attn_linear_v_weight_to_fp16_palettized, x = input_559_cast_fp16)[name = string("linear_95_cast_fp16")]; + tensor var_2629 = const()[name = string("op_2629"), val = tensor([1, -1, 8, 128])]; + tensor v_21_cast_fp16 = reshape(shape = var_2629, x = linear_95_cast_fp16)[name = string("v_21_cast_fp16")]; + tensor value_29_perm_0 = const()[name = string("value_29_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_10_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_10_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217250496)))]; + tensor var_2642_cast_fp16 = add(x = q_61_cast_fp16, y = encoder_layers_10_self_attn_pos_bias_u_to_fp16)[name = string("op_2642_cast_fp16")]; + tensor encoder_layers_10_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_10_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217252608)))]; + tensor var_2644_cast_fp16 = add(x = q_61_cast_fp16, y = encoder_layers_10_self_attn_pos_bias_v_to_fp16)[name = string("op_2644_cast_fp16")]; + tensor q_with_bias_v_21_perm_0 = const()[name = string("q_with_bias_v_21_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_267_transpose_x_0 = const()[name = string("x_267_transpose_x_0"), val = bool(false)]; + bool x_267_transpose_y_0 = const()[name = string("x_267_transpose_y_0"), val = bool(false)]; + tensor op_2646_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217254720))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217354112))))[name = string("op_2646_to_fp16_quantized")]; + tensor q_with_bias_v_21_cast_fp16 = transpose(perm = q_with_bias_v_21_perm_0, x = var_2644_cast_fp16)[name = string("transpose_272")]; + tensor x_267_cast_fp16 = matmul(transpose_x = x_267_transpose_x_0, transpose_y = x_267_transpose_y_0, x = q_with_bias_v_21_cast_fp16, y = op_2646_to_fp16_quantized)[name = string("x_267_cast_fp16")]; + tensor x_269_pad_0 = const()[name = string("x_269_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_269_mode_0 = const()[name = string("x_269_mode_0"), val = string("constant")]; + fp16 const_209_to_fp16 = const()[name = string("const_209_to_fp16"), val = fp16(0x0p+0)]; + tensor x_269_cast_fp16 = pad(constant_val = const_209_to_fp16, mode = x_269_mode_0, pad = x_269_pad_0, x = x_267_cast_fp16)[name = string("x_269_cast_fp16")]; + tensor var_2654 = const()[name = string("op_2654"), val = tensor([1, 8, -1, 7])]; + tensor x_271_cast_fp16 = reshape(shape = var_2654, x = x_269_cast_fp16)[name = string("x_271_cast_fp16")]; + tensor var_2658_begin_0 = const()[name = string("op_2658_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_2658_end_0 = const()[name = string("op_2658_end_0"), val = tensor([1, 8, 98, 7])]; + tensor var_2658_end_mask_0 = const()[name = string("op_2658_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_2658_cast_fp16 = slice_by_index(begin = var_2658_begin_0, end = var_2658_end_0, end_mask = var_2658_end_mask_0, x = x_271_cast_fp16)[name = string("op_2658_cast_fp16")]; + tensor var_2659 = const()[name = string("op_2659"), val = tensor([1, 8, 7, 97])]; + tensor matrix_bd_41_cast_fp16 = reshape(shape = var_2659, x = var_2658_cast_fp16)[name = string("matrix_bd_41_cast_fp16")]; + bool matrix_ac_21_transpose_x_0 = const()[name = string("matrix_ac_21_transpose_x_0"), val = bool(false)]; + bool matrix_ac_21_transpose_y_0 = const()[name = string("matrix_ac_21_transpose_y_0"), val = bool(false)]; + tensor transpose_116_perm_0 = const()[name = string("transpose_116_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_117_perm_0 = const()[name = string("transpose_117_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_117 = transpose(perm = transpose_117_perm_0, x = k_41_cast_fp16)[name = string("transpose_270")]; + tensor transpose_116 = transpose(perm = transpose_116_perm_0, x = var_2642_cast_fp16)[name = string("transpose_271")]; + tensor matrix_ac_21_cast_fp16 = matmul(transpose_x = matrix_ac_21_transpose_x_0, transpose_y = matrix_ac_21_transpose_y_0, x = transpose_116, y = transpose_117)[name = string("matrix_ac_21_cast_fp16")]; + tensor matrix_bd_43_begin_0 = const()[name = string("matrix_bd_43_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_43_end_0 = const()[name = string("matrix_bd_43_end_0"), val = tensor([1, 8, 7, 49])]; + tensor matrix_bd_43_end_mask_0 = const()[name = string("matrix_bd_43_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_43_cast_fp16 = slice_by_index(begin = matrix_bd_43_begin_0, end = matrix_bd_43_end_0, end_mask = matrix_bd_43_end_mask_0, x = matrix_bd_41_cast_fp16)[name = string("matrix_bd_43_cast_fp16")]; + tensor var_2668_cast_fp16 = add(x = matrix_ac_21_cast_fp16, y = matrix_bd_43_cast_fp16)[name = string("op_2668_cast_fp16")]; + fp16 _inversed_scores_41_y_0_to_fp16 = const()[name = string("_inversed_scores_41_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_41_cast_fp16 = mul(x = var_2668_cast_fp16, y = _inversed_scores_41_y_0_to_fp16)[name = string("_inversed_scores_41_cast_fp16")]; + tensor scores_43_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_41_cast_fp16, cond = mask_11)[name = string("scores_43_cast_fp16")]; + tensor var_2674_cast_fp16 = softmax(axis = var_59, x = scores_43_cast_fp16)[name = string("op_2674_cast_fp16")]; + tensor input_561_cast_fp16 = select(a = var_44_to_fp16, b = var_2674_cast_fp16, cond = mask_11)[name = string("input_561_cast_fp16")]; + bool x_273_transpose_x_0 = const()[name = string("x_273_transpose_x_0"), val = bool(false)]; + bool x_273_transpose_y_0 = const()[name = string("x_273_transpose_y_0"), val = bool(false)]; + tensor value_29_cast_fp16 = transpose(perm = value_29_perm_0, x = v_21_cast_fp16)[name = string("transpose_269")]; + tensor x_273_cast_fp16 = matmul(transpose_x = x_273_transpose_x_0, transpose_y = x_273_transpose_y_0, x = input_561_cast_fp16, y = value_29_cast_fp16)[name = string("x_273_cast_fp16")]; + tensor var_2678_perm_0 = const()[name = string("op_2678_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_2679 = const()[name = string("op_2679"), val = tensor([1, -1, 1024])]; + tensor var_2678_cast_fp16 = transpose(perm = var_2678_perm_0, x = x_273_cast_fp16)[name = string("transpose_268")]; + tensor input_563_cast_fp16 = reshape(shape = var_2679, x = var_2678_cast_fp16)[name = string("input_563_cast_fp16")]; + tensor encoder_layers_10_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217354432))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218140928))))[name = string("encoder_layers_10_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_10_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_10_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218141120)))]; + tensor linear_97_cast_fp16 = linear(bias = encoder_layers_10_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_10_self_attn_linear_out_weight_to_fp16_palettized, x = input_563_cast_fp16)[name = string("linear_97_cast_fp16")]; + tensor input_567_cast_fp16 = add(x = input_557_cast_fp16, y = linear_97_cast_fp16)[name = string("input_567_cast_fp16")]; + tensor x_277_axes_0 = const()[name = string("x_277_axes_0"), val = tensor([-1])]; + tensor encoder_layers_10_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_10_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218143232)))]; + tensor encoder_layers_10_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_10_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218145344)))]; + tensor x_277_cast_fp16 = layer_norm(axes = x_277_axes_0, beta = encoder_layers_10_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_10_norm_conv_weight_to_fp16, x = input_567_cast_fp16)[name = string("x_277_cast_fp16")]; + tensor input_569_perm_0 = const()[name = string("input_569_perm_0"), val = tensor([0, 2, 1])]; + string input_571_pad_type_0 = const()[name = string("input_571_pad_type_0"), val = string("valid")]; + tensor input_571_strides_0 = const()[name = string("input_571_strides_0"), val = tensor([1])]; + tensor input_571_pad_0 = const()[name = string("input_571_pad_0"), val = tensor([0, 0])]; + tensor input_571_dilations_0 = const()[name = string("input_571_dilations_0"), val = tensor([1])]; + int32 input_571_groups_0 = const()[name = string("input_571_groups_0"), val = int32(1)]; + tensor encoder_layers_10_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218147456))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220244672))))[name = string("encoder_layers_10_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_569_cast_fp16 = transpose(perm = input_569_perm_0, x = x_277_cast_fp16)[name = string("transpose_267")]; + tensor input_571_cast_fp16 = conv(dilations = input_571_dilations_0, groups = input_571_groups_0, pad = input_571_pad_0, pad_type = input_571_pad_type_0, strides = input_571_strides_0, weight = encoder_layers_10_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_569_cast_fp16)[name = string("input_571_cast_fp16")]; + int32 x_279_split_num_splits_0 = const()[name = string("x_279_split_num_splits_0"), val = int32(2)]; + int32 x_279_split_axis_0 = const()[name = string("x_279_split_axis_0"), val = int32(1)]; + tensor x_279_split_cast_fp16_0, tensor x_279_split_cast_fp16_1 = split(axis = x_279_split_axis_0, num_splits = x_279_split_num_splits_0, x = input_571_cast_fp16)[name = string("x_279_split_cast_fp16")]; + tensor x_279_split_1_sigmoid_cast_fp16 = sigmoid(x = x_279_split_cast_fp16_1)[name = string("x_279_split_1_sigmoid_cast_fp16")]; + tensor x_279_cast_fp16 = mul(x = x_279_split_cast_fp16_0, y = x_279_split_1_sigmoid_cast_fp16)[name = string("x_279_cast_fp16")]; + tensor input_573_cast_fp16 = select(a = var_44_to_fp16, b = x_279_cast_fp16, cond = var_575)[name = string("input_573_cast_fp16")]; + bool new_x_43_interleave_0 = const()[name = string("new_x_43_interleave_0"), val = bool(false)]; + tensor new_x_43_cast_fp16 = concat(axis = var_59, interleave = new_x_43_interleave_0, values = (cache_43_cast_fp16, input_573_cast_fp16))[name = string("new_x_43_cast_fp16")]; + tensor var_2718_begin_0 = const()[name = string("op_2718_begin_0"), val = tensor([0, 0, 7])]; + tensor var_2718_end_0 = const()[name = string("op_2718_end_0"), val = tensor([1, 1024, 15])]; + tensor var_2718_end_mask_0 = const()[name = string("op_2718_end_mask_0"), val = tensor([true, true, true])]; + tensor var_2718_cast_fp16 = slice_by_index(begin = var_2718_begin_0, end = var_2718_end_0, end_mask = var_2718_end_mask_0, x = new_x_43_cast_fp16)[name = string("op_2718_cast_fp16")]; + string x_281_pad_type_0 = const()[name = string("x_281_pad_type_0"), val = string("valid")]; + int32 x_281_groups_0 = const()[name = string("x_281_groups_0"), val = int32(1024)]; + tensor x_281_strides_0 = const()[name = string("x_281_strides_0"), val = tensor([1])]; + tensor x_281_pad_0 = const()[name = string("x_281_pad_0"), val = tensor([0, 0])]; + tensor x_281_dilations_0 = const()[name = string("x_281_dilations_0"), val = tensor([1])]; + tensor encoder_layers_10_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220248832))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220258112))))[name = string("encoder_layers_10_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_281_cast_fp16 = conv(dilations = x_281_dilations_0, groups = x_281_groups_0, pad = x_281_pad_0, pad_type = x_281_pad_type_0, strides = x_281_strides_0, weight = encoder_layers_10_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_43_cast_fp16)[name = string("x_281_cast_fp16")]; + tensor input_575_perm_0 = const()[name = string("input_575_perm_0"), val = tensor([0, 2, 1])]; + tensor x_283_axes_0 = const()[name = string("x_283_axes_0"), val = tensor([-1])]; + tensor encoder_layers_10_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_10_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220260224)))]; + tensor encoder_layers_10_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_10_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220262336)))]; + tensor input_575_cast_fp16 = transpose(perm = input_575_perm_0, x = x_281_cast_fp16)[name = string("transpose_266")]; + tensor x_283_cast_fp16 = layer_norm(axes = x_283_axes_0, beta = encoder_layers_10_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_10_conv_batch_norm_weight_to_fp16, x = input_575_cast_fp16)[name = string("x_283_cast_fp16")]; + tensor input_577_perm_0 = const()[name = string("input_577_perm_0"), val = tensor([0, 2, 1])]; + tensor input_577_cast_fp16 = transpose(perm = input_577_perm_0, x = x_283_cast_fp16)[name = string("transpose_265")]; + tensor input_579_cast_fp16 = silu(x = input_577_cast_fp16)[name = string("input_579_cast_fp16")]; + string x_285_pad_type_0 = const()[name = string("x_285_pad_type_0"), val = string("valid")]; + tensor x_285_strides_0 = const()[name = string("x_285_strides_0"), val = tensor([1])]; + tensor x_285_pad_0 = const()[name = string("x_285_pad_0"), val = tensor([0, 0])]; + tensor x_285_dilations_0 = const()[name = string("x_285_dilations_0"), val = tensor([1])]; + int32 x_285_groups_0 = const()[name = string("x_285_groups_0"), val = int32(1)]; + tensor encoder_layers_10_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220264448))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221313088))))[name = string("encoder_layers_10_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_285_cast_fp16 = conv(dilations = x_285_dilations_0, groups = x_285_groups_0, pad = x_285_pad_0, pad_type = x_285_pad_type_0, strides = x_285_strides_0, weight = encoder_layers_10_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_579_cast_fp16)[name = string("x_285_cast_fp16")]; + tensor input_581_perm_0 = const()[name = string("input_581_perm_0"), val = tensor([0, 2, 1])]; + tensor input_581_cast_fp16 = transpose(perm = input_581_perm_0, x = x_285_cast_fp16)[name = string("transpose_264")]; + tensor input_583_cast_fp16 = add(x = input_567_cast_fp16, y = input_581_cast_fp16)[name = string("input_583_cast_fp16")]; + tensor input_585_axes_0 = const()[name = string("input_585_axes_0"), val = tensor([-1])]; + tensor encoder_layers_10_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_10_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221315200)))]; + tensor encoder_layers_10_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_10_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221317312)))]; + tensor input_585_cast_fp16 = layer_norm(axes = input_585_axes_0, beta = encoder_layers_10_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_10_norm_feed_forward2_weight_to_fp16, x = input_583_cast_fp16)[name = string("input_585_cast_fp16")]; + tensor encoder_layers_10_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221319424))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(224465216))))[name = string("encoder_layers_10_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_10_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_10_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(224465408)))]; + tensor linear_98_cast_fp16 = linear(bias = encoder_layers_10_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_10_feed_forward2_linear1_weight_to_fp16_palettized, x = input_585_cast_fp16)[name = string("linear_98_cast_fp16")]; + tensor input_589_cast_fp16 = silu(x = linear_98_cast_fp16)[name = string("input_589_cast_fp16")]; + tensor encoder_layers_10_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(224473664))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227619456))))[name = string("encoder_layers_10_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_10_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_10_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227619648)))]; + tensor linear_99_cast_fp16 = linear(bias = encoder_layers_10_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_10_feed_forward2_linear2_weight_to_fp16_palettized, x = input_589_cast_fp16)[name = string("linear_99_cast_fp16")]; + fp16 var_2761_to_fp16 = const()[name = string("op_2761_to_fp16"), val = fp16(0x1p-1)]; + tensor var_2762_cast_fp16 = mul(x = linear_99_cast_fp16, y = var_2761_to_fp16)[name = string("op_2762_cast_fp16")]; + tensor input_595_cast_fp16 = add(x = input_583_cast_fp16, y = var_2762_cast_fp16)[name = string("input_595_cast_fp16")]; + tensor input_597_axes_0 = const()[name = string("input_597_axes_0"), val = tensor([-1])]; + tensor encoder_layers_10_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_10_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227621760)))]; + tensor encoder_layers_10_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_10_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227623872)))]; + tensor input_597_cast_fp16 = layer_norm(axes = input_597_axes_0, beta = encoder_layers_10_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_10_norm_out_weight_to_fp16, x = input_595_cast_fp16)[name = string("input_597_cast_fp16")]; + tensor cache_45_begin_0 = const()[name = string("cache_45_begin_0"), val = tensor([11, 0, 0, 0])]; + tensor cache_45_end_0 = const()[name = string("cache_45_end_0"), val = tensor([12, 1, 42, 1024])]; + tensor cache_45_end_mask_0 = const()[name = string("cache_45_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_45_squeeze_mask_0 = const()[name = string("cache_45_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_45_cast_fp16 = slice_by_index(begin = cache_45_begin_0, end = cache_45_end_0, end_mask = cache_45_end_mask_0, squeeze_mask = cache_45_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_45_cast_fp16")]; + tensor cache_47_begin_0 = const()[name = string("cache_47_begin_0"), val = tensor([11, 0, 0, 0])]; + tensor cache_47_end_0 = const()[name = string("cache_47_end_0"), val = tensor([12, 1, 1024, 8])]; + tensor cache_47_end_mask_0 = const()[name = string("cache_47_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_47_squeeze_mask_0 = const()[name = string("cache_47_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_47_cast_fp16 = slice_by_index(begin = cache_47_begin_0, end = cache_47_end_0, end_mask = cache_47_end_mask_0, squeeze_mask = cache_47_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_47_cast_fp16")]; + tensor input_599_axes_0 = const()[name = string("input_599_axes_0"), val = tensor([-1])]; + tensor encoder_layers_11_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_11_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227625984)))]; + tensor encoder_layers_11_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_11_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227628096)))]; + tensor input_599_cast_fp16 = layer_norm(axes = input_599_axes_0, beta = encoder_layers_11_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_11_norm_feed_forward1_weight_to_fp16, x = input_597_cast_fp16)[name = string("input_599_cast_fp16")]; + tensor encoder_layers_11_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227630208))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(230776000))))[name = string("encoder_layers_11_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_11_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_11_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(230776192)))]; + tensor linear_100_cast_fp16 = linear(bias = encoder_layers_11_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_11_feed_forward1_linear1_weight_to_fp16_palettized, x = input_599_cast_fp16)[name = string("linear_100_cast_fp16")]; + tensor input_603_cast_fp16 = silu(x = linear_100_cast_fp16)[name = string("input_603_cast_fp16")]; + tensor encoder_layers_11_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(230784448))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(233930240))))[name = string("encoder_layers_11_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_11_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_11_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(233930432)))]; + tensor linear_101_cast_fp16 = linear(bias = encoder_layers_11_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_11_feed_forward1_linear2_weight_to_fp16_palettized, x = input_603_cast_fp16)[name = string("linear_101_cast_fp16")]; + fp16 var_2798_to_fp16 = const()[name = string("op_2798_to_fp16"), val = fp16(0x1p-1)]; + tensor var_2799_cast_fp16 = mul(x = linear_101_cast_fp16, y = var_2798_to_fp16)[name = string("op_2799_cast_fp16")]; + tensor input_609_cast_fp16 = add(x = input_597_cast_fp16, y = var_2799_cast_fp16)[name = string("input_609_cast_fp16")]; + tensor key_23_axes_0 = const()[name = string("key_23_axes_0"), val = tensor([-1])]; + tensor encoder_layers_11_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_11_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(233932544)))]; + tensor encoder_layers_11_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_11_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(233934656)))]; + tensor key_23_cast_fp16 = layer_norm(axes = key_23_axes_0, beta = encoder_layers_11_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_11_norm_self_att_weight_to_fp16, x = input_609_cast_fp16)[name = string("key_23_cast_fp16")]; + bool input_611_interleave_0 = const()[name = string("input_611_interleave_0"), val = bool(false)]; + tensor input_611_cast_fp16 = concat(axis = var_68, interleave = input_611_interleave_0, values = (cache_45_cast_fp16, key_23_cast_fp16))[name = string("input_611_cast_fp16")]; + tensor var_2821_begin_0 = const()[name = string("op_2821_begin_0"), val = tensor([0, 7, 0])]; + tensor var_2821_end_0 = const()[name = string("op_2821_end_0"), val = tensor([1, 42, 1024])]; + tensor var_2821_end_mask_0 = const()[name = string("op_2821_end_mask_0"), val = tensor([true, true, true])]; + tensor var_2821_cast_fp16 = slice_by_index(begin = var_2821_begin_0, end = var_2821_end_0, end_mask = var_2821_end_mask_0, x = cache_45_cast_fp16)[name = string("op_2821_cast_fp16")]; + bool var_2827_interleave_0 = const()[name = string("op_2827_interleave_0"), val = bool(false)]; + tensor var_2827_cast_fp16 = concat(axis = var_68, interleave = var_2827_interleave_0, values = (var_2821_cast_fp16, key_23_cast_fp16))[name = string("op_2827_cast_fp16")]; + tensor encoder_layers_11_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(233936768))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234723264))))[name = string("encoder_layers_11_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_11_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_11_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234723456)))]; + tensor linear_102_cast_fp16 = linear(bias = encoder_layers_11_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_11_self_attn_linear_q_weight_to_fp16_palettized, x = key_23_cast_fp16)[name = string("linear_102_cast_fp16")]; + tensor var_2832 = const()[name = string("op_2832"), val = tensor([1, -1, 8, 128])]; + tensor q_67_cast_fp16 = reshape(shape = var_2832, x = linear_102_cast_fp16)[name = string("q_67_cast_fp16")]; + tensor encoder_layers_11_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234725568))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(235512064))))[name = string("encoder_layers_11_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_11_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_11_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(235512256)))]; + tensor linear_103_cast_fp16 = linear(bias = encoder_layers_11_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_11_self_attn_linear_k_weight_to_fp16_palettized, x = input_611_cast_fp16)[name = string("linear_103_cast_fp16")]; + tensor var_2837 = const()[name = string("op_2837"), val = tensor([1, -1, 8, 128])]; + tensor k_45_cast_fp16 = reshape(shape = var_2837, x = linear_103_cast_fp16)[name = string("k_45_cast_fp16")]; + tensor encoder_layers_11_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(235514368))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236300864))))[name = string("encoder_layers_11_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_11_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_11_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236301056)))]; + tensor linear_104_cast_fp16 = linear(bias = encoder_layers_11_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_11_self_attn_linear_v_weight_to_fp16_palettized, x = input_611_cast_fp16)[name = string("linear_104_cast_fp16")]; + tensor var_2842 = const()[name = string("op_2842"), val = tensor([1, -1, 8, 128])]; + tensor v_23_cast_fp16 = reshape(shape = var_2842, x = linear_104_cast_fp16)[name = string("v_23_cast_fp16")]; + tensor value_31_perm_0 = const()[name = string("value_31_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_11_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_11_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236303168)))]; + tensor var_2855_cast_fp16 = add(x = q_67_cast_fp16, y = encoder_layers_11_self_attn_pos_bias_u_to_fp16)[name = string("op_2855_cast_fp16")]; + tensor encoder_layers_11_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_11_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236305280)))]; + tensor var_2857_cast_fp16 = add(x = q_67_cast_fp16, y = encoder_layers_11_self_attn_pos_bias_v_to_fp16)[name = string("op_2857_cast_fp16")]; + tensor q_with_bias_v_23_perm_0 = const()[name = string("q_with_bias_v_23_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_293_transpose_x_0 = const()[name = string("x_293_transpose_x_0"), val = bool(false)]; + bool x_293_transpose_y_0 = const()[name = string("x_293_transpose_y_0"), val = bool(false)]; + tensor op_2859_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236307392))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236406784))))[name = string("op_2859_to_fp16_quantized")]; + tensor q_with_bias_v_23_cast_fp16 = transpose(perm = q_with_bias_v_23_perm_0, x = var_2857_cast_fp16)[name = string("transpose_263")]; + tensor x_293_cast_fp16 = matmul(transpose_x = x_293_transpose_x_0, transpose_y = x_293_transpose_y_0, x = q_with_bias_v_23_cast_fp16, y = op_2859_to_fp16_quantized)[name = string("x_293_cast_fp16")]; + tensor x_295_pad_0 = const()[name = string("x_295_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_295_mode_0 = const()[name = string("x_295_mode_0"), val = string("constant")]; + fp16 const_222_to_fp16 = const()[name = string("const_222_to_fp16"), val = fp16(0x0p+0)]; + tensor x_295_cast_fp16 = pad(constant_val = const_222_to_fp16, mode = x_295_mode_0, pad = x_295_pad_0, x = x_293_cast_fp16)[name = string("x_295_cast_fp16")]; + tensor var_2867 = const()[name = string("op_2867"), val = tensor([1, 8, -1, 7])]; + tensor x_297_cast_fp16 = reshape(shape = var_2867, x = x_295_cast_fp16)[name = string("x_297_cast_fp16")]; + tensor var_2871_begin_0 = const()[name = string("op_2871_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_2871_end_0 = const()[name = string("op_2871_end_0"), val = tensor([1, 8, 98, 7])]; + tensor var_2871_end_mask_0 = const()[name = string("op_2871_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_2871_cast_fp16 = slice_by_index(begin = var_2871_begin_0, end = var_2871_end_0, end_mask = var_2871_end_mask_0, x = x_297_cast_fp16)[name = string("op_2871_cast_fp16")]; + tensor var_2872 = const()[name = string("op_2872"), val = tensor([1, 8, 7, 97])]; + tensor matrix_bd_45_cast_fp16 = reshape(shape = var_2872, x = var_2871_cast_fp16)[name = string("matrix_bd_45_cast_fp16")]; + bool matrix_ac_23_transpose_x_0 = const()[name = string("matrix_ac_23_transpose_x_0"), val = bool(false)]; + bool matrix_ac_23_transpose_y_0 = const()[name = string("matrix_ac_23_transpose_y_0"), val = bool(false)]; + tensor transpose_118_perm_0 = const()[name = string("transpose_118_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_119_perm_0 = const()[name = string("transpose_119_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_119 = transpose(perm = transpose_119_perm_0, x = k_45_cast_fp16)[name = string("transpose_261")]; + tensor transpose_118 = transpose(perm = transpose_118_perm_0, x = var_2855_cast_fp16)[name = string("transpose_262")]; + tensor matrix_ac_23_cast_fp16 = matmul(transpose_x = matrix_ac_23_transpose_x_0, transpose_y = matrix_ac_23_transpose_y_0, x = transpose_118, y = transpose_119)[name = string("matrix_ac_23_cast_fp16")]; + tensor matrix_bd_47_begin_0 = const()[name = string("matrix_bd_47_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_47_end_0 = const()[name = string("matrix_bd_47_end_0"), val = tensor([1, 8, 7, 49])]; + tensor matrix_bd_47_end_mask_0 = const()[name = string("matrix_bd_47_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_47_cast_fp16 = slice_by_index(begin = matrix_bd_47_begin_0, end = matrix_bd_47_end_0, end_mask = matrix_bd_47_end_mask_0, x = matrix_bd_45_cast_fp16)[name = string("matrix_bd_47_cast_fp16")]; + tensor var_2881_cast_fp16 = add(x = matrix_ac_23_cast_fp16, y = matrix_bd_47_cast_fp16)[name = string("op_2881_cast_fp16")]; + fp16 _inversed_scores_45_y_0_to_fp16 = const()[name = string("_inversed_scores_45_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_45_cast_fp16 = mul(x = var_2881_cast_fp16, y = _inversed_scores_45_y_0_to_fp16)[name = string("_inversed_scores_45_cast_fp16")]; + tensor scores_47_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_45_cast_fp16, cond = mask_11)[name = string("scores_47_cast_fp16")]; + tensor var_2887_cast_fp16 = softmax(axis = var_59, x = scores_47_cast_fp16)[name = string("op_2887_cast_fp16")]; + tensor input_613_cast_fp16 = select(a = var_44_to_fp16, b = var_2887_cast_fp16, cond = mask_11)[name = string("input_613_cast_fp16")]; + bool x_299_transpose_x_0 = const()[name = string("x_299_transpose_x_0"), val = bool(false)]; + bool x_299_transpose_y_0 = const()[name = string("x_299_transpose_y_0"), val = bool(false)]; + tensor value_31_cast_fp16 = transpose(perm = value_31_perm_0, x = v_23_cast_fp16)[name = string("transpose_260")]; + tensor x_299_cast_fp16 = matmul(transpose_x = x_299_transpose_x_0, transpose_y = x_299_transpose_y_0, x = input_613_cast_fp16, y = value_31_cast_fp16)[name = string("x_299_cast_fp16")]; + tensor var_2891_perm_0 = const()[name = string("op_2891_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_2892 = const()[name = string("op_2892"), val = tensor([1, -1, 1024])]; + tensor var_2891_cast_fp16 = transpose(perm = var_2891_perm_0, x = x_299_cast_fp16)[name = string("transpose_259")]; + tensor input_615_cast_fp16 = reshape(shape = var_2892, x = var_2891_cast_fp16)[name = string("input_615_cast_fp16")]; + tensor encoder_layers_11_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236407104))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(237193600))))[name = string("encoder_layers_11_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_11_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_11_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(237193792)))]; + tensor linear_106_cast_fp16 = linear(bias = encoder_layers_11_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_11_self_attn_linear_out_weight_to_fp16_palettized, x = input_615_cast_fp16)[name = string("linear_106_cast_fp16")]; + tensor input_619_cast_fp16 = add(x = input_609_cast_fp16, y = linear_106_cast_fp16)[name = string("input_619_cast_fp16")]; + tensor x_303_axes_0 = const()[name = string("x_303_axes_0"), val = tensor([-1])]; + tensor encoder_layers_11_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_11_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(237195904)))]; + tensor encoder_layers_11_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_11_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(237198016)))]; + tensor x_303_cast_fp16 = layer_norm(axes = x_303_axes_0, beta = encoder_layers_11_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_11_norm_conv_weight_to_fp16, x = input_619_cast_fp16)[name = string("x_303_cast_fp16")]; + tensor input_621_perm_0 = const()[name = string("input_621_perm_0"), val = tensor([0, 2, 1])]; + string input_623_pad_type_0 = const()[name = string("input_623_pad_type_0"), val = string("valid")]; + tensor input_623_strides_0 = const()[name = string("input_623_strides_0"), val = tensor([1])]; + tensor input_623_pad_0 = const()[name = string("input_623_pad_0"), val = tensor([0, 0])]; + tensor input_623_dilations_0 = const()[name = string("input_623_dilations_0"), val = tensor([1])]; + int32 input_623_groups_0 = const()[name = string("input_623_groups_0"), val = int32(1)]; + tensor encoder_layers_11_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(237200128))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(239297344))))[name = string("encoder_layers_11_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_621_cast_fp16 = transpose(perm = input_621_perm_0, x = x_303_cast_fp16)[name = string("transpose_258")]; + tensor input_623_cast_fp16 = conv(dilations = input_623_dilations_0, groups = input_623_groups_0, pad = input_623_pad_0, pad_type = input_623_pad_type_0, strides = input_623_strides_0, weight = encoder_layers_11_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_621_cast_fp16)[name = string("input_623_cast_fp16")]; + int32 x_305_split_num_splits_0 = const()[name = string("x_305_split_num_splits_0"), val = int32(2)]; + int32 x_305_split_axis_0 = const()[name = string("x_305_split_axis_0"), val = int32(1)]; + tensor x_305_split_cast_fp16_0, tensor x_305_split_cast_fp16_1 = split(axis = x_305_split_axis_0, num_splits = x_305_split_num_splits_0, x = input_623_cast_fp16)[name = string("x_305_split_cast_fp16")]; + tensor x_305_split_1_sigmoid_cast_fp16 = sigmoid(x = x_305_split_cast_fp16_1)[name = string("x_305_split_1_sigmoid_cast_fp16")]; + tensor x_305_cast_fp16 = mul(x = x_305_split_cast_fp16_0, y = x_305_split_1_sigmoid_cast_fp16)[name = string("x_305_cast_fp16")]; + tensor input_625_cast_fp16 = select(a = var_44_to_fp16, b = x_305_cast_fp16, cond = var_575)[name = string("input_625_cast_fp16")]; + bool new_x_47_interleave_0 = const()[name = string("new_x_47_interleave_0"), val = bool(false)]; + tensor new_x_47_cast_fp16 = concat(axis = var_59, interleave = new_x_47_interleave_0, values = (cache_47_cast_fp16, input_625_cast_fp16))[name = string("new_x_47_cast_fp16")]; + tensor var_2931_begin_0 = const()[name = string("op_2931_begin_0"), val = tensor([0, 0, 7])]; + tensor var_2931_end_0 = const()[name = string("op_2931_end_0"), val = tensor([1, 1024, 15])]; + tensor var_2931_end_mask_0 = const()[name = string("op_2931_end_mask_0"), val = tensor([true, true, true])]; + tensor var_2931_cast_fp16 = slice_by_index(begin = var_2931_begin_0, end = var_2931_end_0, end_mask = var_2931_end_mask_0, x = new_x_47_cast_fp16)[name = string("op_2931_cast_fp16")]; + string x_307_pad_type_0 = const()[name = string("x_307_pad_type_0"), val = string("valid")]; + int32 x_307_groups_0 = const()[name = string("x_307_groups_0"), val = int32(1024)]; + tensor x_307_strides_0 = const()[name = string("x_307_strides_0"), val = tensor([1])]; + tensor x_307_pad_0 = const()[name = string("x_307_pad_0"), val = tensor([0, 0])]; + tensor x_307_dilations_0 = const()[name = string("x_307_dilations_0"), val = tensor([1])]; + tensor encoder_layers_11_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(239301504))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(239310784))))[name = string("encoder_layers_11_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_307_cast_fp16 = conv(dilations = x_307_dilations_0, groups = x_307_groups_0, pad = x_307_pad_0, pad_type = x_307_pad_type_0, strides = x_307_strides_0, weight = encoder_layers_11_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_47_cast_fp16)[name = string("x_307_cast_fp16")]; + tensor input_627_perm_0 = const()[name = string("input_627_perm_0"), val = tensor([0, 2, 1])]; + tensor x_309_axes_0 = const()[name = string("x_309_axes_0"), val = tensor([-1])]; + tensor encoder_layers_11_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_11_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(239312896)))]; + tensor encoder_layers_11_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_11_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(239315008)))]; + tensor input_627_cast_fp16 = transpose(perm = input_627_perm_0, x = x_307_cast_fp16)[name = string("transpose_257")]; + tensor x_309_cast_fp16 = layer_norm(axes = x_309_axes_0, beta = encoder_layers_11_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_11_conv_batch_norm_weight_to_fp16, x = input_627_cast_fp16)[name = string("x_309_cast_fp16")]; + tensor input_629_perm_0 = const()[name = string("input_629_perm_0"), val = tensor([0, 2, 1])]; + tensor input_629_cast_fp16 = transpose(perm = input_629_perm_0, x = x_309_cast_fp16)[name = string("transpose_256")]; + tensor input_631_cast_fp16 = silu(x = input_629_cast_fp16)[name = string("input_631_cast_fp16")]; + string x_311_pad_type_0 = const()[name = string("x_311_pad_type_0"), val = string("valid")]; + tensor x_311_strides_0 = const()[name = string("x_311_strides_0"), val = tensor([1])]; + tensor x_311_pad_0 = const()[name = string("x_311_pad_0"), val = tensor([0, 0])]; + tensor x_311_dilations_0 = const()[name = string("x_311_dilations_0"), val = tensor([1])]; + int32 x_311_groups_0 = const()[name = string("x_311_groups_0"), val = int32(1)]; + tensor encoder_layers_11_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(239317120))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240365760))))[name = string("encoder_layers_11_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_311_cast_fp16 = conv(dilations = x_311_dilations_0, groups = x_311_groups_0, pad = x_311_pad_0, pad_type = x_311_pad_type_0, strides = x_311_strides_0, weight = encoder_layers_11_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_631_cast_fp16)[name = string("x_311_cast_fp16")]; + tensor input_633_perm_0 = const()[name = string("input_633_perm_0"), val = tensor([0, 2, 1])]; + tensor input_633_cast_fp16 = transpose(perm = input_633_perm_0, x = x_311_cast_fp16)[name = string("transpose_255")]; + tensor input_635_cast_fp16 = add(x = input_619_cast_fp16, y = input_633_cast_fp16)[name = string("input_635_cast_fp16")]; + tensor input_637_axes_0 = const()[name = string("input_637_axes_0"), val = tensor([-1])]; + tensor encoder_layers_11_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_11_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240367872)))]; + tensor encoder_layers_11_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_11_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240369984)))]; + tensor input_637_cast_fp16 = layer_norm(axes = input_637_axes_0, beta = encoder_layers_11_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_11_norm_feed_forward2_weight_to_fp16, x = input_635_cast_fp16)[name = string("input_637_cast_fp16")]; + tensor encoder_layers_11_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240372096))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(243517888))))[name = string("encoder_layers_11_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_11_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_11_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(243518080)))]; + tensor linear_107_cast_fp16 = linear(bias = encoder_layers_11_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_11_feed_forward2_linear1_weight_to_fp16_palettized, x = input_637_cast_fp16)[name = string("linear_107_cast_fp16")]; + tensor input_641_cast_fp16 = silu(x = linear_107_cast_fp16)[name = string("input_641_cast_fp16")]; + tensor encoder_layers_11_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(243526336))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(246672128))))[name = string("encoder_layers_11_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_11_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_11_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(246672320)))]; + tensor linear_108_cast_fp16 = linear(bias = encoder_layers_11_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_11_feed_forward2_linear2_weight_to_fp16_palettized, x = input_641_cast_fp16)[name = string("linear_108_cast_fp16")]; + fp16 var_2974_to_fp16 = const()[name = string("op_2974_to_fp16"), val = fp16(0x1p-1)]; + tensor var_2975_cast_fp16 = mul(x = linear_108_cast_fp16, y = var_2974_to_fp16)[name = string("op_2975_cast_fp16")]; + tensor input_647_cast_fp16 = add(x = input_635_cast_fp16, y = var_2975_cast_fp16)[name = string("input_647_cast_fp16")]; + tensor input_649_axes_0 = const()[name = string("input_649_axes_0"), val = tensor([-1])]; + tensor encoder_layers_11_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_11_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(246674432)))]; + tensor encoder_layers_11_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_11_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(246676544)))]; + tensor input_649_cast_fp16 = layer_norm(axes = input_649_axes_0, beta = encoder_layers_11_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_11_norm_out_weight_to_fp16, x = input_647_cast_fp16)[name = string("input_649_cast_fp16")]; + tensor cache_49_begin_0 = const()[name = string("cache_49_begin_0"), val = tensor([12, 0, 0, 0])]; + tensor cache_49_end_0 = const()[name = string("cache_49_end_0"), val = tensor([13, 1, 42, 1024])]; + tensor cache_49_end_mask_0 = const()[name = string("cache_49_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_49_squeeze_mask_0 = const()[name = string("cache_49_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_49_cast_fp16 = slice_by_index(begin = cache_49_begin_0, end = cache_49_end_0, end_mask = cache_49_end_mask_0, squeeze_mask = cache_49_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_49_cast_fp16")]; + tensor cache_51_begin_0 = const()[name = string("cache_51_begin_0"), val = tensor([12, 0, 0, 0])]; + tensor cache_51_end_0 = const()[name = string("cache_51_end_0"), val = tensor([13, 1, 1024, 8])]; + tensor cache_51_end_mask_0 = const()[name = string("cache_51_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_51_squeeze_mask_0 = const()[name = string("cache_51_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_51_cast_fp16 = slice_by_index(begin = cache_51_begin_0, end = cache_51_end_0, end_mask = cache_51_end_mask_0, squeeze_mask = cache_51_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_51_cast_fp16")]; + tensor input_651_axes_0 = const()[name = string("input_651_axes_0"), val = tensor([-1])]; + tensor encoder_layers_12_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_12_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(246678656)))]; + tensor encoder_layers_12_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_12_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(246680768)))]; + tensor input_651_cast_fp16 = layer_norm(axes = input_651_axes_0, beta = encoder_layers_12_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_12_norm_feed_forward1_weight_to_fp16, x = input_649_cast_fp16)[name = string("input_651_cast_fp16")]; + tensor encoder_layers_12_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(246682880))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(249828672))))[name = string("encoder_layers_12_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_12_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_12_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(249828864)))]; + tensor linear_109_cast_fp16 = linear(bias = encoder_layers_12_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_12_feed_forward1_linear1_weight_to_fp16_palettized, x = input_651_cast_fp16)[name = string("linear_109_cast_fp16")]; + tensor input_655_cast_fp16 = silu(x = linear_109_cast_fp16)[name = string("input_655_cast_fp16")]; + tensor encoder_layers_12_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(249837120))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(252982912))))[name = string("encoder_layers_12_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_12_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_12_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(252983104)))]; + tensor linear_110_cast_fp16 = linear(bias = encoder_layers_12_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_12_feed_forward1_linear2_weight_to_fp16_palettized, x = input_655_cast_fp16)[name = string("linear_110_cast_fp16")]; + fp16 var_3011_to_fp16 = const()[name = string("op_3011_to_fp16"), val = fp16(0x1p-1)]; + tensor var_3012_cast_fp16 = mul(x = linear_110_cast_fp16, y = var_3011_to_fp16)[name = string("op_3012_cast_fp16")]; + tensor input_661_cast_fp16 = add(x = input_649_cast_fp16, y = var_3012_cast_fp16)[name = string("input_661_cast_fp16")]; + tensor key_25_axes_0 = const()[name = string("key_25_axes_0"), val = tensor([-1])]; + tensor encoder_layers_12_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_12_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(252985216)))]; + tensor encoder_layers_12_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_12_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(252987328)))]; + tensor key_25_cast_fp16 = layer_norm(axes = key_25_axes_0, beta = encoder_layers_12_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_12_norm_self_att_weight_to_fp16, x = input_661_cast_fp16)[name = string("key_25_cast_fp16")]; + bool input_663_interleave_0 = const()[name = string("input_663_interleave_0"), val = bool(false)]; + tensor input_663_cast_fp16 = concat(axis = var_68, interleave = input_663_interleave_0, values = (cache_49_cast_fp16, key_25_cast_fp16))[name = string("input_663_cast_fp16")]; + tensor var_3034_begin_0 = const()[name = string("op_3034_begin_0"), val = tensor([0, 7, 0])]; + tensor var_3034_end_0 = const()[name = string("op_3034_end_0"), val = tensor([1, 42, 1024])]; + tensor var_3034_end_mask_0 = const()[name = string("op_3034_end_mask_0"), val = tensor([true, true, true])]; + tensor var_3034_cast_fp16 = slice_by_index(begin = var_3034_begin_0, end = var_3034_end_0, end_mask = var_3034_end_mask_0, x = cache_49_cast_fp16)[name = string("op_3034_cast_fp16")]; + bool var_3040_interleave_0 = const()[name = string("op_3040_interleave_0"), val = bool(false)]; + tensor var_3040_cast_fp16 = concat(axis = var_68, interleave = var_3040_interleave_0, values = (var_3034_cast_fp16, key_25_cast_fp16))[name = string("op_3040_cast_fp16")]; + tensor encoder_layers_12_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(252989440))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253775936))))[name = string("encoder_layers_12_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_12_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_12_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253776128)))]; + tensor linear_111_cast_fp16 = linear(bias = encoder_layers_12_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_12_self_attn_linear_q_weight_to_fp16_palettized, x = key_25_cast_fp16)[name = string("linear_111_cast_fp16")]; + tensor var_3045 = const()[name = string("op_3045"), val = tensor([1, -1, 8, 128])]; + tensor q_73_cast_fp16 = reshape(shape = var_3045, x = linear_111_cast_fp16)[name = string("q_73_cast_fp16")]; + tensor encoder_layers_12_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253778240))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(254564736))))[name = string("encoder_layers_12_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_12_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_12_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(254564928)))]; + tensor linear_112_cast_fp16 = linear(bias = encoder_layers_12_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_12_self_attn_linear_k_weight_to_fp16_palettized, x = input_663_cast_fp16)[name = string("linear_112_cast_fp16")]; + tensor var_3050 = const()[name = string("op_3050"), val = tensor([1, -1, 8, 128])]; + tensor k_49_cast_fp16 = reshape(shape = var_3050, x = linear_112_cast_fp16)[name = string("k_49_cast_fp16")]; + tensor encoder_layers_12_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(254567040))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255353536))))[name = string("encoder_layers_12_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_12_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_12_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255353728)))]; + tensor linear_113_cast_fp16 = linear(bias = encoder_layers_12_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_12_self_attn_linear_v_weight_to_fp16_palettized, x = input_663_cast_fp16)[name = string("linear_113_cast_fp16")]; + tensor var_3055 = const()[name = string("op_3055"), val = tensor([1, -1, 8, 128])]; + tensor v_25_cast_fp16 = reshape(shape = var_3055, x = linear_113_cast_fp16)[name = string("v_25_cast_fp16")]; + tensor value_33_perm_0 = const()[name = string("value_33_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_12_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_12_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255355840)))]; + tensor var_3068_cast_fp16 = add(x = q_73_cast_fp16, y = encoder_layers_12_self_attn_pos_bias_u_to_fp16)[name = string("op_3068_cast_fp16")]; + tensor encoder_layers_12_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_12_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255357952)))]; + tensor var_3070_cast_fp16 = add(x = q_73_cast_fp16, y = encoder_layers_12_self_attn_pos_bias_v_to_fp16)[name = string("op_3070_cast_fp16")]; + tensor q_with_bias_v_25_perm_0 = const()[name = string("q_with_bias_v_25_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_319_transpose_x_0 = const()[name = string("x_319_transpose_x_0"), val = bool(false)]; + bool x_319_transpose_y_0 = const()[name = string("x_319_transpose_y_0"), val = bool(false)]; + tensor op_3072_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255360064))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255459456))))[name = string("op_3072_to_fp16_quantized")]; + tensor q_with_bias_v_25_cast_fp16 = transpose(perm = q_with_bias_v_25_perm_0, x = var_3070_cast_fp16)[name = string("transpose_254")]; + tensor x_319_cast_fp16 = matmul(transpose_x = x_319_transpose_x_0, transpose_y = x_319_transpose_y_0, x = q_with_bias_v_25_cast_fp16, y = op_3072_to_fp16_quantized)[name = string("x_319_cast_fp16")]; + tensor x_321_pad_0 = const()[name = string("x_321_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_321_mode_0 = const()[name = string("x_321_mode_0"), val = string("constant")]; + fp16 const_235_to_fp16 = const()[name = string("const_235_to_fp16"), val = fp16(0x0p+0)]; + tensor x_321_cast_fp16 = pad(constant_val = const_235_to_fp16, mode = x_321_mode_0, pad = x_321_pad_0, x = x_319_cast_fp16)[name = string("x_321_cast_fp16")]; + tensor var_3080 = const()[name = string("op_3080"), val = tensor([1, 8, -1, 7])]; + tensor x_323_cast_fp16 = reshape(shape = var_3080, x = x_321_cast_fp16)[name = string("x_323_cast_fp16")]; + tensor var_3084_begin_0 = const()[name = string("op_3084_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_3084_end_0 = const()[name = string("op_3084_end_0"), val = tensor([1, 8, 98, 7])]; + tensor var_3084_end_mask_0 = const()[name = string("op_3084_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_3084_cast_fp16 = slice_by_index(begin = var_3084_begin_0, end = var_3084_end_0, end_mask = var_3084_end_mask_0, x = x_323_cast_fp16)[name = string("op_3084_cast_fp16")]; + tensor var_3085 = const()[name = string("op_3085"), val = tensor([1, 8, 7, 97])]; + tensor matrix_bd_49_cast_fp16 = reshape(shape = var_3085, x = var_3084_cast_fp16)[name = string("matrix_bd_49_cast_fp16")]; + bool matrix_ac_25_transpose_x_0 = const()[name = string("matrix_ac_25_transpose_x_0"), val = bool(false)]; + bool matrix_ac_25_transpose_y_0 = const()[name = string("matrix_ac_25_transpose_y_0"), val = bool(false)]; + tensor transpose_120_perm_0 = const()[name = string("transpose_120_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_121_perm_0 = const()[name = string("transpose_121_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_121 = transpose(perm = transpose_121_perm_0, x = k_49_cast_fp16)[name = string("transpose_252")]; + tensor transpose_120 = transpose(perm = transpose_120_perm_0, x = var_3068_cast_fp16)[name = string("transpose_253")]; + tensor matrix_ac_25_cast_fp16 = matmul(transpose_x = matrix_ac_25_transpose_x_0, transpose_y = matrix_ac_25_transpose_y_0, x = transpose_120, y = transpose_121)[name = string("matrix_ac_25_cast_fp16")]; + tensor matrix_bd_51_begin_0 = const()[name = string("matrix_bd_51_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_51_end_0 = const()[name = string("matrix_bd_51_end_0"), val = tensor([1, 8, 7, 49])]; + tensor matrix_bd_51_end_mask_0 = const()[name = string("matrix_bd_51_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_51_cast_fp16 = slice_by_index(begin = matrix_bd_51_begin_0, end = matrix_bd_51_end_0, end_mask = matrix_bd_51_end_mask_0, x = matrix_bd_49_cast_fp16)[name = string("matrix_bd_51_cast_fp16")]; + tensor var_3094_cast_fp16 = add(x = matrix_ac_25_cast_fp16, y = matrix_bd_51_cast_fp16)[name = string("op_3094_cast_fp16")]; + fp16 _inversed_scores_49_y_0_to_fp16 = const()[name = string("_inversed_scores_49_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_49_cast_fp16 = mul(x = var_3094_cast_fp16, y = _inversed_scores_49_y_0_to_fp16)[name = string("_inversed_scores_49_cast_fp16")]; + tensor scores_51_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_49_cast_fp16, cond = mask_11)[name = string("scores_51_cast_fp16")]; + tensor var_3100_cast_fp16 = softmax(axis = var_59, x = scores_51_cast_fp16)[name = string("op_3100_cast_fp16")]; + tensor input_665_cast_fp16 = select(a = var_44_to_fp16, b = var_3100_cast_fp16, cond = mask_11)[name = string("input_665_cast_fp16")]; + bool x_325_transpose_x_0 = const()[name = string("x_325_transpose_x_0"), val = bool(false)]; + bool x_325_transpose_y_0 = const()[name = string("x_325_transpose_y_0"), val = bool(false)]; + tensor value_33_cast_fp16 = transpose(perm = value_33_perm_0, x = v_25_cast_fp16)[name = string("transpose_251")]; + tensor x_325_cast_fp16 = matmul(transpose_x = x_325_transpose_x_0, transpose_y = x_325_transpose_y_0, x = input_665_cast_fp16, y = value_33_cast_fp16)[name = string("x_325_cast_fp16")]; + tensor var_3104_perm_0 = const()[name = string("op_3104_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_3105 = const()[name = string("op_3105"), val = tensor([1, -1, 1024])]; + tensor var_3104_cast_fp16 = transpose(perm = var_3104_perm_0, x = x_325_cast_fp16)[name = string("transpose_250")]; + tensor input_667_cast_fp16 = reshape(shape = var_3105, x = var_3104_cast_fp16)[name = string("input_667_cast_fp16")]; + tensor encoder_layers_12_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255459776))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256246272))))[name = string("encoder_layers_12_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_12_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_12_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256246464)))]; + tensor linear_115_cast_fp16 = linear(bias = encoder_layers_12_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_12_self_attn_linear_out_weight_to_fp16_palettized, x = input_667_cast_fp16)[name = string("linear_115_cast_fp16")]; + tensor input_671_cast_fp16 = add(x = input_661_cast_fp16, y = linear_115_cast_fp16)[name = string("input_671_cast_fp16")]; + tensor x_329_axes_0 = const()[name = string("x_329_axes_0"), val = tensor([-1])]; + tensor encoder_layers_12_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_12_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256248576)))]; + tensor encoder_layers_12_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_12_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256250688)))]; + tensor x_329_cast_fp16 = layer_norm(axes = x_329_axes_0, beta = encoder_layers_12_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_12_norm_conv_weight_to_fp16, x = input_671_cast_fp16)[name = string("x_329_cast_fp16")]; + tensor input_673_perm_0 = const()[name = string("input_673_perm_0"), val = tensor([0, 2, 1])]; + string input_675_pad_type_0 = const()[name = string("input_675_pad_type_0"), val = string("valid")]; + tensor input_675_strides_0 = const()[name = string("input_675_strides_0"), val = tensor([1])]; + tensor input_675_pad_0 = const()[name = string("input_675_pad_0"), val = tensor([0, 0])]; + tensor input_675_dilations_0 = const()[name = string("input_675_dilations_0"), val = tensor([1])]; + int32 input_675_groups_0 = const()[name = string("input_675_groups_0"), val = int32(1)]; + tensor encoder_layers_12_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256252800))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(258350016))))[name = string("encoder_layers_12_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_673_cast_fp16 = transpose(perm = input_673_perm_0, x = x_329_cast_fp16)[name = string("transpose_249")]; + tensor input_675_cast_fp16 = conv(dilations = input_675_dilations_0, groups = input_675_groups_0, pad = input_675_pad_0, pad_type = input_675_pad_type_0, strides = input_675_strides_0, weight = encoder_layers_12_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_673_cast_fp16)[name = string("input_675_cast_fp16")]; + int32 x_331_split_num_splits_0 = const()[name = string("x_331_split_num_splits_0"), val = int32(2)]; + int32 x_331_split_axis_0 = const()[name = string("x_331_split_axis_0"), val = int32(1)]; + tensor x_331_split_cast_fp16_0, tensor x_331_split_cast_fp16_1 = split(axis = x_331_split_axis_0, num_splits = x_331_split_num_splits_0, x = input_675_cast_fp16)[name = string("x_331_split_cast_fp16")]; + tensor x_331_split_1_sigmoid_cast_fp16 = sigmoid(x = x_331_split_cast_fp16_1)[name = string("x_331_split_1_sigmoid_cast_fp16")]; + tensor x_331_cast_fp16 = mul(x = x_331_split_cast_fp16_0, y = x_331_split_1_sigmoid_cast_fp16)[name = string("x_331_cast_fp16")]; + tensor input_677_cast_fp16 = select(a = var_44_to_fp16, b = x_331_cast_fp16, cond = var_575)[name = string("input_677_cast_fp16")]; + bool new_x_51_interleave_0 = const()[name = string("new_x_51_interleave_0"), val = bool(false)]; + tensor new_x_51_cast_fp16 = concat(axis = var_59, interleave = new_x_51_interleave_0, values = (cache_51_cast_fp16, input_677_cast_fp16))[name = string("new_x_51_cast_fp16")]; + tensor var_3144_begin_0 = const()[name = string("op_3144_begin_0"), val = tensor([0, 0, 7])]; + tensor var_3144_end_0 = const()[name = string("op_3144_end_0"), val = tensor([1, 1024, 15])]; + tensor var_3144_end_mask_0 = const()[name = string("op_3144_end_mask_0"), val = tensor([true, true, true])]; + tensor var_3144_cast_fp16 = slice_by_index(begin = var_3144_begin_0, end = var_3144_end_0, end_mask = var_3144_end_mask_0, x = new_x_51_cast_fp16)[name = string("op_3144_cast_fp16")]; + string x_333_pad_type_0 = const()[name = string("x_333_pad_type_0"), val = string("valid")]; + int32 x_333_groups_0 = const()[name = string("x_333_groups_0"), val = int32(1024)]; + tensor x_333_strides_0 = const()[name = string("x_333_strides_0"), val = tensor([1])]; + tensor x_333_pad_0 = const()[name = string("x_333_pad_0"), val = tensor([0, 0])]; + tensor x_333_dilations_0 = const()[name = string("x_333_dilations_0"), val = tensor([1])]; + tensor encoder_layers_12_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(258354176))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(258363456))))[name = string("encoder_layers_12_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_333_cast_fp16 = conv(dilations = x_333_dilations_0, groups = x_333_groups_0, pad = x_333_pad_0, pad_type = x_333_pad_type_0, strides = x_333_strides_0, weight = encoder_layers_12_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_51_cast_fp16)[name = string("x_333_cast_fp16")]; + tensor input_679_perm_0 = const()[name = string("input_679_perm_0"), val = tensor([0, 2, 1])]; + tensor x_335_axes_0 = const()[name = string("x_335_axes_0"), val = tensor([-1])]; + tensor encoder_layers_12_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_12_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(258365568)))]; + tensor encoder_layers_12_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_12_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(258367680)))]; + tensor input_679_cast_fp16 = transpose(perm = input_679_perm_0, x = x_333_cast_fp16)[name = string("transpose_248")]; + tensor x_335_cast_fp16 = layer_norm(axes = x_335_axes_0, beta = encoder_layers_12_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_12_conv_batch_norm_weight_to_fp16, x = input_679_cast_fp16)[name = string("x_335_cast_fp16")]; + tensor input_681_perm_0 = const()[name = string("input_681_perm_0"), val = tensor([0, 2, 1])]; + tensor input_681_cast_fp16 = transpose(perm = input_681_perm_0, x = x_335_cast_fp16)[name = string("transpose_247")]; + tensor input_683_cast_fp16 = silu(x = input_681_cast_fp16)[name = string("input_683_cast_fp16")]; + string x_337_pad_type_0 = const()[name = string("x_337_pad_type_0"), val = string("valid")]; + tensor x_337_strides_0 = const()[name = string("x_337_strides_0"), val = tensor([1])]; + tensor x_337_pad_0 = const()[name = string("x_337_pad_0"), val = tensor([0, 0])]; + tensor x_337_dilations_0 = const()[name = string("x_337_dilations_0"), val = tensor([1])]; + int32 x_337_groups_0 = const()[name = string("x_337_groups_0"), val = int32(1)]; + tensor encoder_layers_12_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(258369792))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259418432))))[name = string("encoder_layers_12_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_337_cast_fp16 = conv(dilations = x_337_dilations_0, groups = x_337_groups_0, pad = x_337_pad_0, pad_type = x_337_pad_type_0, strides = x_337_strides_0, weight = encoder_layers_12_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_683_cast_fp16)[name = string("x_337_cast_fp16")]; + tensor input_685_perm_0 = const()[name = string("input_685_perm_0"), val = tensor([0, 2, 1])]; + tensor input_685_cast_fp16 = transpose(perm = input_685_perm_0, x = x_337_cast_fp16)[name = string("transpose_246")]; + tensor input_687_cast_fp16 = add(x = input_671_cast_fp16, y = input_685_cast_fp16)[name = string("input_687_cast_fp16")]; + tensor input_689_axes_0 = const()[name = string("input_689_axes_0"), val = tensor([-1])]; + tensor encoder_layers_12_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_12_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259420544)))]; + tensor encoder_layers_12_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_12_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259422656)))]; + tensor input_689_cast_fp16 = layer_norm(axes = input_689_axes_0, beta = encoder_layers_12_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_12_norm_feed_forward2_weight_to_fp16, x = input_687_cast_fp16)[name = string("input_689_cast_fp16")]; + tensor encoder_layers_12_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259424768))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(262570560))))[name = string("encoder_layers_12_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_12_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_12_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(262570752)))]; + tensor linear_116_cast_fp16 = linear(bias = encoder_layers_12_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_12_feed_forward2_linear1_weight_to_fp16_palettized, x = input_689_cast_fp16)[name = string("linear_116_cast_fp16")]; + tensor input_693_cast_fp16 = silu(x = linear_116_cast_fp16)[name = string("input_693_cast_fp16")]; + tensor encoder_layers_12_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(262579008))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(265724800))))[name = string("encoder_layers_12_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_12_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_12_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(265724992)))]; + tensor linear_117_cast_fp16 = linear(bias = encoder_layers_12_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_12_feed_forward2_linear2_weight_to_fp16_palettized, x = input_693_cast_fp16)[name = string("linear_117_cast_fp16")]; + fp16 var_3187_to_fp16 = const()[name = string("op_3187_to_fp16"), val = fp16(0x1p-1)]; + tensor var_3188_cast_fp16 = mul(x = linear_117_cast_fp16, y = var_3187_to_fp16)[name = string("op_3188_cast_fp16")]; + tensor input_699_cast_fp16 = add(x = input_687_cast_fp16, y = var_3188_cast_fp16)[name = string("input_699_cast_fp16")]; + tensor input_701_axes_0 = const()[name = string("input_701_axes_0"), val = tensor([-1])]; + tensor encoder_layers_12_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_12_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(265727104)))]; + tensor encoder_layers_12_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_12_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(265729216)))]; + tensor input_701_cast_fp16 = layer_norm(axes = input_701_axes_0, beta = encoder_layers_12_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_12_norm_out_weight_to_fp16, x = input_699_cast_fp16)[name = string("input_701_cast_fp16")]; + tensor cache_53_begin_0 = const()[name = string("cache_53_begin_0"), val = tensor([13, 0, 0, 0])]; + tensor cache_53_end_0 = const()[name = string("cache_53_end_0"), val = tensor([14, 1, 42, 1024])]; + tensor cache_53_end_mask_0 = const()[name = string("cache_53_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_53_squeeze_mask_0 = const()[name = string("cache_53_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_53_cast_fp16 = slice_by_index(begin = cache_53_begin_0, end = cache_53_end_0, end_mask = cache_53_end_mask_0, squeeze_mask = cache_53_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_53_cast_fp16")]; + tensor cache_55_begin_0 = const()[name = string("cache_55_begin_0"), val = tensor([13, 0, 0, 0])]; + tensor cache_55_end_0 = const()[name = string("cache_55_end_0"), val = tensor([14, 1, 1024, 8])]; + tensor cache_55_end_mask_0 = const()[name = string("cache_55_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_55_squeeze_mask_0 = const()[name = string("cache_55_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_55_cast_fp16 = slice_by_index(begin = cache_55_begin_0, end = cache_55_end_0, end_mask = cache_55_end_mask_0, squeeze_mask = cache_55_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_55_cast_fp16")]; + tensor input_703_axes_0 = const()[name = string("input_703_axes_0"), val = tensor([-1])]; + tensor encoder_layers_13_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_13_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(265731328)))]; + tensor encoder_layers_13_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_13_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(265733440)))]; + tensor input_703_cast_fp16 = layer_norm(axes = input_703_axes_0, beta = encoder_layers_13_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_13_norm_feed_forward1_weight_to_fp16, x = input_701_cast_fp16)[name = string("input_703_cast_fp16")]; + tensor encoder_layers_13_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(265735552))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(268881344))))[name = string("encoder_layers_13_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_13_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_13_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(268881536)))]; + tensor linear_118_cast_fp16 = linear(bias = encoder_layers_13_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_13_feed_forward1_linear1_weight_to_fp16_palettized, x = input_703_cast_fp16)[name = string("linear_118_cast_fp16")]; + tensor input_707_cast_fp16 = silu(x = linear_118_cast_fp16)[name = string("input_707_cast_fp16")]; + tensor encoder_layers_13_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(268889792))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272035584))))[name = string("encoder_layers_13_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_13_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_13_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272035776)))]; + tensor linear_119_cast_fp16 = linear(bias = encoder_layers_13_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_13_feed_forward1_linear2_weight_to_fp16_palettized, x = input_707_cast_fp16)[name = string("linear_119_cast_fp16")]; + fp16 var_3224_to_fp16 = const()[name = string("op_3224_to_fp16"), val = fp16(0x1p-1)]; + tensor var_3225_cast_fp16 = mul(x = linear_119_cast_fp16, y = var_3224_to_fp16)[name = string("op_3225_cast_fp16")]; + tensor input_713_cast_fp16 = add(x = input_701_cast_fp16, y = var_3225_cast_fp16)[name = string("input_713_cast_fp16")]; + tensor key_27_axes_0 = const()[name = string("key_27_axes_0"), val = tensor([-1])]; + tensor encoder_layers_13_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_13_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272037888)))]; + tensor encoder_layers_13_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_13_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272040000)))]; + tensor key_27_cast_fp16 = layer_norm(axes = key_27_axes_0, beta = encoder_layers_13_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_13_norm_self_att_weight_to_fp16, x = input_713_cast_fp16)[name = string("key_27_cast_fp16")]; + bool input_715_interleave_0 = const()[name = string("input_715_interleave_0"), val = bool(false)]; + tensor input_715_cast_fp16 = concat(axis = var_68, interleave = input_715_interleave_0, values = (cache_53_cast_fp16, key_27_cast_fp16))[name = string("input_715_cast_fp16")]; + tensor var_3247_begin_0 = const()[name = string("op_3247_begin_0"), val = tensor([0, 7, 0])]; + tensor var_3247_end_0 = const()[name = string("op_3247_end_0"), val = tensor([1, 42, 1024])]; + tensor var_3247_end_mask_0 = const()[name = string("op_3247_end_mask_0"), val = tensor([true, true, true])]; + tensor var_3247_cast_fp16 = slice_by_index(begin = var_3247_begin_0, end = var_3247_end_0, end_mask = var_3247_end_mask_0, x = cache_53_cast_fp16)[name = string("op_3247_cast_fp16")]; + bool var_3253_interleave_0 = const()[name = string("op_3253_interleave_0"), val = bool(false)]; + tensor var_3253_cast_fp16 = concat(axis = var_68, interleave = var_3253_interleave_0, values = (var_3247_cast_fp16, key_27_cast_fp16))[name = string("op_3253_cast_fp16")]; + tensor encoder_layers_13_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272042112))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272828608))))[name = string("encoder_layers_13_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_13_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_13_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272828800)))]; + tensor linear_120_cast_fp16 = linear(bias = encoder_layers_13_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_13_self_attn_linear_q_weight_to_fp16_palettized, x = key_27_cast_fp16)[name = string("linear_120_cast_fp16")]; + tensor var_3258 = const()[name = string("op_3258"), val = tensor([1, -1, 8, 128])]; + tensor q_79_cast_fp16 = reshape(shape = var_3258, x = linear_120_cast_fp16)[name = string("q_79_cast_fp16")]; + tensor encoder_layers_13_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272830912))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(273617408))))[name = string("encoder_layers_13_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_13_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_13_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(273617600)))]; + tensor linear_121_cast_fp16 = linear(bias = encoder_layers_13_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_13_self_attn_linear_k_weight_to_fp16_palettized, x = input_715_cast_fp16)[name = string("linear_121_cast_fp16")]; + tensor var_3263 = const()[name = string("op_3263"), val = tensor([1, -1, 8, 128])]; + tensor k_53_cast_fp16 = reshape(shape = var_3263, x = linear_121_cast_fp16)[name = string("k_53_cast_fp16")]; + tensor encoder_layers_13_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(273619712))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274406208))))[name = string("encoder_layers_13_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_13_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_13_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274406400)))]; + tensor linear_122_cast_fp16 = linear(bias = encoder_layers_13_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_13_self_attn_linear_v_weight_to_fp16_palettized, x = input_715_cast_fp16)[name = string("linear_122_cast_fp16")]; + tensor var_3268 = const()[name = string("op_3268"), val = tensor([1, -1, 8, 128])]; + tensor v_27_cast_fp16 = reshape(shape = var_3268, x = linear_122_cast_fp16)[name = string("v_27_cast_fp16")]; + tensor value_35_perm_0 = const()[name = string("value_35_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_13_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_13_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274408512)))]; + tensor var_3281_cast_fp16 = add(x = q_79_cast_fp16, y = encoder_layers_13_self_attn_pos_bias_u_to_fp16)[name = string("op_3281_cast_fp16")]; + tensor encoder_layers_13_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_13_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274410624)))]; + tensor var_3283_cast_fp16 = add(x = q_79_cast_fp16, y = encoder_layers_13_self_attn_pos_bias_v_to_fp16)[name = string("op_3283_cast_fp16")]; + tensor q_with_bias_v_27_perm_0 = const()[name = string("q_with_bias_v_27_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_345_transpose_x_0 = const()[name = string("x_345_transpose_x_0"), val = bool(false)]; + bool x_345_transpose_y_0 = const()[name = string("x_345_transpose_y_0"), val = bool(false)]; + tensor op_3285_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274412736))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274512128))))[name = string("op_3285_to_fp16_quantized")]; + tensor q_with_bias_v_27_cast_fp16 = transpose(perm = q_with_bias_v_27_perm_0, x = var_3283_cast_fp16)[name = string("transpose_245")]; + tensor x_345_cast_fp16 = matmul(transpose_x = x_345_transpose_x_0, transpose_y = x_345_transpose_y_0, x = q_with_bias_v_27_cast_fp16, y = op_3285_to_fp16_quantized)[name = string("x_345_cast_fp16")]; + tensor x_347_pad_0 = const()[name = string("x_347_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_347_mode_0 = const()[name = string("x_347_mode_0"), val = string("constant")]; + fp16 const_248_to_fp16 = const()[name = string("const_248_to_fp16"), val = fp16(0x0p+0)]; + tensor x_347_cast_fp16 = pad(constant_val = const_248_to_fp16, mode = x_347_mode_0, pad = x_347_pad_0, x = x_345_cast_fp16)[name = string("x_347_cast_fp16")]; + tensor var_3293 = const()[name = string("op_3293"), val = tensor([1, 8, -1, 7])]; + tensor x_349_cast_fp16 = reshape(shape = var_3293, x = x_347_cast_fp16)[name = string("x_349_cast_fp16")]; + tensor var_3297_begin_0 = const()[name = string("op_3297_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_3297_end_0 = const()[name = string("op_3297_end_0"), val = tensor([1, 8, 98, 7])]; + tensor var_3297_end_mask_0 = const()[name = string("op_3297_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_3297_cast_fp16 = slice_by_index(begin = var_3297_begin_0, end = var_3297_end_0, end_mask = var_3297_end_mask_0, x = x_349_cast_fp16)[name = string("op_3297_cast_fp16")]; + tensor var_3298 = const()[name = string("op_3298"), val = tensor([1, 8, 7, 97])]; + tensor matrix_bd_53_cast_fp16 = reshape(shape = var_3298, x = var_3297_cast_fp16)[name = string("matrix_bd_53_cast_fp16")]; + bool matrix_ac_27_transpose_x_0 = const()[name = string("matrix_ac_27_transpose_x_0"), val = bool(false)]; + bool matrix_ac_27_transpose_y_0 = const()[name = string("matrix_ac_27_transpose_y_0"), val = bool(false)]; + tensor transpose_122_perm_0 = const()[name = string("transpose_122_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_123_perm_0 = const()[name = string("transpose_123_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_123 = transpose(perm = transpose_123_perm_0, x = k_53_cast_fp16)[name = string("transpose_243")]; + tensor transpose_122 = transpose(perm = transpose_122_perm_0, x = var_3281_cast_fp16)[name = string("transpose_244")]; + tensor matrix_ac_27_cast_fp16 = matmul(transpose_x = matrix_ac_27_transpose_x_0, transpose_y = matrix_ac_27_transpose_y_0, x = transpose_122, y = transpose_123)[name = string("matrix_ac_27_cast_fp16")]; + tensor matrix_bd_55_begin_0 = const()[name = string("matrix_bd_55_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_55_end_0 = const()[name = string("matrix_bd_55_end_0"), val = tensor([1, 8, 7, 49])]; + tensor matrix_bd_55_end_mask_0 = const()[name = string("matrix_bd_55_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_55_cast_fp16 = slice_by_index(begin = matrix_bd_55_begin_0, end = matrix_bd_55_end_0, end_mask = matrix_bd_55_end_mask_0, x = matrix_bd_53_cast_fp16)[name = string("matrix_bd_55_cast_fp16")]; + tensor var_3307_cast_fp16 = add(x = matrix_ac_27_cast_fp16, y = matrix_bd_55_cast_fp16)[name = string("op_3307_cast_fp16")]; + fp16 _inversed_scores_53_y_0_to_fp16 = const()[name = string("_inversed_scores_53_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_53_cast_fp16 = mul(x = var_3307_cast_fp16, y = _inversed_scores_53_y_0_to_fp16)[name = string("_inversed_scores_53_cast_fp16")]; + tensor scores_55_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_53_cast_fp16, cond = mask_11)[name = string("scores_55_cast_fp16")]; + tensor var_3313_cast_fp16 = softmax(axis = var_59, x = scores_55_cast_fp16)[name = string("op_3313_cast_fp16")]; + tensor input_717_cast_fp16 = select(a = var_44_to_fp16, b = var_3313_cast_fp16, cond = mask_11)[name = string("input_717_cast_fp16")]; + bool x_351_transpose_x_0 = const()[name = string("x_351_transpose_x_0"), val = bool(false)]; + bool x_351_transpose_y_0 = const()[name = string("x_351_transpose_y_0"), val = bool(false)]; + tensor value_35_cast_fp16 = transpose(perm = value_35_perm_0, x = v_27_cast_fp16)[name = string("transpose_242")]; + tensor x_351_cast_fp16 = matmul(transpose_x = x_351_transpose_x_0, transpose_y = x_351_transpose_y_0, x = input_717_cast_fp16, y = value_35_cast_fp16)[name = string("x_351_cast_fp16")]; + tensor var_3317_perm_0 = const()[name = string("op_3317_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_3318 = const()[name = string("op_3318"), val = tensor([1, -1, 1024])]; + tensor var_3317_cast_fp16 = transpose(perm = var_3317_perm_0, x = x_351_cast_fp16)[name = string("transpose_241")]; + tensor input_719_cast_fp16 = reshape(shape = var_3318, x = var_3317_cast_fp16)[name = string("input_719_cast_fp16")]; + tensor encoder_layers_13_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274512448))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275298944))))[name = string("encoder_layers_13_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_13_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_13_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275299136)))]; + tensor linear_124_cast_fp16 = linear(bias = encoder_layers_13_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_13_self_attn_linear_out_weight_to_fp16_palettized, x = input_719_cast_fp16)[name = string("linear_124_cast_fp16")]; + tensor input_723_cast_fp16 = add(x = input_713_cast_fp16, y = linear_124_cast_fp16)[name = string("input_723_cast_fp16")]; + tensor x_355_axes_0 = const()[name = string("x_355_axes_0"), val = tensor([-1])]; + tensor encoder_layers_13_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_13_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275301248)))]; + tensor encoder_layers_13_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_13_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275303360)))]; + tensor x_355_cast_fp16 = layer_norm(axes = x_355_axes_0, beta = encoder_layers_13_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_13_norm_conv_weight_to_fp16, x = input_723_cast_fp16)[name = string("x_355_cast_fp16")]; + tensor input_725_perm_0 = const()[name = string("input_725_perm_0"), val = tensor([0, 2, 1])]; + string input_727_pad_type_0 = const()[name = string("input_727_pad_type_0"), val = string("valid")]; + tensor input_727_strides_0 = const()[name = string("input_727_strides_0"), val = tensor([1])]; + tensor input_727_pad_0 = const()[name = string("input_727_pad_0"), val = tensor([0, 0])]; + tensor input_727_dilations_0 = const()[name = string("input_727_dilations_0"), val = tensor([1])]; + int32 input_727_groups_0 = const()[name = string("input_727_groups_0"), val = int32(1)]; + tensor encoder_layers_13_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275305472))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(277402688))))[name = string("encoder_layers_13_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_725_cast_fp16 = transpose(perm = input_725_perm_0, x = x_355_cast_fp16)[name = string("transpose_240")]; + tensor input_727_cast_fp16 = conv(dilations = input_727_dilations_0, groups = input_727_groups_0, pad = input_727_pad_0, pad_type = input_727_pad_type_0, strides = input_727_strides_0, weight = encoder_layers_13_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_725_cast_fp16)[name = string("input_727_cast_fp16")]; + int32 x_357_split_num_splits_0 = const()[name = string("x_357_split_num_splits_0"), val = int32(2)]; + int32 x_357_split_axis_0 = const()[name = string("x_357_split_axis_0"), val = int32(1)]; + tensor x_357_split_cast_fp16_0, tensor x_357_split_cast_fp16_1 = split(axis = x_357_split_axis_0, num_splits = x_357_split_num_splits_0, x = input_727_cast_fp16)[name = string("x_357_split_cast_fp16")]; + tensor x_357_split_1_sigmoid_cast_fp16 = sigmoid(x = x_357_split_cast_fp16_1)[name = string("x_357_split_1_sigmoid_cast_fp16")]; + tensor x_357_cast_fp16 = mul(x = x_357_split_cast_fp16_0, y = x_357_split_1_sigmoid_cast_fp16)[name = string("x_357_cast_fp16")]; + tensor input_729_cast_fp16 = select(a = var_44_to_fp16, b = x_357_cast_fp16, cond = var_575)[name = string("input_729_cast_fp16")]; + bool new_x_55_interleave_0 = const()[name = string("new_x_55_interleave_0"), val = bool(false)]; + tensor new_x_55_cast_fp16 = concat(axis = var_59, interleave = new_x_55_interleave_0, values = (cache_55_cast_fp16, input_729_cast_fp16))[name = string("new_x_55_cast_fp16")]; + tensor var_3357_begin_0 = const()[name = string("op_3357_begin_0"), val = tensor([0, 0, 7])]; + tensor var_3357_end_0 = const()[name = string("op_3357_end_0"), val = tensor([1, 1024, 15])]; + tensor var_3357_end_mask_0 = const()[name = string("op_3357_end_mask_0"), val = tensor([true, true, true])]; + tensor var_3357_cast_fp16 = slice_by_index(begin = var_3357_begin_0, end = var_3357_end_0, end_mask = var_3357_end_mask_0, x = new_x_55_cast_fp16)[name = string("op_3357_cast_fp16")]; + string x_359_pad_type_0 = const()[name = string("x_359_pad_type_0"), val = string("valid")]; + int32 x_359_groups_0 = const()[name = string("x_359_groups_0"), val = int32(1024)]; + tensor x_359_strides_0 = const()[name = string("x_359_strides_0"), val = tensor([1])]; + tensor x_359_pad_0 = const()[name = string("x_359_pad_0"), val = tensor([0, 0])]; + tensor x_359_dilations_0 = const()[name = string("x_359_dilations_0"), val = tensor([1])]; + tensor encoder_layers_13_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(277406848))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(277416128))))[name = string("encoder_layers_13_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_359_cast_fp16 = conv(dilations = x_359_dilations_0, groups = x_359_groups_0, pad = x_359_pad_0, pad_type = x_359_pad_type_0, strides = x_359_strides_0, weight = encoder_layers_13_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_55_cast_fp16)[name = string("x_359_cast_fp16")]; + tensor input_731_perm_0 = const()[name = string("input_731_perm_0"), val = tensor([0, 2, 1])]; + tensor x_361_axes_0 = const()[name = string("x_361_axes_0"), val = tensor([-1])]; + tensor encoder_layers_13_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_13_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(277418240)))]; + tensor encoder_layers_13_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_13_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(277420352)))]; + tensor input_731_cast_fp16 = transpose(perm = input_731_perm_0, x = x_359_cast_fp16)[name = string("transpose_239")]; + tensor x_361_cast_fp16 = layer_norm(axes = x_361_axes_0, beta = encoder_layers_13_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_13_conv_batch_norm_weight_to_fp16, x = input_731_cast_fp16)[name = string("x_361_cast_fp16")]; + tensor input_733_perm_0 = const()[name = string("input_733_perm_0"), val = tensor([0, 2, 1])]; + tensor input_733_cast_fp16 = transpose(perm = input_733_perm_0, x = x_361_cast_fp16)[name = string("transpose_238")]; + tensor input_735_cast_fp16 = silu(x = input_733_cast_fp16)[name = string("input_735_cast_fp16")]; + string x_363_pad_type_0 = const()[name = string("x_363_pad_type_0"), val = string("valid")]; + tensor x_363_strides_0 = const()[name = string("x_363_strides_0"), val = tensor([1])]; + tensor x_363_pad_0 = const()[name = string("x_363_pad_0"), val = tensor([0, 0])]; + tensor x_363_dilations_0 = const()[name = string("x_363_dilations_0"), val = tensor([1])]; + int32 x_363_groups_0 = const()[name = string("x_363_groups_0"), val = int32(1)]; + tensor encoder_layers_13_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(277422464))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278471104))))[name = string("encoder_layers_13_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_363_cast_fp16 = conv(dilations = x_363_dilations_0, groups = x_363_groups_0, pad = x_363_pad_0, pad_type = x_363_pad_type_0, strides = x_363_strides_0, weight = encoder_layers_13_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_735_cast_fp16)[name = string("x_363_cast_fp16")]; + tensor input_737_perm_0 = const()[name = string("input_737_perm_0"), val = tensor([0, 2, 1])]; + tensor input_737_cast_fp16 = transpose(perm = input_737_perm_0, x = x_363_cast_fp16)[name = string("transpose_237")]; + tensor input_739_cast_fp16 = add(x = input_723_cast_fp16, y = input_737_cast_fp16)[name = string("input_739_cast_fp16")]; + tensor input_741_axes_0 = const()[name = string("input_741_axes_0"), val = tensor([-1])]; + tensor encoder_layers_13_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_13_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278473216)))]; + tensor encoder_layers_13_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_13_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278475328)))]; + tensor input_741_cast_fp16 = layer_norm(axes = input_741_axes_0, beta = encoder_layers_13_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_13_norm_feed_forward2_weight_to_fp16, x = input_739_cast_fp16)[name = string("input_741_cast_fp16")]; + tensor encoder_layers_13_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278477440))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(281623232))))[name = string("encoder_layers_13_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_13_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_13_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(281623424)))]; + tensor linear_125_cast_fp16 = linear(bias = encoder_layers_13_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_13_feed_forward2_linear1_weight_to_fp16_palettized, x = input_741_cast_fp16)[name = string("linear_125_cast_fp16")]; + tensor input_745_cast_fp16 = silu(x = linear_125_cast_fp16)[name = string("input_745_cast_fp16")]; + tensor encoder_layers_13_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(281631680))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(284777472))))[name = string("encoder_layers_13_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_13_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_13_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(284777664)))]; + tensor linear_126_cast_fp16 = linear(bias = encoder_layers_13_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_13_feed_forward2_linear2_weight_to_fp16_palettized, x = input_745_cast_fp16)[name = string("linear_126_cast_fp16")]; + fp16 var_3400_to_fp16 = const()[name = string("op_3400_to_fp16"), val = fp16(0x1p-1)]; + tensor var_3401_cast_fp16 = mul(x = linear_126_cast_fp16, y = var_3400_to_fp16)[name = string("op_3401_cast_fp16")]; + tensor input_751_cast_fp16 = add(x = input_739_cast_fp16, y = var_3401_cast_fp16)[name = string("input_751_cast_fp16")]; + tensor input_753_axes_0 = const()[name = string("input_753_axes_0"), val = tensor([-1])]; + tensor encoder_layers_13_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_13_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(284779776)))]; + tensor encoder_layers_13_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_13_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(284781888)))]; + tensor input_753_cast_fp16 = layer_norm(axes = input_753_axes_0, beta = encoder_layers_13_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_13_norm_out_weight_to_fp16, x = input_751_cast_fp16)[name = string("input_753_cast_fp16")]; + tensor cache_57_begin_0 = const()[name = string("cache_57_begin_0"), val = tensor([14, 0, 0, 0])]; + tensor cache_57_end_0 = const()[name = string("cache_57_end_0"), val = tensor([15, 1, 42, 1024])]; + tensor cache_57_end_mask_0 = const()[name = string("cache_57_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_57_squeeze_mask_0 = const()[name = string("cache_57_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_57_cast_fp16 = slice_by_index(begin = cache_57_begin_0, end = cache_57_end_0, end_mask = cache_57_end_mask_0, squeeze_mask = cache_57_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_57_cast_fp16")]; + tensor cache_59_begin_0 = const()[name = string("cache_59_begin_0"), val = tensor([14, 0, 0, 0])]; + tensor cache_59_end_0 = const()[name = string("cache_59_end_0"), val = tensor([15, 1, 1024, 8])]; + tensor cache_59_end_mask_0 = const()[name = string("cache_59_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_59_squeeze_mask_0 = const()[name = string("cache_59_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_59_cast_fp16 = slice_by_index(begin = cache_59_begin_0, end = cache_59_end_0, end_mask = cache_59_end_mask_0, squeeze_mask = cache_59_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_59_cast_fp16")]; + tensor input_755_axes_0 = const()[name = string("input_755_axes_0"), val = tensor([-1])]; + tensor encoder_layers_14_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_14_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(284784000)))]; + tensor encoder_layers_14_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_14_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(284786112)))]; + tensor input_755_cast_fp16 = layer_norm(axes = input_755_axes_0, beta = encoder_layers_14_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_14_norm_feed_forward1_weight_to_fp16, x = input_753_cast_fp16)[name = string("input_755_cast_fp16")]; + tensor encoder_layers_14_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(284788224))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(287934016))))[name = string("encoder_layers_14_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_14_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_14_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(287934208)))]; + tensor linear_127_cast_fp16 = linear(bias = encoder_layers_14_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_14_feed_forward1_linear1_weight_to_fp16_palettized, x = input_755_cast_fp16)[name = string("linear_127_cast_fp16")]; + tensor input_759_cast_fp16 = silu(x = linear_127_cast_fp16)[name = string("input_759_cast_fp16")]; + tensor encoder_layers_14_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(287942464))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291088256))))[name = string("encoder_layers_14_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_14_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_14_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291088448)))]; + tensor linear_128_cast_fp16 = linear(bias = encoder_layers_14_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_14_feed_forward1_linear2_weight_to_fp16_palettized, x = input_759_cast_fp16)[name = string("linear_128_cast_fp16")]; + fp16 var_3437_to_fp16 = const()[name = string("op_3437_to_fp16"), val = fp16(0x1p-1)]; + tensor var_3438_cast_fp16 = mul(x = linear_128_cast_fp16, y = var_3437_to_fp16)[name = string("op_3438_cast_fp16")]; + tensor input_765_cast_fp16 = add(x = input_753_cast_fp16, y = var_3438_cast_fp16)[name = string("input_765_cast_fp16")]; + tensor key_29_axes_0 = const()[name = string("key_29_axes_0"), val = tensor([-1])]; + tensor encoder_layers_14_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_14_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291090560)))]; + tensor encoder_layers_14_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_14_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291092672)))]; + tensor key_29_cast_fp16 = layer_norm(axes = key_29_axes_0, beta = encoder_layers_14_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_14_norm_self_att_weight_to_fp16, x = input_765_cast_fp16)[name = string("key_29_cast_fp16")]; + bool input_767_interleave_0 = const()[name = string("input_767_interleave_0"), val = bool(false)]; + tensor input_767_cast_fp16 = concat(axis = var_68, interleave = input_767_interleave_0, values = (cache_57_cast_fp16, key_29_cast_fp16))[name = string("input_767_cast_fp16")]; + tensor var_3460_begin_0 = const()[name = string("op_3460_begin_0"), val = tensor([0, 7, 0])]; + tensor var_3460_end_0 = const()[name = string("op_3460_end_0"), val = tensor([1, 42, 1024])]; + tensor var_3460_end_mask_0 = const()[name = string("op_3460_end_mask_0"), val = tensor([true, true, true])]; + tensor var_3460_cast_fp16 = slice_by_index(begin = var_3460_begin_0, end = var_3460_end_0, end_mask = var_3460_end_mask_0, x = cache_57_cast_fp16)[name = string("op_3460_cast_fp16")]; + bool var_3466_interleave_0 = const()[name = string("op_3466_interleave_0"), val = bool(false)]; + tensor var_3466_cast_fp16 = concat(axis = var_68, interleave = var_3466_interleave_0, values = (var_3460_cast_fp16, key_29_cast_fp16))[name = string("op_3466_cast_fp16")]; + tensor encoder_layers_14_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291094784))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291881280))))[name = string("encoder_layers_14_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_14_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_14_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291881472)))]; + tensor linear_129_cast_fp16 = linear(bias = encoder_layers_14_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_14_self_attn_linear_q_weight_to_fp16_palettized, x = key_29_cast_fp16)[name = string("linear_129_cast_fp16")]; + tensor var_3471 = const()[name = string("op_3471"), val = tensor([1, -1, 8, 128])]; + tensor q_85_cast_fp16 = reshape(shape = var_3471, x = linear_129_cast_fp16)[name = string("q_85_cast_fp16")]; + tensor encoder_layers_14_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291883584))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(292670080))))[name = string("encoder_layers_14_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_14_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_14_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(292670272)))]; + tensor linear_130_cast_fp16 = linear(bias = encoder_layers_14_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_14_self_attn_linear_k_weight_to_fp16_palettized, x = input_767_cast_fp16)[name = string("linear_130_cast_fp16")]; + tensor var_3476 = const()[name = string("op_3476"), val = tensor([1, -1, 8, 128])]; + tensor k_57_cast_fp16 = reshape(shape = var_3476, x = linear_130_cast_fp16)[name = string("k_57_cast_fp16")]; + tensor encoder_layers_14_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(292672384))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293458880))))[name = string("encoder_layers_14_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_14_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_14_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293459072)))]; + tensor linear_131_cast_fp16 = linear(bias = encoder_layers_14_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_14_self_attn_linear_v_weight_to_fp16_palettized, x = input_767_cast_fp16)[name = string("linear_131_cast_fp16")]; + tensor var_3481 = const()[name = string("op_3481"), val = tensor([1, -1, 8, 128])]; + tensor v_29_cast_fp16 = reshape(shape = var_3481, x = linear_131_cast_fp16)[name = string("v_29_cast_fp16")]; + tensor value_37_perm_0 = const()[name = string("value_37_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_14_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_14_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293461184)))]; + tensor var_3494_cast_fp16 = add(x = q_85_cast_fp16, y = encoder_layers_14_self_attn_pos_bias_u_to_fp16)[name = string("op_3494_cast_fp16")]; + tensor encoder_layers_14_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_14_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293463296)))]; + tensor var_3496_cast_fp16 = add(x = q_85_cast_fp16, y = encoder_layers_14_self_attn_pos_bias_v_to_fp16)[name = string("op_3496_cast_fp16")]; + tensor q_with_bias_v_29_perm_0 = const()[name = string("q_with_bias_v_29_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_371_transpose_x_0 = const()[name = string("x_371_transpose_x_0"), val = bool(false)]; + bool x_371_transpose_y_0 = const()[name = string("x_371_transpose_y_0"), val = bool(false)]; + tensor op_3498_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293465408))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293564800))))[name = string("op_3498_to_fp16_quantized")]; + tensor q_with_bias_v_29_cast_fp16 = transpose(perm = q_with_bias_v_29_perm_0, x = var_3496_cast_fp16)[name = string("transpose_236")]; + tensor x_371_cast_fp16 = matmul(transpose_x = x_371_transpose_x_0, transpose_y = x_371_transpose_y_0, x = q_with_bias_v_29_cast_fp16, y = op_3498_to_fp16_quantized)[name = string("x_371_cast_fp16")]; + tensor x_373_pad_0 = const()[name = string("x_373_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_373_mode_0 = const()[name = string("x_373_mode_0"), val = string("constant")]; + fp16 const_261_to_fp16 = const()[name = string("const_261_to_fp16"), val = fp16(0x0p+0)]; + tensor x_373_cast_fp16 = pad(constant_val = const_261_to_fp16, mode = x_373_mode_0, pad = x_373_pad_0, x = x_371_cast_fp16)[name = string("x_373_cast_fp16")]; + tensor var_3506 = const()[name = string("op_3506"), val = tensor([1, 8, -1, 7])]; + tensor x_375_cast_fp16 = reshape(shape = var_3506, x = x_373_cast_fp16)[name = string("x_375_cast_fp16")]; + tensor var_3510_begin_0 = const()[name = string("op_3510_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_3510_end_0 = const()[name = string("op_3510_end_0"), val = tensor([1, 8, 98, 7])]; + tensor var_3510_end_mask_0 = const()[name = string("op_3510_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_3510_cast_fp16 = slice_by_index(begin = var_3510_begin_0, end = var_3510_end_0, end_mask = var_3510_end_mask_0, x = x_375_cast_fp16)[name = string("op_3510_cast_fp16")]; + tensor var_3511 = const()[name = string("op_3511"), val = tensor([1, 8, 7, 97])]; + tensor matrix_bd_57_cast_fp16 = reshape(shape = var_3511, x = var_3510_cast_fp16)[name = string("matrix_bd_57_cast_fp16")]; + bool matrix_ac_29_transpose_x_0 = const()[name = string("matrix_ac_29_transpose_x_0"), val = bool(false)]; + bool matrix_ac_29_transpose_y_0 = const()[name = string("matrix_ac_29_transpose_y_0"), val = bool(false)]; + tensor transpose_124_perm_0 = const()[name = string("transpose_124_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_125_perm_0 = const()[name = string("transpose_125_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_125 = transpose(perm = transpose_125_perm_0, x = k_57_cast_fp16)[name = string("transpose_234")]; + tensor transpose_124 = transpose(perm = transpose_124_perm_0, x = var_3494_cast_fp16)[name = string("transpose_235")]; + tensor matrix_ac_29_cast_fp16 = matmul(transpose_x = matrix_ac_29_transpose_x_0, transpose_y = matrix_ac_29_transpose_y_0, x = transpose_124, y = transpose_125)[name = string("matrix_ac_29_cast_fp16")]; + tensor matrix_bd_59_begin_0 = const()[name = string("matrix_bd_59_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_59_end_0 = const()[name = string("matrix_bd_59_end_0"), val = tensor([1, 8, 7, 49])]; + tensor matrix_bd_59_end_mask_0 = const()[name = string("matrix_bd_59_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_59_cast_fp16 = slice_by_index(begin = matrix_bd_59_begin_0, end = matrix_bd_59_end_0, end_mask = matrix_bd_59_end_mask_0, x = matrix_bd_57_cast_fp16)[name = string("matrix_bd_59_cast_fp16")]; + tensor var_3520_cast_fp16 = add(x = matrix_ac_29_cast_fp16, y = matrix_bd_59_cast_fp16)[name = string("op_3520_cast_fp16")]; + fp16 _inversed_scores_57_y_0_to_fp16 = const()[name = string("_inversed_scores_57_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_57_cast_fp16 = mul(x = var_3520_cast_fp16, y = _inversed_scores_57_y_0_to_fp16)[name = string("_inversed_scores_57_cast_fp16")]; + tensor scores_59_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_57_cast_fp16, cond = mask_11)[name = string("scores_59_cast_fp16")]; + tensor var_3526_cast_fp16 = softmax(axis = var_59, x = scores_59_cast_fp16)[name = string("op_3526_cast_fp16")]; + tensor input_769_cast_fp16 = select(a = var_44_to_fp16, b = var_3526_cast_fp16, cond = mask_11)[name = string("input_769_cast_fp16")]; + bool x_377_transpose_x_0 = const()[name = string("x_377_transpose_x_0"), val = bool(false)]; + bool x_377_transpose_y_0 = const()[name = string("x_377_transpose_y_0"), val = bool(false)]; + tensor value_37_cast_fp16 = transpose(perm = value_37_perm_0, x = v_29_cast_fp16)[name = string("transpose_233")]; + tensor x_377_cast_fp16 = matmul(transpose_x = x_377_transpose_x_0, transpose_y = x_377_transpose_y_0, x = input_769_cast_fp16, y = value_37_cast_fp16)[name = string("x_377_cast_fp16")]; + tensor var_3530_perm_0 = const()[name = string("op_3530_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_3531 = const()[name = string("op_3531"), val = tensor([1, -1, 1024])]; + tensor var_3530_cast_fp16 = transpose(perm = var_3530_perm_0, x = x_377_cast_fp16)[name = string("transpose_232")]; + tensor input_771_cast_fp16 = reshape(shape = var_3531, x = var_3530_cast_fp16)[name = string("input_771_cast_fp16")]; + tensor encoder_layers_14_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293565120))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294351616))))[name = string("encoder_layers_14_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_14_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_14_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294351808)))]; + tensor linear_133_cast_fp16 = linear(bias = encoder_layers_14_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_14_self_attn_linear_out_weight_to_fp16_palettized, x = input_771_cast_fp16)[name = string("linear_133_cast_fp16")]; + tensor input_775_cast_fp16 = add(x = input_765_cast_fp16, y = linear_133_cast_fp16)[name = string("input_775_cast_fp16")]; + tensor x_381_axes_0 = const()[name = string("x_381_axes_0"), val = tensor([-1])]; + tensor encoder_layers_14_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_14_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294353920)))]; + tensor encoder_layers_14_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_14_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294356032)))]; + tensor x_381_cast_fp16 = layer_norm(axes = x_381_axes_0, beta = encoder_layers_14_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_14_norm_conv_weight_to_fp16, x = input_775_cast_fp16)[name = string("x_381_cast_fp16")]; + tensor input_777_perm_0 = const()[name = string("input_777_perm_0"), val = tensor([0, 2, 1])]; + string input_779_pad_type_0 = const()[name = string("input_779_pad_type_0"), val = string("valid")]; + tensor input_779_strides_0 = const()[name = string("input_779_strides_0"), val = tensor([1])]; + tensor input_779_pad_0 = const()[name = string("input_779_pad_0"), val = tensor([0, 0])]; + tensor input_779_dilations_0 = const()[name = string("input_779_dilations_0"), val = tensor([1])]; + int32 input_779_groups_0 = const()[name = string("input_779_groups_0"), val = int32(1)]; + tensor encoder_layers_14_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294358144))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296455360))))[name = string("encoder_layers_14_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_777_cast_fp16 = transpose(perm = input_777_perm_0, x = x_381_cast_fp16)[name = string("transpose_231")]; + tensor input_779_cast_fp16 = conv(dilations = input_779_dilations_0, groups = input_779_groups_0, pad = input_779_pad_0, pad_type = input_779_pad_type_0, strides = input_779_strides_0, weight = encoder_layers_14_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_777_cast_fp16)[name = string("input_779_cast_fp16")]; + int32 x_383_split_num_splits_0 = const()[name = string("x_383_split_num_splits_0"), val = int32(2)]; + int32 x_383_split_axis_0 = const()[name = string("x_383_split_axis_0"), val = int32(1)]; + tensor x_383_split_cast_fp16_0, tensor x_383_split_cast_fp16_1 = split(axis = x_383_split_axis_0, num_splits = x_383_split_num_splits_0, x = input_779_cast_fp16)[name = string("x_383_split_cast_fp16")]; + tensor x_383_split_1_sigmoid_cast_fp16 = sigmoid(x = x_383_split_cast_fp16_1)[name = string("x_383_split_1_sigmoid_cast_fp16")]; + tensor x_383_cast_fp16 = mul(x = x_383_split_cast_fp16_0, y = x_383_split_1_sigmoid_cast_fp16)[name = string("x_383_cast_fp16")]; + tensor input_781_cast_fp16 = select(a = var_44_to_fp16, b = x_383_cast_fp16, cond = var_575)[name = string("input_781_cast_fp16")]; + bool new_x_59_interleave_0 = const()[name = string("new_x_59_interleave_0"), val = bool(false)]; + tensor new_x_59_cast_fp16 = concat(axis = var_59, interleave = new_x_59_interleave_0, values = (cache_59_cast_fp16, input_781_cast_fp16))[name = string("new_x_59_cast_fp16")]; + tensor var_3570_begin_0 = const()[name = string("op_3570_begin_0"), val = tensor([0, 0, 7])]; + tensor var_3570_end_0 = const()[name = string("op_3570_end_0"), val = tensor([1, 1024, 15])]; + tensor var_3570_end_mask_0 = const()[name = string("op_3570_end_mask_0"), val = tensor([true, true, true])]; + tensor var_3570_cast_fp16 = slice_by_index(begin = var_3570_begin_0, end = var_3570_end_0, end_mask = var_3570_end_mask_0, x = new_x_59_cast_fp16)[name = string("op_3570_cast_fp16")]; + string x_385_pad_type_0 = const()[name = string("x_385_pad_type_0"), val = string("valid")]; + int32 x_385_groups_0 = const()[name = string("x_385_groups_0"), val = int32(1024)]; + tensor x_385_strides_0 = const()[name = string("x_385_strides_0"), val = tensor([1])]; + tensor x_385_pad_0 = const()[name = string("x_385_pad_0"), val = tensor([0, 0])]; + tensor x_385_dilations_0 = const()[name = string("x_385_dilations_0"), val = tensor([1])]; + tensor encoder_layers_14_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296459520))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296468800))))[name = string("encoder_layers_14_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_385_cast_fp16 = conv(dilations = x_385_dilations_0, groups = x_385_groups_0, pad = x_385_pad_0, pad_type = x_385_pad_type_0, strides = x_385_strides_0, weight = encoder_layers_14_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_59_cast_fp16)[name = string("x_385_cast_fp16")]; + tensor input_783_perm_0 = const()[name = string("input_783_perm_0"), val = tensor([0, 2, 1])]; + tensor x_387_axes_0 = const()[name = string("x_387_axes_0"), val = tensor([-1])]; + tensor encoder_layers_14_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_14_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296470912)))]; + tensor encoder_layers_14_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_14_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296473024)))]; + tensor input_783_cast_fp16 = transpose(perm = input_783_perm_0, x = x_385_cast_fp16)[name = string("transpose_230")]; + tensor x_387_cast_fp16 = layer_norm(axes = x_387_axes_0, beta = encoder_layers_14_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_14_conv_batch_norm_weight_to_fp16, x = input_783_cast_fp16)[name = string("x_387_cast_fp16")]; + tensor input_785_perm_0 = const()[name = string("input_785_perm_0"), val = tensor([0, 2, 1])]; + tensor input_785_cast_fp16 = transpose(perm = input_785_perm_0, x = x_387_cast_fp16)[name = string("transpose_229")]; + tensor input_787_cast_fp16 = silu(x = input_785_cast_fp16)[name = string("input_787_cast_fp16")]; + string x_389_pad_type_0 = const()[name = string("x_389_pad_type_0"), val = string("valid")]; + tensor x_389_strides_0 = const()[name = string("x_389_strides_0"), val = tensor([1])]; + tensor x_389_pad_0 = const()[name = string("x_389_pad_0"), val = tensor([0, 0])]; + tensor x_389_dilations_0 = const()[name = string("x_389_dilations_0"), val = tensor([1])]; + int32 x_389_groups_0 = const()[name = string("x_389_groups_0"), val = int32(1)]; + tensor encoder_layers_14_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296475136))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297523776))))[name = string("encoder_layers_14_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_389_cast_fp16 = conv(dilations = x_389_dilations_0, groups = x_389_groups_0, pad = x_389_pad_0, pad_type = x_389_pad_type_0, strides = x_389_strides_0, weight = encoder_layers_14_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_787_cast_fp16)[name = string("x_389_cast_fp16")]; + tensor input_789_perm_0 = const()[name = string("input_789_perm_0"), val = tensor([0, 2, 1])]; + tensor input_789_cast_fp16 = transpose(perm = input_789_perm_0, x = x_389_cast_fp16)[name = string("transpose_228")]; + tensor input_791_cast_fp16 = add(x = input_775_cast_fp16, y = input_789_cast_fp16)[name = string("input_791_cast_fp16")]; + tensor input_793_axes_0 = const()[name = string("input_793_axes_0"), val = tensor([-1])]; + tensor encoder_layers_14_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_14_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297525888)))]; + tensor encoder_layers_14_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_14_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297528000)))]; + tensor input_793_cast_fp16 = layer_norm(axes = input_793_axes_0, beta = encoder_layers_14_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_14_norm_feed_forward2_weight_to_fp16, x = input_791_cast_fp16)[name = string("input_793_cast_fp16")]; + tensor encoder_layers_14_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297530112))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(300675904))))[name = string("encoder_layers_14_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_14_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_14_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(300676096)))]; + tensor linear_134_cast_fp16 = linear(bias = encoder_layers_14_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_14_feed_forward2_linear1_weight_to_fp16_palettized, x = input_793_cast_fp16)[name = string("linear_134_cast_fp16")]; + tensor input_797_cast_fp16 = silu(x = linear_134_cast_fp16)[name = string("input_797_cast_fp16")]; + tensor encoder_layers_14_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(300684352))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(303830144))))[name = string("encoder_layers_14_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_14_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_14_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(303830336)))]; + tensor linear_135_cast_fp16 = linear(bias = encoder_layers_14_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_14_feed_forward2_linear2_weight_to_fp16_palettized, x = input_797_cast_fp16)[name = string("linear_135_cast_fp16")]; + fp16 var_3613_to_fp16 = const()[name = string("op_3613_to_fp16"), val = fp16(0x1p-1)]; + tensor var_3614_cast_fp16 = mul(x = linear_135_cast_fp16, y = var_3613_to_fp16)[name = string("op_3614_cast_fp16")]; + tensor input_803_cast_fp16 = add(x = input_791_cast_fp16, y = var_3614_cast_fp16)[name = string("input_803_cast_fp16")]; + tensor input_805_axes_0 = const()[name = string("input_805_axes_0"), val = tensor([-1])]; + tensor encoder_layers_14_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_14_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(303832448)))]; + tensor encoder_layers_14_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_14_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(303834560)))]; + tensor input_805_cast_fp16 = layer_norm(axes = input_805_axes_0, beta = encoder_layers_14_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_14_norm_out_weight_to_fp16, x = input_803_cast_fp16)[name = string("input_805_cast_fp16")]; + tensor cache_61_begin_0 = const()[name = string("cache_61_begin_0"), val = tensor([15, 0, 0, 0])]; + tensor cache_61_end_0 = const()[name = string("cache_61_end_0"), val = tensor([16, 1, 42, 1024])]; + tensor cache_61_end_mask_0 = const()[name = string("cache_61_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_61_squeeze_mask_0 = const()[name = string("cache_61_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_61_cast_fp16 = slice_by_index(begin = cache_61_begin_0, end = cache_61_end_0, end_mask = cache_61_end_mask_0, squeeze_mask = cache_61_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_61_cast_fp16")]; + tensor cache_63_begin_0 = const()[name = string("cache_63_begin_0"), val = tensor([15, 0, 0, 0])]; + tensor cache_63_end_0 = const()[name = string("cache_63_end_0"), val = tensor([16, 1, 1024, 8])]; + tensor cache_63_end_mask_0 = const()[name = string("cache_63_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_63_squeeze_mask_0 = const()[name = string("cache_63_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_63_cast_fp16 = slice_by_index(begin = cache_63_begin_0, end = cache_63_end_0, end_mask = cache_63_end_mask_0, squeeze_mask = cache_63_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_63_cast_fp16")]; + tensor input_807_axes_0 = const()[name = string("input_807_axes_0"), val = tensor([-1])]; + tensor encoder_layers_15_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_15_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(303836672)))]; + tensor encoder_layers_15_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_15_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(303838784)))]; + tensor input_807_cast_fp16 = layer_norm(axes = input_807_axes_0, beta = encoder_layers_15_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_15_norm_feed_forward1_weight_to_fp16, x = input_805_cast_fp16)[name = string("input_807_cast_fp16")]; + tensor encoder_layers_15_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(303840896))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(306986688))))[name = string("encoder_layers_15_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_15_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_15_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(306986880)))]; + tensor linear_136_cast_fp16 = linear(bias = encoder_layers_15_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_15_feed_forward1_linear1_weight_to_fp16_palettized, x = input_807_cast_fp16)[name = string("linear_136_cast_fp16")]; + tensor input_811_cast_fp16 = silu(x = linear_136_cast_fp16)[name = string("input_811_cast_fp16")]; + tensor encoder_layers_15_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(306995136))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310140928))))[name = string("encoder_layers_15_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_15_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_15_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310141120)))]; + tensor linear_137_cast_fp16 = linear(bias = encoder_layers_15_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_15_feed_forward1_linear2_weight_to_fp16_palettized, x = input_811_cast_fp16)[name = string("linear_137_cast_fp16")]; + fp16 var_3650_to_fp16 = const()[name = string("op_3650_to_fp16"), val = fp16(0x1p-1)]; + tensor var_3651_cast_fp16 = mul(x = linear_137_cast_fp16, y = var_3650_to_fp16)[name = string("op_3651_cast_fp16")]; + tensor input_817_cast_fp16 = add(x = input_805_cast_fp16, y = var_3651_cast_fp16)[name = string("input_817_cast_fp16")]; + tensor key_31_axes_0 = const()[name = string("key_31_axes_0"), val = tensor([-1])]; + tensor encoder_layers_15_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_15_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310143232)))]; + tensor encoder_layers_15_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_15_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310145344)))]; + tensor key_31_cast_fp16 = layer_norm(axes = key_31_axes_0, beta = encoder_layers_15_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_15_norm_self_att_weight_to_fp16, x = input_817_cast_fp16)[name = string("key_31_cast_fp16")]; + bool input_819_interleave_0 = const()[name = string("input_819_interleave_0"), val = bool(false)]; + tensor input_819_cast_fp16 = concat(axis = var_68, interleave = input_819_interleave_0, values = (cache_61_cast_fp16, key_31_cast_fp16))[name = string("input_819_cast_fp16")]; + tensor var_3673_begin_0 = const()[name = string("op_3673_begin_0"), val = tensor([0, 7, 0])]; + tensor var_3673_end_0 = const()[name = string("op_3673_end_0"), val = tensor([1, 42, 1024])]; + tensor var_3673_end_mask_0 = const()[name = string("op_3673_end_mask_0"), val = tensor([true, true, true])]; + tensor var_3673_cast_fp16 = slice_by_index(begin = var_3673_begin_0, end = var_3673_end_0, end_mask = var_3673_end_mask_0, x = cache_61_cast_fp16)[name = string("op_3673_cast_fp16")]; + bool var_3679_interleave_0 = const()[name = string("op_3679_interleave_0"), val = bool(false)]; + tensor var_3679_cast_fp16 = concat(axis = var_68, interleave = var_3679_interleave_0, values = (var_3673_cast_fp16, key_31_cast_fp16))[name = string("op_3679_cast_fp16")]; + tensor encoder_layers_15_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310147456))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310933952))))[name = string("encoder_layers_15_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_15_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_15_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310934144)))]; + tensor linear_138_cast_fp16 = linear(bias = encoder_layers_15_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_15_self_attn_linear_q_weight_to_fp16_palettized, x = key_31_cast_fp16)[name = string("linear_138_cast_fp16")]; + tensor var_3684 = const()[name = string("op_3684"), val = tensor([1, -1, 8, 128])]; + tensor q_91_cast_fp16 = reshape(shape = var_3684, x = linear_138_cast_fp16)[name = string("q_91_cast_fp16")]; + tensor encoder_layers_15_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310936256))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311722752))))[name = string("encoder_layers_15_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_15_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_15_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311722944)))]; + tensor linear_139_cast_fp16 = linear(bias = encoder_layers_15_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_15_self_attn_linear_k_weight_to_fp16_palettized, x = input_819_cast_fp16)[name = string("linear_139_cast_fp16")]; + tensor var_3689 = const()[name = string("op_3689"), val = tensor([1, -1, 8, 128])]; + tensor k_61_cast_fp16 = reshape(shape = var_3689, x = linear_139_cast_fp16)[name = string("k_61_cast_fp16")]; + tensor encoder_layers_15_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311725056))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(312511552))))[name = string("encoder_layers_15_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_15_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_15_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(312511744)))]; + tensor linear_140_cast_fp16 = linear(bias = encoder_layers_15_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_15_self_attn_linear_v_weight_to_fp16_palettized, x = input_819_cast_fp16)[name = string("linear_140_cast_fp16")]; + tensor var_3694 = const()[name = string("op_3694"), val = tensor([1, -1, 8, 128])]; + tensor v_31_cast_fp16 = reshape(shape = var_3694, x = linear_140_cast_fp16)[name = string("v_31_cast_fp16")]; + tensor value_39_perm_0 = const()[name = string("value_39_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_15_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_15_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(312513856)))]; + tensor var_3707_cast_fp16 = add(x = q_91_cast_fp16, y = encoder_layers_15_self_attn_pos_bias_u_to_fp16)[name = string("op_3707_cast_fp16")]; + tensor encoder_layers_15_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_15_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(312515968)))]; + tensor var_3709_cast_fp16 = add(x = q_91_cast_fp16, y = encoder_layers_15_self_attn_pos_bias_v_to_fp16)[name = string("op_3709_cast_fp16")]; + tensor q_with_bias_v_31_perm_0 = const()[name = string("q_with_bias_v_31_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_397_transpose_x_0 = const()[name = string("x_397_transpose_x_0"), val = bool(false)]; + bool x_397_transpose_y_0 = const()[name = string("x_397_transpose_y_0"), val = bool(false)]; + tensor op_3711_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(312518080))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(312617472))))[name = string("op_3711_to_fp16_quantized")]; + tensor q_with_bias_v_31_cast_fp16 = transpose(perm = q_with_bias_v_31_perm_0, x = var_3709_cast_fp16)[name = string("transpose_227")]; + tensor x_397_cast_fp16 = matmul(transpose_x = x_397_transpose_x_0, transpose_y = x_397_transpose_y_0, x = q_with_bias_v_31_cast_fp16, y = op_3711_to_fp16_quantized)[name = string("x_397_cast_fp16")]; + tensor x_399_pad_0 = const()[name = string("x_399_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_399_mode_0 = const()[name = string("x_399_mode_0"), val = string("constant")]; + fp16 const_274_to_fp16 = const()[name = string("const_274_to_fp16"), val = fp16(0x0p+0)]; + tensor x_399_cast_fp16 = pad(constant_val = const_274_to_fp16, mode = x_399_mode_0, pad = x_399_pad_0, x = x_397_cast_fp16)[name = string("x_399_cast_fp16")]; + tensor var_3719 = const()[name = string("op_3719"), val = tensor([1, 8, -1, 7])]; + tensor x_401_cast_fp16 = reshape(shape = var_3719, x = x_399_cast_fp16)[name = string("x_401_cast_fp16")]; + tensor var_3723_begin_0 = const()[name = string("op_3723_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_3723_end_0 = const()[name = string("op_3723_end_0"), val = tensor([1, 8, 98, 7])]; + tensor var_3723_end_mask_0 = const()[name = string("op_3723_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_3723_cast_fp16 = slice_by_index(begin = var_3723_begin_0, end = var_3723_end_0, end_mask = var_3723_end_mask_0, x = x_401_cast_fp16)[name = string("op_3723_cast_fp16")]; + tensor var_3724 = const()[name = string("op_3724"), val = tensor([1, 8, 7, 97])]; + tensor matrix_bd_61_cast_fp16 = reshape(shape = var_3724, x = var_3723_cast_fp16)[name = string("matrix_bd_61_cast_fp16")]; + bool matrix_ac_31_transpose_x_0 = const()[name = string("matrix_ac_31_transpose_x_0"), val = bool(false)]; + bool matrix_ac_31_transpose_y_0 = const()[name = string("matrix_ac_31_transpose_y_0"), val = bool(false)]; + tensor transpose_126_perm_0 = const()[name = string("transpose_126_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_127_perm_0 = const()[name = string("transpose_127_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_127 = transpose(perm = transpose_127_perm_0, x = k_61_cast_fp16)[name = string("transpose_225")]; + tensor transpose_126 = transpose(perm = transpose_126_perm_0, x = var_3707_cast_fp16)[name = string("transpose_226")]; + tensor matrix_ac_31_cast_fp16 = matmul(transpose_x = matrix_ac_31_transpose_x_0, transpose_y = matrix_ac_31_transpose_y_0, x = transpose_126, y = transpose_127)[name = string("matrix_ac_31_cast_fp16")]; + tensor matrix_bd_63_begin_0 = const()[name = string("matrix_bd_63_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_63_end_0 = const()[name = string("matrix_bd_63_end_0"), val = tensor([1, 8, 7, 49])]; + tensor matrix_bd_63_end_mask_0 = const()[name = string("matrix_bd_63_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_63_cast_fp16 = slice_by_index(begin = matrix_bd_63_begin_0, end = matrix_bd_63_end_0, end_mask = matrix_bd_63_end_mask_0, x = matrix_bd_61_cast_fp16)[name = string("matrix_bd_63_cast_fp16")]; + tensor var_3733_cast_fp16 = add(x = matrix_ac_31_cast_fp16, y = matrix_bd_63_cast_fp16)[name = string("op_3733_cast_fp16")]; + fp16 _inversed_scores_61_y_0_to_fp16 = const()[name = string("_inversed_scores_61_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_61_cast_fp16 = mul(x = var_3733_cast_fp16, y = _inversed_scores_61_y_0_to_fp16)[name = string("_inversed_scores_61_cast_fp16")]; + tensor scores_63_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_61_cast_fp16, cond = mask_11)[name = string("scores_63_cast_fp16")]; + tensor var_3739_cast_fp16 = softmax(axis = var_59, x = scores_63_cast_fp16)[name = string("op_3739_cast_fp16")]; + tensor input_821_cast_fp16 = select(a = var_44_to_fp16, b = var_3739_cast_fp16, cond = mask_11)[name = string("input_821_cast_fp16")]; + bool x_403_transpose_x_0 = const()[name = string("x_403_transpose_x_0"), val = bool(false)]; + bool x_403_transpose_y_0 = const()[name = string("x_403_transpose_y_0"), val = bool(false)]; + tensor value_39_cast_fp16 = transpose(perm = value_39_perm_0, x = v_31_cast_fp16)[name = string("transpose_224")]; + tensor x_403_cast_fp16 = matmul(transpose_x = x_403_transpose_x_0, transpose_y = x_403_transpose_y_0, x = input_821_cast_fp16, y = value_39_cast_fp16)[name = string("x_403_cast_fp16")]; + tensor var_3743_perm_0 = const()[name = string("op_3743_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_3744 = const()[name = string("op_3744"), val = tensor([1, -1, 1024])]; + tensor var_3743_cast_fp16 = transpose(perm = var_3743_perm_0, x = x_403_cast_fp16)[name = string("transpose_223")]; + tensor input_823_cast_fp16 = reshape(shape = var_3744, x = var_3743_cast_fp16)[name = string("input_823_cast_fp16")]; + tensor encoder_layers_15_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(312617792))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(313404288))))[name = string("encoder_layers_15_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_15_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_15_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(313404480)))]; + tensor linear_142_cast_fp16 = linear(bias = encoder_layers_15_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_15_self_attn_linear_out_weight_to_fp16_palettized, x = input_823_cast_fp16)[name = string("linear_142_cast_fp16")]; + tensor input_827_cast_fp16 = add(x = input_817_cast_fp16, y = linear_142_cast_fp16)[name = string("input_827_cast_fp16")]; + tensor x_407_axes_0 = const()[name = string("x_407_axes_0"), val = tensor([-1])]; + tensor encoder_layers_15_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_15_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(313406592)))]; + tensor encoder_layers_15_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_15_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(313408704)))]; + tensor x_407_cast_fp16 = layer_norm(axes = x_407_axes_0, beta = encoder_layers_15_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_15_norm_conv_weight_to_fp16, x = input_827_cast_fp16)[name = string("x_407_cast_fp16")]; + tensor input_829_perm_0 = const()[name = string("input_829_perm_0"), val = tensor([0, 2, 1])]; + string input_831_pad_type_0 = const()[name = string("input_831_pad_type_0"), val = string("valid")]; + tensor input_831_strides_0 = const()[name = string("input_831_strides_0"), val = tensor([1])]; + tensor input_831_pad_0 = const()[name = string("input_831_pad_0"), val = tensor([0, 0])]; + tensor input_831_dilations_0 = const()[name = string("input_831_dilations_0"), val = tensor([1])]; + int32 input_831_groups_0 = const()[name = string("input_831_groups_0"), val = int32(1)]; + tensor encoder_layers_15_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(313410816))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315508032))))[name = string("encoder_layers_15_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_829_cast_fp16 = transpose(perm = input_829_perm_0, x = x_407_cast_fp16)[name = string("transpose_222")]; + tensor input_831_cast_fp16 = conv(dilations = input_831_dilations_0, groups = input_831_groups_0, pad = input_831_pad_0, pad_type = input_831_pad_type_0, strides = input_831_strides_0, weight = encoder_layers_15_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_829_cast_fp16)[name = string("input_831_cast_fp16")]; + int32 x_409_split_num_splits_0 = const()[name = string("x_409_split_num_splits_0"), val = int32(2)]; + int32 x_409_split_axis_0 = const()[name = string("x_409_split_axis_0"), val = int32(1)]; + tensor x_409_split_cast_fp16_0, tensor x_409_split_cast_fp16_1 = split(axis = x_409_split_axis_0, num_splits = x_409_split_num_splits_0, x = input_831_cast_fp16)[name = string("x_409_split_cast_fp16")]; + tensor x_409_split_1_sigmoid_cast_fp16 = sigmoid(x = x_409_split_cast_fp16_1)[name = string("x_409_split_1_sigmoid_cast_fp16")]; + tensor x_409_cast_fp16 = mul(x = x_409_split_cast_fp16_0, y = x_409_split_1_sigmoid_cast_fp16)[name = string("x_409_cast_fp16")]; + tensor input_833_cast_fp16 = select(a = var_44_to_fp16, b = x_409_cast_fp16, cond = var_575)[name = string("input_833_cast_fp16")]; + bool new_x_63_interleave_0 = const()[name = string("new_x_63_interleave_0"), val = bool(false)]; + tensor new_x_63_cast_fp16 = concat(axis = var_59, interleave = new_x_63_interleave_0, values = (cache_63_cast_fp16, input_833_cast_fp16))[name = string("new_x_63_cast_fp16")]; + tensor var_3783_begin_0 = const()[name = string("op_3783_begin_0"), val = tensor([0, 0, 7])]; + tensor var_3783_end_0 = const()[name = string("op_3783_end_0"), val = tensor([1, 1024, 15])]; + tensor var_3783_end_mask_0 = const()[name = string("op_3783_end_mask_0"), val = tensor([true, true, true])]; + tensor var_3783_cast_fp16 = slice_by_index(begin = var_3783_begin_0, end = var_3783_end_0, end_mask = var_3783_end_mask_0, x = new_x_63_cast_fp16)[name = string("op_3783_cast_fp16")]; + string x_411_pad_type_0 = const()[name = string("x_411_pad_type_0"), val = string("valid")]; + int32 x_411_groups_0 = const()[name = string("x_411_groups_0"), val = int32(1024)]; + tensor x_411_strides_0 = const()[name = string("x_411_strides_0"), val = tensor([1])]; + tensor x_411_pad_0 = const()[name = string("x_411_pad_0"), val = tensor([0, 0])]; + tensor x_411_dilations_0 = const()[name = string("x_411_dilations_0"), val = tensor([1])]; + tensor encoder_layers_15_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315512192))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315521472))))[name = string("encoder_layers_15_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_411_cast_fp16 = conv(dilations = x_411_dilations_0, groups = x_411_groups_0, pad = x_411_pad_0, pad_type = x_411_pad_type_0, strides = x_411_strides_0, weight = encoder_layers_15_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_63_cast_fp16)[name = string("x_411_cast_fp16")]; + tensor input_835_perm_0 = const()[name = string("input_835_perm_0"), val = tensor([0, 2, 1])]; + tensor x_413_axes_0 = const()[name = string("x_413_axes_0"), val = tensor([-1])]; + tensor encoder_layers_15_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_15_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315523584)))]; + tensor encoder_layers_15_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_15_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315525696)))]; + tensor input_835_cast_fp16 = transpose(perm = input_835_perm_0, x = x_411_cast_fp16)[name = string("transpose_221")]; + tensor x_413_cast_fp16 = layer_norm(axes = x_413_axes_0, beta = encoder_layers_15_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_15_conv_batch_norm_weight_to_fp16, x = input_835_cast_fp16)[name = string("x_413_cast_fp16")]; + tensor input_837_perm_0 = const()[name = string("input_837_perm_0"), val = tensor([0, 2, 1])]; + tensor input_837_cast_fp16 = transpose(perm = input_837_perm_0, x = x_413_cast_fp16)[name = string("transpose_220")]; + tensor input_839_cast_fp16 = silu(x = input_837_cast_fp16)[name = string("input_839_cast_fp16")]; + string x_415_pad_type_0 = const()[name = string("x_415_pad_type_0"), val = string("valid")]; + tensor x_415_strides_0 = const()[name = string("x_415_strides_0"), val = tensor([1])]; + tensor x_415_pad_0 = const()[name = string("x_415_pad_0"), val = tensor([0, 0])]; + tensor x_415_dilations_0 = const()[name = string("x_415_dilations_0"), val = tensor([1])]; + int32 x_415_groups_0 = const()[name = string("x_415_groups_0"), val = int32(1)]; + tensor encoder_layers_15_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315527808))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316576448))))[name = string("encoder_layers_15_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_415_cast_fp16 = conv(dilations = x_415_dilations_0, groups = x_415_groups_0, pad = x_415_pad_0, pad_type = x_415_pad_type_0, strides = x_415_strides_0, weight = encoder_layers_15_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_839_cast_fp16)[name = string("x_415_cast_fp16")]; + tensor input_841_perm_0 = const()[name = string("input_841_perm_0"), val = tensor([0, 2, 1])]; + tensor input_841_cast_fp16 = transpose(perm = input_841_perm_0, x = x_415_cast_fp16)[name = string("transpose_219")]; + tensor input_843_cast_fp16 = add(x = input_827_cast_fp16, y = input_841_cast_fp16)[name = string("input_843_cast_fp16")]; + tensor input_845_axes_0 = const()[name = string("input_845_axes_0"), val = tensor([-1])]; + tensor encoder_layers_15_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_15_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316578560)))]; + tensor encoder_layers_15_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_15_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316580672)))]; + tensor input_845_cast_fp16 = layer_norm(axes = input_845_axes_0, beta = encoder_layers_15_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_15_norm_feed_forward2_weight_to_fp16, x = input_843_cast_fp16)[name = string("input_845_cast_fp16")]; + tensor encoder_layers_15_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316582784))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(319728576))))[name = string("encoder_layers_15_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_15_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_15_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(319728768)))]; + tensor linear_143_cast_fp16 = linear(bias = encoder_layers_15_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_15_feed_forward2_linear1_weight_to_fp16_palettized, x = input_845_cast_fp16)[name = string("linear_143_cast_fp16")]; + tensor input_849_cast_fp16 = silu(x = linear_143_cast_fp16)[name = string("input_849_cast_fp16")]; + tensor encoder_layers_15_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(319737024))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(322882816))))[name = string("encoder_layers_15_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_15_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_15_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(322883008)))]; + tensor linear_144_cast_fp16 = linear(bias = encoder_layers_15_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_15_feed_forward2_linear2_weight_to_fp16_palettized, x = input_849_cast_fp16)[name = string("linear_144_cast_fp16")]; + fp16 var_3826_to_fp16 = const()[name = string("op_3826_to_fp16"), val = fp16(0x1p-1)]; + tensor var_3827_cast_fp16 = mul(x = linear_144_cast_fp16, y = var_3826_to_fp16)[name = string("op_3827_cast_fp16")]; + tensor input_855_cast_fp16 = add(x = input_843_cast_fp16, y = var_3827_cast_fp16)[name = string("input_855_cast_fp16")]; + tensor input_857_axes_0 = const()[name = string("input_857_axes_0"), val = tensor([-1])]; + tensor encoder_layers_15_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_15_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(322885120)))]; + tensor encoder_layers_15_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_15_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(322887232)))]; + tensor input_857_cast_fp16 = layer_norm(axes = input_857_axes_0, beta = encoder_layers_15_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_15_norm_out_weight_to_fp16, x = input_855_cast_fp16)[name = string("input_857_cast_fp16")]; + tensor cache_65_begin_0 = const()[name = string("cache_65_begin_0"), val = tensor([16, 0, 0, 0])]; + tensor cache_65_end_0 = const()[name = string("cache_65_end_0"), val = tensor([17, 1, 42, 1024])]; + tensor cache_65_end_mask_0 = const()[name = string("cache_65_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_65_squeeze_mask_0 = const()[name = string("cache_65_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_65_cast_fp16 = slice_by_index(begin = cache_65_begin_0, end = cache_65_end_0, end_mask = cache_65_end_mask_0, squeeze_mask = cache_65_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_65_cast_fp16")]; + tensor cache_67_begin_0 = const()[name = string("cache_67_begin_0"), val = tensor([16, 0, 0, 0])]; + tensor cache_67_end_0 = const()[name = string("cache_67_end_0"), val = tensor([17, 1, 1024, 8])]; + tensor cache_67_end_mask_0 = const()[name = string("cache_67_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_67_squeeze_mask_0 = const()[name = string("cache_67_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_67_cast_fp16 = slice_by_index(begin = cache_67_begin_0, end = cache_67_end_0, end_mask = cache_67_end_mask_0, squeeze_mask = cache_67_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_67_cast_fp16")]; + tensor input_859_axes_0 = const()[name = string("input_859_axes_0"), val = tensor([-1])]; + tensor encoder_layers_16_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_16_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(322889344)))]; + tensor encoder_layers_16_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_16_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(322891456)))]; + tensor input_859_cast_fp16 = layer_norm(axes = input_859_axes_0, beta = encoder_layers_16_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_16_norm_feed_forward1_weight_to_fp16, x = input_857_cast_fp16)[name = string("input_859_cast_fp16")]; + tensor encoder_layers_16_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(322893568))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(326039360))))[name = string("encoder_layers_16_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_16_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_16_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(326039552)))]; + tensor linear_145_cast_fp16 = linear(bias = encoder_layers_16_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_16_feed_forward1_linear1_weight_to_fp16_palettized, x = input_859_cast_fp16)[name = string("linear_145_cast_fp16")]; + tensor input_863_cast_fp16 = silu(x = linear_145_cast_fp16)[name = string("input_863_cast_fp16")]; + tensor encoder_layers_16_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(326047808))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(329193600))))[name = string("encoder_layers_16_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_16_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_16_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(329193792)))]; + tensor linear_146_cast_fp16 = linear(bias = encoder_layers_16_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_16_feed_forward1_linear2_weight_to_fp16_palettized, x = input_863_cast_fp16)[name = string("linear_146_cast_fp16")]; + fp16 var_3863_to_fp16 = const()[name = string("op_3863_to_fp16"), val = fp16(0x1p-1)]; + tensor var_3864_cast_fp16 = mul(x = linear_146_cast_fp16, y = var_3863_to_fp16)[name = string("op_3864_cast_fp16")]; + tensor input_869_cast_fp16 = add(x = input_857_cast_fp16, y = var_3864_cast_fp16)[name = string("input_869_cast_fp16")]; + tensor key_33_axes_0 = const()[name = string("key_33_axes_0"), val = tensor([-1])]; + tensor encoder_layers_16_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_16_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(329195904)))]; + tensor encoder_layers_16_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_16_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(329198016)))]; + tensor key_33_cast_fp16 = layer_norm(axes = key_33_axes_0, beta = encoder_layers_16_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_16_norm_self_att_weight_to_fp16, x = input_869_cast_fp16)[name = string("key_33_cast_fp16")]; + bool input_871_interleave_0 = const()[name = string("input_871_interleave_0"), val = bool(false)]; + tensor input_871_cast_fp16 = concat(axis = var_68, interleave = input_871_interleave_0, values = (cache_65_cast_fp16, key_33_cast_fp16))[name = string("input_871_cast_fp16")]; + tensor var_3886_begin_0 = const()[name = string("op_3886_begin_0"), val = tensor([0, 7, 0])]; + tensor var_3886_end_0 = const()[name = string("op_3886_end_0"), val = tensor([1, 42, 1024])]; + tensor var_3886_end_mask_0 = const()[name = string("op_3886_end_mask_0"), val = tensor([true, true, true])]; + tensor var_3886_cast_fp16 = slice_by_index(begin = var_3886_begin_0, end = var_3886_end_0, end_mask = var_3886_end_mask_0, x = cache_65_cast_fp16)[name = string("op_3886_cast_fp16")]; + bool var_3892_interleave_0 = const()[name = string("op_3892_interleave_0"), val = bool(false)]; + tensor var_3892_cast_fp16 = concat(axis = var_68, interleave = var_3892_interleave_0, values = (var_3886_cast_fp16, key_33_cast_fp16))[name = string("op_3892_cast_fp16")]; + tensor encoder_layers_16_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(329200128))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(329986624))))[name = string("encoder_layers_16_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_16_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_16_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(329986816)))]; + tensor linear_147_cast_fp16 = linear(bias = encoder_layers_16_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_16_self_attn_linear_q_weight_to_fp16_palettized, x = key_33_cast_fp16)[name = string("linear_147_cast_fp16")]; + tensor var_3897 = const()[name = string("op_3897"), val = tensor([1, -1, 8, 128])]; + tensor q_97_cast_fp16 = reshape(shape = var_3897, x = linear_147_cast_fp16)[name = string("q_97_cast_fp16")]; + tensor encoder_layers_16_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(329988928))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(330775424))))[name = string("encoder_layers_16_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_16_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_16_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(330775616)))]; + tensor linear_148_cast_fp16 = linear(bias = encoder_layers_16_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_16_self_attn_linear_k_weight_to_fp16_palettized, x = input_871_cast_fp16)[name = string("linear_148_cast_fp16")]; + tensor var_3902 = const()[name = string("op_3902"), val = tensor([1, -1, 8, 128])]; + tensor k_65_cast_fp16 = reshape(shape = var_3902, x = linear_148_cast_fp16)[name = string("k_65_cast_fp16")]; + tensor encoder_layers_16_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(330777728))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331564224))))[name = string("encoder_layers_16_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_16_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_16_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331564416)))]; + tensor linear_149_cast_fp16 = linear(bias = encoder_layers_16_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_16_self_attn_linear_v_weight_to_fp16_palettized, x = input_871_cast_fp16)[name = string("linear_149_cast_fp16")]; + tensor var_3907 = const()[name = string("op_3907"), val = tensor([1, -1, 8, 128])]; + tensor v_33_cast_fp16 = reshape(shape = var_3907, x = linear_149_cast_fp16)[name = string("v_33_cast_fp16")]; + tensor value_41_perm_0 = const()[name = string("value_41_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_16_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_16_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331566528)))]; + tensor var_3920_cast_fp16 = add(x = q_97_cast_fp16, y = encoder_layers_16_self_attn_pos_bias_u_to_fp16)[name = string("op_3920_cast_fp16")]; + tensor encoder_layers_16_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_16_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331568640)))]; + tensor var_3922_cast_fp16 = add(x = q_97_cast_fp16, y = encoder_layers_16_self_attn_pos_bias_v_to_fp16)[name = string("op_3922_cast_fp16")]; + tensor q_with_bias_v_33_perm_0 = const()[name = string("q_with_bias_v_33_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_423_transpose_x_0 = const()[name = string("x_423_transpose_x_0"), val = bool(false)]; + bool x_423_transpose_y_0 = const()[name = string("x_423_transpose_y_0"), val = bool(false)]; + tensor op_3924_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331570752))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331670144))))[name = string("op_3924_to_fp16_quantized")]; + tensor q_with_bias_v_33_cast_fp16 = transpose(perm = q_with_bias_v_33_perm_0, x = var_3922_cast_fp16)[name = string("transpose_218")]; + tensor x_423_cast_fp16 = matmul(transpose_x = x_423_transpose_x_0, transpose_y = x_423_transpose_y_0, x = q_with_bias_v_33_cast_fp16, y = op_3924_to_fp16_quantized)[name = string("x_423_cast_fp16")]; + tensor x_425_pad_0 = const()[name = string("x_425_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_425_mode_0 = const()[name = string("x_425_mode_0"), val = string("constant")]; + fp16 const_287_to_fp16 = const()[name = string("const_287_to_fp16"), val = fp16(0x0p+0)]; + tensor x_425_cast_fp16 = pad(constant_val = const_287_to_fp16, mode = x_425_mode_0, pad = x_425_pad_0, x = x_423_cast_fp16)[name = string("x_425_cast_fp16")]; + tensor var_3932 = const()[name = string("op_3932"), val = tensor([1, 8, -1, 7])]; + tensor x_427_cast_fp16 = reshape(shape = var_3932, x = x_425_cast_fp16)[name = string("x_427_cast_fp16")]; + tensor var_3936_begin_0 = const()[name = string("op_3936_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_3936_end_0 = const()[name = string("op_3936_end_0"), val = tensor([1, 8, 98, 7])]; + tensor var_3936_end_mask_0 = const()[name = string("op_3936_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_3936_cast_fp16 = slice_by_index(begin = var_3936_begin_0, end = var_3936_end_0, end_mask = var_3936_end_mask_0, x = x_427_cast_fp16)[name = string("op_3936_cast_fp16")]; + tensor var_3937 = const()[name = string("op_3937"), val = tensor([1, 8, 7, 97])]; + tensor matrix_bd_65_cast_fp16 = reshape(shape = var_3937, x = var_3936_cast_fp16)[name = string("matrix_bd_65_cast_fp16")]; + bool matrix_ac_33_transpose_x_0 = const()[name = string("matrix_ac_33_transpose_x_0"), val = bool(false)]; + bool matrix_ac_33_transpose_y_0 = const()[name = string("matrix_ac_33_transpose_y_0"), val = bool(false)]; + tensor transpose_128_perm_0 = const()[name = string("transpose_128_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_129_perm_0 = const()[name = string("transpose_129_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_129 = transpose(perm = transpose_129_perm_0, x = k_65_cast_fp16)[name = string("transpose_216")]; + tensor transpose_128 = transpose(perm = transpose_128_perm_0, x = var_3920_cast_fp16)[name = string("transpose_217")]; + tensor matrix_ac_33_cast_fp16 = matmul(transpose_x = matrix_ac_33_transpose_x_0, transpose_y = matrix_ac_33_transpose_y_0, x = transpose_128, y = transpose_129)[name = string("matrix_ac_33_cast_fp16")]; + tensor matrix_bd_67_begin_0 = const()[name = string("matrix_bd_67_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_67_end_0 = const()[name = string("matrix_bd_67_end_0"), val = tensor([1, 8, 7, 49])]; + tensor matrix_bd_67_end_mask_0 = const()[name = string("matrix_bd_67_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_67_cast_fp16 = slice_by_index(begin = matrix_bd_67_begin_0, end = matrix_bd_67_end_0, end_mask = matrix_bd_67_end_mask_0, x = matrix_bd_65_cast_fp16)[name = string("matrix_bd_67_cast_fp16")]; + tensor var_3946_cast_fp16 = add(x = matrix_ac_33_cast_fp16, y = matrix_bd_67_cast_fp16)[name = string("op_3946_cast_fp16")]; + fp16 _inversed_scores_65_y_0_to_fp16 = const()[name = string("_inversed_scores_65_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_65_cast_fp16 = mul(x = var_3946_cast_fp16, y = _inversed_scores_65_y_0_to_fp16)[name = string("_inversed_scores_65_cast_fp16")]; + tensor scores_67_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_65_cast_fp16, cond = mask_11)[name = string("scores_67_cast_fp16")]; + tensor var_3952_cast_fp16 = softmax(axis = var_59, x = scores_67_cast_fp16)[name = string("op_3952_cast_fp16")]; + tensor input_873_cast_fp16 = select(a = var_44_to_fp16, b = var_3952_cast_fp16, cond = mask_11)[name = string("input_873_cast_fp16")]; + bool x_429_transpose_x_0 = const()[name = string("x_429_transpose_x_0"), val = bool(false)]; + bool x_429_transpose_y_0 = const()[name = string("x_429_transpose_y_0"), val = bool(false)]; + tensor value_41_cast_fp16 = transpose(perm = value_41_perm_0, x = v_33_cast_fp16)[name = string("transpose_215")]; + tensor x_429_cast_fp16 = matmul(transpose_x = x_429_transpose_x_0, transpose_y = x_429_transpose_y_0, x = input_873_cast_fp16, y = value_41_cast_fp16)[name = string("x_429_cast_fp16")]; + tensor var_3956_perm_0 = const()[name = string("op_3956_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_3957 = const()[name = string("op_3957"), val = tensor([1, -1, 1024])]; + tensor var_3956_cast_fp16 = transpose(perm = var_3956_perm_0, x = x_429_cast_fp16)[name = string("transpose_214")]; + tensor input_875_cast_fp16 = reshape(shape = var_3957, x = var_3956_cast_fp16)[name = string("input_875_cast_fp16")]; + tensor encoder_layers_16_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331670464))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332456960))))[name = string("encoder_layers_16_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_16_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_16_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332457152)))]; + tensor linear_151_cast_fp16 = linear(bias = encoder_layers_16_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_16_self_attn_linear_out_weight_to_fp16_palettized, x = input_875_cast_fp16)[name = string("linear_151_cast_fp16")]; + tensor input_879_cast_fp16 = add(x = input_869_cast_fp16, y = linear_151_cast_fp16)[name = string("input_879_cast_fp16")]; + tensor x_433_axes_0 = const()[name = string("x_433_axes_0"), val = tensor([-1])]; + tensor encoder_layers_16_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_16_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332459264)))]; + tensor encoder_layers_16_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_16_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332461376)))]; + tensor x_433_cast_fp16 = layer_norm(axes = x_433_axes_0, beta = encoder_layers_16_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_16_norm_conv_weight_to_fp16, x = input_879_cast_fp16)[name = string("x_433_cast_fp16")]; + tensor input_881_perm_0 = const()[name = string("input_881_perm_0"), val = tensor([0, 2, 1])]; + string input_883_pad_type_0 = const()[name = string("input_883_pad_type_0"), val = string("valid")]; + tensor input_883_strides_0 = const()[name = string("input_883_strides_0"), val = tensor([1])]; + tensor input_883_pad_0 = const()[name = string("input_883_pad_0"), val = tensor([0, 0])]; + tensor input_883_dilations_0 = const()[name = string("input_883_dilations_0"), val = tensor([1])]; + int32 input_883_groups_0 = const()[name = string("input_883_groups_0"), val = int32(1)]; + tensor encoder_layers_16_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332463488))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(334560704))))[name = string("encoder_layers_16_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_881_cast_fp16 = transpose(perm = input_881_perm_0, x = x_433_cast_fp16)[name = string("transpose_213")]; + tensor input_883_cast_fp16 = conv(dilations = input_883_dilations_0, groups = input_883_groups_0, pad = input_883_pad_0, pad_type = input_883_pad_type_0, strides = input_883_strides_0, weight = encoder_layers_16_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_881_cast_fp16)[name = string("input_883_cast_fp16")]; + int32 x_435_split_num_splits_0 = const()[name = string("x_435_split_num_splits_0"), val = int32(2)]; + int32 x_435_split_axis_0 = const()[name = string("x_435_split_axis_0"), val = int32(1)]; + tensor x_435_split_cast_fp16_0, tensor x_435_split_cast_fp16_1 = split(axis = x_435_split_axis_0, num_splits = x_435_split_num_splits_0, x = input_883_cast_fp16)[name = string("x_435_split_cast_fp16")]; + tensor x_435_split_1_sigmoid_cast_fp16 = sigmoid(x = x_435_split_cast_fp16_1)[name = string("x_435_split_1_sigmoid_cast_fp16")]; + tensor x_435_cast_fp16 = mul(x = x_435_split_cast_fp16_0, y = x_435_split_1_sigmoid_cast_fp16)[name = string("x_435_cast_fp16")]; + tensor input_885_cast_fp16 = select(a = var_44_to_fp16, b = x_435_cast_fp16, cond = var_575)[name = string("input_885_cast_fp16")]; + bool new_x_67_interleave_0 = const()[name = string("new_x_67_interleave_0"), val = bool(false)]; + tensor new_x_67_cast_fp16 = concat(axis = var_59, interleave = new_x_67_interleave_0, values = (cache_67_cast_fp16, input_885_cast_fp16))[name = string("new_x_67_cast_fp16")]; + tensor var_3996_begin_0 = const()[name = string("op_3996_begin_0"), val = tensor([0, 0, 7])]; + tensor var_3996_end_0 = const()[name = string("op_3996_end_0"), val = tensor([1, 1024, 15])]; + tensor var_3996_end_mask_0 = const()[name = string("op_3996_end_mask_0"), val = tensor([true, true, true])]; + tensor var_3996_cast_fp16 = slice_by_index(begin = var_3996_begin_0, end = var_3996_end_0, end_mask = var_3996_end_mask_0, x = new_x_67_cast_fp16)[name = string("op_3996_cast_fp16")]; + string x_437_pad_type_0 = const()[name = string("x_437_pad_type_0"), val = string("valid")]; + int32 x_437_groups_0 = const()[name = string("x_437_groups_0"), val = int32(1024)]; + tensor x_437_strides_0 = const()[name = string("x_437_strides_0"), val = tensor([1])]; + tensor x_437_pad_0 = const()[name = string("x_437_pad_0"), val = tensor([0, 0])]; + tensor x_437_dilations_0 = const()[name = string("x_437_dilations_0"), val = tensor([1])]; + tensor encoder_layers_16_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(334564864))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(334574144))))[name = string("encoder_layers_16_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_437_cast_fp16 = conv(dilations = x_437_dilations_0, groups = x_437_groups_0, pad = x_437_pad_0, pad_type = x_437_pad_type_0, strides = x_437_strides_0, weight = encoder_layers_16_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_67_cast_fp16)[name = string("x_437_cast_fp16")]; + tensor input_887_perm_0 = const()[name = string("input_887_perm_0"), val = tensor([0, 2, 1])]; + tensor x_439_axes_0 = const()[name = string("x_439_axes_0"), val = tensor([-1])]; + tensor encoder_layers_16_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_16_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(334576256)))]; + tensor encoder_layers_16_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_16_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(334578368)))]; + tensor input_887_cast_fp16 = transpose(perm = input_887_perm_0, x = x_437_cast_fp16)[name = string("transpose_212")]; + tensor x_439_cast_fp16 = layer_norm(axes = x_439_axes_0, beta = encoder_layers_16_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_16_conv_batch_norm_weight_to_fp16, x = input_887_cast_fp16)[name = string("x_439_cast_fp16")]; + tensor input_889_perm_0 = const()[name = string("input_889_perm_0"), val = tensor([0, 2, 1])]; + tensor input_889_cast_fp16 = transpose(perm = input_889_perm_0, x = x_439_cast_fp16)[name = string("transpose_211")]; + tensor input_891_cast_fp16 = silu(x = input_889_cast_fp16)[name = string("input_891_cast_fp16")]; + string x_441_pad_type_0 = const()[name = string("x_441_pad_type_0"), val = string("valid")]; + tensor x_441_strides_0 = const()[name = string("x_441_strides_0"), val = tensor([1])]; + tensor x_441_pad_0 = const()[name = string("x_441_pad_0"), val = tensor([0, 0])]; + tensor x_441_dilations_0 = const()[name = string("x_441_dilations_0"), val = tensor([1])]; + int32 x_441_groups_0 = const()[name = string("x_441_groups_0"), val = int32(1)]; + tensor encoder_layers_16_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(334580480))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(335629120))))[name = string("encoder_layers_16_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_441_cast_fp16 = conv(dilations = x_441_dilations_0, groups = x_441_groups_0, pad = x_441_pad_0, pad_type = x_441_pad_type_0, strides = x_441_strides_0, weight = encoder_layers_16_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_891_cast_fp16)[name = string("x_441_cast_fp16")]; + tensor input_893_perm_0 = const()[name = string("input_893_perm_0"), val = tensor([0, 2, 1])]; + tensor input_893_cast_fp16 = transpose(perm = input_893_perm_0, x = x_441_cast_fp16)[name = string("transpose_210")]; + tensor input_895_cast_fp16 = add(x = input_879_cast_fp16, y = input_893_cast_fp16)[name = string("input_895_cast_fp16")]; + tensor input_897_axes_0 = const()[name = string("input_897_axes_0"), val = tensor([-1])]; + tensor encoder_layers_16_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_16_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(335631232)))]; + tensor encoder_layers_16_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_16_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(335633344)))]; + tensor input_897_cast_fp16 = layer_norm(axes = input_897_axes_0, beta = encoder_layers_16_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_16_norm_feed_forward2_weight_to_fp16, x = input_895_cast_fp16)[name = string("input_897_cast_fp16")]; + tensor encoder_layers_16_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(335635456))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(338781248))))[name = string("encoder_layers_16_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_16_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_16_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(338781440)))]; + tensor linear_152_cast_fp16 = linear(bias = encoder_layers_16_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_16_feed_forward2_linear1_weight_to_fp16_palettized, x = input_897_cast_fp16)[name = string("linear_152_cast_fp16")]; + tensor input_901_cast_fp16 = silu(x = linear_152_cast_fp16)[name = string("input_901_cast_fp16")]; + tensor encoder_layers_16_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(338789696))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(341935488))))[name = string("encoder_layers_16_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_16_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_16_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(341935680)))]; + tensor linear_153_cast_fp16 = linear(bias = encoder_layers_16_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_16_feed_forward2_linear2_weight_to_fp16_palettized, x = input_901_cast_fp16)[name = string("linear_153_cast_fp16")]; + fp16 var_4039_to_fp16 = const()[name = string("op_4039_to_fp16"), val = fp16(0x1p-1)]; + tensor var_4040_cast_fp16 = mul(x = linear_153_cast_fp16, y = var_4039_to_fp16)[name = string("op_4040_cast_fp16")]; + tensor input_907_cast_fp16 = add(x = input_895_cast_fp16, y = var_4040_cast_fp16)[name = string("input_907_cast_fp16")]; + tensor input_909_axes_0 = const()[name = string("input_909_axes_0"), val = tensor([-1])]; + tensor encoder_layers_16_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_16_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(341937792)))]; + tensor encoder_layers_16_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_16_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(341939904)))]; + tensor input_909_cast_fp16 = layer_norm(axes = input_909_axes_0, beta = encoder_layers_16_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_16_norm_out_weight_to_fp16, x = input_907_cast_fp16)[name = string("input_909_cast_fp16")]; + tensor cache_69_begin_0 = const()[name = string("cache_69_begin_0"), val = tensor([17, 0, 0, 0])]; + tensor cache_69_end_0 = const()[name = string("cache_69_end_0"), val = tensor([18, 1, 42, 1024])]; + tensor cache_69_end_mask_0 = const()[name = string("cache_69_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_69_squeeze_mask_0 = const()[name = string("cache_69_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_69_cast_fp16 = slice_by_index(begin = cache_69_begin_0, end = cache_69_end_0, end_mask = cache_69_end_mask_0, squeeze_mask = cache_69_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_69_cast_fp16")]; + tensor cache_71_begin_0 = const()[name = string("cache_71_begin_0"), val = tensor([17, 0, 0, 0])]; + tensor cache_71_end_0 = const()[name = string("cache_71_end_0"), val = tensor([18, 1, 1024, 8])]; + tensor cache_71_end_mask_0 = const()[name = string("cache_71_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_71_squeeze_mask_0 = const()[name = string("cache_71_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_71_cast_fp16 = slice_by_index(begin = cache_71_begin_0, end = cache_71_end_0, end_mask = cache_71_end_mask_0, squeeze_mask = cache_71_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_71_cast_fp16")]; + tensor input_911_axes_0 = const()[name = string("input_911_axes_0"), val = tensor([-1])]; + tensor encoder_layers_17_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_17_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(341942016)))]; + tensor encoder_layers_17_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_17_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(341944128)))]; + tensor input_911_cast_fp16 = layer_norm(axes = input_911_axes_0, beta = encoder_layers_17_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_17_norm_feed_forward1_weight_to_fp16, x = input_909_cast_fp16)[name = string("input_911_cast_fp16")]; + tensor encoder_layers_17_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(341946240))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(345092032))))[name = string("encoder_layers_17_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_17_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_17_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(345092224)))]; + tensor linear_154_cast_fp16 = linear(bias = encoder_layers_17_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_17_feed_forward1_linear1_weight_to_fp16_palettized, x = input_911_cast_fp16)[name = string("linear_154_cast_fp16")]; + tensor input_915_cast_fp16 = silu(x = linear_154_cast_fp16)[name = string("input_915_cast_fp16")]; + tensor encoder_layers_17_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(345100480))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(348246272))))[name = string("encoder_layers_17_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_17_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_17_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(348246464)))]; + tensor linear_155_cast_fp16 = linear(bias = encoder_layers_17_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_17_feed_forward1_linear2_weight_to_fp16_palettized, x = input_915_cast_fp16)[name = string("linear_155_cast_fp16")]; + fp16 var_4076_to_fp16 = const()[name = string("op_4076_to_fp16"), val = fp16(0x1p-1)]; + tensor var_4077_cast_fp16 = mul(x = linear_155_cast_fp16, y = var_4076_to_fp16)[name = string("op_4077_cast_fp16")]; + tensor input_921_cast_fp16 = add(x = input_909_cast_fp16, y = var_4077_cast_fp16)[name = string("input_921_cast_fp16")]; + tensor key_35_axes_0 = const()[name = string("key_35_axes_0"), val = tensor([-1])]; + tensor encoder_layers_17_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_17_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(348248576)))]; + tensor encoder_layers_17_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_17_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(348250688)))]; + tensor key_35_cast_fp16 = layer_norm(axes = key_35_axes_0, beta = encoder_layers_17_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_17_norm_self_att_weight_to_fp16, x = input_921_cast_fp16)[name = string("key_35_cast_fp16")]; + bool input_923_interleave_0 = const()[name = string("input_923_interleave_0"), val = bool(false)]; + tensor input_923_cast_fp16 = concat(axis = var_68, interleave = input_923_interleave_0, values = (cache_69_cast_fp16, key_35_cast_fp16))[name = string("input_923_cast_fp16")]; + tensor var_4099_begin_0 = const()[name = string("op_4099_begin_0"), val = tensor([0, 7, 0])]; + tensor var_4099_end_0 = const()[name = string("op_4099_end_0"), val = tensor([1, 42, 1024])]; + tensor var_4099_end_mask_0 = const()[name = string("op_4099_end_mask_0"), val = tensor([true, true, true])]; + tensor var_4099_cast_fp16 = slice_by_index(begin = var_4099_begin_0, end = var_4099_end_0, end_mask = var_4099_end_mask_0, x = cache_69_cast_fp16)[name = string("op_4099_cast_fp16")]; + bool var_4105_interleave_0 = const()[name = string("op_4105_interleave_0"), val = bool(false)]; + tensor var_4105_cast_fp16 = concat(axis = var_68, interleave = var_4105_interleave_0, values = (var_4099_cast_fp16, key_35_cast_fp16))[name = string("op_4105_cast_fp16")]; + tensor encoder_layers_17_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(348252800))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(349039296))))[name = string("encoder_layers_17_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_17_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_17_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(349039488)))]; + tensor linear_156_cast_fp16 = linear(bias = encoder_layers_17_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_17_self_attn_linear_q_weight_to_fp16_palettized, x = key_35_cast_fp16)[name = string("linear_156_cast_fp16")]; + tensor var_4110 = const()[name = string("op_4110"), val = tensor([1, -1, 8, 128])]; + tensor q_103_cast_fp16 = reshape(shape = var_4110, x = linear_156_cast_fp16)[name = string("q_103_cast_fp16")]; + tensor encoder_layers_17_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(349041600))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(349828096))))[name = string("encoder_layers_17_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_17_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_17_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(349828288)))]; + tensor linear_157_cast_fp16 = linear(bias = encoder_layers_17_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_17_self_attn_linear_k_weight_to_fp16_palettized, x = input_923_cast_fp16)[name = string("linear_157_cast_fp16")]; + tensor var_4115 = const()[name = string("op_4115"), val = tensor([1, -1, 8, 128])]; + tensor k_69_cast_fp16 = reshape(shape = var_4115, x = linear_157_cast_fp16)[name = string("k_69_cast_fp16")]; + tensor encoder_layers_17_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(349830400))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350616896))))[name = string("encoder_layers_17_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_17_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_17_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350617088)))]; + tensor linear_158_cast_fp16 = linear(bias = encoder_layers_17_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_17_self_attn_linear_v_weight_to_fp16_palettized, x = input_923_cast_fp16)[name = string("linear_158_cast_fp16")]; + tensor var_4120 = const()[name = string("op_4120"), val = tensor([1, -1, 8, 128])]; + tensor v_35_cast_fp16 = reshape(shape = var_4120, x = linear_158_cast_fp16)[name = string("v_35_cast_fp16")]; + tensor value_43_perm_0 = const()[name = string("value_43_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_17_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_17_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350619200)))]; + tensor var_4133_cast_fp16 = add(x = q_103_cast_fp16, y = encoder_layers_17_self_attn_pos_bias_u_to_fp16)[name = string("op_4133_cast_fp16")]; + tensor encoder_layers_17_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_17_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350621312)))]; + tensor var_4135_cast_fp16 = add(x = q_103_cast_fp16, y = encoder_layers_17_self_attn_pos_bias_v_to_fp16)[name = string("op_4135_cast_fp16")]; + tensor q_with_bias_v_35_perm_0 = const()[name = string("q_with_bias_v_35_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_449_transpose_x_0 = const()[name = string("x_449_transpose_x_0"), val = bool(false)]; + bool x_449_transpose_y_0 = const()[name = string("x_449_transpose_y_0"), val = bool(false)]; + tensor op_4137_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350623424))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350722816))))[name = string("op_4137_to_fp16_quantized")]; + tensor q_with_bias_v_35_cast_fp16 = transpose(perm = q_with_bias_v_35_perm_0, x = var_4135_cast_fp16)[name = string("transpose_209")]; + tensor x_449_cast_fp16 = matmul(transpose_x = x_449_transpose_x_0, transpose_y = x_449_transpose_y_0, x = q_with_bias_v_35_cast_fp16, y = op_4137_to_fp16_quantized)[name = string("x_449_cast_fp16")]; + tensor x_451_pad_0 = const()[name = string("x_451_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_451_mode_0 = const()[name = string("x_451_mode_0"), val = string("constant")]; + fp16 const_300_to_fp16 = const()[name = string("const_300_to_fp16"), val = fp16(0x0p+0)]; + tensor x_451_cast_fp16 = pad(constant_val = const_300_to_fp16, mode = x_451_mode_0, pad = x_451_pad_0, x = x_449_cast_fp16)[name = string("x_451_cast_fp16")]; + tensor var_4145 = const()[name = string("op_4145"), val = tensor([1, 8, -1, 7])]; + tensor x_453_cast_fp16 = reshape(shape = var_4145, x = x_451_cast_fp16)[name = string("x_453_cast_fp16")]; + tensor var_4149_begin_0 = const()[name = string("op_4149_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_4149_end_0 = const()[name = string("op_4149_end_0"), val = tensor([1, 8, 98, 7])]; + tensor var_4149_end_mask_0 = const()[name = string("op_4149_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_4149_cast_fp16 = slice_by_index(begin = var_4149_begin_0, end = var_4149_end_0, end_mask = var_4149_end_mask_0, x = x_453_cast_fp16)[name = string("op_4149_cast_fp16")]; + tensor var_4150 = const()[name = string("op_4150"), val = tensor([1, 8, 7, 97])]; + tensor matrix_bd_69_cast_fp16 = reshape(shape = var_4150, x = var_4149_cast_fp16)[name = string("matrix_bd_69_cast_fp16")]; + bool matrix_ac_35_transpose_x_0 = const()[name = string("matrix_ac_35_transpose_x_0"), val = bool(false)]; + bool matrix_ac_35_transpose_y_0 = const()[name = string("matrix_ac_35_transpose_y_0"), val = bool(false)]; + tensor transpose_130_perm_0 = const()[name = string("transpose_130_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_131_perm_0 = const()[name = string("transpose_131_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_131 = transpose(perm = transpose_131_perm_0, x = k_69_cast_fp16)[name = string("transpose_207")]; + tensor transpose_130 = transpose(perm = transpose_130_perm_0, x = var_4133_cast_fp16)[name = string("transpose_208")]; + tensor matrix_ac_35_cast_fp16 = matmul(transpose_x = matrix_ac_35_transpose_x_0, transpose_y = matrix_ac_35_transpose_y_0, x = transpose_130, y = transpose_131)[name = string("matrix_ac_35_cast_fp16")]; + tensor matrix_bd_71_begin_0 = const()[name = string("matrix_bd_71_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_71_end_0 = const()[name = string("matrix_bd_71_end_0"), val = tensor([1, 8, 7, 49])]; + tensor matrix_bd_71_end_mask_0 = const()[name = string("matrix_bd_71_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_71_cast_fp16 = slice_by_index(begin = matrix_bd_71_begin_0, end = matrix_bd_71_end_0, end_mask = matrix_bd_71_end_mask_0, x = matrix_bd_69_cast_fp16)[name = string("matrix_bd_71_cast_fp16")]; + tensor var_4159_cast_fp16 = add(x = matrix_ac_35_cast_fp16, y = matrix_bd_71_cast_fp16)[name = string("op_4159_cast_fp16")]; + fp16 _inversed_scores_69_y_0_to_fp16 = const()[name = string("_inversed_scores_69_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_69_cast_fp16 = mul(x = var_4159_cast_fp16, y = _inversed_scores_69_y_0_to_fp16)[name = string("_inversed_scores_69_cast_fp16")]; + tensor scores_71_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_69_cast_fp16, cond = mask_11)[name = string("scores_71_cast_fp16")]; + tensor var_4165_cast_fp16 = softmax(axis = var_59, x = scores_71_cast_fp16)[name = string("op_4165_cast_fp16")]; + tensor input_925_cast_fp16 = select(a = var_44_to_fp16, b = var_4165_cast_fp16, cond = mask_11)[name = string("input_925_cast_fp16")]; + bool x_455_transpose_x_0 = const()[name = string("x_455_transpose_x_0"), val = bool(false)]; + bool x_455_transpose_y_0 = const()[name = string("x_455_transpose_y_0"), val = bool(false)]; + tensor value_43_cast_fp16 = transpose(perm = value_43_perm_0, x = v_35_cast_fp16)[name = string("transpose_206")]; + tensor x_455_cast_fp16 = matmul(transpose_x = x_455_transpose_x_0, transpose_y = x_455_transpose_y_0, x = input_925_cast_fp16, y = value_43_cast_fp16)[name = string("x_455_cast_fp16")]; + tensor var_4169_perm_0 = const()[name = string("op_4169_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_4170 = const()[name = string("op_4170"), val = tensor([1, -1, 1024])]; + tensor var_4169_cast_fp16 = transpose(perm = var_4169_perm_0, x = x_455_cast_fp16)[name = string("transpose_205")]; + tensor input_927_cast_fp16 = reshape(shape = var_4170, x = var_4169_cast_fp16)[name = string("input_927_cast_fp16")]; + tensor encoder_layers_17_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350723136))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351509632))))[name = string("encoder_layers_17_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_17_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_17_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351509824)))]; + tensor linear_160_cast_fp16 = linear(bias = encoder_layers_17_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_17_self_attn_linear_out_weight_to_fp16_palettized, x = input_927_cast_fp16)[name = string("linear_160_cast_fp16")]; + tensor input_931_cast_fp16 = add(x = input_921_cast_fp16, y = linear_160_cast_fp16)[name = string("input_931_cast_fp16")]; + tensor x_459_axes_0 = const()[name = string("x_459_axes_0"), val = tensor([-1])]; + tensor encoder_layers_17_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_17_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351511936)))]; + tensor encoder_layers_17_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_17_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351514048)))]; + tensor x_459_cast_fp16 = layer_norm(axes = x_459_axes_0, beta = encoder_layers_17_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_17_norm_conv_weight_to_fp16, x = input_931_cast_fp16)[name = string("x_459_cast_fp16")]; + tensor input_933_perm_0 = const()[name = string("input_933_perm_0"), val = tensor([0, 2, 1])]; + string input_935_pad_type_0 = const()[name = string("input_935_pad_type_0"), val = string("valid")]; + tensor input_935_strides_0 = const()[name = string("input_935_strides_0"), val = tensor([1])]; + tensor input_935_pad_0 = const()[name = string("input_935_pad_0"), val = tensor([0, 0])]; + tensor input_935_dilations_0 = const()[name = string("input_935_dilations_0"), val = tensor([1])]; + int32 input_935_groups_0 = const()[name = string("input_935_groups_0"), val = int32(1)]; + tensor encoder_layers_17_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351516160))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(353613376))))[name = string("encoder_layers_17_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_933_cast_fp16 = transpose(perm = input_933_perm_0, x = x_459_cast_fp16)[name = string("transpose_204")]; + tensor input_935_cast_fp16 = conv(dilations = input_935_dilations_0, groups = input_935_groups_0, pad = input_935_pad_0, pad_type = input_935_pad_type_0, strides = input_935_strides_0, weight = encoder_layers_17_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_933_cast_fp16)[name = string("input_935_cast_fp16")]; + int32 x_461_split_num_splits_0 = const()[name = string("x_461_split_num_splits_0"), val = int32(2)]; + int32 x_461_split_axis_0 = const()[name = string("x_461_split_axis_0"), val = int32(1)]; + tensor x_461_split_cast_fp16_0, tensor x_461_split_cast_fp16_1 = split(axis = x_461_split_axis_0, num_splits = x_461_split_num_splits_0, x = input_935_cast_fp16)[name = string("x_461_split_cast_fp16")]; + tensor x_461_split_1_sigmoid_cast_fp16 = sigmoid(x = x_461_split_cast_fp16_1)[name = string("x_461_split_1_sigmoid_cast_fp16")]; + tensor x_461_cast_fp16 = mul(x = x_461_split_cast_fp16_0, y = x_461_split_1_sigmoid_cast_fp16)[name = string("x_461_cast_fp16")]; + tensor input_937_cast_fp16 = select(a = var_44_to_fp16, b = x_461_cast_fp16, cond = var_575)[name = string("input_937_cast_fp16")]; + bool new_x_71_interleave_0 = const()[name = string("new_x_71_interleave_0"), val = bool(false)]; + tensor new_x_71_cast_fp16 = concat(axis = var_59, interleave = new_x_71_interleave_0, values = (cache_71_cast_fp16, input_937_cast_fp16))[name = string("new_x_71_cast_fp16")]; + tensor var_4209_begin_0 = const()[name = string("op_4209_begin_0"), val = tensor([0, 0, 7])]; + tensor var_4209_end_0 = const()[name = string("op_4209_end_0"), val = tensor([1, 1024, 15])]; + tensor var_4209_end_mask_0 = const()[name = string("op_4209_end_mask_0"), val = tensor([true, true, true])]; + tensor var_4209_cast_fp16 = slice_by_index(begin = var_4209_begin_0, end = var_4209_end_0, end_mask = var_4209_end_mask_0, x = new_x_71_cast_fp16)[name = string("op_4209_cast_fp16")]; + string x_463_pad_type_0 = const()[name = string("x_463_pad_type_0"), val = string("valid")]; + int32 x_463_groups_0 = const()[name = string("x_463_groups_0"), val = int32(1024)]; + tensor x_463_strides_0 = const()[name = string("x_463_strides_0"), val = tensor([1])]; + tensor x_463_pad_0 = const()[name = string("x_463_pad_0"), val = tensor([0, 0])]; + tensor x_463_dilations_0 = const()[name = string("x_463_dilations_0"), val = tensor([1])]; + tensor encoder_layers_17_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(353617536))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(353626816))))[name = string("encoder_layers_17_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_463_cast_fp16 = conv(dilations = x_463_dilations_0, groups = x_463_groups_0, pad = x_463_pad_0, pad_type = x_463_pad_type_0, strides = x_463_strides_0, weight = encoder_layers_17_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_71_cast_fp16)[name = string("x_463_cast_fp16")]; + tensor input_939_perm_0 = const()[name = string("input_939_perm_0"), val = tensor([0, 2, 1])]; + tensor x_465_axes_0 = const()[name = string("x_465_axes_0"), val = tensor([-1])]; + tensor encoder_layers_17_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_17_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(353628928)))]; + tensor encoder_layers_17_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_17_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(353631040)))]; + tensor input_939_cast_fp16 = transpose(perm = input_939_perm_0, x = x_463_cast_fp16)[name = string("transpose_203")]; + tensor x_465_cast_fp16 = layer_norm(axes = x_465_axes_0, beta = encoder_layers_17_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_17_conv_batch_norm_weight_to_fp16, x = input_939_cast_fp16)[name = string("x_465_cast_fp16")]; + tensor input_941_perm_0 = const()[name = string("input_941_perm_0"), val = tensor([0, 2, 1])]; + tensor input_941_cast_fp16 = transpose(perm = input_941_perm_0, x = x_465_cast_fp16)[name = string("transpose_202")]; + tensor input_943_cast_fp16 = silu(x = input_941_cast_fp16)[name = string("input_943_cast_fp16")]; + string x_467_pad_type_0 = const()[name = string("x_467_pad_type_0"), val = string("valid")]; + tensor x_467_strides_0 = const()[name = string("x_467_strides_0"), val = tensor([1])]; + tensor x_467_pad_0 = const()[name = string("x_467_pad_0"), val = tensor([0, 0])]; + tensor x_467_dilations_0 = const()[name = string("x_467_dilations_0"), val = tensor([1])]; + int32 x_467_groups_0 = const()[name = string("x_467_groups_0"), val = int32(1)]; + tensor encoder_layers_17_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(353633152))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(354681792))))[name = string("encoder_layers_17_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_467_cast_fp16 = conv(dilations = x_467_dilations_0, groups = x_467_groups_0, pad = x_467_pad_0, pad_type = x_467_pad_type_0, strides = x_467_strides_0, weight = encoder_layers_17_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_943_cast_fp16)[name = string("x_467_cast_fp16")]; + tensor input_945_perm_0 = const()[name = string("input_945_perm_0"), val = tensor([0, 2, 1])]; + tensor input_945_cast_fp16 = transpose(perm = input_945_perm_0, x = x_467_cast_fp16)[name = string("transpose_201")]; + tensor input_947_cast_fp16 = add(x = input_931_cast_fp16, y = input_945_cast_fp16)[name = string("input_947_cast_fp16")]; + tensor input_949_axes_0 = const()[name = string("input_949_axes_0"), val = tensor([-1])]; + tensor encoder_layers_17_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_17_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(354683904)))]; + tensor encoder_layers_17_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_17_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(354686016)))]; + tensor input_949_cast_fp16 = layer_norm(axes = input_949_axes_0, beta = encoder_layers_17_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_17_norm_feed_forward2_weight_to_fp16, x = input_947_cast_fp16)[name = string("input_949_cast_fp16")]; + tensor encoder_layers_17_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(354688128))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(357833920))))[name = string("encoder_layers_17_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_17_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_17_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(357834112)))]; + tensor linear_161_cast_fp16 = linear(bias = encoder_layers_17_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_17_feed_forward2_linear1_weight_to_fp16_palettized, x = input_949_cast_fp16)[name = string("linear_161_cast_fp16")]; + tensor input_953_cast_fp16 = silu(x = linear_161_cast_fp16)[name = string("input_953_cast_fp16")]; + tensor encoder_layers_17_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(357842368))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(360988160))))[name = string("encoder_layers_17_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_17_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_17_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(360988352)))]; + tensor linear_162_cast_fp16 = linear(bias = encoder_layers_17_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_17_feed_forward2_linear2_weight_to_fp16_palettized, x = input_953_cast_fp16)[name = string("linear_162_cast_fp16")]; + fp16 var_4252_to_fp16 = const()[name = string("op_4252_to_fp16"), val = fp16(0x1p-1)]; + tensor var_4253_cast_fp16 = mul(x = linear_162_cast_fp16, y = var_4252_to_fp16)[name = string("op_4253_cast_fp16")]; + tensor input_959_cast_fp16 = add(x = input_947_cast_fp16, y = var_4253_cast_fp16)[name = string("input_959_cast_fp16")]; + tensor input_961_axes_0 = const()[name = string("input_961_axes_0"), val = tensor([-1])]; + tensor encoder_layers_17_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_17_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(360990464)))]; + tensor encoder_layers_17_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_17_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(360992576)))]; + tensor input_961_cast_fp16 = layer_norm(axes = input_961_axes_0, beta = encoder_layers_17_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_17_norm_out_weight_to_fp16, x = input_959_cast_fp16)[name = string("input_961_cast_fp16")]; + tensor cache_73_begin_0 = const()[name = string("cache_73_begin_0"), val = tensor([18, 0, 0, 0])]; + tensor cache_73_end_0 = const()[name = string("cache_73_end_0"), val = tensor([19, 1, 42, 1024])]; + tensor cache_73_end_mask_0 = const()[name = string("cache_73_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_73_squeeze_mask_0 = const()[name = string("cache_73_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_73_cast_fp16 = slice_by_index(begin = cache_73_begin_0, end = cache_73_end_0, end_mask = cache_73_end_mask_0, squeeze_mask = cache_73_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_73_cast_fp16")]; + tensor cache_75_begin_0 = const()[name = string("cache_75_begin_0"), val = tensor([18, 0, 0, 0])]; + tensor cache_75_end_0 = const()[name = string("cache_75_end_0"), val = tensor([19, 1, 1024, 8])]; + tensor cache_75_end_mask_0 = const()[name = string("cache_75_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_75_squeeze_mask_0 = const()[name = string("cache_75_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_75_cast_fp16 = slice_by_index(begin = cache_75_begin_0, end = cache_75_end_0, end_mask = cache_75_end_mask_0, squeeze_mask = cache_75_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_75_cast_fp16")]; + tensor input_963_axes_0 = const()[name = string("input_963_axes_0"), val = tensor([-1])]; + tensor encoder_layers_18_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_18_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(360994688)))]; + tensor encoder_layers_18_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_18_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(360996800)))]; + tensor input_963_cast_fp16 = layer_norm(axes = input_963_axes_0, beta = encoder_layers_18_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_18_norm_feed_forward1_weight_to_fp16, x = input_961_cast_fp16)[name = string("input_963_cast_fp16")]; + tensor encoder_layers_18_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(360998912))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(364144704))))[name = string("encoder_layers_18_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_18_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_18_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(364144896)))]; + tensor linear_163_cast_fp16 = linear(bias = encoder_layers_18_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_18_feed_forward1_linear1_weight_to_fp16_palettized, x = input_963_cast_fp16)[name = string("linear_163_cast_fp16")]; + tensor input_967_cast_fp16 = silu(x = linear_163_cast_fp16)[name = string("input_967_cast_fp16")]; + tensor encoder_layers_18_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(364153152))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(367298944))))[name = string("encoder_layers_18_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_18_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_18_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(367299136)))]; + tensor linear_164_cast_fp16 = linear(bias = encoder_layers_18_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_18_feed_forward1_linear2_weight_to_fp16_palettized, x = input_967_cast_fp16)[name = string("linear_164_cast_fp16")]; + fp16 var_4289_to_fp16 = const()[name = string("op_4289_to_fp16"), val = fp16(0x1p-1)]; + tensor var_4290_cast_fp16 = mul(x = linear_164_cast_fp16, y = var_4289_to_fp16)[name = string("op_4290_cast_fp16")]; + tensor input_973_cast_fp16 = add(x = input_961_cast_fp16, y = var_4290_cast_fp16)[name = string("input_973_cast_fp16")]; + tensor key_37_axes_0 = const()[name = string("key_37_axes_0"), val = tensor([-1])]; + tensor encoder_layers_18_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_18_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(367301248)))]; + tensor encoder_layers_18_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_18_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(367303360)))]; + tensor key_37_cast_fp16 = layer_norm(axes = key_37_axes_0, beta = encoder_layers_18_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_18_norm_self_att_weight_to_fp16, x = input_973_cast_fp16)[name = string("key_37_cast_fp16")]; + bool input_975_interleave_0 = const()[name = string("input_975_interleave_0"), val = bool(false)]; + tensor input_975_cast_fp16 = concat(axis = var_68, interleave = input_975_interleave_0, values = (cache_73_cast_fp16, key_37_cast_fp16))[name = string("input_975_cast_fp16")]; + tensor var_4312_begin_0 = const()[name = string("op_4312_begin_0"), val = tensor([0, 7, 0])]; + tensor var_4312_end_0 = const()[name = string("op_4312_end_0"), val = tensor([1, 42, 1024])]; + tensor var_4312_end_mask_0 = const()[name = string("op_4312_end_mask_0"), val = tensor([true, true, true])]; + tensor var_4312_cast_fp16 = slice_by_index(begin = var_4312_begin_0, end = var_4312_end_0, end_mask = var_4312_end_mask_0, x = cache_73_cast_fp16)[name = string("op_4312_cast_fp16")]; + bool var_4318_interleave_0 = const()[name = string("op_4318_interleave_0"), val = bool(false)]; + tensor var_4318_cast_fp16 = concat(axis = var_68, interleave = var_4318_interleave_0, values = (var_4312_cast_fp16, key_37_cast_fp16))[name = string("op_4318_cast_fp16")]; + tensor encoder_layers_18_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(367305472))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368091968))))[name = string("encoder_layers_18_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_18_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_18_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368092160)))]; + tensor linear_165_cast_fp16 = linear(bias = encoder_layers_18_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_18_self_attn_linear_q_weight_to_fp16_palettized, x = key_37_cast_fp16)[name = string("linear_165_cast_fp16")]; + tensor var_4323 = const()[name = string("op_4323"), val = tensor([1, -1, 8, 128])]; + tensor q_109_cast_fp16 = reshape(shape = var_4323, x = linear_165_cast_fp16)[name = string("q_109_cast_fp16")]; + tensor encoder_layers_18_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368094272))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368880768))))[name = string("encoder_layers_18_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_18_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_18_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368880960)))]; + tensor linear_166_cast_fp16 = linear(bias = encoder_layers_18_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_18_self_attn_linear_k_weight_to_fp16_palettized, x = input_975_cast_fp16)[name = string("linear_166_cast_fp16")]; + tensor var_4328 = const()[name = string("op_4328"), val = tensor([1, -1, 8, 128])]; + tensor k_73_cast_fp16 = reshape(shape = var_4328, x = linear_166_cast_fp16)[name = string("k_73_cast_fp16")]; + tensor encoder_layers_18_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368883072))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369669568))))[name = string("encoder_layers_18_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_18_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_18_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369669760)))]; + tensor linear_167_cast_fp16 = linear(bias = encoder_layers_18_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_18_self_attn_linear_v_weight_to_fp16_palettized, x = input_975_cast_fp16)[name = string("linear_167_cast_fp16")]; + tensor var_4333 = const()[name = string("op_4333"), val = tensor([1, -1, 8, 128])]; + tensor v_37_cast_fp16 = reshape(shape = var_4333, x = linear_167_cast_fp16)[name = string("v_37_cast_fp16")]; + tensor value_45_perm_0 = const()[name = string("value_45_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_18_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_18_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369671872)))]; + tensor var_4346_cast_fp16 = add(x = q_109_cast_fp16, y = encoder_layers_18_self_attn_pos_bias_u_to_fp16)[name = string("op_4346_cast_fp16")]; + tensor encoder_layers_18_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_18_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369673984)))]; + tensor var_4348_cast_fp16 = add(x = q_109_cast_fp16, y = encoder_layers_18_self_attn_pos_bias_v_to_fp16)[name = string("op_4348_cast_fp16")]; + tensor q_with_bias_v_37_perm_0 = const()[name = string("q_with_bias_v_37_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_475_transpose_x_0 = const()[name = string("x_475_transpose_x_0"), val = bool(false)]; + bool x_475_transpose_y_0 = const()[name = string("x_475_transpose_y_0"), val = bool(false)]; + tensor op_4350_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369676096))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369775488))))[name = string("op_4350_to_fp16_quantized")]; + tensor q_with_bias_v_37_cast_fp16 = transpose(perm = q_with_bias_v_37_perm_0, x = var_4348_cast_fp16)[name = string("transpose_200")]; + tensor x_475_cast_fp16 = matmul(transpose_x = x_475_transpose_x_0, transpose_y = x_475_transpose_y_0, x = q_with_bias_v_37_cast_fp16, y = op_4350_to_fp16_quantized)[name = string("x_475_cast_fp16")]; + tensor x_477_pad_0 = const()[name = string("x_477_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_477_mode_0 = const()[name = string("x_477_mode_0"), val = string("constant")]; + fp16 const_313_to_fp16 = const()[name = string("const_313_to_fp16"), val = fp16(0x0p+0)]; + tensor x_477_cast_fp16 = pad(constant_val = const_313_to_fp16, mode = x_477_mode_0, pad = x_477_pad_0, x = x_475_cast_fp16)[name = string("x_477_cast_fp16")]; + tensor var_4358 = const()[name = string("op_4358"), val = tensor([1, 8, -1, 7])]; + tensor x_479_cast_fp16 = reshape(shape = var_4358, x = x_477_cast_fp16)[name = string("x_479_cast_fp16")]; + tensor var_4362_begin_0 = const()[name = string("op_4362_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_4362_end_0 = const()[name = string("op_4362_end_0"), val = tensor([1, 8, 98, 7])]; + tensor var_4362_end_mask_0 = const()[name = string("op_4362_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_4362_cast_fp16 = slice_by_index(begin = var_4362_begin_0, end = var_4362_end_0, end_mask = var_4362_end_mask_0, x = x_479_cast_fp16)[name = string("op_4362_cast_fp16")]; + tensor var_4363 = const()[name = string("op_4363"), val = tensor([1, 8, 7, 97])]; + tensor matrix_bd_73_cast_fp16 = reshape(shape = var_4363, x = var_4362_cast_fp16)[name = string("matrix_bd_73_cast_fp16")]; + bool matrix_ac_37_transpose_x_0 = const()[name = string("matrix_ac_37_transpose_x_0"), val = bool(false)]; + bool matrix_ac_37_transpose_y_0 = const()[name = string("matrix_ac_37_transpose_y_0"), val = bool(false)]; + tensor transpose_132_perm_0 = const()[name = string("transpose_132_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_133_perm_0 = const()[name = string("transpose_133_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_133 = transpose(perm = transpose_133_perm_0, x = k_73_cast_fp16)[name = string("transpose_198")]; + tensor transpose_132 = transpose(perm = transpose_132_perm_0, x = var_4346_cast_fp16)[name = string("transpose_199")]; + tensor matrix_ac_37_cast_fp16 = matmul(transpose_x = matrix_ac_37_transpose_x_0, transpose_y = matrix_ac_37_transpose_y_0, x = transpose_132, y = transpose_133)[name = string("matrix_ac_37_cast_fp16")]; + tensor matrix_bd_75_begin_0 = const()[name = string("matrix_bd_75_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_75_end_0 = const()[name = string("matrix_bd_75_end_0"), val = tensor([1, 8, 7, 49])]; + tensor matrix_bd_75_end_mask_0 = const()[name = string("matrix_bd_75_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_75_cast_fp16 = slice_by_index(begin = matrix_bd_75_begin_0, end = matrix_bd_75_end_0, end_mask = matrix_bd_75_end_mask_0, x = matrix_bd_73_cast_fp16)[name = string("matrix_bd_75_cast_fp16")]; + tensor var_4372_cast_fp16 = add(x = matrix_ac_37_cast_fp16, y = matrix_bd_75_cast_fp16)[name = string("op_4372_cast_fp16")]; + fp16 _inversed_scores_73_y_0_to_fp16 = const()[name = string("_inversed_scores_73_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_73_cast_fp16 = mul(x = var_4372_cast_fp16, y = _inversed_scores_73_y_0_to_fp16)[name = string("_inversed_scores_73_cast_fp16")]; + tensor scores_75_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_73_cast_fp16, cond = mask_11)[name = string("scores_75_cast_fp16")]; + tensor var_4378_cast_fp16 = softmax(axis = var_59, x = scores_75_cast_fp16)[name = string("op_4378_cast_fp16")]; + tensor input_977_cast_fp16 = select(a = var_44_to_fp16, b = var_4378_cast_fp16, cond = mask_11)[name = string("input_977_cast_fp16")]; + bool x_481_transpose_x_0 = const()[name = string("x_481_transpose_x_0"), val = bool(false)]; + bool x_481_transpose_y_0 = const()[name = string("x_481_transpose_y_0"), val = bool(false)]; + tensor value_45_cast_fp16 = transpose(perm = value_45_perm_0, x = v_37_cast_fp16)[name = string("transpose_197")]; + tensor x_481_cast_fp16 = matmul(transpose_x = x_481_transpose_x_0, transpose_y = x_481_transpose_y_0, x = input_977_cast_fp16, y = value_45_cast_fp16)[name = string("x_481_cast_fp16")]; + tensor var_4382_perm_0 = const()[name = string("op_4382_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_4383 = const()[name = string("op_4383"), val = tensor([1, -1, 1024])]; + tensor var_4382_cast_fp16 = transpose(perm = var_4382_perm_0, x = x_481_cast_fp16)[name = string("transpose_196")]; + tensor input_979_cast_fp16 = reshape(shape = var_4383, x = var_4382_cast_fp16)[name = string("input_979_cast_fp16")]; + tensor encoder_layers_18_self_attn_linear_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369775808))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(370824448))))[name = string("encoder_layers_18_self_attn_linear_out_weight_to_fp16_quantized")]; + tensor encoder_layers_18_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_18_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(370826560)))]; + tensor linear_169_cast_fp16 = linear(bias = encoder_layers_18_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_18_self_attn_linear_out_weight_to_fp16_quantized, x = input_979_cast_fp16)[name = string("linear_169_cast_fp16")]; + tensor input_983_cast_fp16 = add(x = input_973_cast_fp16, y = linear_169_cast_fp16)[name = string("input_983_cast_fp16")]; + tensor x_485_axes_0 = const()[name = string("x_485_axes_0"), val = tensor([-1])]; + tensor encoder_layers_18_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_18_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(370828672)))]; + tensor encoder_layers_18_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_18_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(370830784)))]; + tensor x_485_cast_fp16 = layer_norm(axes = x_485_axes_0, beta = encoder_layers_18_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_18_norm_conv_weight_to_fp16, x = input_983_cast_fp16)[name = string("x_485_cast_fp16")]; + tensor input_985_perm_0 = const()[name = string("input_985_perm_0"), val = tensor([0, 2, 1])]; + string input_987_pad_type_0 = const()[name = string("input_987_pad_type_0"), val = string("valid")]; + tensor input_987_strides_0 = const()[name = string("input_987_strides_0"), val = tensor([1])]; + tensor input_987_pad_0 = const()[name = string("input_987_pad_0"), val = tensor([0, 0])]; + tensor input_987_dilations_0 = const()[name = string("input_987_dilations_0"), val = tensor([1])]; + int32 input_987_groups_0 = const()[name = string("input_987_groups_0"), val = int32(1)]; + tensor encoder_layers_18_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(370832896))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(372930112))))[name = string("encoder_layers_18_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_985_cast_fp16 = transpose(perm = input_985_perm_0, x = x_485_cast_fp16)[name = string("transpose_195")]; + tensor input_987_cast_fp16 = conv(dilations = input_987_dilations_0, groups = input_987_groups_0, pad = input_987_pad_0, pad_type = input_987_pad_type_0, strides = input_987_strides_0, weight = encoder_layers_18_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_985_cast_fp16)[name = string("input_987_cast_fp16")]; + int32 x_487_split_num_splits_0 = const()[name = string("x_487_split_num_splits_0"), val = int32(2)]; + int32 x_487_split_axis_0 = const()[name = string("x_487_split_axis_0"), val = int32(1)]; + tensor x_487_split_cast_fp16_0, tensor x_487_split_cast_fp16_1 = split(axis = x_487_split_axis_0, num_splits = x_487_split_num_splits_0, x = input_987_cast_fp16)[name = string("x_487_split_cast_fp16")]; + tensor x_487_split_1_sigmoid_cast_fp16 = sigmoid(x = x_487_split_cast_fp16_1)[name = string("x_487_split_1_sigmoid_cast_fp16")]; + tensor x_487_cast_fp16 = mul(x = x_487_split_cast_fp16_0, y = x_487_split_1_sigmoid_cast_fp16)[name = string("x_487_cast_fp16")]; + tensor input_989_cast_fp16 = select(a = var_44_to_fp16, b = x_487_cast_fp16, cond = var_575)[name = string("input_989_cast_fp16")]; + bool new_x_75_interleave_0 = const()[name = string("new_x_75_interleave_0"), val = bool(false)]; + tensor new_x_75_cast_fp16 = concat(axis = var_59, interleave = new_x_75_interleave_0, values = (cache_75_cast_fp16, input_989_cast_fp16))[name = string("new_x_75_cast_fp16")]; + tensor var_4422_begin_0 = const()[name = string("op_4422_begin_0"), val = tensor([0, 0, 7])]; + tensor var_4422_end_0 = const()[name = string("op_4422_end_0"), val = tensor([1, 1024, 15])]; + tensor var_4422_end_mask_0 = const()[name = string("op_4422_end_mask_0"), val = tensor([true, true, true])]; + tensor var_4422_cast_fp16 = slice_by_index(begin = var_4422_begin_0, end = var_4422_end_0, end_mask = var_4422_end_mask_0, x = new_x_75_cast_fp16)[name = string("op_4422_cast_fp16")]; + string x_489_pad_type_0 = const()[name = string("x_489_pad_type_0"), val = string("valid")]; + int32 x_489_groups_0 = const()[name = string("x_489_groups_0"), val = int32(1024)]; + tensor x_489_strides_0 = const()[name = string("x_489_strides_0"), val = tensor([1])]; + tensor x_489_pad_0 = const()[name = string("x_489_pad_0"), val = tensor([0, 0])]; + tensor x_489_dilations_0 = const()[name = string("x_489_dilations_0"), val = tensor([1])]; + tensor encoder_layers_18_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(372934272))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(372943552))))[name = string("encoder_layers_18_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_489_cast_fp16 = conv(dilations = x_489_dilations_0, groups = x_489_groups_0, pad = x_489_pad_0, pad_type = x_489_pad_type_0, strides = x_489_strides_0, weight = encoder_layers_18_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_75_cast_fp16)[name = string("x_489_cast_fp16")]; + tensor input_991_perm_0 = const()[name = string("input_991_perm_0"), val = tensor([0, 2, 1])]; + tensor x_491_axes_0 = const()[name = string("x_491_axes_0"), val = tensor([-1])]; + tensor encoder_layers_18_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_18_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(372945664)))]; + tensor encoder_layers_18_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_18_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(372947776)))]; + tensor input_991_cast_fp16 = transpose(perm = input_991_perm_0, x = x_489_cast_fp16)[name = string("transpose_194")]; + tensor x_491_cast_fp16 = layer_norm(axes = x_491_axes_0, beta = encoder_layers_18_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_18_conv_batch_norm_weight_to_fp16, x = input_991_cast_fp16)[name = string("x_491_cast_fp16")]; + tensor input_993_perm_0 = const()[name = string("input_993_perm_0"), val = tensor([0, 2, 1])]; + tensor input_993_cast_fp16 = transpose(perm = input_993_perm_0, x = x_491_cast_fp16)[name = string("transpose_193")]; + tensor input_995_cast_fp16 = silu(x = input_993_cast_fp16)[name = string("input_995_cast_fp16")]; + string x_493_pad_type_0 = const()[name = string("x_493_pad_type_0"), val = string("valid")]; + tensor x_493_strides_0 = const()[name = string("x_493_strides_0"), val = tensor([1])]; + tensor x_493_pad_0 = const()[name = string("x_493_pad_0"), val = tensor([0, 0])]; + tensor x_493_dilations_0 = const()[name = string("x_493_dilations_0"), val = tensor([1])]; + int32 x_493_groups_0 = const()[name = string("x_493_groups_0"), val = int32(1)]; + tensor encoder_layers_18_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(372949888))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(373998528))))[name = string("encoder_layers_18_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_493_cast_fp16 = conv(dilations = x_493_dilations_0, groups = x_493_groups_0, pad = x_493_pad_0, pad_type = x_493_pad_type_0, strides = x_493_strides_0, weight = encoder_layers_18_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_995_cast_fp16)[name = string("x_493_cast_fp16")]; + tensor input_997_perm_0 = const()[name = string("input_997_perm_0"), val = tensor([0, 2, 1])]; + tensor input_997_cast_fp16 = transpose(perm = input_997_perm_0, x = x_493_cast_fp16)[name = string("transpose_192")]; + tensor input_999_cast_fp16 = add(x = input_983_cast_fp16, y = input_997_cast_fp16)[name = string("input_999_cast_fp16")]; + tensor input_1001_axes_0 = const()[name = string("input_1001_axes_0"), val = tensor([-1])]; + tensor encoder_layers_18_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_18_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(374000640)))]; + tensor encoder_layers_18_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_18_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(374002752)))]; + tensor input_1001_cast_fp16 = layer_norm(axes = input_1001_axes_0, beta = encoder_layers_18_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_18_norm_feed_forward2_weight_to_fp16, x = input_999_cast_fp16)[name = string("input_1001_cast_fp16")]; + tensor encoder_layers_18_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(374004864))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(378199232))))[name = string("encoder_layers_18_feed_forward2_linear1_weight_to_fp16_quantized")]; + tensor encoder_layers_18_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_18_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(378207488)))]; + tensor linear_170_cast_fp16 = linear(bias = encoder_layers_18_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_18_feed_forward2_linear1_weight_to_fp16_quantized, x = input_1001_cast_fp16)[name = string("linear_170_cast_fp16")]; + tensor input_1005_cast_fp16 = silu(x = linear_170_cast_fp16)[name = string("input_1005_cast_fp16")]; + tensor encoder_layers_18_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(378215744))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(382410112))))[name = string("encoder_layers_18_feed_forward2_linear2_weight_to_fp16_quantized")]; + tensor encoder_layers_18_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_18_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(382412224)))]; + tensor linear_171_cast_fp16 = linear(bias = encoder_layers_18_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_18_feed_forward2_linear2_weight_to_fp16_quantized, x = input_1005_cast_fp16)[name = string("linear_171_cast_fp16")]; + fp16 var_4465_to_fp16 = const()[name = string("op_4465_to_fp16"), val = fp16(0x1p-1)]; + tensor var_4466_cast_fp16 = mul(x = linear_171_cast_fp16, y = var_4465_to_fp16)[name = string("op_4466_cast_fp16")]; + tensor input_1011_cast_fp16 = add(x = input_999_cast_fp16, y = var_4466_cast_fp16)[name = string("input_1011_cast_fp16")]; + tensor input_1013_axes_0 = const()[name = string("input_1013_axes_0"), val = tensor([-1])]; + tensor encoder_layers_18_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_18_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(382414336)))]; + tensor encoder_layers_18_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_18_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(382416448)))]; + tensor input_1013_cast_fp16 = layer_norm(axes = input_1013_axes_0, beta = encoder_layers_18_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_18_norm_out_weight_to_fp16, x = input_1011_cast_fp16)[name = string("input_1013_cast_fp16")]; + tensor cache_77_begin_0 = const()[name = string("cache_77_begin_0"), val = tensor([19, 0, 0, 0])]; + tensor cache_77_end_0 = const()[name = string("cache_77_end_0"), val = tensor([20, 1, 42, 1024])]; + tensor cache_77_end_mask_0 = const()[name = string("cache_77_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_77_squeeze_mask_0 = const()[name = string("cache_77_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_77_cast_fp16 = slice_by_index(begin = cache_77_begin_0, end = cache_77_end_0, end_mask = cache_77_end_mask_0, squeeze_mask = cache_77_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_77_cast_fp16")]; + tensor cache_79_begin_0 = const()[name = string("cache_79_begin_0"), val = tensor([19, 0, 0, 0])]; + tensor cache_79_end_0 = const()[name = string("cache_79_end_0"), val = tensor([20, 1, 1024, 8])]; + tensor cache_79_end_mask_0 = const()[name = string("cache_79_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_79_squeeze_mask_0 = const()[name = string("cache_79_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_79_cast_fp16 = slice_by_index(begin = cache_79_begin_0, end = cache_79_end_0, end_mask = cache_79_end_mask_0, squeeze_mask = cache_79_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_79_cast_fp16")]; + tensor input_1015_axes_0 = const()[name = string("input_1015_axes_0"), val = tensor([-1])]; + tensor encoder_layers_19_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_19_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(382418560)))]; + tensor encoder_layers_19_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_19_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(382420672)))]; + tensor input_1015_cast_fp16 = layer_norm(axes = input_1015_axes_0, beta = encoder_layers_19_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_19_norm_feed_forward1_weight_to_fp16, x = input_1013_cast_fp16)[name = string("input_1015_cast_fp16")]; + tensor encoder_layers_19_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(382422784))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(386617152))))[name = string("encoder_layers_19_feed_forward1_linear1_weight_to_fp16_quantized")]; + tensor encoder_layers_19_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_19_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(386625408)))]; + tensor linear_172_cast_fp16 = linear(bias = encoder_layers_19_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_19_feed_forward1_linear1_weight_to_fp16_quantized, x = input_1015_cast_fp16)[name = string("linear_172_cast_fp16")]; + tensor input_1019_cast_fp16 = silu(x = linear_172_cast_fp16)[name = string("input_1019_cast_fp16")]; + tensor encoder_layers_19_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(386633664))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(390828032))))[name = string("encoder_layers_19_feed_forward1_linear2_weight_to_fp16_quantized")]; + tensor encoder_layers_19_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_19_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(390830144)))]; + tensor linear_173_cast_fp16 = linear(bias = encoder_layers_19_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_19_feed_forward1_linear2_weight_to_fp16_quantized, x = input_1019_cast_fp16)[name = string("linear_173_cast_fp16")]; + fp16 var_4502_to_fp16 = const()[name = string("op_4502_to_fp16"), val = fp16(0x1p-1)]; + tensor var_4503_cast_fp16 = mul(x = linear_173_cast_fp16, y = var_4502_to_fp16)[name = string("op_4503_cast_fp16")]; + tensor input_1025_cast_fp16 = add(x = input_1013_cast_fp16, y = var_4503_cast_fp16)[name = string("input_1025_cast_fp16")]; + tensor key_39_axes_0 = const()[name = string("key_39_axes_0"), val = tensor([-1])]; + tensor encoder_layers_19_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_19_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(390832256)))]; + tensor encoder_layers_19_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_19_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(390834368)))]; + tensor key_39_cast_fp16 = layer_norm(axes = key_39_axes_0, beta = encoder_layers_19_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_19_norm_self_att_weight_to_fp16, x = input_1025_cast_fp16)[name = string("key_39_cast_fp16")]; + bool input_1027_interleave_0 = const()[name = string("input_1027_interleave_0"), val = bool(false)]; + tensor input_1027_cast_fp16 = concat(axis = var_68, interleave = input_1027_interleave_0, values = (cache_77_cast_fp16, key_39_cast_fp16))[name = string("input_1027_cast_fp16")]; + tensor var_4525_begin_0 = const()[name = string("op_4525_begin_0"), val = tensor([0, 7, 0])]; + tensor var_4525_end_0 = const()[name = string("op_4525_end_0"), val = tensor([1, 42, 1024])]; + tensor var_4525_end_mask_0 = const()[name = string("op_4525_end_mask_0"), val = tensor([true, true, true])]; + tensor var_4525_cast_fp16 = slice_by_index(begin = var_4525_begin_0, end = var_4525_end_0, end_mask = var_4525_end_mask_0, x = cache_77_cast_fp16)[name = string("op_4525_cast_fp16")]; + bool var_4531_interleave_0 = const()[name = string("op_4531_interleave_0"), val = bool(false)]; + tensor var_4531_cast_fp16 = concat(axis = var_68, interleave = var_4531_interleave_0, values = (var_4525_cast_fp16, key_39_cast_fp16))[name = string("op_4531_cast_fp16")]; + tensor encoder_layers_19_self_attn_linear_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(390836480))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391885120))))[name = string("encoder_layers_19_self_attn_linear_q_weight_to_fp16_quantized")]; + tensor encoder_layers_19_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_19_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391887232)))]; + tensor linear_174_cast_fp16 = linear(bias = encoder_layers_19_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_19_self_attn_linear_q_weight_to_fp16_quantized, x = key_39_cast_fp16)[name = string("linear_174_cast_fp16")]; + tensor var_4536 = const()[name = string("op_4536"), val = tensor([1, -1, 8, 128])]; + tensor q_115_cast_fp16 = reshape(shape = var_4536, x = linear_174_cast_fp16)[name = string("q_115_cast_fp16")]; + tensor encoder_layers_19_self_attn_linear_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391889344))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(392937984))))[name = string("encoder_layers_19_self_attn_linear_k_weight_to_fp16_quantized")]; + tensor encoder_layers_19_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_19_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(392940096)))]; + tensor linear_175_cast_fp16 = linear(bias = encoder_layers_19_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_19_self_attn_linear_k_weight_to_fp16_quantized, x = input_1027_cast_fp16)[name = string("linear_175_cast_fp16")]; + tensor var_4541 = const()[name = string("op_4541"), val = tensor([1, -1, 8, 128])]; + tensor k_77_cast_fp16 = reshape(shape = var_4541, x = linear_175_cast_fp16)[name = string("k_77_cast_fp16")]; + tensor encoder_layers_19_self_attn_linear_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(392942208))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(393990848))))[name = string("encoder_layers_19_self_attn_linear_v_weight_to_fp16_quantized")]; + tensor encoder_layers_19_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_19_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(393992960)))]; + tensor linear_176_cast_fp16 = linear(bias = encoder_layers_19_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_19_self_attn_linear_v_weight_to_fp16_quantized, x = input_1027_cast_fp16)[name = string("linear_176_cast_fp16")]; + tensor var_4546 = const()[name = string("op_4546"), val = tensor([1, -1, 8, 128])]; + tensor v_39_cast_fp16 = reshape(shape = var_4546, x = linear_176_cast_fp16)[name = string("v_39_cast_fp16")]; + tensor value_47_perm_0 = const()[name = string("value_47_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_19_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_19_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(393995072)))]; + tensor var_4559_cast_fp16 = add(x = q_115_cast_fp16, y = encoder_layers_19_self_attn_pos_bias_u_to_fp16)[name = string("op_4559_cast_fp16")]; + tensor encoder_layers_19_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_19_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(393997184)))]; + tensor var_4561_cast_fp16 = add(x = q_115_cast_fp16, y = encoder_layers_19_self_attn_pos_bias_v_to_fp16)[name = string("op_4561_cast_fp16")]; + tensor q_with_bias_v_39_perm_0 = const()[name = string("q_with_bias_v_39_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_501_transpose_x_0 = const()[name = string("x_501_transpose_x_0"), val = bool(false)]; + bool x_501_transpose_y_0 = const()[name = string("x_501_transpose_y_0"), val = bool(false)]; + tensor op_4563_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(393999296))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(394098688))))[name = string("op_4563_to_fp16_quantized")]; + tensor q_with_bias_v_39_cast_fp16 = transpose(perm = q_with_bias_v_39_perm_0, x = var_4561_cast_fp16)[name = string("transpose_191")]; + tensor x_501_cast_fp16 = matmul(transpose_x = x_501_transpose_x_0, transpose_y = x_501_transpose_y_0, x = q_with_bias_v_39_cast_fp16, y = op_4563_to_fp16_quantized)[name = string("x_501_cast_fp16")]; + tensor x_503_pad_0 = const()[name = string("x_503_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_503_mode_0 = const()[name = string("x_503_mode_0"), val = string("constant")]; + fp16 const_326_to_fp16 = const()[name = string("const_326_to_fp16"), val = fp16(0x0p+0)]; + tensor x_503_cast_fp16 = pad(constant_val = const_326_to_fp16, mode = x_503_mode_0, pad = x_503_pad_0, x = x_501_cast_fp16)[name = string("x_503_cast_fp16")]; + tensor var_4571 = const()[name = string("op_4571"), val = tensor([1, 8, -1, 7])]; + tensor x_505_cast_fp16 = reshape(shape = var_4571, x = x_503_cast_fp16)[name = string("x_505_cast_fp16")]; + tensor var_4575_begin_0 = const()[name = string("op_4575_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_4575_end_0 = const()[name = string("op_4575_end_0"), val = tensor([1, 8, 98, 7])]; + tensor var_4575_end_mask_0 = const()[name = string("op_4575_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_4575_cast_fp16 = slice_by_index(begin = var_4575_begin_0, end = var_4575_end_0, end_mask = var_4575_end_mask_0, x = x_505_cast_fp16)[name = string("op_4575_cast_fp16")]; + tensor var_4576 = const()[name = string("op_4576"), val = tensor([1, 8, 7, 97])]; + tensor matrix_bd_77_cast_fp16 = reshape(shape = var_4576, x = var_4575_cast_fp16)[name = string("matrix_bd_77_cast_fp16")]; + bool matrix_ac_39_transpose_x_0 = const()[name = string("matrix_ac_39_transpose_x_0"), val = bool(false)]; + bool matrix_ac_39_transpose_y_0 = const()[name = string("matrix_ac_39_transpose_y_0"), val = bool(false)]; + tensor transpose_134_perm_0 = const()[name = string("transpose_134_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_135_perm_0 = const()[name = string("transpose_135_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_135 = transpose(perm = transpose_135_perm_0, x = k_77_cast_fp16)[name = string("transpose_189")]; + tensor transpose_134 = transpose(perm = transpose_134_perm_0, x = var_4559_cast_fp16)[name = string("transpose_190")]; + tensor matrix_ac_39_cast_fp16 = matmul(transpose_x = matrix_ac_39_transpose_x_0, transpose_y = matrix_ac_39_transpose_y_0, x = transpose_134, y = transpose_135)[name = string("matrix_ac_39_cast_fp16")]; + tensor matrix_bd_79_begin_0 = const()[name = string("matrix_bd_79_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_79_end_0 = const()[name = string("matrix_bd_79_end_0"), val = tensor([1, 8, 7, 49])]; + tensor matrix_bd_79_end_mask_0 = const()[name = string("matrix_bd_79_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_79_cast_fp16 = slice_by_index(begin = matrix_bd_79_begin_0, end = matrix_bd_79_end_0, end_mask = matrix_bd_79_end_mask_0, x = matrix_bd_77_cast_fp16)[name = string("matrix_bd_79_cast_fp16")]; + tensor var_4585_cast_fp16 = add(x = matrix_ac_39_cast_fp16, y = matrix_bd_79_cast_fp16)[name = string("op_4585_cast_fp16")]; + fp16 _inversed_scores_77_y_0_to_fp16 = const()[name = string("_inversed_scores_77_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_77_cast_fp16 = mul(x = var_4585_cast_fp16, y = _inversed_scores_77_y_0_to_fp16)[name = string("_inversed_scores_77_cast_fp16")]; + tensor scores_79_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_77_cast_fp16, cond = mask_11)[name = string("scores_79_cast_fp16")]; + tensor var_4591_cast_fp16 = softmax(axis = var_59, x = scores_79_cast_fp16)[name = string("op_4591_cast_fp16")]; + tensor input_1029_cast_fp16 = select(a = var_44_to_fp16, b = var_4591_cast_fp16, cond = mask_11)[name = string("input_1029_cast_fp16")]; + bool x_507_transpose_x_0 = const()[name = string("x_507_transpose_x_0"), val = bool(false)]; + bool x_507_transpose_y_0 = const()[name = string("x_507_transpose_y_0"), val = bool(false)]; + tensor value_47_cast_fp16 = transpose(perm = value_47_perm_0, x = v_39_cast_fp16)[name = string("transpose_188")]; + tensor x_507_cast_fp16 = matmul(transpose_x = x_507_transpose_x_0, transpose_y = x_507_transpose_y_0, x = input_1029_cast_fp16, y = value_47_cast_fp16)[name = string("x_507_cast_fp16")]; + tensor var_4595_perm_0 = const()[name = string("op_4595_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_4596 = const()[name = string("op_4596"), val = tensor([1, -1, 1024])]; + tensor var_4595_cast_fp16 = transpose(perm = var_4595_perm_0, x = x_507_cast_fp16)[name = string("transpose_187")]; + tensor input_1031_cast_fp16 = reshape(shape = var_4596, x = var_4595_cast_fp16)[name = string("input_1031_cast_fp16")]; + tensor encoder_layers_19_self_attn_linear_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(394099008))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(395147648))))[name = string("encoder_layers_19_self_attn_linear_out_weight_to_fp16_quantized")]; + tensor encoder_layers_19_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_19_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(395149760)))]; + tensor linear_178_cast_fp16 = linear(bias = encoder_layers_19_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_19_self_attn_linear_out_weight_to_fp16_quantized, x = input_1031_cast_fp16)[name = string("linear_178_cast_fp16")]; + tensor input_1035_cast_fp16 = add(x = input_1025_cast_fp16, y = linear_178_cast_fp16)[name = string("input_1035_cast_fp16")]; + tensor x_511_axes_0 = const()[name = string("x_511_axes_0"), val = tensor([-1])]; + tensor encoder_layers_19_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_19_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(395151872)))]; + tensor encoder_layers_19_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_19_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(395153984)))]; + tensor x_511_cast_fp16 = layer_norm(axes = x_511_axes_0, beta = encoder_layers_19_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_19_norm_conv_weight_to_fp16, x = input_1035_cast_fp16)[name = string("x_511_cast_fp16")]; + tensor input_1037_perm_0 = const()[name = string("input_1037_perm_0"), val = tensor([0, 2, 1])]; + string input_1039_pad_type_0 = const()[name = string("input_1039_pad_type_0"), val = string("valid")]; + tensor input_1039_strides_0 = const()[name = string("input_1039_strides_0"), val = tensor([1])]; + tensor input_1039_pad_0 = const()[name = string("input_1039_pad_0"), val = tensor([0, 0])]; + tensor input_1039_dilations_0 = const()[name = string("input_1039_dilations_0"), val = tensor([1])]; + int32 input_1039_groups_0 = const()[name = string("input_1039_groups_0"), val = int32(1)]; + tensor encoder_layers_19_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(395156096))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397253312))))[name = string("encoder_layers_19_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_1037_cast_fp16 = transpose(perm = input_1037_perm_0, x = x_511_cast_fp16)[name = string("transpose_186")]; + tensor input_1039_cast_fp16 = conv(dilations = input_1039_dilations_0, groups = input_1039_groups_0, pad = input_1039_pad_0, pad_type = input_1039_pad_type_0, strides = input_1039_strides_0, weight = encoder_layers_19_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_1037_cast_fp16)[name = string("input_1039_cast_fp16")]; + int32 x_513_split_num_splits_0 = const()[name = string("x_513_split_num_splits_0"), val = int32(2)]; + int32 x_513_split_axis_0 = const()[name = string("x_513_split_axis_0"), val = int32(1)]; + tensor x_513_split_cast_fp16_0, tensor x_513_split_cast_fp16_1 = split(axis = x_513_split_axis_0, num_splits = x_513_split_num_splits_0, x = input_1039_cast_fp16)[name = string("x_513_split_cast_fp16")]; + tensor x_513_split_1_sigmoid_cast_fp16 = sigmoid(x = x_513_split_cast_fp16_1)[name = string("x_513_split_1_sigmoid_cast_fp16")]; + tensor x_513_cast_fp16 = mul(x = x_513_split_cast_fp16_0, y = x_513_split_1_sigmoid_cast_fp16)[name = string("x_513_cast_fp16")]; + tensor input_1041_cast_fp16 = select(a = var_44_to_fp16, b = x_513_cast_fp16, cond = var_575)[name = string("input_1041_cast_fp16")]; + bool new_x_79_interleave_0 = const()[name = string("new_x_79_interleave_0"), val = bool(false)]; + tensor new_x_79_cast_fp16 = concat(axis = var_59, interleave = new_x_79_interleave_0, values = (cache_79_cast_fp16, input_1041_cast_fp16))[name = string("new_x_79_cast_fp16")]; + tensor var_4635_begin_0 = const()[name = string("op_4635_begin_0"), val = tensor([0, 0, 7])]; + tensor var_4635_end_0 = const()[name = string("op_4635_end_0"), val = tensor([1, 1024, 15])]; + tensor var_4635_end_mask_0 = const()[name = string("op_4635_end_mask_0"), val = tensor([true, true, true])]; + tensor var_4635_cast_fp16 = slice_by_index(begin = var_4635_begin_0, end = var_4635_end_0, end_mask = var_4635_end_mask_0, x = new_x_79_cast_fp16)[name = string("op_4635_cast_fp16")]; + string x_515_pad_type_0 = const()[name = string("x_515_pad_type_0"), val = string("valid")]; + int32 x_515_groups_0 = const()[name = string("x_515_groups_0"), val = int32(1024)]; + tensor x_515_strides_0 = const()[name = string("x_515_strides_0"), val = tensor([1])]; + tensor x_515_pad_0 = const()[name = string("x_515_pad_0"), val = tensor([0, 0])]; + tensor x_515_dilations_0 = const()[name = string("x_515_dilations_0"), val = tensor([1])]; + tensor encoder_layers_19_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397257472))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397266752))))[name = string("encoder_layers_19_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_515_cast_fp16 = conv(dilations = x_515_dilations_0, groups = x_515_groups_0, pad = x_515_pad_0, pad_type = x_515_pad_type_0, strides = x_515_strides_0, weight = encoder_layers_19_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_79_cast_fp16)[name = string("x_515_cast_fp16")]; + tensor input_1043_perm_0 = const()[name = string("input_1043_perm_0"), val = tensor([0, 2, 1])]; + tensor x_517_axes_0 = const()[name = string("x_517_axes_0"), val = tensor([-1])]; + tensor encoder_layers_19_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_19_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397268864)))]; + tensor encoder_layers_19_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_19_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397270976)))]; + tensor input_1043_cast_fp16 = transpose(perm = input_1043_perm_0, x = x_515_cast_fp16)[name = string("transpose_185")]; + tensor x_517_cast_fp16 = layer_norm(axes = x_517_axes_0, beta = encoder_layers_19_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_19_conv_batch_norm_weight_to_fp16, x = input_1043_cast_fp16)[name = string("x_517_cast_fp16")]; + tensor input_1045_perm_0 = const()[name = string("input_1045_perm_0"), val = tensor([0, 2, 1])]; + tensor input_1045_cast_fp16 = transpose(perm = input_1045_perm_0, x = x_517_cast_fp16)[name = string("transpose_184")]; + tensor input_1047_cast_fp16 = silu(x = input_1045_cast_fp16)[name = string("input_1047_cast_fp16")]; + string x_519_pad_type_0 = const()[name = string("x_519_pad_type_0"), val = string("valid")]; + tensor x_519_strides_0 = const()[name = string("x_519_strides_0"), val = tensor([1])]; + tensor x_519_pad_0 = const()[name = string("x_519_pad_0"), val = tensor([0, 0])]; + tensor x_519_dilations_0 = const()[name = string("x_519_dilations_0"), val = tensor([1])]; + int32 x_519_groups_0 = const()[name = string("x_519_groups_0"), val = int32(1)]; + tensor encoder_layers_19_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397273088))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(398321728))))[name = string("encoder_layers_19_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_519_cast_fp16 = conv(dilations = x_519_dilations_0, groups = x_519_groups_0, pad = x_519_pad_0, pad_type = x_519_pad_type_0, strides = x_519_strides_0, weight = encoder_layers_19_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_1047_cast_fp16)[name = string("x_519_cast_fp16")]; + tensor input_1049_perm_0 = const()[name = string("input_1049_perm_0"), val = tensor([0, 2, 1])]; + tensor input_1049_cast_fp16 = transpose(perm = input_1049_perm_0, x = x_519_cast_fp16)[name = string("transpose_183")]; + tensor input_1051_cast_fp16 = add(x = input_1035_cast_fp16, y = input_1049_cast_fp16)[name = string("input_1051_cast_fp16")]; + tensor input_1053_axes_0 = const()[name = string("input_1053_axes_0"), val = tensor([-1])]; + tensor encoder_layers_19_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_19_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(398323840)))]; + tensor encoder_layers_19_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_19_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(398325952)))]; + tensor input_1053_cast_fp16 = layer_norm(axes = input_1053_axes_0, beta = encoder_layers_19_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_19_norm_feed_forward2_weight_to_fp16, x = input_1051_cast_fp16)[name = string("input_1053_cast_fp16")]; + tensor encoder_layers_19_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(398328064))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(402522432))))[name = string("encoder_layers_19_feed_forward2_linear1_weight_to_fp16_quantized")]; + tensor encoder_layers_19_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_19_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(402530688)))]; + tensor linear_179_cast_fp16 = linear(bias = encoder_layers_19_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_19_feed_forward2_linear1_weight_to_fp16_quantized, x = input_1053_cast_fp16)[name = string("linear_179_cast_fp16")]; + tensor input_1057_cast_fp16 = silu(x = linear_179_cast_fp16)[name = string("input_1057_cast_fp16")]; + tensor encoder_layers_19_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(402538944))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406733312))))[name = string("encoder_layers_19_feed_forward2_linear2_weight_to_fp16_quantized")]; + tensor encoder_layers_19_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_19_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406735424)))]; + tensor linear_180_cast_fp16 = linear(bias = encoder_layers_19_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_19_feed_forward2_linear2_weight_to_fp16_quantized, x = input_1057_cast_fp16)[name = string("linear_180_cast_fp16")]; + fp16 var_4678_to_fp16 = const()[name = string("op_4678_to_fp16"), val = fp16(0x1p-1)]; + tensor var_4679_cast_fp16 = mul(x = linear_180_cast_fp16, y = var_4678_to_fp16)[name = string("op_4679_cast_fp16")]; + tensor input_1063_cast_fp16 = add(x = input_1051_cast_fp16, y = var_4679_cast_fp16)[name = string("input_1063_cast_fp16")]; + tensor input_1065_axes_0 = const()[name = string("input_1065_axes_0"), val = tensor([-1])]; + tensor encoder_layers_19_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_19_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406737536)))]; + tensor encoder_layers_19_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_19_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406739648)))]; + tensor input_1065_cast_fp16 = layer_norm(axes = input_1065_axes_0, beta = encoder_layers_19_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_19_norm_out_weight_to_fp16, x = input_1063_cast_fp16)[name = string("input_1065_cast_fp16")]; + tensor cache_81_begin_0 = const()[name = string("cache_81_begin_0"), val = tensor([20, 0, 0, 0])]; + tensor cache_81_end_0 = const()[name = string("cache_81_end_0"), val = tensor([21, 1, 42, 1024])]; + tensor cache_81_end_mask_0 = const()[name = string("cache_81_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_81_squeeze_mask_0 = const()[name = string("cache_81_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_81_cast_fp16 = slice_by_index(begin = cache_81_begin_0, end = cache_81_end_0, end_mask = cache_81_end_mask_0, squeeze_mask = cache_81_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_81_cast_fp16")]; + tensor cache_83_begin_0 = const()[name = string("cache_83_begin_0"), val = tensor([20, 0, 0, 0])]; + tensor cache_83_end_0 = const()[name = string("cache_83_end_0"), val = tensor([21, 1, 1024, 8])]; + tensor cache_83_end_mask_0 = const()[name = string("cache_83_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_83_squeeze_mask_0 = const()[name = string("cache_83_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_83_cast_fp16 = slice_by_index(begin = cache_83_begin_0, end = cache_83_end_0, end_mask = cache_83_end_mask_0, squeeze_mask = cache_83_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_83_cast_fp16")]; + tensor input_1067_axes_0 = const()[name = string("input_1067_axes_0"), val = tensor([-1])]; + tensor encoder_layers_20_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_20_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406741760)))]; + tensor encoder_layers_20_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_20_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406743872)))]; + tensor input_1067_cast_fp16 = layer_norm(axes = input_1067_axes_0, beta = encoder_layers_20_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_20_norm_feed_forward1_weight_to_fp16, x = input_1065_cast_fp16)[name = string("input_1067_cast_fp16")]; + tensor encoder_layers_20_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406745984))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(410940352))))[name = string("encoder_layers_20_feed_forward1_linear1_weight_to_fp16_quantized")]; + tensor encoder_layers_20_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_20_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(410948608)))]; + tensor linear_181_cast_fp16 = linear(bias = encoder_layers_20_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_20_feed_forward1_linear1_weight_to_fp16_quantized, x = input_1067_cast_fp16)[name = string("linear_181_cast_fp16")]; + tensor input_1071_cast_fp16 = silu(x = linear_181_cast_fp16)[name = string("input_1071_cast_fp16")]; + tensor encoder_layers_20_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(410956864))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(415151232))))[name = string("encoder_layers_20_feed_forward1_linear2_weight_to_fp16_quantized")]; + tensor encoder_layers_20_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_20_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(415153344)))]; + tensor linear_182_cast_fp16 = linear(bias = encoder_layers_20_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_20_feed_forward1_linear2_weight_to_fp16_quantized, x = input_1071_cast_fp16)[name = string("linear_182_cast_fp16")]; + fp16 var_4715_to_fp16 = const()[name = string("op_4715_to_fp16"), val = fp16(0x1p-1)]; + tensor var_4716_cast_fp16 = mul(x = linear_182_cast_fp16, y = var_4715_to_fp16)[name = string("op_4716_cast_fp16")]; + tensor input_1077_cast_fp16 = add(x = input_1065_cast_fp16, y = var_4716_cast_fp16)[name = string("input_1077_cast_fp16")]; + tensor key_41_axes_0 = const()[name = string("key_41_axes_0"), val = tensor([-1])]; + tensor encoder_layers_20_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_20_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(415155456)))]; + tensor encoder_layers_20_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_20_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(415157568)))]; + tensor key_41_cast_fp16 = layer_norm(axes = key_41_axes_0, beta = encoder_layers_20_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_20_norm_self_att_weight_to_fp16, x = input_1077_cast_fp16)[name = string("key_41_cast_fp16")]; + bool input_1079_interleave_0 = const()[name = string("input_1079_interleave_0"), val = bool(false)]; + tensor input_1079_cast_fp16 = concat(axis = var_68, interleave = input_1079_interleave_0, values = (cache_81_cast_fp16, key_41_cast_fp16))[name = string("input_1079_cast_fp16")]; + tensor var_4738_begin_0 = const()[name = string("op_4738_begin_0"), val = tensor([0, 7, 0])]; + tensor var_4738_end_0 = const()[name = string("op_4738_end_0"), val = tensor([1, 42, 1024])]; + tensor var_4738_end_mask_0 = const()[name = string("op_4738_end_mask_0"), val = tensor([true, true, true])]; + tensor var_4738_cast_fp16 = slice_by_index(begin = var_4738_begin_0, end = var_4738_end_0, end_mask = var_4738_end_mask_0, x = cache_81_cast_fp16)[name = string("op_4738_cast_fp16")]; + bool var_4744_interleave_0 = const()[name = string("op_4744_interleave_0"), val = bool(false)]; + tensor var_4744_cast_fp16 = concat(axis = var_68, interleave = var_4744_interleave_0, values = (var_4738_cast_fp16, key_41_cast_fp16))[name = string("op_4744_cast_fp16")]; + tensor encoder_layers_20_self_attn_linear_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(415159680))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(416208320))))[name = string("encoder_layers_20_self_attn_linear_q_weight_to_fp16_quantized")]; + tensor encoder_layers_20_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_20_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(416210432)))]; + tensor linear_183_cast_fp16 = linear(bias = encoder_layers_20_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_20_self_attn_linear_q_weight_to_fp16_quantized, x = key_41_cast_fp16)[name = string("linear_183_cast_fp16")]; + tensor var_4749 = const()[name = string("op_4749"), val = tensor([1, -1, 8, 128])]; + tensor q_121_cast_fp16 = reshape(shape = var_4749, x = linear_183_cast_fp16)[name = string("q_121_cast_fp16")]; + tensor encoder_layers_20_self_attn_linear_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(416212544))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(417261184))))[name = string("encoder_layers_20_self_attn_linear_k_weight_to_fp16_quantized")]; + tensor encoder_layers_20_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_20_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(417263296)))]; + tensor linear_184_cast_fp16 = linear(bias = encoder_layers_20_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_20_self_attn_linear_k_weight_to_fp16_quantized, x = input_1079_cast_fp16)[name = string("linear_184_cast_fp16")]; + tensor var_4754 = const()[name = string("op_4754"), val = tensor([1, -1, 8, 128])]; + tensor k_81_cast_fp16 = reshape(shape = var_4754, x = linear_184_cast_fp16)[name = string("k_81_cast_fp16")]; + tensor encoder_layers_20_self_attn_linear_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(417265408))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(418314048))))[name = string("encoder_layers_20_self_attn_linear_v_weight_to_fp16_quantized")]; + tensor encoder_layers_20_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_20_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(418316160)))]; + tensor linear_185_cast_fp16 = linear(bias = encoder_layers_20_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_20_self_attn_linear_v_weight_to_fp16_quantized, x = input_1079_cast_fp16)[name = string("linear_185_cast_fp16")]; + tensor var_4759 = const()[name = string("op_4759"), val = tensor([1, -1, 8, 128])]; + tensor v_41_cast_fp16 = reshape(shape = var_4759, x = linear_185_cast_fp16)[name = string("v_41_cast_fp16")]; + tensor value_49_perm_0 = const()[name = string("value_49_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_20_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_20_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(418318272)))]; + tensor var_4772_cast_fp16 = add(x = q_121_cast_fp16, y = encoder_layers_20_self_attn_pos_bias_u_to_fp16)[name = string("op_4772_cast_fp16")]; + tensor encoder_layers_20_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_20_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(418320384)))]; + tensor var_4774_cast_fp16 = add(x = q_121_cast_fp16, y = encoder_layers_20_self_attn_pos_bias_v_to_fp16)[name = string("op_4774_cast_fp16")]; + tensor q_with_bias_v_41_perm_0 = const()[name = string("q_with_bias_v_41_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_527_transpose_x_0 = const()[name = string("x_527_transpose_x_0"), val = bool(false)]; + bool x_527_transpose_y_0 = const()[name = string("x_527_transpose_y_0"), val = bool(false)]; + tensor op_4776_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(418322496))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(418421888))))[name = string("op_4776_to_fp16_quantized")]; + tensor q_with_bias_v_41_cast_fp16 = transpose(perm = q_with_bias_v_41_perm_0, x = var_4774_cast_fp16)[name = string("transpose_182")]; + tensor x_527_cast_fp16 = matmul(transpose_x = x_527_transpose_x_0, transpose_y = x_527_transpose_y_0, x = q_with_bias_v_41_cast_fp16, y = op_4776_to_fp16_quantized)[name = string("x_527_cast_fp16")]; + tensor x_529_pad_0 = const()[name = string("x_529_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_529_mode_0 = const()[name = string("x_529_mode_0"), val = string("constant")]; + fp16 const_339_to_fp16 = const()[name = string("const_339_to_fp16"), val = fp16(0x0p+0)]; + tensor x_529_cast_fp16 = pad(constant_val = const_339_to_fp16, mode = x_529_mode_0, pad = x_529_pad_0, x = x_527_cast_fp16)[name = string("x_529_cast_fp16")]; + tensor var_4784 = const()[name = string("op_4784"), val = tensor([1, 8, -1, 7])]; + tensor x_531_cast_fp16 = reshape(shape = var_4784, x = x_529_cast_fp16)[name = string("x_531_cast_fp16")]; + tensor var_4788_begin_0 = const()[name = string("op_4788_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_4788_end_0 = const()[name = string("op_4788_end_0"), val = tensor([1, 8, 98, 7])]; + tensor var_4788_end_mask_0 = const()[name = string("op_4788_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_4788_cast_fp16 = slice_by_index(begin = var_4788_begin_0, end = var_4788_end_0, end_mask = var_4788_end_mask_0, x = x_531_cast_fp16)[name = string("op_4788_cast_fp16")]; + tensor var_4789 = const()[name = string("op_4789"), val = tensor([1, 8, 7, 97])]; + tensor matrix_bd_81_cast_fp16 = reshape(shape = var_4789, x = var_4788_cast_fp16)[name = string("matrix_bd_81_cast_fp16")]; + bool matrix_ac_41_transpose_x_0 = const()[name = string("matrix_ac_41_transpose_x_0"), val = bool(false)]; + bool matrix_ac_41_transpose_y_0 = const()[name = string("matrix_ac_41_transpose_y_0"), val = bool(false)]; + tensor transpose_136_perm_0 = const()[name = string("transpose_136_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_137_perm_0 = const()[name = string("transpose_137_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_137 = transpose(perm = transpose_137_perm_0, x = k_81_cast_fp16)[name = string("transpose_180")]; + tensor transpose_136 = transpose(perm = transpose_136_perm_0, x = var_4772_cast_fp16)[name = string("transpose_181")]; + tensor matrix_ac_41_cast_fp16 = matmul(transpose_x = matrix_ac_41_transpose_x_0, transpose_y = matrix_ac_41_transpose_y_0, x = transpose_136, y = transpose_137)[name = string("matrix_ac_41_cast_fp16")]; + tensor matrix_bd_83_begin_0 = const()[name = string("matrix_bd_83_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_83_end_0 = const()[name = string("matrix_bd_83_end_0"), val = tensor([1, 8, 7, 49])]; + tensor matrix_bd_83_end_mask_0 = const()[name = string("matrix_bd_83_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_83_cast_fp16 = slice_by_index(begin = matrix_bd_83_begin_0, end = matrix_bd_83_end_0, end_mask = matrix_bd_83_end_mask_0, x = matrix_bd_81_cast_fp16)[name = string("matrix_bd_83_cast_fp16")]; + tensor var_4798_cast_fp16 = add(x = matrix_ac_41_cast_fp16, y = matrix_bd_83_cast_fp16)[name = string("op_4798_cast_fp16")]; + fp16 _inversed_scores_81_y_0_to_fp16 = const()[name = string("_inversed_scores_81_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_81_cast_fp16 = mul(x = var_4798_cast_fp16, y = _inversed_scores_81_y_0_to_fp16)[name = string("_inversed_scores_81_cast_fp16")]; + tensor scores_83_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_81_cast_fp16, cond = mask_11)[name = string("scores_83_cast_fp16")]; + tensor var_4804_cast_fp16 = softmax(axis = var_59, x = scores_83_cast_fp16)[name = string("op_4804_cast_fp16")]; + tensor input_1081_cast_fp16 = select(a = var_44_to_fp16, b = var_4804_cast_fp16, cond = mask_11)[name = string("input_1081_cast_fp16")]; + bool x_533_transpose_x_0 = const()[name = string("x_533_transpose_x_0"), val = bool(false)]; + bool x_533_transpose_y_0 = const()[name = string("x_533_transpose_y_0"), val = bool(false)]; + tensor value_49_cast_fp16 = transpose(perm = value_49_perm_0, x = v_41_cast_fp16)[name = string("transpose_179")]; + tensor x_533_cast_fp16 = matmul(transpose_x = x_533_transpose_x_0, transpose_y = x_533_transpose_y_0, x = input_1081_cast_fp16, y = value_49_cast_fp16)[name = string("x_533_cast_fp16")]; + tensor var_4808_perm_0 = const()[name = string("op_4808_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_4809 = const()[name = string("op_4809"), val = tensor([1, -1, 1024])]; + tensor var_4808_cast_fp16 = transpose(perm = var_4808_perm_0, x = x_533_cast_fp16)[name = string("transpose_178")]; + tensor input_1083_cast_fp16 = reshape(shape = var_4809, x = var_4808_cast_fp16)[name = string("input_1083_cast_fp16")]; + tensor encoder_layers_20_self_attn_linear_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(418422208))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(419470848))))[name = string("encoder_layers_20_self_attn_linear_out_weight_to_fp16_quantized")]; + tensor encoder_layers_20_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_20_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(419472960)))]; + tensor linear_187_cast_fp16 = linear(bias = encoder_layers_20_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_20_self_attn_linear_out_weight_to_fp16_quantized, x = input_1083_cast_fp16)[name = string("linear_187_cast_fp16")]; + tensor input_1087_cast_fp16 = add(x = input_1077_cast_fp16, y = linear_187_cast_fp16)[name = string("input_1087_cast_fp16")]; + tensor x_537_axes_0 = const()[name = string("x_537_axes_0"), val = tensor([-1])]; + tensor encoder_layers_20_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_20_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(419475072)))]; + tensor encoder_layers_20_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_20_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(419477184)))]; + tensor x_537_cast_fp16 = layer_norm(axes = x_537_axes_0, beta = encoder_layers_20_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_20_norm_conv_weight_to_fp16, x = input_1087_cast_fp16)[name = string("x_537_cast_fp16")]; + tensor input_1089_perm_0 = const()[name = string("input_1089_perm_0"), val = tensor([0, 2, 1])]; + string input_1091_pad_type_0 = const()[name = string("input_1091_pad_type_0"), val = string("valid")]; + tensor input_1091_strides_0 = const()[name = string("input_1091_strides_0"), val = tensor([1])]; + tensor input_1091_pad_0 = const()[name = string("input_1091_pad_0"), val = tensor([0, 0])]; + tensor input_1091_dilations_0 = const()[name = string("input_1091_dilations_0"), val = tensor([1])]; + int32 input_1091_groups_0 = const()[name = string("input_1091_groups_0"), val = int32(1)]; + tensor encoder_layers_20_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(419479296))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421576512))))[name = string("encoder_layers_20_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_1089_cast_fp16 = transpose(perm = input_1089_perm_0, x = x_537_cast_fp16)[name = string("transpose_177")]; + tensor input_1091_cast_fp16 = conv(dilations = input_1091_dilations_0, groups = input_1091_groups_0, pad = input_1091_pad_0, pad_type = input_1091_pad_type_0, strides = input_1091_strides_0, weight = encoder_layers_20_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_1089_cast_fp16)[name = string("input_1091_cast_fp16")]; + int32 x_539_split_num_splits_0 = const()[name = string("x_539_split_num_splits_0"), val = int32(2)]; + int32 x_539_split_axis_0 = const()[name = string("x_539_split_axis_0"), val = int32(1)]; + tensor x_539_split_cast_fp16_0, tensor x_539_split_cast_fp16_1 = split(axis = x_539_split_axis_0, num_splits = x_539_split_num_splits_0, x = input_1091_cast_fp16)[name = string("x_539_split_cast_fp16")]; + tensor x_539_split_1_sigmoid_cast_fp16 = sigmoid(x = x_539_split_cast_fp16_1)[name = string("x_539_split_1_sigmoid_cast_fp16")]; + tensor x_539_cast_fp16 = mul(x = x_539_split_cast_fp16_0, y = x_539_split_1_sigmoid_cast_fp16)[name = string("x_539_cast_fp16")]; + tensor input_1093_cast_fp16 = select(a = var_44_to_fp16, b = x_539_cast_fp16, cond = var_575)[name = string("input_1093_cast_fp16")]; + bool new_x_83_interleave_0 = const()[name = string("new_x_83_interleave_0"), val = bool(false)]; + tensor new_x_83_cast_fp16 = concat(axis = var_59, interleave = new_x_83_interleave_0, values = (cache_83_cast_fp16, input_1093_cast_fp16))[name = string("new_x_83_cast_fp16")]; + tensor var_4848_begin_0 = const()[name = string("op_4848_begin_0"), val = tensor([0, 0, 7])]; + tensor var_4848_end_0 = const()[name = string("op_4848_end_0"), val = tensor([1, 1024, 15])]; + tensor var_4848_end_mask_0 = const()[name = string("op_4848_end_mask_0"), val = tensor([true, true, true])]; + tensor var_4848_cast_fp16 = slice_by_index(begin = var_4848_begin_0, end = var_4848_end_0, end_mask = var_4848_end_mask_0, x = new_x_83_cast_fp16)[name = string("op_4848_cast_fp16")]; + string x_541_pad_type_0 = const()[name = string("x_541_pad_type_0"), val = string("valid")]; + int32 x_541_groups_0 = const()[name = string("x_541_groups_0"), val = int32(1024)]; + tensor x_541_strides_0 = const()[name = string("x_541_strides_0"), val = tensor([1])]; + tensor x_541_pad_0 = const()[name = string("x_541_pad_0"), val = tensor([0, 0])]; + tensor x_541_dilations_0 = const()[name = string("x_541_dilations_0"), val = tensor([1])]; + tensor encoder_layers_20_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421580672))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421589952))))[name = string("encoder_layers_20_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_541_cast_fp16 = conv(dilations = x_541_dilations_0, groups = x_541_groups_0, pad = x_541_pad_0, pad_type = x_541_pad_type_0, strides = x_541_strides_0, weight = encoder_layers_20_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_83_cast_fp16)[name = string("x_541_cast_fp16")]; + tensor input_1095_perm_0 = const()[name = string("input_1095_perm_0"), val = tensor([0, 2, 1])]; + tensor x_543_axes_0 = const()[name = string("x_543_axes_0"), val = tensor([-1])]; + tensor encoder_layers_20_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_20_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421592064)))]; + tensor encoder_layers_20_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_20_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421594176)))]; + tensor input_1095_cast_fp16 = transpose(perm = input_1095_perm_0, x = x_541_cast_fp16)[name = string("transpose_176")]; + tensor x_543_cast_fp16 = layer_norm(axes = x_543_axes_0, beta = encoder_layers_20_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_20_conv_batch_norm_weight_to_fp16, x = input_1095_cast_fp16)[name = string("x_543_cast_fp16")]; + tensor input_1097_perm_0 = const()[name = string("input_1097_perm_0"), val = tensor([0, 2, 1])]; + tensor input_1097_cast_fp16 = transpose(perm = input_1097_perm_0, x = x_543_cast_fp16)[name = string("transpose_175")]; + tensor input_1099_cast_fp16 = silu(x = input_1097_cast_fp16)[name = string("input_1099_cast_fp16")]; + string x_545_pad_type_0 = const()[name = string("x_545_pad_type_0"), val = string("valid")]; + tensor x_545_strides_0 = const()[name = string("x_545_strides_0"), val = tensor([1])]; + tensor x_545_pad_0 = const()[name = string("x_545_pad_0"), val = tensor([0, 0])]; + tensor x_545_dilations_0 = const()[name = string("x_545_dilations_0"), val = tensor([1])]; + int32 x_545_groups_0 = const()[name = string("x_545_groups_0"), val = int32(1)]; + tensor encoder_layers_20_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421596288))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(422644928))))[name = string("encoder_layers_20_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_545_cast_fp16 = conv(dilations = x_545_dilations_0, groups = x_545_groups_0, pad = x_545_pad_0, pad_type = x_545_pad_type_0, strides = x_545_strides_0, weight = encoder_layers_20_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_1099_cast_fp16)[name = string("x_545_cast_fp16")]; + tensor input_1101_perm_0 = const()[name = string("input_1101_perm_0"), val = tensor([0, 2, 1])]; + tensor input_1101_cast_fp16 = transpose(perm = input_1101_perm_0, x = x_545_cast_fp16)[name = string("transpose_174")]; + tensor input_1103_cast_fp16 = add(x = input_1087_cast_fp16, y = input_1101_cast_fp16)[name = string("input_1103_cast_fp16")]; + tensor input_1105_axes_0 = const()[name = string("input_1105_axes_0"), val = tensor([-1])]; + tensor encoder_layers_20_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_20_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(422647040)))]; + tensor encoder_layers_20_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_20_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(422649152)))]; + tensor input_1105_cast_fp16 = layer_norm(axes = input_1105_axes_0, beta = encoder_layers_20_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_20_norm_feed_forward2_weight_to_fp16, x = input_1103_cast_fp16)[name = string("input_1105_cast_fp16")]; + tensor encoder_layers_20_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(422651264))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(426845632))))[name = string("encoder_layers_20_feed_forward2_linear1_weight_to_fp16_quantized")]; + tensor encoder_layers_20_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_20_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(426853888)))]; + tensor linear_188_cast_fp16 = linear(bias = encoder_layers_20_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_20_feed_forward2_linear1_weight_to_fp16_quantized, x = input_1105_cast_fp16)[name = string("linear_188_cast_fp16")]; + tensor input_1109_cast_fp16 = silu(x = linear_188_cast_fp16)[name = string("input_1109_cast_fp16")]; + tensor encoder_layers_20_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(426862144))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(431056512))))[name = string("encoder_layers_20_feed_forward2_linear2_weight_to_fp16_quantized")]; + tensor encoder_layers_20_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_20_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(431058624)))]; + tensor linear_189_cast_fp16 = linear(bias = encoder_layers_20_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_20_feed_forward2_linear2_weight_to_fp16_quantized, x = input_1109_cast_fp16)[name = string("linear_189_cast_fp16")]; + fp16 var_4891_to_fp16 = const()[name = string("op_4891_to_fp16"), val = fp16(0x1p-1)]; + tensor var_4892_cast_fp16 = mul(x = linear_189_cast_fp16, y = var_4891_to_fp16)[name = string("op_4892_cast_fp16")]; + tensor input_1115_cast_fp16 = add(x = input_1103_cast_fp16, y = var_4892_cast_fp16)[name = string("input_1115_cast_fp16")]; + tensor input_1117_axes_0 = const()[name = string("input_1117_axes_0"), val = tensor([-1])]; + tensor encoder_layers_20_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_20_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(431060736)))]; + tensor encoder_layers_20_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_20_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(431062848)))]; + tensor input_1117_cast_fp16 = layer_norm(axes = input_1117_axes_0, beta = encoder_layers_20_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_20_norm_out_weight_to_fp16, x = input_1115_cast_fp16)[name = string("input_1117_cast_fp16")]; + tensor cache_85_begin_0 = const()[name = string("cache_85_begin_0"), val = tensor([21, 0, 0, 0])]; + tensor cache_85_end_0 = const()[name = string("cache_85_end_0"), val = tensor([22, 1, 42, 1024])]; + tensor cache_85_end_mask_0 = const()[name = string("cache_85_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_85_squeeze_mask_0 = const()[name = string("cache_85_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_85_cast_fp16 = slice_by_index(begin = cache_85_begin_0, end = cache_85_end_0, end_mask = cache_85_end_mask_0, squeeze_mask = cache_85_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_85_cast_fp16")]; + tensor cache_87_begin_0 = const()[name = string("cache_87_begin_0"), val = tensor([21, 0, 0, 0])]; + tensor cache_87_end_0 = const()[name = string("cache_87_end_0"), val = tensor([22, 1, 1024, 8])]; + tensor cache_87_end_mask_0 = const()[name = string("cache_87_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_87_squeeze_mask_0 = const()[name = string("cache_87_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_87_cast_fp16 = slice_by_index(begin = cache_87_begin_0, end = cache_87_end_0, end_mask = cache_87_end_mask_0, squeeze_mask = cache_87_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_87_cast_fp16")]; + tensor input_1119_axes_0 = const()[name = string("input_1119_axes_0"), val = tensor([-1])]; + tensor encoder_layers_21_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_21_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(431064960)))]; + tensor encoder_layers_21_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_21_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(431067072)))]; + tensor input_1119_cast_fp16 = layer_norm(axes = input_1119_axes_0, beta = encoder_layers_21_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_21_norm_feed_forward1_weight_to_fp16, x = input_1117_cast_fp16)[name = string("input_1119_cast_fp16")]; + tensor encoder_layers_21_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(431069184))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(435263552))))[name = string("encoder_layers_21_feed_forward1_linear1_weight_to_fp16_quantized")]; + tensor encoder_layers_21_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_21_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(435271808)))]; + tensor linear_190_cast_fp16 = linear(bias = encoder_layers_21_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_21_feed_forward1_linear1_weight_to_fp16_quantized, x = input_1119_cast_fp16)[name = string("linear_190_cast_fp16")]; + tensor input_1123_cast_fp16 = silu(x = linear_190_cast_fp16)[name = string("input_1123_cast_fp16")]; + tensor encoder_layers_21_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(435280064))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(439474432))))[name = string("encoder_layers_21_feed_forward1_linear2_weight_to_fp16_quantized")]; + tensor encoder_layers_21_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_21_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(439476544)))]; + tensor linear_191_cast_fp16 = linear(bias = encoder_layers_21_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_21_feed_forward1_linear2_weight_to_fp16_quantized, x = input_1123_cast_fp16)[name = string("linear_191_cast_fp16")]; + fp16 var_4928_to_fp16 = const()[name = string("op_4928_to_fp16"), val = fp16(0x1p-1)]; + tensor var_4929_cast_fp16 = mul(x = linear_191_cast_fp16, y = var_4928_to_fp16)[name = string("op_4929_cast_fp16")]; + tensor input_1129_cast_fp16 = add(x = input_1117_cast_fp16, y = var_4929_cast_fp16)[name = string("input_1129_cast_fp16")]; + tensor key_43_axes_0 = const()[name = string("key_43_axes_0"), val = tensor([-1])]; + tensor encoder_layers_21_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_21_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(439478656)))]; + tensor encoder_layers_21_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_21_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(439480768)))]; + tensor key_43_cast_fp16 = layer_norm(axes = key_43_axes_0, beta = encoder_layers_21_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_21_norm_self_att_weight_to_fp16, x = input_1129_cast_fp16)[name = string("key_43_cast_fp16")]; + bool input_1131_interleave_0 = const()[name = string("input_1131_interleave_0"), val = bool(false)]; + tensor input_1131_cast_fp16 = concat(axis = var_68, interleave = input_1131_interleave_0, values = (cache_85_cast_fp16, key_43_cast_fp16))[name = string("input_1131_cast_fp16")]; + tensor var_4951_begin_0 = const()[name = string("op_4951_begin_0"), val = tensor([0, 7, 0])]; + tensor var_4951_end_0 = const()[name = string("op_4951_end_0"), val = tensor([1, 42, 1024])]; + tensor var_4951_end_mask_0 = const()[name = string("op_4951_end_mask_0"), val = tensor([true, true, true])]; + tensor var_4951_cast_fp16 = slice_by_index(begin = var_4951_begin_0, end = var_4951_end_0, end_mask = var_4951_end_mask_0, x = cache_85_cast_fp16)[name = string("op_4951_cast_fp16")]; + bool var_4957_interleave_0 = const()[name = string("op_4957_interleave_0"), val = bool(false)]; + tensor var_4957_cast_fp16 = concat(axis = var_68, interleave = var_4957_interleave_0, values = (var_4951_cast_fp16, key_43_cast_fp16))[name = string("op_4957_cast_fp16")]; + tensor encoder_layers_21_self_attn_linear_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(439482880))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(440531520))))[name = string("encoder_layers_21_self_attn_linear_q_weight_to_fp16_quantized")]; + tensor encoder_layers_21_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_21_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(440533632)))]; + tensor linear_192_cast_fp16 = linear(bias = encoder_layers_21_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_21_self_attn_linear_q_weight_to_fp16_quantized, x = key_43_cast_fp16)[name = string("linear_192_cast_fp16")]; + tensor var_4962 = const()[name = string("op_4962"), val = tensor([1, -1, 8, 128])]; + tensor q_127_cast_fp16 = reshape(shape = var_4962, x = linear_192_cast_fp16)[name = string("q_127_cast_fp16")]; + tensor encoder_layers_21_self_attn_linear_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(440535744))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(441584384))))[name = string("encoder_layers_21_self_attn_linear_k_weight_to_fp16_quantized")]; + tensor encoder_layers_21_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_21_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(441586496)))]; + tensor linear_193_cast_fp16 = linear(bias = encoder_layers_21_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_21_self_attn_linear_k_weight_to_fp16_quantized, x = input_1131_cast_fp16)[name = string("linear_193_cast_fp16")]; + tensor var_4967 = const()[name = string("op_4967"), val = tensor([1, -1, 8, 128])]; + tensor k_85_cast_fp16 = reshape(shape = var_4967, x = linear_193_cast_fp16)[name = string("k_85_cast_fp16")]; + tensor encoder_layers_21_self_attn_linear_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(441588608))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(442637248))))[name = string("encoder_layers_21_self_attn_linear_v_weight_to_fp16_quantized")]; + tensor encoder_layers_21_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_21_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(442639360)))]; + tensor linear_194_cast_fp16 = linear(bias = encoder_layers_21_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_21_self_attn_linear_v_weight_to_fp16_quantized, x = input_1131_cast_fp16)[name = string("linear_194_cast_fp16")]; + tensor var_4972 = const()[name = string("op_4972"), val = tensor([1, -1, 8, 128])]; + tensor v_43_cast_fp16 = reshape(shape = var_4972, x = linear_194_cast_fp16)[name = string("v_43_cast_fp16")]; + tensor value_51_perm_0 = const()[name = string("value_51_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_21_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_21_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(442641472)))]; + tensor var_4985_cast_fp16 = add(x = q_127_cast_fp16, y = encoder_layers_21_self_attn_pos_bias_u_to_fp16)[name = string("op_4985_cast_fp16")]; + tensor encoder_layers_21_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_21_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(442643584)))]; + tensor var_4987_cast_fp16 = add(x = q_127_cast_fp16, y = encoder_layers_21_self_attn_pos_bias_v_to_fp16)[name = string("op_4987_cast_fp16")]; + tensor q_with_bias_v_43_perm_0 = const()[name = string("q_with_bias_v_43_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_553_transpose_x_0 = const()[name = string("x_553_transpose_x_0"), val = bool(false)]; + bool x_553_transpose_y_0 = const()[name = string("x_553_transpose_y_0"), val = bool(false)]; + tensor op_4989_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(442645696))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(442745088))))[name = string("op_4989_to_fp16_quantized")]; + tensor q_with_bias_v_43_cast_fp16 = transpose(perm = q_with_bias_v_43_perm_0, x = var_4987_cast_fp16)[name = string("transpose_173")]; + tensor x_553_cast_fp16 = matmul(transpose_x = x_553_transpose_x_0, transpose_y = x_553_transpose_y_0, x = q_with_bias_v_43_cast_fp16, y = op_4989_to_fp16_quantized)[name = string("x_553_cast_fp16")]; + tensor x_555_pad_0 = const()[name = string("x_555_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_555_mode_0 = const()[name = string("x_555_mode_0"), val = string("constant")]; + fp16 const_352_to_fp16 = const()[name = string("const_352_to_fp16"), val = fp16(0x0p+0)]; + tensor x_555_cast_fp16 = pad(constant_val = const_352_to_fp16, mode = x_555_mode_0, pad = x_555_pad_0, x = x_553_cast_fp16)[name = string("x_555_cast_fp16")]; + tensor var_4997 = const()[name = string("op_4997"), val = tensor([1, 8, -1, 7])]; + tensor x_557_cast_fp16 = reshape(shape = var_4997, x = x_555_cast_fp16)[name = string("x_557_cast_fp16")]; + tensor var_5001_begin_0 = const()[name = string("op_5001_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_5001_end_0 = const()[name = string("op_5001_end_0"), val = tensor([1, 8, 98, 7])]; + tensor var_5001_end_mask_0 = const()[name = string("op_5001_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_5001_cast_fp16 = slice_by_index(begin = var_5001_begin_0, end = var_5001_end_0, end_mask = var_5001_end_mask_0, x = x_557_cast_fp16)[name = string("op_5001_cast_fp16")]; + tensor var_5002 = const()[name = string("op_5002"), val = tensor([1, 8, 7, 97])]; + tensor matrix_bd_85_cast_fp16 = reshape(shape = var_5002, x = var_5001_cast_fp16)[name = string("matrix_bd_85_cast_fp16")]; + bool matrix_ac_43_transpose_x_0 = const()[name = string("matrix_ac_43_transpose_x_0"), val = bool(false)]; + bool matrix_ac_43_transpose_y_0 = const()[name = string("matrix_ac_43_transpose_y_0"), val = bool(false)]; + tensor transpose_138_perm_0 = const()[name = string("transpose_138_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_139_perm_0 = const()[name = string("transpose_139_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_139 = transpose(perm = transpose_139_perm_0, x = k_85_cast_fp16)[name = string("transpose_171")]; + tensor transpose_138 = transpose(perm = transpose_138_perm_0, x = var_4985_cast_fp16)[name = string("transpose_172")]; + tensor matrix_ac_43_cast_fp16 = matmul(transpose_x = matrix_ac_43_transpose_x_0, transpose_y = matrix_ac_43_transpose_y_0, x = transpose_138, y = transpose_139)[name = string("matrix_ac_43_cast_fp16")]; + tensor matrix_bd_87_begin_0 = const()[name = string("matrix_bd_87_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_87_end_0 = const()[name = string("matrix_bd_87_end_0"), val = tensor([1, 8, 7, 49])]; + tensor matrix_bd_87_end_mask_0 = const()[name = string("matrix_bd_87_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_87_cast_fp16 = slice_by_index(begin = matrix_bd_87_begin_0, end = matrix_bd_87_end_0, end_mask = matrix_bd_87_end_mask_0, x = matrix_bd_85_cast_fp16)[name = string("matrix_bd_87_cast_fp16")]; + tensor var_5011_cast_fp16 = add(x = matrix_ac_43_cast_fp16, y = matrix_bd_87_cast_fp16)[name = string("op_5011_cast_fp16")]; + fp16 _inversed_scores_85_y_0_to_fp16 = const()[name = string("_inversed_scores_85_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_85_cast_fp16 = mul(x = var_5011_cast_fp16, y = _inversed_scores_85_y_0_to_fp16)[name = string("_inversed_scores_85_cast_fp16")]; + tensor scores_87_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_85_cast_fp16, cond = mask_11)[name = string("scores_87_cast_fp16")]; + tensor var_5017_cast_fp16 = softmax(axis = var_59, x = scores_87_cast_fp16)[name = string("op_5017_cast_fp16")]; + tensor input_1133_cast_fp16 = select(a = var_44_to_fp16, b = var_5017_cast_fp16, cond = mask_11)[name = string("input_1133_cast_fp16")]; + bool x_559_transpose_x_0 = const()[name = string("x_559_transpose_x_0"), val = bool(false)]; + bool x_559_transpose_y_0 = const()[name = string("x_559_transpose_y_0"), val = bool(false)]; + tensor value_51_cast_fp16 = transpose(perm = value_51_perm_0, x = v_43_cast_fp16)[name = string("transpose_170")]; + tensor x_559_cast_fp16 = matmul(transpose_x = x_559_transpose_x_0, transpose_y = x_559_transpose_y_0, x = input_1133_cast_fp16, y = value_51_cast_fp16)[name = string("x_559_cast_fp16")]; + tensor var_5021_perm_0 = const()[name = string("op_5021_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_5022 = const()[name = string("op_5022"), val = tensor([1, -1, 1024])]; + tensor var_5021_cast_fp16 = transpose(perm = var_5021_perm_0, x = x_559_cast_fp16)[name = string("transpose_169")]; + tensor input_1135_cast_fp16 = reshape(shape = var_5022, x = var_5021_cast_fp16)[name = string("input_1135_cast_fp16")]; + tensor encoder_layers_21_self_attn_linear_out_weight_to_fp16 = const()[name = string("encoder_layers_21_self_attn_linear_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(442745408)))]; + tensor encoder_layers_21_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_21_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(444842624)))]; + tensor linear_196_cast_fp16 = linear(bias = encoder_layers_21_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_21_self_attn_linear_out_weight_to_fp16, x = input_1135_cast_fp16)[name = string("linear_196_cast_fp16")]; + tensor input_1139_cast_fp16 = add(x = input_1129_cast_fp16, y = linear_196_cast_fp16)[name = string("input_1139_cast_fp16")]; + tensor x_563_axes_0 = const()[name = string("x_563_axes_0"), val = tensor([-1])]; + tensor encoder_layers_21_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_21_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(444844736)))]; + tensor encoder_layers_21_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_21_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(444846848)))]; + tensor x_563_cast_fp16 = layer_norm(axes = x_563_axes_0, beta = encoder_layers_21_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_21_norm_conv_weight_to_fp16, x = input_1139_cast_fp16)[name = string("x_563_cast_fp16")]; + tensor input_1141_perm_0 = const()[name = string("input_1141_perm_0"), val = tensor([0, 2, 1])]; + string input_1143_pad_type_0 = const()[name = string("input_1143_pad_type_0"), val = string("valid")]; + tensor input_1143_strides_0 = const()[name = string("input_1143_strides_0"), val = tensor([1])]; + tensor input_1143_pad_0 = const()[name = string("input_1143_pad_0"), val = tensor([0, 0])]; + tensor input_1143_dilations_0 = const()[name = string("input_1143_dilations_0"), val = tensor([1])]; + int32 input_1143_groups_0 = const()[name = string("input_1143_groups_0"), val = int32(1)]; + tensor encoder_layers_21_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(444848960))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(446946176))))[name = string("encoder_layers_21_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_1141_cast_fp16 = transpose(perm = input_1141_perm_0, x = x_563_cast_fp16)[name = string("transpose_168")]; + tensor input_1143_cast_fp16 = conv(dilations = input_1143_dilations_0, groups = input_1143_groups_0, pad = input_1143_pad_0, pad_type = input_1143_pad_type_0, strides = input_1143_strides_0, weight = encoder_layers_21_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_1141_cast_fp16)[name = string("input_1143_cast_fp16")]; + int32 x_565_split_num_splits_0 = const()[name = string("x_565_split_num_splits_0"), val = int32(2)]; + int32 x_565_split_axis_0 = const()[name = string("x_565_split_axis_0"), val = int32(1)]; + tensor x_565_split_cast_fp16_0, tensor x_565_split_cast_fp16_1 = split(axis = x_565_split_axis_0, num_splits = x_565_split_num_splits_0, x = input_1143_cast_fp16)[name = string("x_565_split_cast_fp16")]; + tensor x_565_split_1_sigmoid_cast_fp16 = sigmoid(x = x_565_split_cast_fp16_1)[name = string("x_565_split_1_sigmoid_cast_fp16")]; + tensor x_565_cast_fp16 = mul(x = x_565_split_cast_fp16_0, y = x_565_split_1_sigmoid_cast_fp16)[name = string("x_565_cast_fp16")]; + tensor input_1145_cast_fp16 = select(a = var_44_to_fp16, b = x_565_cast_fp16, cond = var_575)[name = string("input_1145_cast_fp16")]; + bool new_x_87_interleave_0 = const()[name = string("new_x_87_interleave_0"), val = bool(false)]; + tensor new_x_87_cast_fp16 = concat(axis = var_59, interleave = new_x_87_interleave_0, values = (cache_87_cast_fp16, input_1145_cast_fp16))[name = string("new_x_87_cast_fp16")]; + tensor var_5061_begin_0 = const()[name = string("op_5061_begin_0"), val = tensor([0, 0, 7])]; + tensor var_5061_end_0 = const()[name = string("op_5061_end_0"), val = tensor([1, 1024, 15])]; + tensor var_5061_end_mask_0 = const()[name = string("op_5061_end_mask_0"), val = tensor([true, true, true])]; + tensor var_5061_cast_fp16 = slice_by_index(begin = var_5061_begin_0, end = var_5061_end_0, end_mask = var_5061_end_mask_0, x = new_x_87_cast_fp16)[name = string("op_5061_cast_fp16")]; + string x_567_pad_type_0 = const()[name = string("x_567_pad_type_0"), val = string("valid")]; + int32 x_567_groups_0 = const()[name = string("x_567_groups_0"), val = int32(1024)]; + tensor x_567_strides_0 = const()[name = string("x_567_strides_0"), val = tensor([1])]; + tensor x_567_pad_0 = const()[name = string("x_567_pad_0"), val = tensor([0, 0])]; + tensor x_567_dilations_0 = const()[name = string("x_567_dilations_0"), val = tensor([1])]; + tensor encoder_layers_21_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(446950336))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(446959616))))[name = string("encoder_layers_21_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_567_cast_fp16 = conv(dilations = x_567_dilations_0, groups = x_567_groups_0, pad = x_567_pad_0, pad_type = x_567_pad_type_0, strides = x_567_strides_0, weight = encoder_layers_21_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_87_cast_fp16)[name = string("x_567_cast_fp16")]; + tensor input_1147_perm_0 = const()[name = string("input_1147_perm_0"), val = tensor([0, 2, 1])]; + tensor x_569_axes_0 = const()[name = string("x_569_axes_0"), val = tensor([-1])]; + tensor encoder_layers_21_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_21_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(446961728)))]; + tensor encoder_layers_21_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_21_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(446963840)))]; + tensor input_1147_cast_fp16 = transpose(perm = input_1147_perm_0, x = x_567_cast_fp16)[name = string("transpose_167")]; + tensor x_569_cast_fp16 = layer_norm(axes = x_569_axes_0, beta = encoder_layers_21_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_21_conv_batch_norm_weight_to_fp16, x = input_1147_cast_fp16)[name = string("x_569_cast_fp16")]; + tensor input_1149_perm_0 = const()[name = string("input_1149_perm_0"), val = tensor([0, 2, 1])]; + tensor input_1149_cast_fp16 = transpose(perm = input_1149_perm_0, x = x_569_cast_fp16)[name = string("transpose_166")]; + tensor input_1151_cast_fp16 = silu(x = input_1149_cast_fp16)[name = string("input_1151_cast_fp16")]; + string x_571_pad_type_0 = const()[name = string("x_571_pad_type_0"), val = string("valid")]; + tensor x_571_strides_0 = const()[name = string("x_571_strides_0"), val = tensor([1])]; + tensor x_571_pad_0 = const()[name = string("x_571_pad_0"), val = tensor([0, 0])]; + tensor x_571_dilations_0 = const()[name = string("x_571_dilations_0"), val = tensor([1])]; + int32 x_571_groups_0 = const()[name = string("x_571_groups_0"), val = int32(1)]; + tensor encoder_layers_21_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(446965952))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(448014592))))[name = string("encoder_layers_21_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_571_cast_fp16 = conv(dilations = x_571_dilations_0, groups = x_571_groups_0, pad = x_571_pad_0, pad_type = x_571_pad_type_0, strides = x_571_strides_0, weight = encoder_layers_21_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_1151_cast_fp16)[name = string("x_571_cast_fp16")]; + tensor input_1153_perm_0 = const()[name = string("input_1153_perm_0"), val = tensor([0, 2, 1])]; + tensor input_1153_cast_fp16 = transpose(perm = input_1153_perm_0, x = x_571_cast_fp16)[name = string("transpose_165")]; + tensor input_1155_cast_fp16 = add(x = input_1139_cast_fp16, y = input_1153_cast_fp16)[name = string("input_1155_cast_fp16")]; + tensor input_1157_axes_0 = const()[name = string("input_1157_axes_0"), val = tensor([-1])]; + tensor encoder_layers_21_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_21_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(448016704)))]; + tensor encoder_layers_21_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_21_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(448018816)))]; + tensor input_1157_cast_fp16 = layer_norm(axes = input_1157_axes_0, beta = encoder_layers_21_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_21_norm_feed_forward2_weight_to_fp16, x = input_1155_cast_fp16)[name = string("input_1157_cast_fp16")]; + tensor encoder_layers_21_feed_forward2_linear1_weight_to_fp16 = const()[name = string("encoder_layers_21_feed_forward2_linear1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(448020928)))]; + tensor encoder_layers_21_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_21_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(456409600)))]; + tensor linear_197_cast_fp16 = linear(bias = encoder_layers_21_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_21_feed_forward2_linear1_weight_to_fp16, x = input_1157_cast_fp16)[name = string("linear_197_cast_fp16")]; + tensor input_1161_cast_fp16 = silu(x = linear_197_cast_fp16)[name = string("input_1161_cast_fp16")]; + tensor encoder_layers_21_feed_forward2_linear2_weight_to_fp16 = const()[name = string("encoder_layers_21_feed_forward2_linear2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(456417856)))]; + tensor encoder_layers_21_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_21_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(464806528)))]; + tensor linear_198_cast_fp16 = linear(bias = encoder_layers_21_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_21_feed_forward2_linear2_weight_to_fp16, x = input_1161_cast_fp16)[name = string("linear_198_cast_fp16")]; + fp16 var_5104_to_fp16 = const()[name = string("op_5104_to_fp16"), val = fp16(0x1p-1)]; + tensor var_5105_cast_fp16 = mul(x = linear_198_cast_fp16, y = var_5104_to_fp16)[name = string("op_5105_cast_fp16")]; + tensor input_1167_cast_fp16 = add(x = input_1155_cast_fp16, y = var_5105_cast_fp16)[name = string("input_1167_cast_fp16")]; + tensor input_1169_axes_0 = const()[name = string("input_1169_axes_0"), val = tensor([-1])]; + tensor encoder_layers_21_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_21_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(464808640)))]; + tensor encoder_layers_21_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_21_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(464810752)))]; + tensor input_1169_cast_fp16 = layer_norm(axes = input_1169_axes_0, beta = encoder_layers_21_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_21_norm_out_weight_to_fp16, x = input_1167_cast_fp16)[name = string("input_1169_cast_fp16")]; + tensor cache_89_begin_0 = const()[name = string("cache_89_begin_0"), val = tensor([22, 0, 0, 0])]; + tensor cache_89_end_0 = const()[name = string("cache_89_end_0"), val = tensor([23, 1, 42, 1024])]; + tensor cache_89_end_mask_0 = const()[name = string("cache_89_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_89_squeeze_mask_0 = const()[name = string("cache_89_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_89_cast_fp16 = slice_by_index(begin = cache_89_begin_0, end = cache_89_end_0, end_mask = cache_89_end_mask_0, squeeze_mask = cache_89_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_89_cast_fp16")]; + tensor cache_91_begin_0 = const()[name = string("cache_91_begin_0"), val = tensor([22, 0, 0, 0])]; + tensor cache_91_end_0 = const()[name = string("cache_91_end_0"), val = tensor([23, 1, 1024, 8])]; + tensor cache_91_end_mask_0 = const()[name = string("cache_91_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_91_squeeze_mask_0 = const()[name = string("cache_91_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_91_cast_fp16 = slice_by_index(begin = cache_91_begin_0, end = cache_91_end_0, end_mask = cache_91_end_mask_0, squeeze_mask = cache_91_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_91_cast_fp16")]; + tensor input_1171_axes_0 = const()[name = string("input_1171_axes_0"), val = tensor([-1])]; + tensor encoder_layers_22_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_22_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(464812864)))]; + tensor encoder_layers_22_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_22_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(464814976)))]; + tensor input_1171_cast_fp16 = layer_norm(axes = input_1171_axes_0, beta = encoder_layers_22_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_22_norm_feed_forward1_weight_to_fp16, x = input_1169_cast_fp16)[name = string("input_1171_cast_fp16")]; + tensor encoder_layers_22_feed_forward1_linear1_weight_to_fp16 = const()[name = string("encoder_layers_22_feed_forward1_linear1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(464817088)))]; + tensor encoder_layers_22_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_22_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(473205760)))]; + tensor linear_199_cast_fp16 = linear(bias = encoder_layers_22_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_22_feed_forward1_linear1_weight_to_fp16, x = input_1171_cast_fp16)[name = string("linear_199_cast_fp16")]; + tensor input_1175_cast_fp16 = silu(x = linear_199_cast_fp16)[name = string("input_1175_cast_fp16")]; + tensor encoder_layers_22_feed_forward1_linear2_weight_to_fp16 = const()[name = string("encoder_layers_22_feed_forward1_linear2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(473214016)))]; + tensor encoder_layers_22_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_22_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481602688)))]; + tensor linear_200_cast_fp16 = linear(bias = encoder_layers_22_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_22_feed_forward1_linear2_weight_to_fp16, x = input_1175_cast_fp16)[name = string("linear_200_cast_fp16")]; + fp16 var_5141_to_fp16 = const()[name = string("op_5141_to_fp16"), val = fp16(0x1p-1)]; + tensor var_5142_cast_fp16 = mul(x = linear_200_cast_fp16, y = var_5141_to_fp16)[name = string("op_5142_cast_fp16")]; + tensor input_1181_cast_fp16 = add(x = input_1169_cast_fp16, y = var_5142_cast_fp16)[name = string("input_1181_cast_fp16")]; + tensor key_45_axes_0 = const()[name = string("key_45_axes_0"), val = tensor([-1])]; + tensor encoder_layers_22_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_22_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481604800)))]; + tensor encoder_layers_22_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_22_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481606912)))]; + tensor key_45_cast_fp16 = layer_norm(axes = key_45_axes_0, beta = encoder_layers_22_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_22_norm_self_att_weight_to_fp16, x = input_1181_cast_fp16)[name = string("key_45_cast_fp16")]; + bool input_1183_interleave_0 = const()[name = string("input_1183_interleave_0"), val = bool(false)]; + tensor input_1183_cast_fp16 = concat(axis = var_68, interleave = input_1183_interleave_0, values = (cache_89_cast_fp16, key_45_cast_fp16))[name = string("input_1183_cast_fp16")]; + tensor var_5164_begin_0 = const()[name = string("op_5164_begin_0"), val = tensor([0, 7, 0])]; + tensor var_5164_end_0 = const()[name = string("op_5164_end_0"), val = tensor([1, 42, 1024])]; + tensor var_5164_end_mask_0 = const()[name = string("op_5164_end_mask_0"), val = tensor([true, true, true])]; + tensor var_5164_cast_fp16 = slice_by_index(begin = var_5164_begin_0, end = var_5164_end_0, end_mask = var_5164_end_mask_0, x = cache_89_cast_fp16)[name = string("op_5164_cast_fp16")]; + bool var_5170_interleave_0 = const()[name = string("op_5170_interleave_0"), val = bool(false)]; + tensor var_5170_cast_fp16 = concat(axis = var_68, interleave = var_5170_interleave_0, values = (var_5164_cast_fp16, key_45_cast_fp16))[name = string("op_5170_cast_fp16")]; + tensor encoder_layers_22_self_attn_linear_q_weight_to_fp16 = const()[name = string("encoder_layers_22_self_attn_linear_q_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481609024)))]; + tensor encoder_layers_22_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_22_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(483706240)))]; + tensor linear_201_cast_fp16 = linear(bias = encoder_layers_22_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_22_self_attn_linear_q_weight_to_fp16, x = key_45_cast_fp16)[name = string("linear_201_cast_fp16")]; + tensor var_5175 = const()[name = string("op_5175"), val = tensor([1, -1, 8, 128])]; + tensor q_133_cast_fp16 = reshape(shape = var_5175, x = linear_201_cast_fp16)[name = string("q_133_cast_fp16")]; + tensor encoder_layers_22_self_attn_linear_k_weight_to_fp16 = const()[name = string("encoder_layers_22_self_attn_linear_k_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(483708352)))]; + tensor encoder_layers_22_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_22_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(485805568)))]; + tensor linear_202_cast_fp16 = linear(bias = encoder_layers_22_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_22_self_attn_linear_k_weight_to_fp16, x = input_1183_cast_fp16)[name = string("linear_202_cast_fp16")]; + tensor var_5180 = const()[name = string("op_5180"), val = tensor([1, -1, 8, 128])]; + tensor k_89_cast_fp16 = reshape(shape = var_5180, x = linear_202_cast_fp16)[name = string("k_89_cast_fp16")]; + tensor encoder_layers_22_self_attn_linear_v_weight_to_fp16 = const()[name = string("encoder_layers_22_self_attn_linear_v_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(485807680)))]; + tensor encoder_layers_22_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_22_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(487904896)))]; + tensor linear_203_cast_fp16 = linear(bias = encoder_layers_22_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_22_self_attn_linear_v_weight_to_fp16, x = input_1183_cast_fp16)[name = string("linear_203_cast_fp16")]; + tensor var_5185 = const()[name = string("op_5185"), val = tensor([1, -1, 8, 128])]; + tensor v_45_cast_fp16 = reshape(shape = var_5185, x = linear_203_cast_fp16)[name = string("v_45_cast_fp16")]; + tensor value_53_perm_0 = const()[name = string("value_53_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_22_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_22_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(487907008)))]; + tensor var_5198_cast_fp16 = add(x = q_133_cast_fp16, y = encoder_layers_22_self_attn_pos_bias_u_to_fp16)[name = string("op_5198_cast_fp16")]; + tensor encoder_layers_22_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_22_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(487909120)))]; + tensor var_5200_cast_fp16 = add(x = q_133_cast_fp16, y = encoder_layers_22_self_attn_pos_bias_v_to_fp16)[name = string("op_5200_cast_fp16")]; + tensor q_with_bias_v_45_perm_0 = const()[name = string("q_with_bias_v_45_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_579_transpose_x_0 = const()[name = string("x_579_transpose_x_0"), val = bool(false)]; + bool x_579_transpose_y_0 = const()[name = string("x_579_transpose_y_0"), val = bool(false)]; + tensor op_5202_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(487911232))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(488010624))))[name = string("op_5202_to_fp16_quantized")]; + tensor q_with_bias_v_45_cast_fp16 = transpose(perm = q_with_bias_v_45_perm_0, x = var_5200_cast_fp16)[name = string("transpose_164")]; + tensor x_579_cast_fp16 = matmul(transpose_x = x_579_transpose_x_0, transpose_y = x_579_transpose_y_0, x = q_with_bias_v_45_cast_fp16, y = op_5202_to_fp16_quantized)[name = string("x_579_cast_fp16")]; + tensor x_581_pad_0 = const()[name = string("x_581_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_581_mode_0 = const()[name = string("x_581_mode_0"), val = string("constant")]; + fp16 const_365_to_fp16 = const()[name = string("const_365_to_fp16"), val = fp16(0x0p+0)]; + tensor x_581_cast_fp16 = pad(constant_val = const_365_to_fp16, mode = x_581_mode_0, pad = x_581_pad_0, x = x_579_cast_fp16)[name = string("x_581_cast_fp16")]; + tensor var_5210 = const()[name = string("op_5210"), val = tensor([1, 8, -1, 7])]; + tensor x_583_cast_fp16 = reshape(shape = var_5210, x = x_581_cast_fp16)[name = string("x_583_cast_fp16")]; + tensor var_5214_begin_0 = const()[name = string("op_5214_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_5214_end_0 = const()[name = string("op_5214_end_0"), val = tensor([1, 8, 98, 7])]; + tensor var_5214_end_mask_0 = const()[name = string("op_5214_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_5214_cast_fp16 = slice_by_index(begin = var_5214_begin_0, end = var_5214_end_0, end_mask = var_5214_end_mask_0, x = x_583_cast_fp16)[name = string("op_5214_cast_fp16")]; + tensor var_5215 = const()[name = string("op_5215"), val = tensor([1, 8, 7, 97])]; + tensor matrix_bd_89_cast_fp16 = reshape(shape = var_5215, x = var_5214_cast_fp16)[name = string("matrix_bd_89_cast_fp16")]; + bool matrix_ac_45_transpose_x_0 = const()[name = string("matrix_ac_45_transpose_x_0"), val = bool(false)]; + bool matrix_ac_45_transpose_y_0 = const()[name = string("matrix_ac_45_transpose_y_0"), val = bool(false)]; + tensor transpose_140_perm_0 = const()[name = string("transpose_140_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_141_perm_0 = const()[name = string("transpose_141_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_141 = transpose(perm = transpose_141_perm_0, x = k_89_cast_fp16)[name = string("transpose_162")]; + tensor transpose_140 = transpose(perm = transpose_140_perm_0, x = var_5198_cast_fp16)[name = string("transpose_163")]; + tensor matrix_ac_45_cast_fp16 = matmul(transpose_x = matrix_ac_45_transpose_x_0, transpose_y = matrix_ac_45_transpose_y_0, x = transpose_140, y = transpose_141)[name = string("matrix_ac_45_cast_fp16")]; + tensor matrix_bd_91_begin_0 = const()[name = string("matrix_bd_91_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_91_end_0 = const()[name = string("matrix_bd_91_end_0"), val = tensor([1, 8, 7, 49])]; + tensor matrix_bd_91_end_mask_0 = const()[name = string("matrix_bd_91_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_91_cast_fp16 = slice_by_index(begin = matrix_bd_91_begin_0, end = matrix_bd_91_end_0, end_mask = matrix_bd_91_end_mask_0, x = matrix_bd_89_cast_fp16)[name = string("matrix_bd_91_cast_fp16")]; + tensor var_5224_cast_fp16 = add(x = matrix_ac_45_cast_fp16, y = matrix_bd_91_cast_fp16)[name = string("op_5224_cast_fp16")]; + fp16 _inversed_scores_89_y_0_to_fp16 = const()[name = string("_inversed_scores_89_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_89_cast_fp16 = mul(x = var_5224_cast_fp16, y = _inversed_scores_89_y_0_to_fp16)[name = string("_inversed_scores_89_cast_fp16")]; + tensor scores_91_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_89_cast_fp16, cond = mask_11)[name = string("scores_91_cast_fp16")]; + tensor var_5230_cast_fp16 = softmax(axis = var_59, x = scores_91_cast_fp16)[name = string("op_5230_cast_fp16")]; + tensor input_1185_cast_fp16 = select(a = var_44_to_fp16, b = var_5230_cast_fp16, cond = mask_11)[name = string("input_1185_cast_fp16")]; + bool x_585_transpose_x_0 = const()[name = string("x_585_transpose_x_0"), val = bool(false)]; + bool x_585_transpose_y_0 = const()[name = string("x_585_transpose_y_0"), val = bool(false)]; + tensor value_53_cast_fp16 = transpose(perm = value_53_perm_0, x = v_45_cast_fp16)[name = string("transpose_161")]; + tensor x_585_cast_fp16 = matmul(transpose_x = x_585_transpose_x_0, transpose_y = x_585_transpose_y_0, x = input_1185_cast_fp16, y = value_53_cast_fp16)[name = string("x_585_cast_fp16")]; + tensor var_5234_perm_0 = const()[name = string("op_5234_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_5235 = const()[name = string("op_5235"), val = tensor([1, -1, 1024])]; + tensor var_5234_cast_fp16 = transpose(perm = var_5234_perm_0, x = x_585_cast_fp16)[name = string("transpose_160")]; + tensor input_1187_cast_fp16 = reshape(shape = var_5235, x = var_5234_cast_fp16)[name = string("input_1187_cast_fp16")]; + tensor encoder_layers_22_self_attn_linear_out_weight_to_fp16 = const()[name = string("encoder_layers_22_self_attn_linear_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(488010944)))]; + tensor encoder_layers_22_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_22_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(490108160)))]; + tensor linear_205_cast_fp16 = linear(bias = encoder_layers_22_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_22_self_attn_linear_out_weight_to_fp16, x = input_1187_cast_fp16)[name = string("linear_205_cast_fp16")]; + tensor input_1191_cast_fp16 = add(x = input_1181_cast_fp16, y = linear_205_cast_fp16)[name = string("input_1191_cast_fp16")]; + tensor x_589_axes_0 = const()[name = string("x_589_axes_0"), val = tensor([-1])]; + tensor encoder_layers_22_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_22_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(490110272)))]; + tensor encoder_layers_22_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_22_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(490112384)))]; + tensor x_589_cast_fp16 = layer_norm(axes = x_589_axes_0, beta = encoder_layers_22_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_22_norm_conv_weight_to_fp16, x = input_1191_cast_fp16)[name = string("x_589_cast_fp16")]; + tensor input_1193_perm_0 = const()[name = string("input_1193_perm_0"), val = tensor([0, 2, 1])]; + string input_1195_pad_type_0 = const()[name = string("input_1195_pad_type_0"), val = string("valid")]; + tensor input_1195_strides_0 = const()[name = string("input_1195_strides_0"), val = tensor([1])]; + tensor input_1195_pad_0 = const()[name = string("input_1195_pad_0"), val = tensor([0, 0])]; + tensor input_1195_dilations_0 = const()[name = string("input_1195_dilations_0"), val = tensor([1])]; + int32 input_1195_groups_0 = const()[name = string("input_1195_groups_0"), val = int32(1)]; + tensor encoder_layers_22_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(490114496))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(492211712))))[name = string("encoder_layers_22_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_1193_cast_fp16 = transpose(perm = input_1193_perm_0, x = x_589_cast_fp16)[name = string("transpose_159")]; + tensor input_1195_cast_fp16 = conv(dilations = input_1195_dilations_0, groups = input_1195_groups_0, pad = input_1195_pad_0, pad_type = input_1195_pad_type_0, strides = input_1195_strides_0, weight = encoder_layers_22_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_1193_cast_fp16)[name = string("input_1195_cast_fp16")]; + int32 x_591_split_num_splits_0 = const()[name = string("x_591_split_num_splits_0"), val = int32(2)]; + int32 x_591_split_axis_0 = const()[name = string("x_591_split_axis_0"), val = int32(1)]; + tensor x_591_split_cast_fp16_0, tensor x_591_split_cast_fp16_1 = split(axis = x_591_split_axis_0, num_splits = x_591_split_num_splits_0, x = input_1195_cast_fp16)[name = string("x_591_split_cast_fp16")]; + tensor x_591_split_1_sigmoid_cast_fp16 = sigmoid(x = x_591_split_cast_fp16_1)[name = string("x_591_split_1_sigmoid_cast_fp16")]; + tensor x_591_cast_fp16 = mul(x = x_591_split_cast_fp16_0, y = x_591_split_1_sigmoid_cast_fp16)[name = string("x_591_cast_fp16")]; + tensor input_1197_cast_fp16 = select(a = var_44_to_fp16, b = x_591_cast_fp16, cond = var_575)[name = string("input_1197_cast_fp16")]; + bool new_x_91_interleave_0 = const()[name = string("new_x_91_interleave_0"), val = bool(false)]; + tensor new_x_91_cast_fp16 = concat(axis = var_59, interleave = new_x_91_interleave_0, values = (cache_91_cast_fp16, input_1197_cast_fp16))[name = string("new_x_91_cast_fp16")]; + tensor var_5274_begin_0 = const()[name = string("op_5274_begin_0"), val = tensor([0, 0, 7])]; + tensor var_5274_end_0 = const()[name = string("op_5274_end_0"), val = tensor([1, 1024, 15])]; + tensor var_5274_end_mask_0 = const()[name = string("op_5274_end_mask_0"), val = tensor([true, true, true])]; + tensor var_5274_cast_fp16 = slice_by_index(begin = var_5274_begin_0, end = var_5274_end_0, end_mask = var_5274_end_mask_0, x = new_x_91_cast_fp16)[name = string("op_5274_cast_fp16")]; + string x_593_pad_type_0 = const()[name = string("x_593_pad_type_0"), val = string("valid")]; + int32 x_593_groups_0 = const()[name = string("x_593_groups_0"), val = int32(1024)]; + tensor x_593_strides_0 = const()[name = string("x_593_strides_0"), val = tensor([1])]; + tensor x_593_pad_0 = const()[name = string("x_593_pad_0"), val = tensor([0, 0])]; + tensor x_593_dilations_0 = const()[name = string("x_593_dilations_0"), val = tensor([1])]; + tensor encoder_layers_22_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(492215872))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(492225152))))[name = string("encoder_layers_22_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_593_cast_fp16 = conv(dilations = x_593_dilations_0, groups = x_593_groups_0, pad = x_593_pad_0, pad_type = x_593_pad_type_0, strides = x_593_strides_0, weight = encoder_layers_22_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_91_cast_fp16)[name = string("x_593_cast_fp16")]; + tensor input_1199_perm_0 = const()[name = string("input_1199_perm_0"), val = tensor([0, 2, 1])]; + tensor x_595_axes_0 = const()[name = string("x_595_axes_0"), val = tensor([-1])]; + tensor encoder_layers_22_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_22_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(492227264)))]; + tensor encoder_layers_22_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_22_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(492229376)))]; + tensor input_1199_cast_fp16 = transpose(perm = input_1199_perm_0, x = x_593_cast_fp16)[name = string("transpose_158")]; + tensor x_595_cast_fp16 = layer_norm(axes = x_595_axes_0, beta = encoder_layers_22_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_22_conv_batch_norm_weight_to_fp16, x = input_1199_cast_fp16)[name = string("x_595_cast_fp16")]; + tensor input_1201_perm_0 = const()[name = string("input_1201_perm_0"), val = tensor([0, 2, 1])]; + tensor input_1201_cast_fp16 = transpose(perm = input_1201_perm_0, x = x_595_cast_fp16)[name = string("transpose_157")]; + tensor input_1203_cast_fp16 = silu(x = input_1201_cast_fp16)[name = string("input_1203_cast_fp16")]; + string x_597_pad_type_0 = const()[name = string("x_597_pad_type_0"), val = string("valid")]; + tensor x_597_strides_0 = const()[name = string("x_597_strides_0"), val = tensor([1])]; + tensor x_597_pad_0 = const()[name = string("x_597_pad_0"), val = tensor([0, 0])]; + tensor x_597_dilations_0 = const()[name = string("x_597_dilations_0"), val = tensor([1])]; + int32 x_597_groups_0 = const()[name = string("x_597_groups_0"), val = int32(1)]; + tensor encoder_layers_22_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(492231488))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(493280128))))[name = string("encoder_layers_22_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_597_cast_fp16 = conv(dilations = x_597_dilations_0, groups = x_597_groups_0, pad = x_597_pad_0, pad_type = x_597_pad_type_0, strides = x_597_strides_0, weight = encoder_layers_22_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_1203_cast_fp16)[name = string("x_597_cast_fp16")]; + tensor input_1205_perm_0 = const()[name = string("input_1205_perm_0"), val = tensor([0, 2, 1])]; + tensor input_1205_cast_fp16 = transpose(perm = input_1205_perm_0, x = x_597_cast_fp16)[name = string("transpose_156")]; + tensor input_1207_cast_fp16 = add(x = input_1191_cast_fp16, y = input_1205_cast_fp16)[name = string("input_1207_cast_fp16")]; + tensor input_1209_axes_0 = const()[name = string("input_1209_axes_0"), val = tensor([-1])]; + tensor encoder_layers_22_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_22_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(493282240)))]; + tensor encoder_layers_22_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_22_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(493284352)))]; + tensor input_1209_cast_fp16 = layer_norm(axes = input_1209_axes_0, beta = encoder_layers_22_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_22_norm_feed_forward2_weight_to_fp16, x = input_1207_cast_fp16)[name = string("input_1209_cast_fp16")]; + tensor encoder_layers_22_feed_forward2_linear1_weight_to_fp16 = const()[name = string("encoder_layers_22_feed_forward2_linear1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(493286464)))]; + tensor encoder_layers_22_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_22_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(501675136)))]; + tensor linear_206_cast_fp16 = linear(bias = encoder_layers_22_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_22_feed_forward2_linear1_weight_to_fp16, x = input_1209_cast_fp16)[name = string("linear_206_cast_fp16")]; + tensor input_1213_cast_fp16 = silu(x = linear_206_cast_fp16)[name = string("input_1213_cast_fp16")]; + tensor encoder_layers_22_feed_forward2_linear2_weight_to_fp16 = const()[name = string("encoder_layers_22_feed_forward2_linear2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(501683392)))]; + tensor encoder_layers_22_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_22_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(510072064)))]; + tensor linear_207_cast_fp16 = linear(bias = encoder_layers_22_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_22_feed_forward2_linear2_weight_to_fp16, x = input_1213_cast_fp16)[name = string("linear_207_cast_fp16")]; + fp16 var_5317_to_fp16 = const()[name = string("op_5317_to_fp16"), val = fp16(0x1p-1)]; + tensor var_5318_cast_fp16 = mul(x = linear_207_cast_fp16, y = var_5317_to_fp16)[name = string("op_5318_cast_fp16")]; + tensor input_1219_cast_fp16 = add(x = input_1207_cast_fp16, y = var_5318_cast_fp16)[name = string("input_1219_cast_fp16")]; + tensor input_1221_axes_0 = const()[name = string("input_1221_axes_0"), val = tensor([-1])]; + tensor encoder_layers_22_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_22_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(510074176)))]; + tensor encoder_layers_22_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_22_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(510076288)))]; + tensor input_1221_cast_fp16 = layer_norm(axes = input_1221_axes_0, beta = encoder_layers_22_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_22_norm_out_weight_to_fp16, x = input_1219_cast_fp16)[name = string("input_1221_cast_fp16")]; + tensor cache_93_begin_0 = const()[name = string("cache_93_begin_0"), val = tensor([23, 0, 0, 0])]; + tensor cache_93_end_0 = const()[name = string("cache_93_end_0"), val = tensor([24, 1, 42, 1024])]; + tensor cache_93_end_mask_0 = const()[name = string("cache_93_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_93_squeeze_mask_0 = const()[name = string("cache_93_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_93_cast_fp16 = slice_by_index(begin = cache_93_begin_0, end = cache_93_end_0, end_mask = cache_93_end_mask_0, squeeze_mask = cache_93_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_93_cast_fp16")]; + tensor cache_begin_0 = const()[name = string("cache_begin_0"), val = tensor([23, 0, 0, 0])]; + tensor cache_end_0 = const()[name = string("cache_end_0"), val = tensor([24, 1, 1024, 8])]; + tensor cache_end_mask_0 = const()[name = string("cache_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_squeeze_mask_0 = const()[name = string("cache_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_cast_fp16 = slice_by_index(begin = cache_begin_0, end = cache_end_0, end_mask = cache_end_mask_0, squeeze_mask = cache_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_cast_fp16")]; + tensor input_1223_axes_0 = const()[name = string("input_1223_axes_0"), val = tensor([-1])]; + tensor encoder_layers_23_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_23_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(510078400)))]; + tensor encoder_layers_23_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_23_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(510080512)))]; + tensor input_1223_cast_fp16 = layer_norm(axes = input_1223_axes_0, beta = encoder_layers_23_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_23_norm_feed_forward1_weight_to_fp16, x = input_1221_cast_fp16)[name = string("input_1223_cast_fp16")]; + tensor encoder_layers_23_feed_forward1_linear1_weight_to_fp16 = const()[name = string("encoder_layers_23_feed_forward1_linear1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(510082624)))]; + tensor encoder_layers_23_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_23_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(518471296)))]; + tensor linear_208_cast_fp16 = linear(bias = encoder_layers_23_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_23_feed_forward1_linear1_weight_to_fp16, x = input_1223_cast_fp16)[name = string("linear_208_cast_fp16")]; + tensor input_1227_cast_fp16 = silu(x = linear_208_cast_fp16)[name = string("input_1227_cast_fp16")]; + tensor encoder_layers_23_feed_forward1_linear2_weight_to_fp16 = const()[name = string("encoder_layers_23_feed_forward1_linear2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(518479552)))]; + tensor encoder_layers_23_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_23_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(526868224)))]; + tensor linear_209_cast_fp16 = linear(bias = encoder_layers_23_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_23_feed_forward1_linear2_weight_to_fp16, x = input_1227_cast_fp16)[name = string("linear_209_cast_fp16")]; + fp16 var_5354_to_fp16 = const()[name = string("op_5354_to_fp16"), val = fp16(0x1p-1)]; + tensor var_5355_cast_fp16 = mul(x = linear_209_cast_fp16, y = var_5354_to_fp16)[name = string("op_5355_cast_fp16")]; + tensor input_1233_cast_fp16 = add(x = input_1221_cast_fp16, y = var_5355_cast_fp16)[name = string("input_1233_cast_fp16")]; + tensor key_axes_0 = const()[name = string("key_axes_0"), val = tensor([-1])]; + tensor encoder_layers_23_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_23_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(526870336)))]; + tensor encoder_layers_23_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_23_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(526872448)))]; + tensor key_cast_fp16 = layer_norm(axes = key_axes_0, beta = encoder_layers_23_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_23_norm_self_att_weight_to_fp16, x = input_1233_cast_fp16)[name = string("key_cast_fp16")]; + bool input_1235_interleave_0 = const()[name = string("input_1235_interleave_0"), val = bool(false)]; + tensor input_1235_cast_fp16 = concat(axis = var_68, interleave = input_1235_interleave_0, values = (cache_93_cast_fp16, key_cast_fp16))[name = string("input_1235_cast_fp16")]; + tensor var_5377_begin_0 = const()[name = string("op_5377_begin_0"), val = tensor([0, 7, 0])]; + tensor var_5377_end_0 = const()[name = string("op_5377_end_0"), val = tensor([1, 42, 1024])]; + tensor var_5377_end_mask_0 = const()[name = string("op_5377_end_mask_0"), val = tensor([true, true, true])]; + tensor var_5377_cast_fp16 = slice_by_index(begin = var_5377_begin_0, end = var_5377_end_0, end_mask = var_5377_end_mask_0, x = cache_93_cast_fp16)[name = string("op_5377_cast_fp16")]; + bool cache_last_channel_cur_interleave_0 = const()[name = string("cache_last_channel_cur_interleave_0"), val = bool(false)]; + tensor cache_last_channel_cur_cast_fp16 = concat(axis = var_68, interleave = cache_last_channel_cur_interleave_0, values = (var_5377_cast_fp16, key_cast_fp16))[name = string("cache_last_channel_cur_cast_fp16")]; + tensor encoder_layers_23_self_attn_linear_q_weight_to_fp16 = const()[name = string("encoder_layers_23_self_attn_linear_q_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(526874560)))]; + tensor encoder_layers_23_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_23_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(528971776)))]; + tensor linear_210_cast_fp16 = linear(bias = encoder_layers_23_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_23_self_attn_linear_q_weight_to_fp16, x = key_cast_fp16)[name = string("linear_210_cast_fp16")]; + tensor var_5388 = const()[name = string("op_5388"), val = tensor([1, -1, 8, 128])]; + tensor q_139_cast_fp16 = reshape(shape = var_5388, x = linear_210_cast_fp16)[name = string("q_139_cast_fp16")]; + tensor encoder_layers_23_self_attn_linear_k_weight_to_fp16 = const()[name = string("encoder_layers_23_self_attn_linear_k_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(528973888)))]; + tensor encoder_layers_23_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_23_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(531071104)))]; + tensor linear_211_cast_fp16 = linear(bias = encoder_layers_23_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_23_self_attn_linear_k_weight_to_fp16, x = input_1235_cast_fp16)[name = string("linear_211_cast_fp16")]; + tensor var_5393 = const()[name = string("op_5393"), val = tensor([1, -1, 8, 128])]; + tensor k_93_cast_fp16 = reshape(shape = var_5393, x = linear_211_cast_fp16)[name = string("k_93_cast_fp16")]; + tensor encoder_layers_23_self_attn_linear_v_weight_to_fp16 = const()[name = string("encoder_layers_23_self_attn_linear_v_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(531073216)))]; + tensor encoder_layers_23_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_23_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(533170432)))]; + tensor linear_212_cast_fp16 = linear(bias = encoder_layers_23_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_23_self_attn_linear_v_weight_to_fp16, x = input_1235_cast_fp16)[name = string("linear_212_cast_fp16")]; + tensor var_5398 = const()[name = string("op_5398"), val = tensor([1, -1, 8, 128])]; + tensor v_cast_fp16 = reshape(shape = var_5398, x = linear_212_cast_fp16)[name = string("v_cast_fp16")]; + tensor value_perm_0 = const()[name = string("value_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_23_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_23_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(533172544)))]; + tensor var_5411_cast_fp16 = add(x = q_139_cast_fp16, y = encoder_layers_23_self_attn_pos_bias_u_to_fp16)[name = string("op_5411_cast_fp16")]; + tensor encoder_layers_23_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_23_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(533174656)))]; + tensor var_5413_cast_fp16 = add(x = q_139_cast_fp16, y = encoder_layers_23_self_attn_pos_bias_v_to_fp16)[name = string("op_5413_cast_fp16")]; + tensor q_with_bias_v_perm_0 = const()[name = string("q_with_bias_v_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_605_transpose_x_0 = const()[name = string("x_605_transpose_x_0"), val = bool(false)]; + bool x_605_transpose_y_0 = const()[name = string("x_605_transpose_y_0"), val = bool(false)]; + tensor op_5415_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(533176768))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(533276160))))[name = string("op_5415_to_fp16_quantized")]; + tensor q_with_bias_v_cast_fp16 = transpose(perm = q_with_bias_v_perm_0, x = var_5413_cast_fp16)[name = string("transpose_155")]; + tensor x_605_cast_fp16 = matmul(transpose_x = x_605_transpose_x_0, transpose_y = x_605_transpose_y_0, x = q_with_bias_v_cast_fp16, y = op_5415_to_fp16_quantized)[name = string("x_605_cast_fp16")]; + tensor x_607_pad_0 = const()[name = string("x_607_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_607_mode_0 = const()[name = string("x_607_mode_0"), val = string("constant")]; + fp16 const_378_to_fp16 = const()[name = string("const_378_to_fp16"), val = fp16(0x0p+0)]; + tensor x_607_cast_fp16 = pad(constant_val = const_378_to_fp16, mode = x_607_mode_0, pad = x_607_pad_0, x = x_605_cast_fp16)[name = string("x_607_cast_fp16")]; + tensor var_5423 = const()[name = string("op_5423"), val = tensor([1, 8, -1, 7])]; + tensor x_609_cast_fp16 = reshape(shape = var_5423, x = x_607_cast_fp16)[name = string("x_609_cast_fp16")]; + tensor var_5427_begin_0 = const()[name = string("op_5427_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_5427_end_0 = const()[name = string("op_5427_end_0"), val = tensor([1, 8, 98, 7])]; + tensor var_5427_end_mask_0 = const()[name = string("op_5427_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_5427_cast_fp16 = slice_by_index(begin = var_5427_begin_0, end = var_5427_end_0, end_mask = var_5427_end_mask_0, x = x_609_cast_fp16)[name = string("op_5427_cast_fp16")]; + tensor var_5428 = const()[name = string("op_5428"), val = tensor([1, 8, 7, 97])]; + tensor matrix_bd_93_cast_fp16 = reshape(shape = var_5428, x = var_5427_cast_fp16)[name = string("matrix_bd_93_cast_fp16")]; + bool matrix_ac_transpose_x_0 = const()[name = string("matrix_ac_transpose_x_0"), val = bool(false)]; + bool matrix_ac_transpose_y_0 = const()[name = string("matrix_ac_transpose_y_0"), val = bool(false)]; + tensor transpose_142_perm_0 = const()[name = string("transpose_142_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_143_perm_0 = const()[name = string("transpose_143_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_143 = transpose(perm = transpose_143_perm_0, x = k_93_cast_fp16)[name = string("transpose_153")]; + tensor transpose_142 = transpose(perm = transpose_142_perm_0, x = var_5411_cast_fp16)[name = string("transpose_154")]; + tensor matrix_ac_cast_fp16 = matmul(transpose_x = matrix_ac_transpose_x_0, transpose_y = matrix_ac_transpose_y_0, x = transpose_142, y = transpose_143)[name = string("matrix_ac_cast_fp16")]; + tensor matrix_bd_begin_0 = const()[name = string("matrix_bd_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_end_0 = const()[name = string("matrix_bd_end_0"), val = tensor([1, 8, 7, 49])]; + tensor matrix_bd_end_mask_0 = const()[name = string("matrix_bd_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_cast_fp16 = slice_by_index(begin = matrix_bd_begin_0, end = matrix_bd_end_0, end_mask = matrix_bd_end_mask_0, x = matrix_bd_93_cast_fp16)[name = string("matrix_bd_cast_fp16")]; + tensor var_5437_cast_fp16 = add(x = matrix_ac_cast_fp16, y = matrix_bd_cast_fp16)[name = string("op_5437_cast_fp16")]; + fp16 _inversed_scores_93_y_0_to_fp16 = const()[name = string("_inversed_scores_93_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_93_cast_fp16 = mul(x = var_5437_cast_fp16, y = _inversed_scores_93_y_0_to_fp16)[name = string("_inversed_scores_93_cast_fp16")]; + tensor scores_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_93_cast_fp16, cond = mask_11)[name = string("scores_cast_fp16")]; + tensor var_5443_cast_fp16 = softmax(axis = var_59, x = scores_cast_fp16)[name = string("op_5443_cast_fp16")]; + tensor input_1237_cast_fp16 = select(a = var_44_to_fp16, b = var_5443_cast_fp16, cond = mask_11)[name = string("input_1237_cast_fp16")]; + bool x_611_transpose_x_0 = const()[name = string("x_611_transpose_x_0"), val = bool(false)]; + bool x_611_transpose_y_0 = const()[name = string("x_611_transpose_y_0"), val = bool(false)]; + tensor value_cast_fp16 = transpose(perm = value_perm_0, x = v_cast_fp16)[name = string("transpose_152")]; + tensor x_611_cast_fp16 = matmul(transpose_x = x_611_transpose_x_0, transpose_y = x_611_transpose_y_0, x = input_1237_cast_fp16, y = value_cast_fp16)[name = string("x_611_cast_fp16")]; + tensor var_5447_perm_0 = const()[name = string("op_5447_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_5448 = const()[name = string("op_5448"), val = tensor([1, -1, 1024])]; + tensor var_5447_cast_fp16 = transpose(perm = var_5447_perm_0, x = x_611_cast_fp16)[name = string("transpose_151")]; + tensor input_1239_cast_fp16 = reshape(shape = var_5448, x = var_5447_cast_fp16)[name = string("input_1239_cast_fp16")]; + tensor encoder_layers_23_self_attn_linear_out_weight_to_fp16 = const()[name = string("encoder_layers_23_self_attn_linear_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(533276480)))]; + tensor encoder_layers_23_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_23_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(535373696)))]; + tensor linear_214_cast_fp16 = linear(bias = encoder_layers_23_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_23_self_attn_linear_out_weight_to_fp16, x = input_1239_cast_fp16)[name = string("linear_214_cast_fp16")]; + tensor input_1243_cast_fp16 = add(x = input_1233_cast_fp16, y = linear_214_cast_fp16)[name = string("input_1243_cast_fp16")]; + tensor x_615_axes_0 = const()[name = string("x_615_axes_0"), val = tensor([-1])]; + tensor encoder_layers_23_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_23_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(535375808)))]; + tensor encoder_layers_23_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_23_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(535377920)))]; + tensor x_615_cast_fp16 = layer_norm(axes = x_615_axes_0, beta = encoder_layers_23_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_23_norm_conv_weight_to_fp16, x = input_1243_cast_fp16)[name = string("x_615_cast_fp16")]; + tensor input_1245_perm_0 = const()[name = string("input_1245_perm_0"), val = tensor([0, 2, 1])]; + string input_1247_pad_type_0 = const()[name = string("input_1247_pad_type_0"), val = string("valid")]; + tensor input_1247_strides_0 = const()[name = string("input_1247_strides_0"), val = tensor([1])]; + tensor input_1247_pad_0 = const()[name = string("input_1247_pad_0"), val = tensor([0, 0])]; + tensor input_1247_dilations_0 = const()[name = string("input_1247_dilations_0"), val = tensor([1])]; + int32 input_1247_groups_0 = const()[name = string("input_1247_groups_0"), val = int32(1)]; + tensor encoder_layers_23_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(535380032))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(537477248))))[name = string("encoder_layers_23_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_1245_cast_fp16 = transpose(perm = input_1245_perm_0, x = x_615_cast_fp16)[name = string("transpose_150")]; + tensor input_1247_cast_fp16 = conv(dilations = input_1247_dilations_0, groups = input_1247_groups_0, pad = input_1247_pad_0, pad_type = input_1247_pad_type_0, strides = input_1247_strides_0, weight = encoder_layers_23_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_1245_cast_fp16)[name = string("input_1247_cast_fp16")]; + int32 x_617_split_num_splits_0 = const()[name = string("x_617_split_num_splits_0"), val = int32(2)]; + int32 x_617_split_axis_0 = const()[name = string("x_617_split_axis_0"), val = int32(1)]; + tensor x_617_split_cast_fp16_0, tensor x_617_split_cast_fp16_1 = split(axis = x_617_split_axis_0, num_splits = x_617_split_num_splits_0, x = input_1247_cast_fp16)[name = string("x_617_split_cast_fp16")]; + tensor x_617_split_1_sigmoid_cast_fp16 = sigmoid(x = x_617_split_cast_fp16_1)[name = string("x_617_split_1_sigmoid_cast_fp16")]; + tensor x_617_cast_fp16 = mul(x = x_617_split_cast_fp16_0, y = x_617_split_1_sigmoid_cast_fp16)[name = string("x_617_cast_fp16")]; + tensor input_1249_cast_fp16 = select(a = var_44_to_fp16, b = x_617_cast_fp16, cond = var_575)[name = string("input_1249_cast_fp16")]; + bool new_x_interleave_0 = const()[name = string("new_x_interleave_0"), val = bool(false)]; + tensor new_x_cast_fp16 = concat(axis = var_59, interleave = new_x_interleave_0, values = (cache_cast_fp16, input_1249_cast_fp16))[name = string("new_x_cast_fp16")]; + tensor cache_last_time_cur_begin_0 = const()[name = string("cache_last_time_cur_begin_0"), val = tensor([0, 0, 7])]; + tensor cache_last_time_cur_end_0 = const()[name = string("cache_last_time_cur_end_0"), val = tensor([1, 1024, 15])]; + tensor cache_last_time_cur_end_mask_0 = const()[name = string("cache_last_time_cur_end_mask_0"), val = tensor([true, true, true])]; + tensor cache_last_time_cur_cast_fp16 = slice_by_index(begin = cache_last_time_cur_begin_0, end = cache_last_time_cur_end_0, end_mask = cache_last_time_cur_end_mask_0, x = new_x_cast_fp16)[name = string("cache_last_time_cur_cast_fp16")]; + string x_619_pad_type_0 = const()[name = string("x_619_pad_type_0"), val = string("valid")]; + int32 x_619_groups_0 = const()[name = string("x_619_groups_0"), val = int32(1024)]; + tensor x_619_strides_0 = const()[name = string("x_619_strides_0"), val = tensor([1])]; + tensor x_619_pad_0 = const()[name = string("x_619_pad_0"), val = tensor([0, 0])]; + tensor x_619_dilations_0 = const()[name = string("x_619_dilations_0"), val = tensor([1])]; + tensor encoder_layers_23_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(537481408))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(537490688))))[name = string("encoder_layers_23_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_619_cast_fp16 = conv(dilations = x_619_dilations_0, groups = x_619_groups_0, pad = x_619_pad_0, pad_type = x_619_pad_type_0, strides = x_619_strides_0, weight = encoder_layers_23_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_cast_fp16)[name = string("x_619_cast_fp16")]; + tensor input_1251_perm_0 = const()[name = string("input_1251_perm_0"), val = tensor([0, 2, 1])]; + tensor x_621_axes_0 = const()[name = string("x_621_axes_0"), val = tensor([-1])]; + tensor encoder_layers_23_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_23_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(537492800)))]; + tensor encoder_layers_23_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_23_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(537494912)))]; + tensor input_1251_cast_fp16 = transpose(perm = input_1251_perm_0, x = x_619_cast_fp16)[name = string("transpose_149")]; + tensor x_621_cast_fp16 = layer_norm(axes = x_621_axes_0, beta = encoder_layers_23_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_23_conv_batch_norm_weight_to_fp16, x = input_1251_cast_fp16)[name = string("x_621_cast_fp16")]; + tensor input_1253_perm_0 = const()[name = string("input_1253_perm_0"), val = tensor([0, 2, 1])]; + tensor input_1253_cast_fp16 = transpose(perm = input_1253_perm_0, x = x_621_cast_fp16)[name = string("transpose_148")]; + tensor input_1255_cast_fp16 = silu(x = input_1253_cast_fp16)[name = string("input_1255_cast_fp16")]; + string x_623_pad_type_0 = const()[name = string("x_623_pad_type_0"), val = string("valid")]; + tensor x_623_strides_0 = const()[name = string("x_623_strides_0"), val = tensor([1])]; + tensor x_623_pad_0 = const()[name = string("x_623_pad_0"), val = tensor([0, 0])]; + tensor x_623_dilations_0 = const()[name = string("x_623_dilations_0"), val = tensor([1])]; + int32 x_623_groups_0 = const()[name = string("x_623_groups_0"), val = int32(1)]; + tensor encoder_layers_23_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(537497024))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(538545664))))[name = string("encoder_layers_23_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_623_cast_fp16 = conv(dilations = x_623_dilations_0, groups = x_623_groups_0, pad = x_623_pad_0, pad_type = x_623_pad_type_0, strides = x_623_strides_0, weight = encoder_layers_23_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_1255_cast_fp16)[name = string("x_623_cast_fp16")]; + tensor input_1257_perm_0 = const()[name = string("input_1257_perm_0"), val = tensor([0, 2, 1])]; + tensor input_1257_cast_fp16 = transpose(perm = input_1257_perm_0, x = x_623_cast_fp16)[name = string("transpose_147")]; + tensor input_1259_cast_fp16 = add(x = input_1243_cast_fp16, y = input_1257_cast_fp16)[name = string("input_1259_cast_fp16")]; + tensor input_1261_axes_0 = const()[name = string("input_1261_axes_0"), val = tensor([-1])]; + tensor encoder_layers_23_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_23_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(538547776)))]; + tensor encoder_layers_23_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_23_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(538549888)))]; + tensor input_1261_cast_fp16 = layer_norm(axes = input_1261_axes_0, beta = encoder_layers_23_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_23_norm_feed_forward2_weight_to_fp16, x = input_1259_cast_fp16)[name = string("input_1261_cast_fp16")]; + tensor encoder_layers_23_feed_forward2_linear1_weight_to_fp16 = const()[name = string("encoder_layers_23_feed_forward2_linear1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(538552000)))]; + tensor encoder_layers_23_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_23_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(546940672)))]; + tensor linear_215_cast_fp16 = linear(bias = encoder_layers_23_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_23_feed_forward2_linear1_weight_to_fp16, x = input_1261_cast_fp16)[name = string("linear_215_cast_fp16")]; + tensor input_1265_cast_fp16 = silu(x = linear_215_cast_fp16)[name = string("input_1265_cast_fp16")]; + tensor encoder_layers_23_feed_forward2_linear2_weight_to_fp16 = const()[name = string("encoder_layers_23_feed_forward2_linear2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(546948928)))]; + tensor encoder_layers_23_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_23_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(555337600)))]; + tensor linear_216_cast_fp16 = linear(bias = encoder_layers_23_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_23_feed_forward2_linear2_weight_to_fp16, x = input_1265_cast_fp16)[name = string("linear_216_cast_fp16")]; + fp16 var_5530_to_fp16 = const()[name = string("op_5530_to_fp16"), val = fp16(0x1p-1)]; + tensor var_5531_cast_fp16 = mul(x = linear_216_cast_fp16, y = var_5530_to_fp16)[name = string("op_5531_cast_fp16")]; + tensor input_1271_cast_fp16 = add(x = input_1259_cast_fp16, y = var_5531_cast_fp16)[name = string("input_1271_cast_fp16")]; + tensor audio_signal_axes_0 = const()[name = string("audio_signal_axes_0"), val = tensor([-1])]; + tensor encoder_layers_23_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_23_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(555339712)))]; + tensor encoder_layers_23_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_23_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(555341824)))]; + tensor audio_signal_cast_fp16 = layer_norm(axes = audio_signal_axes_0, beta = encoder_layers_23_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_23_norm_out_weight_to_fp16, x = input_1271_cast_fp16)[name = string("audio_signal_cast_fp16")]; + int32 obj_5_axis_0 = const()[name = string("obj_5_axis_0"), val = int32(0)]; + tensor obj_5_cast_fp16 = stack(axis = obj_5_axis_0, values = (var_484_cast_fp16, var_697_cast_fp16, var_910_cast_fp16, var_1123_cast_fp16, var_1336_cast_fp16, var_1549_cast_fp16, var_1762_cast_fp16, var_1975_cast_fp16, var_2188_cast_fp16, var_2401_cast_fp16, var_2614_cast_fp16, var_2827_cast_fp16, var_3040_cast_fp16, var_3253_cast_fp16, var_3466_cast_fp16, var_3679_cast_fp16, var_3892_cast_fp16, var_4105_cast_fp16, var_4318_cast_fp16, var_4531_cast_fp16, var_4744_cast_fp16, var_4957_cast_fp16, var_5170_cast_fp16, cache_last_channel_cur_cast_fp16))[name = string("obj_5_cast_fp16")]; + int32 obj_7_axis_0 = const()[name = string("obj_7_axis_0"), val = int32(0)]; + tensor obj_7_cast_fp16 = stack(axis = obj_7_axis_0, values = (var_588_cast_fp16, var_801_cast_fp16, var_1014_cast_fp16, var_1227_cast_fp16, var_1440_cast_fp16, var_1653_cast_fp16, var_1866_cast_fp16, var_2079_cast_fp16, var_2292_cast_fp16, var_2505_cast_fp16, var_2718_cast_fp16, var_2931_cast_fp16, var_3144_cast_fp16, var_3357_cast_fp16, var_3570_cast_fp16, var_3783_cast_fp16, var_3996_cast_fp16, var_4209_cast_fp16, var_4422_cast_fp16, var_4635_cast_fp16, var_4848_cast_fp16, var_5061_cast_fp16, var_5274_cast_fp16, cache_last_time_cur_cast_fp16))[name = string("obj_7_cast_fp16")]; + tensor var_5547 = add(x = cache_len, y = max_audio_length_1)[name = string("op_5547")]; + string var_5547_promoted_to_fp16_dtype_0 = const()[name = string("op_5547_promoted_to_fp16_dtype_0"), val = string("fp16")]; + fp16 const_384_to_fp16 = const()[name = string("const_384_to_fp16"), val = fp16(-inf)]; + fp16 var_49_promoted_to_fp16 = const()[name = string("op_49_promoted_to_fp16"), val = fp16(0x1.5p+5)]; + tensor var_5547_to_fp16 = cast(dtype = var_5547_promoted_to_fp16_dtype_0, x = var_5547)[name = string("cast_9")]; + tensor clip_1_cast_fp16 = clip(alpha = const_384_to_fp16, beta = var_49_promoted_to_fp16, x = var_5547_to_fp16)[name = string("clip_1_cast_fp16")]; + int32 one_hot_1_batch_dims_0 = const()[name = string("one_hot_1_batch_dims_0"), val = int32(0)]; + bool one_hot_1_validate_indices_0 = const()[name = string("one_hot_1_validate_indices_0"), val = bool(false)]; + tensor to_onehot_identity_table_to_fp16 = const()[name = string("to_onehot_identity_table_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(555343936)))]; + string prompt_id_to_int16_dtype_0 = const()[name = string("prompt_id_to_int16_dtype_0"), val = string("int16")]; + string cast_230_dtype_0 = const()[name = string("cast_230_dtype_0"), val = string("int32")]; + int32 greater_equal_0_y_0 = const()[name = string("greater_equal_0_y_0"), val = int32(0)]; + tensor prompt_id_to_int16 = cast(dtype = prompt_id_to_int16_dtype_0, x = prompt_id)[name = string("cast_8")]; + tensor cast_230 = cast(dtype = cast_230_dtype_0, x = prompt_id_to_int16)[name = string("cast_7")]; + tensor greater_equal_0 = greater_equal(x = cast_230, y = greater_equal_0_y_0)[name = string("greater_equal_0")]; + int32 slice_by_index_2 = const()[name = string("slice_by_index_2"), val = int32(128)]; + tensor add_0 = add(x = cast_230, y = slice_by_index_2)[name = string("add_0")]; + tensor select_0 = select(a = cast_230, b = add_0, cond = greater_equal_0)[name = string("select_0")]; + string select_0_to_int16_dtype_0 = const()[name = string("select_0_to_int16_dtype_0"), val = string("int16")]; + string cast_0_dtype_0 = const()[name = string("cast_0_dtype_0"), val = string("int32")]; + int32 greater_equal_0_y_0_1 = const()[name = string("greater_equal_0_y_0_1"), val = int32(0)]; + tensor select_0_to_int16 = cast(dtype = select_0_to_int16_dtype_0, x = select_0)[name = string("cast_6")]; + tensor cast_0 = cast(dtype = cast_0_dtype_0, x = select_0_to_int16)[name = string("cast_5")]; + tensor greater_equal_0_1 = greater_equal(x = cast_0, y = greater_equal_0_y_0_1)[name = string("greater_equal_0_1")]; + int32 slice_by_index_0 = const()[name = string("slice_by_index_0"), val = int32(128)]; + tensor add_0_1 = add(x = cast_0, y = slice_by_index_0)[name = string("add_0_1")]; + tensor select_0_1 = select(a = cast_0, b = add_0_1, cond = greater_equal_0_1)[name = string("select_0_1")]; + int32 greater_equal_0_y_0_2 = const()[name = string("greater_equal_0_y_0_2"), val = int32(0)]; + tensor greater_equal_0_2 = greater_equal(x = select_0_1, y = greater_equal_0_y_0_2)[name = string("greater_equal_0_2")]; + int32 slice_by_index_0_1 = const()[name = string("slice_by_index_0_1"), val = int32(128)]; + tensor add_0_2 = add(x = select_0_1, y = slice_by_index_0_1)[name = string("add_0_2")]; + tensor select_0_2 = select(a = select_0_1, b = add_0_2, cond = greater_equal_0_2)[name = string("select_0_2")]; + int32 one_hot_1_cast_fp16_cast_uint16_cast_uint16_axis_0 = const()[name = string("one_hot_1_cast_fp16_cast_uint16_cast_uint16_axis_0"), val = int32(0)]; + tensor one_hot_1_cast_fp16_cast_uint16_cast_uint16 = gather(axis = one_hot_1_cast_fp16_cast_uint16_cast_uint16_axis_0, batch_dims = one_hot_1_batch_dims_0, indices = select_0_2, validate_indices = one_hot_1_validate_indices_0, x = to_onehot_identity_table_to_fp16)[name = string("one_hot_1_cast_fp16_cast_uint16_cast_uint16")]; + tensor var_5593_axes_0 = const()[name = string("op_5593_axes_0"), val = tensor([1])]; + tensor var_5593_cast_fp16 = expand_dims(axes = var_5593_axes_0, x = one_hot_1_cast_fp16_cast_uint16_cast_uint16)[name = string("op_5593_cast_fp16")]; + tensor one_hot_reps_0 = const()[name = string("one_hot_reps_0"), val = tensor([1, 7, 1])]; + tensor one_hot_cast_fp16 = tile(reps = one_hot_reps_0, x = var_5593_cast_fp16)[name = string("one_hot_cast_fp16")]; + int32 var_5602 = const()[name = string("op_5602"), val = int32(-1)]; + bool input_1273_interleave_0 = const()[name = string("input_1273_interleave_0"), val = bool(false)]; + tensor input_1273_cast_fp16 = concat(axis = var_5602, interleave = input_1273_interleave_0, values = (audio_signal_cast_fp16, one_hot_cast_fp16))[name = string("input_1273_cast_fp16")]; + tensor prompt_kernel_0_weight_to_fp16 = const()[name = string("prompt_kernel_0_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(555376768)))]; + tensor prompt_kernel_0_bias_to_fp16 = const()[name = string("prompt_kernel_0_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(560095424)))]; + tensor linear_217_cast_fp16 = linear(bias = prompt_kernel_0_bias_to_fp16, weight = prompt_kernel_0_weight_to_fp16, x = input_1273_cast_fp16)[name = string("linear_217_cast_fp16")]; + tensor input_cast_fp16 = relu(x = linear_217_cast_fp16)[name = string("input_cast_fp16")]; + tensor prompt_kernel_2_weight_to_fp16 = const()[name = string("prompt_kernel_2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(560099584)))]; + tensor prompt_kernel_2_bias_to_fp16 = const()[name = string("prompt_kernel_2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(564293952)))]; + tensor linear_218_cast_fp16 = linear(bias = prompt_kernel_2_bias_to_fp16, weight = prompt_kernel_2_weight_to_fp16, x = input_cast_fp16)[name = string("linear_218_cast_fp16")]; + tensor var_5615_perm_0 = const()[name = string("op_5615_perm_0"), val = tensor([0, 2, 1])]; + string var_5615_cast_fp16_to_fp32_dtype_0 = const()[name = string("op_5615_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + string var_5620_dtype_0 = const()[name = string("op_5620_dtype_0"), val = string("int32")]; + tensor var_5623_perm_0 = const()[name = string("op_5623_perm_0"), val = tensor([1, 0, 2, 3])]; + string var_5623_cast_fp16_to_fp32_dtype_0 = const()[name = string("op_5623_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + tensor var_5626_perm_0 = const()[name = string("op_5626_perm_0"), val = tensor([1, 0, 2, 3])]; + string var_5626_cast_fp16_to_fp32_dtype_0 = const()[name = string("op_5626_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + string var_5631_dtype_0 = const()[name = string("op_5631_dtype_0"), val = string("int32")]; + tensor cache_len_out = cast(dtype = var_5631_dtype_0, x = clip_1_cast_fp16)[name = string("cast_0")]; + tensor var_5626_cast_fp16 = transpose(perm = var_5626_perm_0, x = obj_7_cast_fp16)[name = string("transpose_144")]; + tensor cache_time_out = cast(dtype = var_5626_cast_fp16_to_fp32_dtype_0, x = var_5626_cast_fp16)[name = string("cast_1")]; + tensor var_5623_cast_fp16 = transpose(perm = var_5623_perm_0, x = obj_5_cast_fp16)[name = string("transpose_145")]; + tensor cache_channel_out = cast(dtype = var_5623_cast_fp16_to_fp32_dtype_0, x = var_5623_cast_fp16)[name = string("cast_2")]; + tensor encoded_length = cast(dtype = var_5620_dtype_0, x = clip_0_cast_fp16)[name = string("cast_3")]; + tensor var_5615_cast_fp16 = transpose(perm = var_5615_perm_0, x = linear_218_cast_fp16)[name = string("transpose_146")]; + tensor encoded = cast(dtype = var_5615_cast_fp16_to_fp32_dtype_0, x = var_5615_cast_fp16)[name = string("cast_4")]; + } -> (encoded, encoded_length, cache_channel_out, cache_time_out, cache_len_out); +} \ No newline at end of file diff --git a/it/560ms/encoder.mlmodelc/weights/weight.bin b/it/560ms/encoder.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..7b1ca90fb42a5ee3e6703488fa2d418b42814a01 --- /dev/null +++ b/it/560ms/encoder.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:beaf958ec4cca237d94bf72f5afded9f1b3f6b93439b5d8e3b309e4ee1560e83 +size 564296064 diff --git a/it/560ms/encoder.mlpackage/Data/com.apple.CoreML/model.mlmodel b/it/560ms/encoder.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..21020cc3a69303cc019d3c163864648766ab26b9 --- /dev/null +++ b/it/560ms/encoder.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1e6d4ec3401a73a4b4c0409170717e2caf7bc829f85aa844dbb5f70309dc61ea +size 802813 diff --git a/it/560ms/encoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/it/560ms/encoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..7b1ca90fb42a5ee3e6703488fa2d418b42814a01 --- /dev/null +++ b/it/560ms/encoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:beaf958ec4cca237d94bf72f5afded9f1b3f6b93439b5d8e3b309e4ee1560e83 +size 564296064 diff --git a/it/560ms/encoder.mlpackage/Manifest.json b/it/560ms/encoder.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..44bfa973b64e3a625c9a52d15b370d5a8baa0bbd --- /dev/null +++ b/it/560ms/encoder.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "5BB04C05-9A8D-40F1-9561-33AB8C6304FC": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + }, + "7CFD7E55-2E07-4224-889C-214DD954BA80": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + } + }, + "rootModelIdentifier": "7CFD7E55-2E07-4224-889C-214DD954BA80" +} diff --git a/it/560ms/joint.mlmodelc/analytics/coremldata.bin b/it/560ms/joint.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..ee72f6da06176879d781b1a46f441fa01048d662 --- /dev/null +++ b/it/560ms/joint.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cae47ef3b8cb372625305f4cdd31159b5ce56be1470c94613ff946be342f6d7a +size 243 diff --git a/it/560ms/joint.mlmodelc/coremldata.bin b/it/560ms/joint.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..07f143747ee2f43103809647f4058203bf60dc56 --- /dev/null +++ b/it/560ms/joint.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:215d8dd2e33da0c37f08e6b0c0a1e997a3e056f2cb9113fdcf17f8027a61216d +size 341 diff --git a/it/560ms/joint.mlmodelc/model.mil b/it/560ms/joint.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..bfde40ec94bf61746424d2d3e196a4fba198de2d --- /dev/null +++ b/it/560ms/joint.mlmodelc/model.mil @@ -0,0 +1,31 @@ +program(1.3) +[buildInfo = dict({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.5.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.3.0"}})] +{ + func main(tensor decoder, tensor encoder) { + tensor input_1_perm_0 = const()[name = string("input_1_perm_0"), val = tensor([0, 2, 1])]; + string encoder_to_fp16_dtype_0 = const()[name = string("encoder_to_fp16_dtype_0"), val = string("fp16")]; + tensor input_3_perm_0 = const()[name = string("input_3_perm_0"), val = tensor([0, 2, 1])]; + string decoder_to_fp16_dtype_0 = const()[name = string("decoder_to_fp16_dtype_0"), val = string("fp16")]; + tensor module_enc_weight_to_fp16 = const()[name = string("module_enc_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))]; + tensor module_enc_bias_to_fp16 = const()[name = string("module_enc_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1310848)))]; + tensor encoder_to_fp16 = cast(dtype = encoder_to_fp16_dtype_0, x = encoder)[name = string("cast_2")]; + tensor input_1_cast_fp16 = transpose(perm = input_1_perm_0, x = encoder_to_fp16)[name = string("transpose_1")]; + tensor linear_0_cast_fp16 = linear(bias = module_enc_bias_to_fp16, weight = module_enc_weight_to_fp16, x = input_1_cast_fp16)[name = string("linear_0_cast_fp16")]; + tensor module_pred_weight_to_fp16 = const()[name = string("module_pred_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1312192)))]; + tensor module_pred_bias_to_fp16 = const()[name = string("module_pred_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2131456)))]; + tensor decoder_to_fp16 = cast(dtype = decoder_to_fp16_dtype_0, x = decoder)[name = string("cast_1")]; + tensor input_3_cast_fp16 = transpose(perm = input_3_perm_0, x = decoder_to_fp16)[name = string("transpose_0")]; + tensor linear_1_cast_fp16 = linear(bias = module_pred_bias_to_fp16, weight = module_pred_weight_to_fp16, x = input_3_cast_fp16)[name = string("linear_1_cast_fp16")]; + tensor var_23_axes_0 = const()[name = string("op_23_axes_0"), val = tensor([2])]; + tensor var_23_cast_fp16 = expand_dims(axes = var_23_axes_0, x = linear_0_cast_fp16)[name = string("op_23_cast_fp16")]; + tensor var_25_axes_0 = const()[name = string("op_25_axes_0"), val = tensor([1])]; + tensor var_25_cast_fp16 = expand_dims(axes = var_25_axes_0, x = linear_1_cast_fp16)[name = string("op_25_cast_fp16")]; + tensor input_5_cast_fp16 = add(x = var_23_cast_fp16, y = var_25_cast_fp16)[name = string("input_5_cast_fp16")]; + tensor input_7_cast_fp16 = relu(x = input_5_cast_fp16)[name = string("input_7_cast_fp16")]; + tensor module_joint_net_2_weight_to_fp16 = const()[name = string("module_joint_net_2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2132800)))]; + tensor module_joint_net_2_bias_to_fp16 = const()[name = string("module_joint_net_2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3164544)))]; + tensor linear_2_cast_fp16 = linear(bias = module_joint_net_2_bias_to_fp16, weight = module_joint_net_2_weight_to_fp16, x = input_7_cast_fp16)[name = string("linear_2_cast_fp16")]; + string linear_2_cast_fp16_to_fp32_dtype_0 = const()[name = string("linear_2_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + tensor logits = cast(dtype = linear_2_cast_fp16_to_fp32_dtype_0, x = linear_2_cast_fp16)[name = string("cast_0")]; + } -> (logits); +} \ No newline at end of file diff --git a/it/560ms/joint.mlmodelc/weights/weight.bin b/it/560ms/joint.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..3673a1d2a1060ecda8626c76ffacdfd89c5f00c1 --- /dev/null +++ b/it/560ms/joint.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1740c99cbe34ebeaa0163c8421135b4586df09960ef07fe02abb2a94b5693411 +size 3166220 diff --git a/it/560ms/joint.mlpackage/Data/com.apple.CoreML/model.mlmodel b/it/560ms/joint.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..f7f468ddd814131e36b8af9ed7a3358576bffcf0 --- /dev/null +++ b/it/560ms/joint.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d902ae83b8a93f0f4240a4a6939466dbd1a6b2291f1615d81d7ac26d9115bc23 +size 4486 diff --git a/it/560ms/joint.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/it/560ms/joint.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..3673a1d2a1060ecda8626c76ffacdfd89c5f00c1 --- /dev/null +++ b/it/560ms/joint.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1740c99cbe34ebeaa0163c8421135b4586df09960ef07fe02abb2a94b5693411 +size 3166220 diff --git a/it/560ms/joint.mlpackage/Manifest.json b/it/560ms/joint.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..fea9f1ee7eee62ace96d28134fe38a74b32b40c9 --- /dev/null +++ b/it/560ms/joint.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "6D4EAD3B-A17D-4807-8572-3E0A513C0C7E": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + }, + "8242AF76-3606-4A98-8B3D-E0BB730201A4": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + } + }, + "rootModelIdentifier": "6D4EAD3B-A17D-4807-8572-3E0A513C0C7E" +} diff --git a/it/560ms/joint_noencproj_batched.mlmodelc/analytics/coremldata.bin b/it/560ms/joint_noencproj_batched.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..d2ffa5c9579dd51ce490b9c8fe47e3c27a05c723 --- /dev/null +++ b/it/560ms/joint_noencproj_batched.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:24510a0ca90eea355a0e446ba54d63004a3f2c51d30eb3e415e4af938ac1acba +size 243 diff --git a/it/560ms/joint_noencproj_batched.mlmodelc/coremldata.bin b/it/560ms/joint_noencproj_batched.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..03999c10ebcacf3633f026f01254ed98ae34fa25 --- /dev/null +++ b/it/560ms/joint_noencproj_batched.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dc93bb8f241754f1359837f597d600b029d687ce0c57a4280fa586f8c8386337 +size 406 diff --git a/it/560ms/joint_noencproj_batched.mlmodelc/model.mil b/it/560ms/joint_noencproj_batched.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..58942374262618031d39a52f7a009b81c7f24c24 --- /dev/null +++ b/it/560ms/joint_noencproj_batched.mlmodelc/model.mil @@ -0,0 +1,26 @@ +program(1.3) +[buildInfo = dict({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.10.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor decoder, tensor encoder_proj) { + tensor input_1_perm_0 = const()[name = string("input_1_perm_0"), val = tensor([0, 2, 1])]; + string decoder_to_fp16_dtype_0 = const()[name = string("decoder_to_fp16_dtype_0"), val = string("fp16")]; + tensor joint_module_pred_weight_to_fp16 = const()[name = string("joint_module_pred_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))]; + tensor joint_module_pred_bias_to_fp16 = const()[name = string("joint_module_pred_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(819328)))]; + tensor decoder_to_fp16 = cast(dtype = decoder_to_fp16_dtype_0, x = decoder)[name = string("cast_2")]; + tensor input_1_cast_fp16 = transpose(perm = input_1_perm_0, x = decoder_to_fp16)[name = string("transpose_0")]; + tensor linear_0_cast_fp16 = linear(bias = joint_module_pred_bias_to_fp16, weight = joint_module_pred_weight_to_fp16, x = input_1_cast_fp16)[name = string("linear_0_cast_fp16")]; + tensor var_15_axes_0 = const()[name = string("op_15_axes_0"), val = tensor([2])]; + string encoder_proj_to_fp16_dtype_0 = const()[name = string("encoder_proj_to_fp16_dtype_0"), val = string("fp16")]; + tensor encoder_proj_to_fp16 = cast(dtype = encoder_proj_to_fp16_dtype_0, x = encoder_proj)[name = string("cast_1")]; + tensor var_15_cast_fp16 = expand_dims(axes = var_15_axes_0, x = encoder_proj_to_fp16)[name = string("op_15_cast_fp16")]; + tensor var_17_axes_0 = const()[name = string("op_17_axes_0"), val = tensor([1])]; + tensor var_17_cast_fp16 = expand_dims(axes = var_17_axes_0, x = linear_0_cast_fp16)[name = string("op_17_cast_fp16")]; + tensor input_3_cast_fp16 = add(x = var_15_cast_fp16, y = var_17_cast_fp16)[name = string("input_3_cast_fp16")]; + tensor input_5_cast_fp16 = relu(x = input_3_cast_fp16)[name = string("input_5_cast_fp16")]; + tensor joint_module_joint_net_2_weight_to_fp16 = const()[name = string("joint_module_joint_net_2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(820672)))]; + tensor joint_module_joint_net_2_bias_to_fp16 = const()[name = string("joint_module_joint_net_2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1852416)))]; + tensor linear_1_cast_fp16 = linear(bias = joint_module_joint_net_2_bias_to_fp16, weight = joint_module_joint_net_2_weight_to_fp16, x = input_5_cast_fp16)[name = string("linear_1_cast_fp16")]; + string linear_1_cast_fp16_to_fp32_dtype_0 = const()[name = string("linear_1_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + tensor logits = cast(dtype = linear_1_cast_fp16_to_fp32_dtype_0, x = linear_1_cast_fp16)[name = string("cast_0")]; + } -> (logits); +} \ No newline at end of file diff --git a/it/560ms/joint_noencproj_batched.mlmodelc/weights/weight.bin b/it/560ms/joint_noencproj_batched.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..40ceadd152241059aa378e2ddb6cc9f649e0b59c --- /dev/null +++ b/it/560ms/joint_noencproj_batched.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cd83e82dcfec315f28c8a8872b0d7f22e668a2c485821de86a0379ae2b3864ad +size 1854092 diff --git a/it/560ms/joint_noencproj_batched.mlpackage/Data/com.apple.CoreML/model.mlmodel b/it/560ms/joint_noencproj_batched.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..49f1fd816a2725ab32d473dfe600f911656f675b --- /dev/null +++ b/it/560ms/joint_noencproj_batched.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:26a19f2e269811070e57eda978e8226ca9dcca6177b46b95f916538b5f815716 +size 3839 diff --git a/it/560ms/joint_noencproj_batched.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/it/560ms/joint_noencproj_batched.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..40ceadd152241059aa378e2ddb6cc9f649e0b59c --- /dev/null +++ b/it/560ms/joint_noencproj_batched.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cd83e82dcfec315f28c8a8872b0d7f22e668a2c485821de86a0379ae2b3864ad +size 1854092 diff --git a/it/560ms/joint_noencproj_batched.mlpackage/Manifest.json b/it/560ms/joint_noencproj_batched.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..e3857a03022f3cec16823ed1d8a741fee56cfce4 --- /dev/null +++ b/it/560ms/joint_noencproj_batched.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "4233CE8E-FB95-4FF9-BCD8-2A834D55C580": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + }, + "96E0F26C-90DC-49EE-B510-D0FB3FC812CC": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + } + }, + "rootModelIdentifier": "96E0F26C-90DC-49EE-B510-D0FB3FC812CC" +} diff --git a/it/560ms/metadata.json b/it/560ms/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..2dafed9cb8456ca449c4a608b0196c87efa3e06e --- /dev/null +++ b/it/560ms/metadata.json @@ -0,0 +1,199 @@ +{ + "model": "nvidia/nemotron-asr-streaming-multilingual-0.6b", + "model_class": "nemo.collections.asr.models.rnnt_bpe_models_prompt.EncDecRNNTBPEModelWithPrompt", + "sample_rate": 16000, + "mel_features": 128, + "chunk_mel_frames": 56, + "pre_encode_cache": 9, + "total_mel_frames": 65, + "att_context_size": [ + 42, + 13 + ], + "vocab_size": 805, + "blank_idx": 805, + "vocab_pruned": true, + "vocab_pruned_original_size": 13087, + "cache_channel_shape": [ + 1, + 24, + 42, + 1024 + ], + "cache_time_shape": [ + 1, + 24, + 1024, + 8 + ], + "decoder_hidden": 640, + "decoder_layers": 2, + "encoder_dim": 1024, + "num_prompts": 128, + "prompt_dictionary": { + "af-ZA": 54, + "am-ET": 49, + "ar": 7, + "ar-AR": 7, + "auto": 101, + "ay-BO": 81, + "az-AZ": 66, + "bg": 30, + "bg-BG": 30, + "bn-IN": 36, + "cs": 22, + "cs-CZ": 22, + "da": 25, + "da-DK": 25, + "de": 9, + "de-DE": 9, + "el": 21, + "el-GR": 21, + "en": 0, + "en-GB": 1, + "en-US": 0, + "enGB": 1, + "es": 3, + "es-ES": 2, + "es-US": 3, + "esES": 2, + "et": 60, + "et-EE": 60, + "fa-IR": 38, + "fi": 26, + "fi-FI": 26, + "fr": 8, + "fr-CA": 100, + "fr-FR": 8, + "gn-PY": 82, + "gu-IN": 42, + "ha-NG": 50, + "haw-US": 97, + "he-IL": 64, + "hi": 6, + "hi-HI": 6, + "hi-IN": 6, + "hr": 29, + "hr-HR": 29, + "hu": 23, + "hu-HU": 23, + "hy-AM": 68, + "id-ID": 34, + "ig-NG": 53, + "it": 15, + "it-IT": 15, + "ja-JA": 10, + "ja-JP": 10, + "ka-GE": 67, + "km-KH": 47, + "kn-IN": 43, + "ko": 14, + "ko-KO": 14, + "ko-KR": 14, + "ku-TR": 65, + "ky-KG": 71, + "ln-CD": 58, + "lt": 31, + "lt-LT": 31, + "lv": 61, + "lv-LV": 61, + "mi-NZ": 96, + "ml-IN": 44, + "mr-IN": 41, + "ms-MY": 35, + "mt-MT": 102, + "nah-MX": 83, + "nb": 103, + "nb-NO": 103, + "ne-NP": 46, + "nl": 16, + "nl-NL": 16, + "nn": 104, + "nn-NO": 104, + "no": 27, + "no-NO": 27, + "ny-MW": 57, + "or-KE": 59, + "pl": 17, + "pl-PL": 17, + "pt": 13, + "pt-BR": 12, + "pt-PT": 13, + "qu-PE": 80, + "ro": 20, + "ro-RO": 20, + "ru": 11, + "ru-RU": 11, + "rw-RW": 55, + "si-LK": 45, + "sk": 28, + "sk-SK": 28, + "sl": 62, + "sl-SI": 62, + "sm-WS": 98, + "so-SO": 56, + "sv": 24, + "sv-SE": 24, + "sw-KE": 48, + "ta-IN": 39, + "te-IN": 40, + "tg-TJ": 70, + "th-TH": 32, + "to-TO": 99, + "tr": 18, + "tr-TR": 18, + "uk": 19, + "uk-UA": 19, + "ur-PK": 37, + "uz-UZ": 69, + "vi-VN": 33, + "yo-NG": 52, + "zh-CN": 4, + "zh-TW": 5, + "zh-ZH": 4, + "zu-ZA": 51 + }, + "default_prompt_id": 101, + "lang_tag_token_ids": [ + 1, + 63, + 115, + 167, + 226, + 227, + 259, + 276, + 328, + 353, + 368, + 462, + 481, + 499, + 518, + 542, + 571, + 602, + 603, + 612, + 624, + 646, + 647, + 667, + 689, + 699, + 720, + 727, + 747, + 748, + 750, + 751, + 752, + 756, + 774, + 787, + 788, + 801, + 802 + ], + "chunk_ms": 560 +} \ No newline at end of file diff --git a/it/560ms/preprocessor.mlmodelc/analytics/coremldata.bin b/it/560ms/preprocessor.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..d4b1561ab413a9d87db506bc842f077779dcbded --- /dev/null +++ b/it/560ms/preprocessor.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3b8033bec5ee01649f325b8f4c5aeef1b31c99b469ce56d46039c1b73f09585d +size 243 diff --git a/it/560ms/preprocessor.mlmodelc/coremldata.bin b/it/560ms/preprocessor.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..671433a4b499506c5255c70ea8b5355488317851 --- /dev/null +++ b/it/560ms/preprocessor.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:262d0b39a4231c9052739617854413a079f7a11ce7ca0ed83226e4094659b2c2 +size 431 diff --git a/it/560ms/preprocessor.mlmodelc/model.mil b/it/560ms/preprocessor.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..b1a0b2b9193c992de42e51245fc1ef433d345afc --- /dev/null +++ b/it/560ms/preprocessor.mlmodelc/model.mil @@ -0,0 +1,122 @@ +program(1.3) +[buildInfo = dict({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.10.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor audio, tensor audio_length) [FlexibleShapeInformation = tuple>>, tuple, ?>>>>((("DefaultShapes", {{"audio", [1, 1]}}), ("RangeDims", {{"audio", [[1, 1], [1, 1280000]]}})))] { + int32 var_9 = const()[name = string("op_9"), val = int32(1)]; + int32 var_10 = const()[name = string("op_10"), val = int32(160)]; + int32 var_12 = const()[name = string("op_12"), val = int32(0)]; + int32 var_33 = const()[name = string("op_33"), val = int32(512)]; + tensor var_34 = add(x = audio_length, y = var_33)[name = string("op_34")]; + int32 var_35 = const()[name = string("op_35"), val = int32(512)]; + tensor var_36 = sub(x = var_34, y = var_35)[name = string("op_36")]; + tensor floor_div_0 = floor_div(x = var_36, y = var_10)[name = string("floor_div_0")]; + tensor var_39 = equal(x = audio_length, y = var_12)[name = string("op_39")]; + tensor var_40 = const()[name = string("op_40"), val = tensor([0])]; + tensor mel_length = select(a = var_40, b = floor_div_0, cond = var_39)[name = string("seq_len")]; + string audio_to_fp16_dtype_0 = const()[name = string("audio_to_fp16_dtype_0"), val = string("fp16")]; + tensor audio_to_fp16 = cast(dtype = audio_to_fp16_dtype_0, x = audio)[name = string("cast_10")]; + tensor var_42_shape_cast_fp16 = shape(x = audio_to_fp16)[name = string("op_42_shape_cast_fp16")]; + int32 gather_0_axis_0 = const()[name = string("gather_0_axis_0"), val = int32(0)]; + int32 gather_0_batch_dims_0 = const()[name = string("gather_0_batch_dims_0"), val = int32(0)]; + bool gather_0_validate_indices_0 = const()[name = string("gather_0_validate_indices_0"), val = bool(false)]; + string var_42_shape_cast_fp16_to_int16_dtype_0 = const()[name = string("op_42_shape_cast_fp16_to_int16_dtype_0"), val = string("int16")]; + uint16 gather_0_indices_0_to_uint16 = const()[name = string("gather_0_indices_0_to_uint16"), val = uint16(1)]; + tensor var_42_shape_cast_fp16_to_int16 = cast(dtype = var_42_shape_cast_fp16_to_int16_dtype_0, x = var_42_shape_cast_fp16)[name = string("cast_9")]; + int16 gather_0_cast_uint16 = gather(axis = gather_0_axis_0, batch_dims = gather_0_batch_dims_0, indices = gather_0_indices_0_to_uint16, validate_indices = gather_0_validate_indices_0, x = var_42_shape_cast_fp16_to_int16)[name = string("gather_0_cast_uint16")]; + string gather_0_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_0_cast_uint16_to_int32_dtype_0"), val = string("int32")]; + int32 const_0 = const()[name = string("const_0"), val = int32(0)]; + int32 const_1 = const()[name = string("const_1"), val = int32(1)]; + int32 gather_0_cast_uint16_to_int32 = cast(dtype = gather_0_cast_uint16_to_int32_dtype_0, x = gather_0_cast_uint16)[name = string("cast_8")]; + tensor var_43 = range_1d(end = gather_0_cast_uint16_to_int32, start = const_0, step = const_1)[name = string("op_43")]; + tensor var_44_axes_0 = const()[name = string("op_44_axes_0"), val = tensor([0])]; + tensor var_44 = expand_dims(axes = var_44_axes_0, x = var_43)[name = string("op_44")]; + tensor var_45_axes_0 = const()[name = string("op_45_axes_0"), val = tensor([1])]; + tensor var_45 = expand_dims(axes = var_45_axes_0, x = audio_length)[name = string("op_45")]; + tensor timemask = less(x = var_44, y = var_45)[name = string("timemask")]; + tensor var_48_begin_0 = const()[name = string("op_48_begin_0"), val = tensor([0, 0])]; + tensor var_48_end_0 = const()[name = string("op_48_end_0"), val = tensor([1, 1])]; + tensor var_48_end_mask_0 = const()[name = string("op_48_end_mask_0"), val = tensor([true, false])]; + tensor var_48_squeeze_mask_0 = const()[name = string("op_48_squeeze_mask_0"), val = tensor([false, true])]; + tensor var_48_cast_fp16 = slice_by_index(begin = var_48_begin_0, end = var_48_end_0, end_mask = var_48_end_mask_0, squeeze_mask = var_48_squeeze_mask_0, x = audio_to_fp16)[name = string("op_48_cast_fp16")]; + tensor var_49_axes_0 = const()[name = string("op_49_axes_0"), val = tensor([1])]; + tensor var_49_cast_fp16 = expand_dims(axes = var_49_axes_0, x = var_48_cast_fp16)[name = string("op_49_cast_fp16")]; + tensor var_51_begin_0 = const()[name = string("op_51_begin_0"), val = tensor([0, 1])]; + tensor var_51_end_0 = const()[name = string("op_51_end_0"), val = tensor([1, 0])]; + tensor var_51_end_mask_0 = const()[name = string("op_51_end_mask_0"), val = tensor([true, true])]; + tensor var_51_cast_fp16 = slice_by_index(begin = var_51_begin_0, end = var_51_end_0, end_mask = var_51_end_mask_0, x = audio_to_fp16)[name = string("op_51_cast_fp16")]; + tensor var_53_begin_0 = const()[name = string("op_53_begin_0"), val = tensor([0, 0])]; + tensor var_53_end_0 = const()[name = string("op_53_end_0"), val = tensor([1, -1])]; + tensor var_53_end_mask_0 = const()[name = string("op_53_end_mask_0"), val = tensor([true, false])]; + tensor var_53_cast_fp16 = slice_by_index(begin = var_53_begin_0, end = var_53_end_0, end_mask = var_53_end_mask_0, x = audio_to_fp16)[name = string("op_53_cast_fp16")]; + fp16 var_54_to_fp16 = const()[name = string("op_54_to_fp16"), val = fp16(0x1.f0cp-1)]; + tensor var_55_cast_fp16 = mul(x = var_53_cast_fp16, y = var_54_to_fp16)[name = string("op_55_cast_fp16")]; + tensor var_56_cast_fp16 = sub(x = var_51_cast_fp16, y = var_55_cast_fp16)[name = string("op_56_cast_fp16")]; + bool x_3_interleave_0 = const()[name = string("x_3_interleave_0"), val = bool(false)]; + tensor x_3_cast_fp16 = concat(axis = var_9, interleave = x_3_interleave_0, values = (var_49_cast_fp16, var_56_cast_fp16))[name = string("x_3_cast_fp16")]; + tensor var_59 = logical_not(x = timemask)[name = string("op_59")]; + fp16 var_16_to_fp16 = const()[name = string("op_16_to_fp16"), val = fp16(0x0p+0)]; + tensor input_1_cast_fp16 = select(a = var_16_to_fp16, b = x_3_cast_fp16, cond = var_59)[name = string("input_1_cast_fp16")]; + tensor concat_1x = const()[name = string("concat_1x"), val = tensor([1, 1, -1])]; + tensor input_3_cast_fp16 = reshape(shape = concat_1x, x = input_1_cast_fp16)[name = string("input_3_cast_fp16")]; + tensor input_5_pad_0 = const()[name = string("input_5_pad_0"), val = tensor([0, 0, 0, 0, 256, 256])]; + string input_5_mode_0 = const()[name = string("input_5_mode_0"), val = string("constant")]; + fp16 const_3_to_fp16 = const()[name = string("const_3_to_fp16"), val = fp16(0x0p+0)]; + tensor input_5_cast_fp16 = pad(constant_val = const_3_to_fp16, mode = input_5_mode_0, pad = input_5_pad_0, x = input_3_cast_fp16)[name = string("input_5_cast_fp16")]; + tensor concat_2x = const()[name = string("concat_2x"), val = tensor([1, -1])]; + tensor input_cast_fp16 = reshape(shape = concat_2x, x = input_5_cast_fp16)[name = string("input_cast_fp16")]; + tensor expand_dims_3 = const()[name = string("expand_dims_3"), val = tensor([160])]; + tensor expand_dims_4_axes_0 = const()[name = string("expand_dims_4_axes_0"), val = tensor([1])]; + tensor expand_dims_4_cast_fp16 = expand_dims(axes = expand_dims_4_axes_0, x = input_cast_fp16)[name = string("expand_dims_4_cast_fp16")]; + string conv_0_pad_type_0 = const()[name = string("conv_0_pad_type_0"), val = string("valid")]; + tensor conv_0_pad_0 = const()[name = string("conv_0_pad_0"), val = tensor([0, 0])]; + tensor conv_0_dilations_0 = const()[name = string("conv_0_dilations_0"), val = tensor([1])]; + int32 conv_0_groups_0 = const()[name = string("conv_0_groups_0"), val = int32(1)]; + tensor expand_dims_1_to_fp16 = const()[name = string("expand_dims_1_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))]; + tensor conv_0_cast_fp16 = conv(dilations = conv_0_dilations_0, groups = conv_0_groups_0, pad = conv_0_pad_0, pad_type = conv_0_pad_type_0, strides = expand_dims_3, weight = expand_dims_1_to_fp16, x = expand_dims_4_cast_fp16)[name = string("conv_0_cast_fp16")]; + string conv_1_pad_type_0 = const()[name = string("conv_1_pad_type_0"), val = string("valid")]; + tensor conv_1_pad_0 = const()[name = string("conv_1_pad_0"), val = tensor([0, 0])]; + tensor conv_1_dilations_0 = const()[name = string("conv_1_dilations_0"), val = tensor([1])]; + int32 conv_1_groups_0 = const()[name = string("conv_1_groups_0"), val = int32(1)]; + tensor expand_dims_2_to_fp16 = const()[name = string("expand_dims_2_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(263296)))]; + tensor conv_1_cast_fp16 = conv(dilations = conv_1_dilations_0, groups = conv_1_groups_0, pad = conv_1_pad_0, pad_type = conv_1_pad_type_0, strides = expand_dims_3, weight = expand_dims_2_to_fp16, x = expand_dims_4_cast_fp16)[name = string("conv_1_cast_fp16")]; + int32 stack_0_axis_0 = const()[name = string("stack_0_axis_0"), val = int32(-1)]; + tensor stack_0_cast_fp16 = stack(axis = stack_0_axis_0, values = (conv_0_cast_fp16, conv_1_cast_fp16))[name = string("stack_0_cast_fp16")]; + fp16 var_19_promoted_to_fp16 = const()[name = string("op_19_promoted_to_fp16"), val = fp16(0x1p+1)]; + tensor var_74_cast_fp16 = pow(x = stack_0_cast_fp16, y = var_19_promoted_to_fp16)[name = string("op_74_cast_fp16")]; + tensor var_76_axes_0 = const()[name = string("op_76_axes_0"), val = tensor([-1])]; + bool var_76_keep_dims_0 = const()[name = string("op_76_keep_dims_0"), val = bool(false)]; + tensor var_76_cast_fp16 = reduce_sum(axes = var_76_axes_0, keep_dims = var_76_keep_dims_0, x = var_74_cast_fp16)[name = string("op_76_cast_fp16")]; + tensor x_11_cast_fp16 = identity(x = var_76_cast_fp16)[name = string("x_11_cast_fp16")]; + bool x_13_transpose_x_0 = const()[name = string("x_13_transpose_x_0"), val = bool(false)]; + bool x_13_transpose_y_0 = const()[name = string("x_13_transpose_y_0"), val = bool(false)]; + tensor const_4_to_fp16 = const()[name = string("const_4_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(526528)))]; + tensor x_13_cast_fp16 = matmul(transpose_x = x_13_transpose_x_0, transpose_y = x_13_transpose_y_0, x = const_4_to_fp16, y = x_11_cast_fp16)[name = string("x_13_cast_fp16")]; + fp16 var_83_to_fp16 = const()[name = string("op_83_to_fp16"), val = fp16(0x1p-24)]; + tensor var_84_cast_fp16 = add(x = x_13_cast_fp16, y = var_83_to_fp16)[name = string("op_84_cast_fp16")]; + fp32 x_epsilon_0 = const()[name = string("x_epsilon_0"), val = fp32(0x1p-149)]; + tensor x_cast_fp16 = log(epsilon = x_epsilon_0, x = var_84_cast_fp16)[name = string("x_cast_fp16")]; + tensor var_86_shape_cast_fp16 = shape(x = x_cast_fp16)[name = string("op_86_shape_cast_fp16")]; + int32 gather_5_batch_dims_0 = const()[name = string("gather_5_batch_dims_0"), val = int32(0)]; + bool gather_5_validate_indices_0 = const()[name = string("gather_5_validate_indices_0"), val = bool(false)]; + string var_86_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_86_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")]; + int32 gather_5_cast_uint16_axis_0 = const()[name = string("gather_5_cast_uint16_axis_0"), val = int32(0)]; + uint16 select_0_to_uint16 = const()[name = string("select_0_to_uint16"), val = uint16(2)]; + tensor var_86_shape_cast_fp16_to_uint16 = cast(dtype = var_86_shape_cast_fp16_to_uint16_dtype_0, x = var_86_shape_cast_fp16)[name = string("cast_7")]; + uint16 gather_5_cast_uint16_cast_uint16 = gather(axis = gather_5_cast_uint16_axis_0, batch_dims = gather_5_batch_dims_0, indices = select_0_to_uint16, validate_indices = gather_5_validate_indices_0, x = var_86_shape_cast_fp16_to_uint16)[name = string("gather_5_cast_uint16_cast_uint16")]; + string gather_5_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_5_cast_uint16_to_int32_dtype_0"), val = string("int32")]; + int32 const_5 = const()[name = string("const_5"), val = int32(0)]; + int32 const_6 = const()[name = string("const_6"), val = int32(1)]; + int32 gather_5_cast_uint16_to_int32 = cast(dtype = gather_5_cast_uint16_to_int32_dtype_0, x = gather_5_cast_uint16_cast_uint16)[name = string("cast_6")]; + tensor mask_1 = range_1d(end = gather_5_cast_uint16_to_int32, start = const_5, step = const_6)[name = string("mask_1")]; + tensor expand_dims_0_axes_0 = const()[name = string("expand_dims_0_axes_0"), val = tensor([0])]; + tensor expand_dims_0 = expand_dims(axes = expand_dims_0_axes_0, x = mask_1)[name = string("expand_dims_0")]; + tensor var_91_axes_0 = const()[name = string("op_91_axes_0"), val = tensor([1])]; + tensor var_91 = expand_dims(axes = var_91_axes_0, x = mel_length)[name = string("op_91")]; + tensor mask = greater_equal(x = expand_dims_0, y = var_91)[name = string("mask")]; + tensor var_93_axes_0 = const()[name = string("op_93_axes_0"), val = tensor([1])]; + tensor var_93 = expand_dims(axes = var_93_axes_0, x = mask)[name = string("op_93")]; + tensor processed_signal_cast_fp16 = select(a = var_16_to_fp16, b = x_cast_fp16, cond = var_93)[name = string("processed_signal_cast_fp16")]; + string processed_signal_cast_fp16_to_fp32_dtype_0 = const()[name = string("processed_signal_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + tensor mel = cast(dtype = processed_signal_cast_fp16_to_fp32_dtype_0, x = processed_signal_cast_fp16)[name = string("cast_5")]; + } -> (mel, mel_length); +} \ No newline at end of file diff --git a/it/560ms/preprocessor.mlmodelc/weights/weight.bin b/it/560ms/preprocessor.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..86dd375f6649d262d58c9cf8b89006ceab216411 --- /dev/null +++ b/it/560ms/preprocessor.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:297514e2b211d14b0e53cb97193d679bb89ead98d28e578f3f1d049ddbcc36b3 +size 592384 diff --git a/it/560ms/preprocessor.mlpackage/Data/com.apple.CoreML/model.mlmodel b/it/560ms/preprocessor.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..050fa97ca7a2aa4b7c4fa318f4fa2a51914287c4 --- /dev/null +++ b/it/560ms/preprocessor.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0feb90f2cf9a5bc749fa763b6eb78a23755b6184b41a262fe08faebe4e709b3e +size 16035 diff --git a/it/560ms/preprocessor.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/it/560ms/preprocessor.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..86dd375f6649d262d58c9cf8b89006ceab216411 --- /dev/null +++ b/it/560ms/preprocessor.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:297514e2b211d14b0e53cb97193d679bb89ead98d28e578f3f1d049ddbcc36b3 +size 592384 diff --git a/it/560ms/preprocessor.mlpackage/Manifest.json b/it/560ms/preprocessor.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..11f4407dd2c2a1fafa63dd69a93da5d4e2866fba --- /dev/null +++ b/it/560ms/preprocessor.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "571B03EE-25FB-448D-B106-8C0685101326": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + }, + "B1E3045B-7172-4628-BD63-F2960CEFBA16": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + } + }, + "rootModelIdentifier": "571B03EE-25FB-448D-B106-8C0685101326" +} diff --git a/it/560ms/tokenizer.json b/it/560ms/tokenizer.json new file mode 100644 index 0000000000000000000000000000000000000000..033007249c80c55ca310e0c90339004b867968f6 --- /dev/null +++ b/it/560ms/tokenizer.json @@ -0,0 +1,808 @@ +{ + "0": "", + "1": "", + "2": "▁", + "3": ".", + "4": ",", + "5": "e", + "6": "t", + "7": "a", + "8": "s", + "9": "o", + "10": "i", + "11": "r", + "12": "l", + "13": "u", + "14": "d", + "15": "c", + "16": "h", + "17": "m", + "18": "p", + "19": "n", + "20": "g", + "21": "f", + "22": "en", + "23": "in", + "24": "on", + "25": "y", + "26": "er", + "27": "an", + "28": "w", + "29": "0", + "30": "b", + "31": "v", + "32": "2", + "33": "1", + "34": "k", + "35": "3", + "36": "5", + "37": "q", + "38": "I", + "39": "é", + "40": "4", + "41": "z", + "42": "A", + "43": "6", + "44": "j", + "45": "E", + "46": "7", + "47": "8", + "48": "9", + "49": "S", + "50": "x", + "51": "C", + "52": "P", + "53": "U", + "54": "N", + "55": "O", + "56": "à", + "57": "L", + "58": "è", + "59": "V", + "60": "R", + "61": "ó", + "62": "J", + "63": "", + "64": "st", + "65": "ch", + "66": "le", + "67": "li", + "68": "▁po", + "69": "no", + "70": "to", + "71": "me", + "72": "te", + "73": "ho", + "74": "▁pro", + "75": "ro", + "76": "▁na", + "77": "ce", + "78": "la", + "79": "ni", + "80": "ra", + "81": "ti", + "82": "lo", + "83": "ko", + "84": "po", + "85": "je", + "86": "de", + "87": "na", + "88": "mi", + "89": "ci", + "90": "▁by", + "91": "ve", + "92": "▁za", + "93": "▁A", + "94": "re", + "95": "ou", + "96": "vo", + "97": "né", + "98": "va", + "99": "mo", + "100": "ze", + "101": "ne", + "102": "ka", + "103": "ky", + "104": "ovat", + "105": "▁jak", + "106": "ny", + "107": "vi", + "108": "ent", + "109": "prav", + "110": "Z", + "111": "▁Je", + "112": "oval", + "113": "ă", + "114": "X", + "115": "", + "116": "▁for", + "117": "▁det", + "118": "▁at", + "119": "et", + "120": "▁og", + "121": "▁vi", + "122": "al", + "123": "▁de", + "124": "▁der", + "125": "or", + "126": "om", + "127": "and", + "128": "▁har", + "129": "at", + "130": "▁af", + "131": "ge", + "132": "ar", + "133": "is", + "134": "▁med", + "135": "▁be", + "136": "un", + "137": "lig", + "138": "▁man", + "139": "ig", + "140": "▁som", + "141": "▁Og", + "142": "el", + "143": "ag", + "144": "erne", + "145": "▁den", + "146": "ste", + "147": "id", + "148": "▁kan", + "149": "ske", + "150": "iv", + "151": "ion", + "152": "am", + "153": "ur", + "154": "for", + "155": "else", + "156": "▁sig", + "157": "▁ind", + "158": "ende", + "159": "▁Vi", + "160": "ation", + "161": "mme", + "162": "▁op", + "163": "▁fra", + "164": "▁alle", + "165": "▁Men", + "166": "▁var", + "167": "", + "168": "▁die", + "169": "▁und", + "170": "sch", + "171": "it", + "172": "gen", + "173": "▁W", + "174": "▁B", + "175": "▁E", + "176": "▁F", + "177": "ll", + "178": "▁es", + "179": "▁K", + "180": "ie", + "181": "au", + "182": "ich", + "183": "ck", + "184": "ten", + "185": "mal", + "186": "ein", + "187": "▁T", + "188": "▁dann", + "189": "▁mit", + "190": "ter", + "191": "tz", + "192": "▁G", + "193": "ben", + "194": "um", + "195": "us", + "196": "il", + "197": "ut", + "198": "▁ver", + "199": "ri", + "200": "ach", + "201": "ol", + "202": "▁Da", + "203": "sp", + "204": "ell", + "205": "▁was", + "206": "▁ja", + "207": "wi", + "208": "rei", + "209": "▁Ge", + "210": "und", + "211": "▁St", + "212": "▁sie", + "213": "▁Ja", + "214": "▁du", + "215": "▁Sch", + "216": "▁Ma", + "217": "▁De", + "218": "▁Sie", + "219": "▁vor", + "220": "▁Le", + "221": "▁In", + "222": "▁Ver", + "223": "▁Re", + "224": "▁Mi", + "225": "▁Ha", + "226": "", + "227": "", + "228": "ma", + "229": "ta", + "230": "se", + "231": "da", + "232": "si", + "233": "ks", + "234": "ga", + "235": "he", + "236": "mu", + "237": "tu", + "238": "ha", + "239": "ja", + "240": "gi", + "241": "gu", + "242": "ju", + "243": "est", + "244": "▁pa", + "245": "tud", + "246": "nda", + "247": "ke", + "248": "sta", + "249": "sed", + "250": "di", + "251": "▁su", + "252": "ide", + "253": "pool", + "254": "val", + "255": "▁Me", + "256": "ment", + "257": "ndus", + "258": "Q", + "259": "", + "260": "ssa", + "261": "lla", + "262": "ki", + "263": "pa", + "264": "lle", + "265": "lu", + "266": "tta", + "267": "isi", + "268": "tte", + "269": "ista", + "270": "llis", + "271": "vu", + "272": "▁voi", + "273": "utta", + "274": "iden", + "275": "parlament", + "276": "", + "277": "▁est", + "278": "▁c", + "279": "▁que", + "280": "es", + "281": "▁un", + "282": "▁pas", + "283": "▁qui", + "284": "▁il", + "285": "▁des", + "286": "▁qu", + "287": "▁par", + "288": "ant", + "289": "▁C", + "290": "tre", + "291": "ir", + "292": "elle", + "293": "eur", + "294": "▁sur", + "295": "▁con", + "296": "ement", + "297": "tion", + "298": "mp", + "299": "▁comme", + "300": "ac", + "301": "▁là", + "302": "che", + "303": "que", + "304": "ul", + "305": "▁bien", + "306": "age", + "307": "▁mon", + "308": "end", + "309": "ver", + "310": "tra", + "311": "ille", + "312": "ff", + "313": "▁ex", + "314": "▁Il", + "315": "im", + "316": "▁dire", + "317": "ance", + "318": "aire", + "319": "▁app", + "320": "onne", + "321": "mb", + "322": "man", + "323": "▁quand", + "324": "port", + "325": "form", + "326": "ture", + "327": "ù", + "328": "", + "329": "ok", + "330": "gy", + "331": "ek", + "332": "em", + "333": "▁is", + "334": "os", + "335": "ak", + "336": "ban", + "337": "ik", + "338": "▁nem", + "339": "oz", + "340": "cs", + "341": "nek", + "342": "bb", + "343": "▁ha", + "344": "ott", + "345": "▁van", + "346": "▁fel", + "347": "leg", + "348": "▁ami", + "349": "tart", + "350": "rend", + "351": "▁fog", + "352": "▁volt", + "353": "", + "354": "▁bi", + "355": "▁sa", + "356": "ru", + "357": "go", + "358": "sti", + "359": "▁pri", + "360": "ima", + "361": "nu", + "362": "▁pre", + "363": "zi", + "364": "vr", + "365": "ca", + "366": "ba", + "367": "▁raz", + "368": "", + "369": "▁di", + "370": "▁che", + "371": "▁è", + "372": "co", + "373": "▁per", + "374": "▁non", + "375": "do", + "376": "gli", + "377": "so", + "378": "amo", + "379": "sa", + "380": "ndo", + "381": "▁una", + "382": "fi", + "383": "pi", + "384": "nti", + "385": "tto", + "386": "tro", + "387": "▁fa", + "388": "chi", + "389": "▁qua", + "390": "zione", + "391": "bi", + "392": "▁del", + "393": "mente", + "394": "pe", + "395": "ssi", + "396": "▁sono", + "397": "▁questo", + "398": "nte", + "399": "tti", + "400": "tà", + "401": "▁nel", + "402": "▁anche", + "403": "sso", + "404": "▁perché", + "405": "▁più", + "406": "nta", + "407": "▁come", + "408": "cu", + "409": "▁quindi", + "410": "ggi", + "411": "nza", + "412": "sto", + "413": "ò", + "414": "▁della", + "415": "gra", + "416": "▁fare", + "417": "spe", + "418": "cco", + "419": "nde", + "420": "mento", + "421": "fe", + "422": "gio", + "423": "pu", + "424": "▁questa", + "425": "zza", + "426": "sci", + "427": "▁dei", + "428": "▁poi", + "429": "sco", + "430": "stra", + "431": "▁quel", + "432": "qui", + "433": "▁delle", + "434": "▁cosa", + "435": "▁molto", + "436": "sse", + "437": "zioni", + "438": "▁inter", + "439": "sce", + "440": "▁fatto", + "441": "▁com", + "442": "▁quello", + "443": "▁essere", + "444": "▁due", + "445": "▁abbiamo", + "446": "▁comp", + "447": "▁tutti", + "448": "ì", + "449": "▁prima", + "450": "▁parte", + "451": "▁così", + "452": "▁sempre", + "453": "▁tutto", + "454": "▁video", + "455": "▁imp", + "456": "▁cui", + "457": "▁dove", + "458": "▁Quindi", + "459": "sione", + "460": "rebbe", + "461": "scri", + "462": "", + "463": "ai", + "464": "▁ir", + "465": "as", + "466": "▁tai", + "467": "uo", + "468": "tin", + "469": "▁vis", + "470": "ly", + "471": "gal", + "472": "tar", + "473": "▁Ir", + "474": "▁turi", + "475": "▁Tai", + "476": "▁nu", + "477": "▁mes", + "478": "išk", + "479": "imas", + "480": "▁pra", + "481": "", + "482": "iem", + "483": "▁pie", + "484": "ies", + "485": "ot", + "486": "▁vien", + "487": "▁Un", + "488": "iet", + "489": "▁Ta", + "490": "dar", + "491": "sim", + "492": "gan", + "493": "▁ap", + "494": "▁nav", + "495": "▁Nu", + "496": "▁cit", + "497": "▁Ne", + "498": "▁20", + "499": "", + "500": "▁dat", + "501": "▁we", + "502": "▁En", + "503": "▁dan", + "504": "▁zo", + "505": "▁met", + "506": "▁wat", + "507": "der", + "508": "ui", + "509": "den", + "510": "op", + "511": "oor", + "512": "▁of", + "513": "ven", + "514": "acht", + "515": "▁even", + "516": "▁wil", + "517": "vol", + "518": "", + "519": "nie", + "520": "wa", + "521": "cie", + "522": "nia", + "523": "wo", + "524": "rze", + "525": "by", + "526": "za", + "527": "dy", + "528": "ry", + "529": "ego", + "530": "mie", + "531": "rz", + "532": "pie", + "533": "▁bo", + "534": "bie", + "535": "▁Po", + "536": "ski", + "537": "nego", + "538": "▁No", + "539": "▁Na", + "540": "▁Was", + "541": "▁musi", + "542": "", + "543": "-", + "544": "▁para", + "545": "▁uma", + "546": "▁pe", + "547": "▁tem", + "548": "▁gente", + "549": "▁O", + "550": "▁ele", + "551": "pre", + "552": "ria", + "553": "▁fo", + "554": "mos", + "555": "bo", + "556": "▁por", + "557": "nto", + "558": "be", + "559": "▁esse", + "560": "ente", + "561": "▁essa", + "562": "▁mas", + "563": "qua", + "564": "fica", + "565": "▁Se", + "566": "▁Por", + "567": "▁Co", + "568": "iza", + "569": "▁sua", + "570": "▁quando", + "571": "", + "572": "▁mai", + "573": "sc", + "574": "are", + "575": "▁din", + "576": "▁este", + "577": "rea", + "578": "ele", + "579": "du", + "580": "▁M", + "581": "▁fac", + "582": "lor", + "583": "▁mult", + "584": "per", + "585": "cur", + "586": "tor", + "587": "inte", + "588": "▁sau", + "589": "tat", + "590": "ori", + "591": "▁prim", + "592": "▁spun", + "593": "▁lui", + "594": "▁sub", + "595": "itate", + "596": "▁prin", + "597": "▁alt", + "598": "stru", + "599": "▁vede", + "600": "fer", + "601": "▁chiar", + "602": "", + "603": "", + "604": "ov", + "605": "ob", + "606": "▁bol", + "607": "ali", + "608": "rov", + "609": "rob", + "610": "▁spo", + "611": "osti", + "612": "", + "613": "sl", + "614": "udi", + "615": "del", + "616": "▁sem", + "617": "▁samo", + "618": "▁pred", + "619": "nost", + "620": "▁Pre", + "621": "▁prot", + "622": "▁internet", + "623": "▁film", + "624": "", + "625": "▁att", + "626": "▁inte", + "627": "▁av", + "628": "all", + "629": "era", + "630": "pp", + "631": "▁upp", + "632": "isk", + "633": "het", + "634": "▁vill", + "635": "erna", + "636": "ande", + "637": "ade", + "638": "bil", + "639": "▁min", + "640": "▁alla", + "641": "lev", + "642": "▁oss", + "643": "land", + "644": "▁Vad", + "645": "person", + "646": "", + "647": "", + "648": "vy", + "649": "ft", + "650": "lige", + "651": "ved", + "652": "'", + "653": "▁H", + "654": "▁D", + "655": "aus", + "656": "▁N", + "657": "▁Be", + "658": "mm", + "659": "ab", + "660": "▁Er", + "661": "ssen", + "662": "rie", + "663": "lei", + "664": "▁An", + "665": "rau", + "666": "▁So", + "667": "", + "668": "▁and", + "669": "▁can", + "670": "ed", + "671": "ay", + "672": "th", + "673": "ic", + "674": "hi", + "675": "▁Oh", + "676": "▁not", + "677": "ight", + "678": "ex", + "679": "▁great", + "680": "ill", + "681": "▁don", + "682": "▁problem", + "683": "▁fine", + "684": "▁month", + "685": "▁check", + "686": "▁zero", + "687": "▁first", + "688": "▁question", + "689": "", + "690": "ive", + "691": "ate", + "692": "ad", + "693": "ng", + "694": "ity", + "695": "ther", + "696": "act", + "697": "side", + "698": "\"", + "699": "", + "700": "ción", + "701": "▁Es", + "702": "res", + "703": "▁La", + "704": "dos", + "705": "▁El", + "706": "▁las", + "707": "men", + "708": "par", + "709": "rio", + "710": "enta", + "711": "▁Ca", + "712": "▁Su", + "713": "▁son", + "714": "ncia", + "715": "▁Con", + "716": "ones", + "717": "▁San", + "718": "▁persona", + "719": "▁Com", + "720": "", + "721": "cia", + "722": "▁Y", + "723": "ron", + "724": "les", + "725": "cio", + "726": "bu", + "727": "", + "728": "ré", + "729": "▁Les", + "730": "our", + "731": "▁Ce", + "732": "com", + "733": "ale", + "734": "if", + "735": "iste", + "736": "▁parti", + "737": "avec", + "738": "app", + "739": "gue", + "740": "▁grand", + "741": "Une", + "742": "È", + "743": "av", + "744": "pri", + "745": "sion", + "746": "ard", + "747": "", + "748": "", + "749": "!", + "750": "", + "751": "", + "752": "", + "753": "ene", + "754": "opp", + "755": "▁han", + "756": "", + "757": "eg", + "758": "kk", + "759": "▁god", + "760": "dde", + "761": "inn", + "762": "dig", + "763": "ord", + "764": "▁tru", + "765": "▁sei", + "766": "ller", + "767": "car", + "768": "ito", + "769": "ram", + "770": "fa", + "771": "▁mil", + "772": "▁passa", + "773": "▁casa", + "774": "", + "775": "▁Pa", + "776": "tura", + "777": "forma", + "778": "tua", + "779": "mar", + "780": "este", + "781": "fun", + "782": "gua", + "783": "▁grande", + "784": "▁nome", + "785": "▁Sua", + "786": "var", + "787": "", + "788": "", + "789": "ş", + "790": "ğ", + "791": "ya", + "792": "▁ve", + "793": "lar", + "794": "ler", + "795": "leri", + "796": "▁bu", + "797": "lan", + "798": "ara", + "799": "▁Bu", + "800": "yo", + "801": "", + "802": "", + "803": "▁t", + "804": "nh", + "805": "" +} \ No newline at end of file diff --git a/ja/1120ms/decoder.mlmodelc/analytics/coremldata.bin b/ja/1120ms/decoder.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..a23f14dd8e4d2bccc2844d3d81c6c9ca86ea3cba --- /dev/null +++ b/ja/1120ms/decoder.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2fcae710f3db79230f47be6daadc8af085539067285a96f89b2a4c0fd0cb3808 +size 243 diff --git a/ja/1120ms/decoder.mlmodelc/coremldata.bin b/ja/1120ms/decoder.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..89ec734bb4645199981d835aff20eb64bd9e3c4e --- /dev/null +++ b/ja/1120ms/decoder.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:298eeaf999bc86b5914efa85450328efc8cf08459f1e8edaefd676f7d5a8410c +size 493 diff --git a/ja/1120ms/decoder.mlmodelc/model.mil b/ja/1120ms/decoder.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..9f4b6b4cebc16f759164ca05a77b06fb57dedbce --- /dev/null +++ b/ja/1120ms/decoder.mlmodelc/model.mil @@ -0,0 +1,73 @@ +program(1.3) +[buildInfo = dict({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.10.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor c_in, tensor h_in, tensor token, tensor token_length) { + int32 y_batch_dims_0 = const()[name = string("y_batch_dims_0"), val = int32(0)]; + bool y_validate_indices_0 = const()[name = string("y_validate_indices_0"), val = bool(false)]; + tensor module_prediction_embed_weight_to_fp16 = const()[name = string("module_prediction_embed_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))]; + string token_to_int16_dtype_0 = const()[name = string("token_to_int16_dtype_0"), val = string("int16")]; + string cast_1_dtype_0 = const()[name = string("cast_1_dtype_0"), val = string("int32")]; + int32 greater_equal_0_y_0 = const()[name = string("greater_equal_0_y_0"), val = int32(0)]; + tensor token_to_int16 = cast(dtype = token_to_int16_dtype_0, x = token)[name = string("cast_9")]; + tensor cast_1 = cast(dtype = cast_1_dtype_0, x = token_to_int16)[name = string("cast_8")]; + tensor greater_equal_0 = greater_equal(x = cast_1, y = greater_equal_0_y_0)[name = string("greater_equal_0")]; + int32 slice_by_index_0 = const()[name = string("slice_by_index_0"), val = int32(1404)]; + tensor add_2 = add(x = cast_1, y = slice_by_index_0)[name = string("add_2")]; + tensor select_0 = select(a = cast_1, b = add_2, cond = greater_equal_0)[name = string("select_0")]; + int32 y_cast_fp16_cast_uint16_axis_0 = const()[name = string("y_cast_fp16_cast_uint16_axis_0"), val = int32(0)]; + string select_0_to_int16_dtype_0 = const()[name = string("select_0_to_int16_dtype_0"), val = string("int16")]; + tensor select_0_to_int16 = cast(dtype = select_0_to_int16_dtype_0, x = select_0)[name = string("cast_7")]; + tensor y_cast_fp16_cast_uint16_cast_uint16 = gather(axis = y_cast_fp16_cast_uint16_axis_0, batch_dims = y_batch_dims_0, indices = select_0_to_int16, validate_indices = y_validate_indices_0, x = module_prediction_embed_weight_to_fp16)[name = string("y_cast_fp16_cast_uint16_cast_uint16")]; + tensor input_3_perm_0 = const()[name = string("input_3_perm_0"), val = tensor([1, 0, 2])]; + int32 split_0_num_splits_0 = const()[name = string("split_0_num_splits_0"), val = int32(2)]; + int32 split_0_axis_0 = const()[name = string("split_0_axis_0"), val = int32(0)]; + string h_in_to_fp16_dtype_0 = const()[name = string("h_in_to_fp16_dtype_0"), val = string("fp16")]; + tensor h_in_to_fp16 = cast(dtype = h_in_to_fp16_dtype_0, x = h_in)[name = string("cast_6")]; + tensor split_0_cast_fp16_0, tensor split_0_cast_fp16_1 = split(axis = split_0_axis_0, num_splits = split_0_num_splits_0, x = h_in_to_fp16)[name = string("split_0_cast_fp16")]; + int32 split_1_num_splits_0 = const()[name = string("split_1_num_splits_0"), val = int32(2)]; + int32 split_1_axis_0 = const()[name = string("split_1_axis_0"), val = int32(0)]; + string c_in_to_fp16_dtype_0 = const()[name = string("c_in_to_fp16_dtype_0"), val = string("fp16")]; + tensor c_in_to_fp16 = cast(dtype = c_in_to_fp16_dtype_0, x = c_in)[name = string("cast_5")]; + tensor split_1_cast_fp16_0, tensor split_1_cast_fp16_1 = split(axis = split_1_axis_0, num_splits = split_1_num_splits_0, x = c_in_to_fp16)[name = string("split_1_cast_fp16")]; + tensor input_lstm_layer_0_lstm_h0_squeeze_axes_0 = const()[name = string("input_lstm_layer_0_lstm_h0_squeeze_axes_0"), val = tensor([0])]; + tensor input_lstm_layer_0_lstm_h0_squeeze_cast_fp16 = squeeze(axes = input_lstm_layer_0_lstm_h0_squeeze_axes_0, x = split_0_cast_fp16_0)[name = string("input_lstm_layer_0_lstm_h0_squeeze_cast_fp16")]; + tensor input_lstm_layer_0_lstm_c0_squeeze_axes_0 = const()[name = string("input_lstm_layer_0_lstm_c0_squeeze_axes_0"), val = tensor([0])]; + tensor input_lstm_layer_0_lstm_c0_squeeze_cast_fp16 = squeeze(axes = input_lstm_layer_0_lstm_c0_squeeze_axes_0, x = split_1_cast_fp16_0)[name = string("input_lstm_layer_0_lstm_c0_squeeze_cast_fp16")]; + string input_lstm_layer_0_direction_0 = const()[name = string("input_lstm_layer_0_direction_0"), val = string("forward")]; + bool input_lstm_layer_0_output_sequence_0 = const()[name = string("input_lstm_layer_0_output_sequence_0"), val = bool(true)]; + string input_lstm_layer_0_recurrent_activation_0 = const()[name = string("input_lstm_layer_0_recurrent_activation_0"), val = string("sigmoid")]; + string input_lstm_layer_0_cell_activation_0 = const()[name = string("input_lstm_layer_0_cell_activation_0"), val = string("tanh")]; + string input_lstm_layer_0_activation_0 = const()[name = string("input_lstm_layer_0_activation_0"), val = string("tanh")]; + tensor concat_1_to_fp16 = const()[name = string("concat_1_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1797248)))]; + tensor concat_2_to_fp16 = const()[name = string("concat_2_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5074112)))]; + tensor concat_0_to_fp16 = const()[name = string("concat_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8350976)))]; + tensor input_3_cast_fp16 = transpose(perm = input_3_perm_0, x = y_cast_fp16_cast_uint16_cast_uint16)[name = string("transpose_2")]; + tensor input_lstm_layer_0_cast_fp16_0, tensor input_lstm_layer_0_cast_fp16_1, tensor input_lstm_layer_0_cast_fp16_2 = lstm(activation = input_lstm_layer_0_activation_0, bias = concat_0_to_fp16, cell_activation = input_lstm_layer_0_cell_activation_0, direction = input_lstm_layer_0_direction_0, initial_c = input_lstm_layer_0_lstm_c0_squeeze_cast_fp16, initial_h = input_lstm_layer_0_lstm_h0_squeeze_cast_fp16, output_sequence = input_lstm_layer_0_output_sequence_0, recurrent_activation = input_lstm_layer_0_recurrent_activation_0, weight_hh = concat_2_to_fp16, weight_ih = concat_1_to_fp16, x = input_3_cast_fp16)[name = string("input_lstm_layer_0_cast_fp16")]; + tensor input_lstm_h0_squeeze_axes_0 = const()[name = string("input_lstm_h0_squeeze_axes_0"), val = tensor([0])]; + tensor input_lstm_h0_squeeze_cast_fp16 = squeeze(axes = input_lstm_h0_squeeze_axes_0, x = split_0_cast_fp16_1)[name = string("input_lstm_h0_squeeze_cast_fp16")]; + tensor input_lstm_c0_squeeze_axes_0 = const()[name = string("input_lstm_c0_squeeze_axes_0"), val = tensor([0])]; + tensor input_lstm_c0_squeeze_cast_fp16 = squeeze(axes = input_lstm_c0_squeeze_axes_0, x = split_1_cast_fp16_1)[name = string("input_lstm_c0_squeeze_cast_fp16")]; + string input_direction_0 = const()[name = string("input_direction_0"), val = string("forward")]; + bool input_output_sequence_0 = const()[name = string("input_output_sequence_0"), val = bool(true)]; + string input_recurrent_activation_0 = const()[name = string("input_recurrent_activation_0"), val = string("sigmoid")]; + string input_cell_activation_0 = const()[name = string("input_cell_activation_0"), val = string("tanh")]; + string input_activation_0 = const()[name = string("input_activation_0"), val = string("tanh")]; + tensor concat_4_to_fp16 = const()[name = string("concat_4_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8356160)))]; + tensor concat_5_to_fp16 = const()[name = string("concat_5_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11633024)))]; + tensor concat_3_to_fp16 = const()[name = string("concat_3_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14909888)))]; + tensor input_cast_fp16_0, tensor input_cast_fp16_1, tensor input_cast_fp16_2 = lstm(activation = input_activation_0, bias = concat_3_to_fp16, cell_activation = input_cell_activation_0, direction = input_direction_0, initial_c = input_lstm_c0_squeeze_cast_fp16, initial_h = input_lstm_h0_squeeze_cast_fp16, output_sequence = input_output_sequence_0, recurrent_activation = input_recurrent_activation_0, weight_hh = concat_5_to_fp16, weight_ih = concat_4_to_fp16, x = input_lstm_layer_0_cast_fp16_0)[name = string("input_cast_fp16")]; + int32 obj_3_axis_0 = const()[name = string("obj_3_axis_0"), val = int32(0)]; + tensor obj_3_cast_fp16 = stack(axis = obj_3_axis_0, values = (input_lstm_layer_0_cast_fp16_1, input_cast_fp16_1))[name = string("obj_3_cast_fp16")]; + string obj_3_cast_fp16_to_fp32_dtype_0 = const()[name = string("obj_3_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + int32 obj_axis_0 = const()[name = string("obj_axis_0"), val = int32(0)]; + tensor obj_cast_fp16 = stack(axis = obj_axis_0, values = (input_lstm_layer_0_cast_fp16_2, input_cast_fp16_2))[name = string("obj_cast_fp16")]; + string obj_cast_fp16_to_fp32_dtype_0 = const()[name = string("obj_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + tensor transpose_0_perm_0 = const()[name = string("transpose_0_perm_0"), val = tensor([1, 2, 0])]; + string transpose_0_cast_fp16_to_fp32_dtype_0 = const()[name = string("transpose_0_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + tensor transpose_0_cast_fp16 = transpose(perm = transpose_0_perm_0, x = input_cast_fp16_0)[name = string("transpose_1")]; + tensor decoder_out = cast(dtype = transpose_0_cast_fp16_to_fp32_dtype_0, x = transpose_0_cast_fp16)[name = string("cast_2")]; + tensor c_out = cast(dtype = obj_cast_fp16_to_fp32_dtype_0, x = obj_cast_fp16)[name = string("cast_3")]; + tensor h_out = cast(dtype = obj_3_cast_fp16_to_fp32_dtype_0, x = obj_3_cast_fp16)[name = string("cast_4")]; + tensor token_length_tmp = identity(x = token_length)[name = string("token_length_tmp")]; + } -> (decoder_out, h_out, c_out); +} \ No newline at end of file diff --git a/ja/1120ms/decoder.mlmodelc/weights/weight.bin b/ja/1120ms/decoder.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..9bcbce3e617e1135c0460941dfc723b71230d030 --- /dev/null +++ b/ja/1120ms/decoder.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d5eeb17b5de822e941f66ce3bd6739aae939dfa054f61b69ff6ce89d0ac9e778 +size 14915072 diff --git a/ja/1120ms/decoder.mlpackage/Data/com.apple.CoreML/model.mlmodel b/ja/1120ms/decoder.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..931ee2253d124627cf1b2689c6e01d5cf3746838 --- /dev/null +++ b/ja/1120ms/decoder.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:25c461bd45595f33022b4ce50bf3d493d5b70ae73c50bd0a98598336bd38864a +size 11598 diff --git a/ja/1120ms/decoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/ja/1120ms/decoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..9bcbce3e617e1135c0460941dfc723b71230d030 --- /dev/null +++ b/ja/1120ms/decoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d5eeb17b5de822e941f66ce3bd6739aae939dfa054f61b69ff6ce89d0ac9e778 +size 14915072 diff --git a/ja/1120ms/decoder.mlpackage/Manifest.json b/ja/1120ms/decoder.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..b4431e1d0753f3273eeff30f89de1232349486a7 --- /dev/null +++ b/ja/1120ms/decoder.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "8C20B369-4E12-4E4E-B3E8-A79B91D9CAFC": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + }, + "9356FC01-CF91-4D74-A142-118AF15703DD": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + } + }, + "rootModelIdentifier": "9356FC01-CF91-4D74-A142-118AF15703DD" +} diff --git a/ja/1120ms/decoder_joint.mlmodelc/analytics/coremldata.bin b/ja/1120ms/decoder_joint.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..30ee1bc4e73ed57bede1d9e6315c983146d06e8c --- /dev/null +++ b/ja/1120ms/decoder_joint.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:adba69d0e8e1547064d062072f64e8a9f1da383a6d09e2986a28268dd78cb23c +size 243 diff --git a/ja/1120ms/decoder_joint.mlmodelc/coremldata.bin b/ja/1120ms/decoder_joint.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..ff735d9d794bd3717f0344022771f29df72a633d --- /dev/null +++ b/ja/1120ms/decoder_joint.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:85c85ecec674a9bee9777b7cf93682fd4cb5ea9bed388a030224a6f90dd72cde +size 514 diff --git a/ja/1120ms/decoder_joint.mlmodelc/model.mil b/ja/1120ms/decoder_joint.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..2cb099751b91d7bd911aeac28092cc495bcaf315 --- /dev/null +++ b/ja/1120ms/decoder_joint.mlmodelc/model.mil @@ -0,0 +1,92 @@ +program(1.3) +[buildInfo = dict({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.10.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor c_in, tensor encoder, tensor h_in, tensor token, tensor token_length) { + int32 y_batch_dims_0 = const()[name = string("y_batch_dims_0"), val = int32(0)]; + bool y_validate_indices_0 = const()[name = string("y_validate_indices_0"), val = bool(false)]; + tensor decoder_module_prediction_embed_weight_to_fp16 = const()[name = string("decoder_module_prediction_embed_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))]; + string token_to_int16_dtype_0 = const()[name = string("token_to_int16_dtype_0"), val = string("int16")]; + string cast_1_dtype_0 = const()[name = string("cast_1_dtype_0"), val = string("int32")]; + int32 greater_equal_0_y_0 = const()[name = string("greater_equal_0_y_0"), val = int32(0)]; + tensor token_to_int16 = cast(dtype = token_to_int16_dtype_0, x = token)[name = string("cast_10")]; + tensor cast_1 = cast(dtype = cast_1_dtype_0, x = token_to_int16)[name = string("cast_9")]; + tensor greater_equal_0 = greater_equal(x = cast_1, y = greater_equal_0_y_0)[name = string("greater_equal_0")]; + int32 slice_by_index_0 = const()[name = string("slice_by_index_0"), val = int32(1404)]; + tensor add_2 = add(x = cast_1, y = slice_by_index_0)[name = string("add_2")]; + tensor select_0 = select(a = cast_1, b = add_2, cond = greater_equal_0)[name = string("select_0")]; + int32 y_cast_fp16_cast_uint16_axis_0 = const()[name = string("y_cast_fp16_cast_uint16_axis_0"), val = int32(0)]; + string select_0_to_int16_dtype_0 = const()[name = string("select_0_to_int16_dtype_0"), val = string("int16")]; + tensor select_0_to_int16 = cast(dtype = select_0_to_int16_dtype_0, x = select_0)[name = string("cast_8")]; + tensor y_cast_fp16_cast_uint16_cast_uint16 = gather(axis = y_cast_fp16_cast_uint16_axis_0, batch_dims = y_batch_dims_0, indices = select_0_to_int16, validate_indices = y_validate_indices_0, x = decoder_module_prediction_embed_weight_to_fp16)[name = string("y_cast_fp16_cast_uint16_cast_uint16")]; + tensor input_3_perm_0 = const()[name = string("input_3_perm_0"), val = tensor([1, 0, 2])]; + int32 split_0_num_splits_0 = const()[name = string("split_0_num_splits_0"), val = int32(2)]; + int32 split_0_axis_0 = const()[name = string("split_0_axis_0"), val = int32(0)]; + string h_in_to_fp16_dtype_0 = const()[name = string("h_in_to_fp16_dtype_0"), val = string("fp16")]; + tensor h_in_to_fp16 = cast(dtype = h_in_to_fp16_dtype_0, x = h_in)[name = string("cast_7")]; + tensor split_0_cast_fp16_0, tensor split_0_cast_fp16_1 = split(axis = split_0_axis_0, num_splits = split_0_num_splits_0, x = h_in_to_fp16)[name = string("split_0_cast_fp16")]; + int32 split_1_num_splits_0 = const()[name = string("split_1_num_splits_0"), val = int32(2)]; + int32 split_1_axis_0 = const()[name = string("split_1_axis_0"), val = int32(0)]; + string c_in_to_fp16_dtype_0 = const()[name = string("c_in_to_fp16_dtype_0"), val = string("fp16")]; + tensor c_in_to_fp16 = cast(dtype = c_in_to_fp16_dtype_0, x = c_in)[name = string("cast_6")]; + tensor split_1_cast_fp16_0, tensor split_1_cast_fp16_1 = split(axis = split_1_axis_0, num_splits = split_1_num_splits_0, x = c_in_to_fp16)[name = string("split_1_cast_fp16")]; + tensor input_5_lstm_layer_0_lstm_h0_squeeze_axes_0 = const()[name = string("input_5_lstm_layer_0_lstm_h0_squeeze_axes_0"), val = tensor([0])]; + tensor input_5_lstm_layer_0_lstm_h0_squeeze_cast_fp16 = squeeze(axes = input_5_lstm_layer_0_lstm_h0_squeeze_axes_0, x = split_0_cast_fp16_0)[name = string("input_5_lstm_layer_0_lstm_h0_squeeze_cast_fp16")]; + tensor input_5_lstm_layer_0_lstm_c0_squeeze_axes_0 = const()[name = string("input_5_lstm_layer_0_lstm_c0_squeeze_axes_0"), val = tensor([0])]; + tensor input_5_lstm_layer_0_lstm_c0_squeeze_cast_fp16 = squeeze(axes = input_5_lstm_layer_0_lstm_c0_squeeze_axes_0, x = split_1_cast_fp16_0)[name = string("input_5_lstm_layer_0_lstm_c0_squeeze_cast_fp16")]; + string input_5_lstm_layer_0_direction_0 = const()[name = string("input_5_lstm_layer_0_direction_0"), val = string("forward")]; + bool input_5_lstm_layer_0_output_sequence_0 = const()[name = string("input_5_lstm_layer_0_output_sequence_0"), val = bool(true)]; + string input_5_lstm_layer_0_recurrent_activation_0 = const()[name = string("input_5_lstm_layer_0_recurrent_activation_0"), val = string("sigmoid")]; + string input_5_lstm_layer_0_cell_activation_0 = const()[name = string("input_5_lstm_layer_0_cell_activation_0"), val = string("tanh")]; + string input_5_lstm_layer_0_activation_0 = const()[name = string("input_5_lstm_layer_0_activation_0"), val = string("tanh")]; + tensor concat_1_to_fp16 = const()[name = string("concat_1_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1797248)))]; + tensor concat_2_to_fp16 = const()[name = string("concat_2_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5074112)))]; + tensor concat_0_to_fp16 = const()[name = string("concat_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8350976)))]; + tensor input_3_cast_fp16 = transpose(perm = input_3_perm_0, x = y_cast_fp16_cast_uint16_cast_uint16)[name = string("transpose_4")]; + tensor input_5_lstm_layer_0_cast_fp16_0, tensor input_5_lstm_layer_0_cast_fp16_1, tensor input_5_lstm_layer_0_cast_fp16_2 = lstm(activation = input_5_lstm_layer_0_activation_0, bias = concat_0_to_fp16, cell_activation = input_5_lstm_layer_0_cell_activation_0, direction = input_5_lstm_layer_0_direction_0, initial_c = input_5_lstm_layer_0_lstm_c0_squeeze_cast_fp16, initial_h = input_5_lstm_layer_0_lstm_h0_squeeze_cast_fp16, output_sequence = input_5_lstm_layer_0_output_sequence_0, recurrent_activation = input_5_lstm_layer_0_recurrent_activation_0, weight_hh = concat_2_to_fp16, weight_ih = concat_1_to_fp16, x = input_3_cast_fp16)[name = string("input_5_lstm_layer_0_cast_fp16")]; + tensor input_5_lstm_h0_squeeze_axes_0 = const()[name = string("input_5_lstm_h0_squeeze_axes_0"), val = tensor([0])]; + tensor input_5_lstm_h0_squeeze_cast_fp16 = squeeze(axes = input_5_lstm_h0_squeeze_axes_0, x = split_0_cast_fp16_1)[name = string("input_5_lstm_h0_squeeze_cast_fp16")]; + tensor input_5_lstm_c0_squeeze_axes_0 = const()[name = string("input_5_lstm_c0_squeeze_axes_0"), val = tensor([0])]; + tensor input_5_lstm_c0_squeeze_cast_fp16 = squeeze(axes = input_5_lstm_c0_squeeze_axes_0, x = split_1_cast_fp16_1)[name = string("input_5_lstm_c0_squeeze_cast_fp16")]; + string input_5_direction_0 = const()[name = string("input_5_direction_0"), val = string("forward")]; + bool input_5_output_sequence_0 = const()[name = string("input_5_output_sequence_0"), val = bool(true)]; + string input_5_recurrent_activation_0 = const()[name = string("input_5_recurrent_activation_0"), val = string("sigmoid")]; + string input_5_cell_activation_0 = const()[name = string("input_5_cell_activation_0"), val = string("tanh")]; + string input_5_activation_0 = const()[name = string("input_5_activation_0"), val = string("tanh")]; + tensor concat_4_to_fp16 = const()[name = string("concat_4_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8356160)))]; + tensor concat_5_to_fp16 = const()[name = string("concat_5_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11633024)))]; + tensor concat_3_to_fp16 = const()[name = string("concat_3_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14909888)))]; + tensor input_5_cast_fp16_0, tensor input_5_cast_fp16_1, tensor input_5_cast_fp16_2 = lstm(activation = input_5_activation_0, bias = concat_3_to_fp16, cell_activation = input_5_cell_activation_0, direction = input_5_direction_0, initial_c = input_5_lstm_c0_squeeze_cast_fp16, initial_h = input_5_lstm_h0_squeeze_cast_fp16, output_sequence = input_5_output_sequence_0, recurrent_activation = input_5_recurrent_activation_0, weight_hh = concat_5_to_fp16, weight_ih = concat_4_to_fp16, x = input_5_lstm_layer_0_cast_fp16_0)[name = string("input_5_cast_fp16")]; + int32 obj_3_axis_0 = const()[name = string("obj_3_axis_0"), val = int32(0)]; + tensor obj_3_cast_fp16 = stack(axis = obj_3_axis_0, values = (input_5_lstm_layer_0_cast_fp16_1, input_5_cast_fp16_1))[name = string("obj_3_cast_fp16")]; + string obj_3_cast_fp16_to_fp32_dtype_0 = const()[name = string("obj_3_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + int32 obj_axis_0 = const()[name = string("obj_axis_0"), val = int32(0)]; + tensor obj_cast_fp16 = stack(axis = obj_axis_0, values = (input_5_lstm_layer_0_cast_fp16_2, input_5_cast_fp16_2))[name = string("obj_cast_fp16")]; + string obj_cast_fp16_to_fp32_dtype_0 = const()[name = string("obj_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + tensor transpose_1_perm_0 = const()[name = string("transpose_1_perm_0"), val = tensor([1, 0, 2])]; + tensor input_7_perm_0 = const()[name = string("input_7_perm_0"), val = tensor([0, 2, 1])]; + string encoder_to_fp16_dtype_0 = const()[name = string("encoder_to_fp16_dtype_0"), val = string("fp16")]; + tensor joint_module_enc_weight_to_fp16 = const()[name = string("joint_module_enc_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14915072)))]; + tensor joint_module_enc_bias_to_fp16 = const()[name = string("joint_module_enc_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16225856)))]; + tensor encoder_to_fp16 = cast(dtype = encoder_to_fp16_dtype_0, x = encoder)[name = string("cast_3")]; + tensor input_7_cast_fp16 = transpose(perm = input_7_perm_0, x = encoder_to_fp16)[name = string("transpose_2")]; + tensor linear_0_cast_fp16 = linear(bias = joint_module_enc_bias_to_fp16, weight = joint_module_enc_weight_to_fp16, x = input_7_cast_fp16)[name = string("linear_0_cast_fp16")]; + tensor joint_module_pred_weight_to_fp16 = const()[name = string("joint_module_pred_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16227200)))]; + tensor joint_module_pred_bias_to_fp16 = const()[name = string("joint_module_pred_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17046464)))]; + tensor transpose_1_cast_fp16 = transpose(perm = transpose_1_perm_0, x = input_5_cast_fp16_0)[name = string("transpose_3")]; + tensor linear_1_cast_fp16 = linear(bias = joint_module_pred_bias_to_fp16, weight = joint_module_pred_weight_to_fp16, x = transpose_1_cast_fp16)[name = string("linear_1_cast_fp16")]; + tensor var_79_axes_0 = const()[name = string("op_79_axes_0"), val = tensor([2])]; + tensor var_79_cast_fp16 = expand_dims(axes = var_79_axes_0, x = linear_0_cast_fp16)[name = string("op_79_cast_fp16")]; + tensor var_80_axes_0 = const()[name = string("op_80_axes_0"), val = tensor([1])]; + tensor var_80_cast_fp16 = expand_dims(axes = var_80_axes_0, x = linear_1_cast_fp16)[name = string("op_80_cast_fp16")]; + tensor input_11_cast_fp16 = add(x = var_79_cast_fp16, y = var_80_cast_fp16)[name = string("input_11_cast_fp16")]; + tensor input_13_cast_fp16 = relu(x = input_11_cast_fp16)[name = string("input_13_cast_fp16")]; + tensor joint_module_joint_net_2_weight_to_fp16 = const()[name = string("joint_module_joint_net_2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17047808)))]; + tensor joint_module_joint_net_2_bias_to_fp16 = const()[name = string("joint_module_joint_net_2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18844992)))]; + tensor linear_2_cast_fp16 = linear(bias = joint_module_joint_net_2_bias_to_fp16, weight = joint_module_joint_net_2_weight_to_fp16, x = input_13_cast_fp16)[name = string("linear_2_cast_fp16")]; + string linear_2_cast_fp16_to_fp32_dtype_0 = const()[name = string("linear_2_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + tensor logits = cast(dtype = linear_2_cast_fp16_to_fp32_dtype_0, x = linear_2_cast_fp16)[name = string("cast_2")]; + tensor c_out = cast(dtype = obj_cast_fp16_to_fp32_dtype_0, x = obj_cast_fp16)[name = string("cast_4")]; + tensor h_out = cast(dtype = obj_3_cast_fp16_to_fp32_dtype_0, x = obj_3_cast_fp16)[name = string("cast_5")]; + tensor token_length_tmp = identity(x = token_length)[name = string("token_length_tmp")]; + } -> (logits, h_out, c_out); +} \ No newline at end of file diff --git a/ja/1120ms/decoder_joint.mlmodelc/weights/weight.bin b/ja/1120ms/decoder_joint.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..c3bb2e494d21dfd602e504fdfe76da274071d914 --- /dev/null +++ b/ja/1120ms/decoder_joint.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bfc1c768dd0e0e61c0ab8806894ecc03902d2a5028e9c30f5d0a5e38d5139fd9 +size 18847864 diff --git a/ja/1120ms/decoder_joint.mlpackage/Data/com.apple.CoreML/model.mlmodel b/ja/1120ms/decoder_joint.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..4b6b70ea10ec445416345cff7ae90b536c38ce69 --- /dev/null +++ b/ja/1120ms/decoder_joint.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dad926552d028aa609be27dde32ede7e7e0687ce48d057ee7bae6d472f15229e +size 14986 diff --git a/ja/1120ms/decoder_joint.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/ja/1120ms/decoder_joint.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..c3bb2e494d21dfd602e504fdfe76da274071d914 --- /dev/null +++ b/ja/1120ms/decoder_joint.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bfc1c768dd0e0e61c0ab8806894ecc03902d2a5028e9c30f5d0a5e38d5139fd9 +size 18847864 diff --git a/ja/1120ms/decoder_joint.mlpackage/Manifest.json b/ja/1120ms/decoder_joint.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..8a2ce629ffcccdfce8e228506afd499877ee4507 --- /dev/null +++ b/ja/1120ms/decoder_joint.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "4DAC4947-89EA-4167-8E25-F5740051D450": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + }, + "D23EE53F-5E4C-405E-B75A-E4553E8229D1": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + } + }, + "rootModelIdentifier": "D23EE53F-5E4C-405E-B75A-E4553E8229D1" +} diff --git a/ja/1120ms/encoder.mlmodelc/analytics/coremldata.bin b/ja/1120ms/encoder.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..cdc6c00cf855a2b827e4760c156a421699a9f39e --- /dev/null +++ b/ja/1120ms/encoder.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0638780288ed25026170632a153dbed5d798954deac843fa58ddaae3190914d4 +size 243 diff --git a/ja/1120ms/encoder.mlmodelc/coremldata.bin b/ja/1120ms/encoder.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..30cc40a7a7368be50bd2b51487f050bfff4a8758 --- /dev/null +++ b/ja/1120ms/encoder.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f68f09c6358cf29875a774c8dee3a5f2b79bf9f85bca6404f10e34a40ac28248 +size 662 diff --git a/ja/1120ms/encoder.mlmodelc/model.mil b/ja/1120ms/encoder.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..8243eb3dcf4d99803bf1e78f45f94ceda08ed8ee --- /dev/null +++ b/ja/1120ms/encoder.mlmodelc/model.mil @@ -0,0 +1,4439 @@ +program(1.3) +[buildInfo = dict({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}})] +{ + func main(tensor cache_channel, tensor cache_len, tensor cache_time, tensor mel, tensor mel_length, tensor prompt_id) { + tensor value_3_perm_0 = const()[name = string("value_3_perm_0"), val = tensor([1, 0, 2, 3])]; + string cache_channel_to_fp16_dtype_0 = const()[name = string("cache_channel_to_fp16_dtype_0"), val = string("fp16")]; + tensor value_5_perm_0 = const()[name = string("value_5_perm_0"), val = tensor([1, 0, 2, 3])]; + string cache_time_to_fp16_dtype_0 = const()[name = string("cache_time_to_fp16_dtype_0"), val = string("fp16")]; + int32 var_60 = const()[name = string("op_60"), val = int32(-1)]; + int32 var_69 = const()[name = string("op_69"), val = int32(1)]; + tensor x_1_perm_0 = const()[name = string("x_1_perm_0"), val = tensor([0, 2, 1])]; + string mel_to_fp16_dtype_0 = const()[name = string("mel_to_fp16_dtype_0"), val = string("fp16")]; + tensor tensor_1_axes_0 = const()[name = string("tensor_1_axes_0"), val = tensor([1])]; + tensor mel_to_fp16 = cast(dtype = mel_to_fp16_dtype_0, x = mel)[name = string("cast_22")]; + tensor x_1_cast_fp16 = transpose(perm = x_1_perm_0, x = mel_to_fp16)[name = string("transpose_367")]; + tensor tensor_1_cast_fp16 = expand_dims(axes = tensor_1_axes_0, x = x_1_cast_fp16)[name = string("tensor_1_cast_fp16")]; + tensor expand_dims_0 = const()[name = string("expand_dims_0"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))]; + tensor var_138_axes_0 = const()[name = string("op_138_axes_0"), val = tensor([1])]; + tensor var_138 = expand_dims(axes = var_138_axes_0, x = mel_length)[name = string("op_138")]; + tensor time_mask_1 = less(x = expand_dims_0, y = var_138)[name = string("time_mask_1")]; + tensor var_140_axes_0 = const()[name = string("op_140_axes_0"), val = tensor([-1])]; + tensor var_140 = expand_dims(axes = var_140_axes_0, x = time_mask_1)[name = string("op_140")]; + tensor var_142_reps_0 = const()[name = string("op_142_reps_0"), val = tensor([1, 1, 128])]; + tensor var_142 = tile(reps = var_142_reps_0, x = var_140)[name = string("op_142")]; + tensor var_148_axes_0 = const()[name = string("op_148_axes_0"), val = tensor([1])]; + string mask_1_to_fp16_dtype_0 = const()[name = string("mask_1_to_fp16_dtype_0"), val = string("fp16")]; + tensor var_142_to_fp16 = cast(dtype = mask_1_to_fp16_dtype_0, x = var_142)[name = string("cast_21")]; + tensor var_148_cast_fp16 = expand_dims(axes = var_148_axes_0, x = var_142_to_fp16)[name = string("op_148_cast_fp16")]; + tensor input_1_cast_fp16 = mul(x = tensor_1_cast_fp16, y = var_148_cast_fp16)[name = string("input_1_cast_fp16")]; + tensor input_3_pad_0 = const()[name = string("input_3_pad_0"), val = tensor([0, 0, 0, 0, 2, 1, 2, 1])]; + string input_3_mode_0 = const()[name = string("input_3_mode_0"), val = string("constant")]; + fp16 const_9_to_fp16 = const()[name = string("const_9_to_fp16"), val = fp16(0x0p+0)]; + tensor input_3_cast_fp16 = pad(constant_val = const_9_to_fp16, mode = input_3_mode_0, pad = input_3_pad_0, x = input_1_cast_fp16)[name = string("input_3_cast_fp16")]; + string tensor_3_pad_type_0 = const()[name = string("tensor_3_pad_type_0"), val = string("valid")]; + tensor tensor_3_strides_0 = const()[name = string("tensor_3_strides_0"), val = tensor([2, 2])]; + tensor tensor_3_pad_0 = const()[name = string("tensor_3_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor tensor_3_dilations_0 = const()[name = string("tensor_3_dilations_0"), val = tensor([1, 1])]; + int32 tensor_3_groups_0 = const()[name = string("tensor_3_groups_0"), val = int32(1)]; + tensor encoder_pre_encode_conv_0_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(640))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3008))))[name = string("encoder_pre_encode_conv_0_weight_to_fp16_quantized")]; + tensor encoder_pre_encode_conv_0_bias_to_fp16 = const()[name = string("encoder_pre_encode_conv_0_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3584)))]; + tensor tensor_3_cast_fp16 = conv(bias = encoder_pre_encode_conv_0_bias_to_fp16, dilations = tensor_3_dilations_0, groups = tensor_3_groups_0, pad = tensor_3_pad_0, pad_type = tensor_3_pad_type_0, strides = tensor_3_strides_0, weight = encoder_pre_encode_conv_0_weight_to_fp16_quantized, x = input_3_cast_fp16)[name = string("tensor_3_cast_fp16")]; + string current_lengths_1_to_fp16_dtype_0 = const()[name = string("current_lengths_1_to_fp16_dtype_0"), val = string("fp16")]; + fp16 var_161_promoted_to_fp16 = const()[name = string("op_161_promoted_to_fp16"), val = fp16(0x1p+1)]; + tensor mel_length_to_fp16 = cast(dtype = current_lengths_1_to_fp16_dtype_0, x = mel_length)[name = string("cast_20")]; + tensor var_162_cast_fp16 = add(x = mel_length_to_fp16, y = var_161_promoted_to_fp16)[name = string("op_162_cast_fp16")]; + fp16 var_163_promoted_to_fp16 = const()[name = string("op_163_promoted_to_fp16"), val = fp16(0x1p+0)]; + tensor var_164_cast_fp16 = add(x = var_162_cast_fp16, y = var_163_promoted_to_fp16)[name = string("op_164_cast_fp16")]; + fp16 var_165_promoted_to_fp16 = const()[name = string("op_165_promoted_to_fp16"), val = fp16(0x1.8p+1)]; + tensor var_166_cast_fp16 = sub(x = var_164_cast_fp16, y = var_165_promoted_to_fp16)[name = string("op_166_cast_fp16")]; + fp16 var_57_promoted_to_fp16 = const()[name = string("op_57_promoted_to_fp16"), val = fp16(0x1p+1)]; + tensor floor_div_0_cast_fp16 = floor_div(x = var_166_cast_fp16, y = var_57_promoted_to_fp16)[name = string("floor_div_0_cast_fp16")]; + fp16 var_168_promoted_to_fp16 = const()[name = string("op_168_promoted_to_fp16"), val = fp16(0x1p+0)]; + tensor current_lengths_3_cast_fp16 = add(x = floor_div_0_cast_fp16, y = var_168_promoted_to_fp16)[name = string("current_lengths_3_cast_fp16")]; + string lengths_19_dtype_0 = const()[name = string("lengths_19_dtype_0"), val = string("int32")]; + tensor expand_dims_1 = const()[name = string("expand_dims_1"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4160)))]; + tensor var_177_axes_0 = const()[name = string("op_177_axes_0"), val = tensor([1])]; + tensor current_lengths_3_cast_fp16_to_int32 = cast(dtype = lengths_19_dtype_0, x = current_lengths_3_cast_fp16)[name = string("cast_19")]; + tensor var_177 = expand_dims(axes = var_177_axes_0, x = current_lengths_3_cast_fp16_to_int32)[name = string("op_177")]; + tensor time_mask_3 = less(x = expand_dims_1, y = var_177)[name = string("time_mask_3")]; + tensor var_179_axes_0 = const()[name = string("op_179_axes_0"), val = tensor([-1])]; + tensor var_179 = expand_dims(axes = var_179_axes_0, x = time_mask_3)[name = string("op_179")]; + tensor var_181_reps_0 = const()[name = string("op_181_reps_0"), val = tensor([1, 1, 65])]; + tensor var_181 = tile(reps = var_181_reps_0, x = var_179)[name = string("op_181")]; + tensor var_187_axes_0 = const()[name = string("op_187_axes_0"), val = tensor([1])]; + string mask_3_to_fp16_dtype_0 = const()[name = string("mask_3_to_fp16_dtype_0"), val = string("fp16")]; + tensor var_181_to_fp16 = cast(dtype = mask_3_to_fp16_dtype_0, x = var_181)[name = string("cast_18")]; + tensor var_187_cast_fp16 = expand_dims(axes = var_187_axes_0, x = var_181_to_fp16)[name = string("op_187_cast_fp16")]; + tensor expanded_mask_3_reps_0 = const()[name = string("expanded_mask_3_reps_0"), val = tensor([1, 256, 1, 1])]; + tensor expanded_mask_3_cast_fp16 = tile(reps = expanded_mask_3_reps_0, x = var_187_cast_fp16)[name = string("expanded_mask_3_cast_fp16")]; + tensor input_5_cast_fp16 = mul(x = tensor_3_cast_fp16, y = expanded_mask_3_cast_fp16)[name = string("input_5_cast_fp16")]; + tensor tensor_5_cast_fp16 = relu(x = input_5_cast_fp16)[name = string("tensor_5_cast_fp16")]; + tensor input_7_cast_fp16 = mul(x = tensor_5_cast_fp16, y = expanded_mask_3_cast_fp16)[name = string("input_7_cast_fp16")]; + tensor input_9_pad_0 = const()[name = string("input_9_pad_0"), val = tensor([0, 0, 0, 0, 2, 1, 2, 1])]; + string input_9_mode_0 = const()[name = string("input_9_mode_0"), val = string("constant")]; + fp16 const_23_to_fp16 = const()[name = string("const_23_to_fp16"), val = fp16(0x0p+0)]; + tensor input_9_cast_fp16 = pad(constant_val = const_23_to_fp16, mode = input_9_mode_0, pad = input_9_pad_0, x = input_7_cast_fp16)[name = string("input_9_cast_fp16")]; + string tensor_7_pad_type_0 = const()[name = string("tensor_7_pad_type_0"), val = string("valid")]; + tensor tensor_7_strides_0 = const()[name = string("tensor_7_strides_0"), val = tensor([2, 2])]; + int32 tensor_7_groups_0 = const()[name = string("tensor_7_groups_0"), val = int32(256)]; + tensor tensor_7_pad_0 = const()[name = string("tensor_7_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor tensor_7_dilations_0 = const()[name = string("tensor_7_dilations_0"), val = tensor([1, 1])]; + tensor encoder_pre_encode_conv_2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4480))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6848))))[name = string("encoder_pre_encode_conv_2_weight_to_fp16_quantized")]; + tensor encoder_pre_encode_conv_2_bias_to_fp16 = const()[name = string("encoder_pre_encode_conv_2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7424)))]; + tensor tensor_7_cast_fp16 = conv(bias = encoder_pre_encode_conv_2_bias_to_fp16, dilations = tensor_7_dilations_0, groups = tensor_7_groups_0, pad = tensor_7_pad_0, pad_type = tensor_7_pad_type_0, strides = tensor_7_strides_0, weight = encoder_pre_encode_conv_2_weight_to_fp16_quantized, x = input_9_cast_fp16)[name = string("tensor_7_cast_fp16")]; + fp16 var_209_promoted_to_fp16 = const()[name = string("op_209_promoted_to_fp16"), val = fp16(0x1p+1)]; + tensor var_210_cast_fp16 = add(x = current_lengths_3_cast_fp16, y = var_209_promoted_to_fp16)[name = string("op_210_cast_fp16")]; + fp16 var_211_promoted_to_fp16 = const()[name = string("op_211_promoted_to_fp16"), val = fp16(0x1p+0)]; + tensor var_212_cast_fp16 = add(x = var_210_cast_fp16, y = var_211_promoted_to_fp16)[name = string("op_212_cast_fp16")]; + fp16 var_213_promoted_to_fp16 = const()[name = string("op_213_promoted_to_fp16"), val = fp16(0x1.8p+1)]; + tensor var_214_cast_fp16 = sub(x = var_212_cast_fp16, y = var_213_promoted_to_fp16)[name = string("op_214_cast_fp16")]; + fp16 var_57_promoted_1_to_fp16 = const()[name = string("op_57_promoted_1_to_fp16"), val = fp16(0x1p+1)]; + tensor floor_div_1_cast_fp16 = floor_div(x = var_214_cast_fp16, y = var_57_promoted_1_to_fp16)[name = string("floor_div_1_cast_fp16")]; + fp16 var_216_promoted_to_fp16 = const()[name = string("op_216_promoted_to_fp16"), val = fp16(0x1p+0)]; + tensor current_lengths_5_cast_fp16 = add(x = floor_div_1_cast_fp16, y = var_216_promoted_to_fp16)[name = string("current_lengths_5_cast_fp16")]; + string lengths_21_dtype_0 = const()[name = string("lengths_21_dtype_0"), val = string("int32")]; + tensor expand_dims_2 = const()[name = string("expand_dims_2"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8000)))]; + tensor var_225_axes_0 = const()[name = string("op_225_axes_0"), val = tensor([1])]; + tensor current_lengths_5_cast_fp16_to_int32 = cast(dtype = lengths_21_dtype_0, x = current_lengths_5_cast_fp16)[name = string("cast_17")]; + tensor var_225 = expand_dims(axes = var_225_axes_0, x = current_lengths_5_cast_fp16_to_int32)[name = string("op_225")]; + tensor time_mask_5 = less(x = expand_dims_2, y = var_225)[name = string("time_mask_5")]; + tensor var_227_axes_0 = const()[name = string("op_227_axes_0"), val = tensor([-1])]; + tensor var_227 = expand_dims(axes = var_227_axes_0, x = time_mask_5)[name = string("op_227")]; + tensor var_229_reps_0 = const()[name = string("op_229_reps_0"), val = tensor([1, 1, 33])]; + tensor var_229 = tile(reps = var_229_reps_0, x = var_227)[name = string("op_229")]; + tensor var_235_axes_0 = const()[name = string("op_235_axes_0"), val = tensor([1])]; + string mask_5_to_fp16_dtype_0 = const()[name = string("mask_5_to_fp16_dtype_0"), val = string("fp16")]; + tensor var_229_to_fp16 = cast(dtype = mask_5_to_fp16_dtype_0, x = var_229)[name = string("cast_16")]; + tensor var_235_cast_fp16 = expand_dims(axes = var_235_axes_0, x = var_229_to_fp16)[name = string("op_235_cast_fp16")]; + tensor expanded_mask_7_reps_0 = const()[name = string("expanded_mask_7_reps_0"), val = tensor([1, 256, 1, 1])]; + tensor expanded_mask_7_cast_fp16 = tile(reps = expanded_mask_7_reps_0, x = var_235_cast_fp16)[name = string("expanded_mask_7_cast_fp16")]; + tensor input_11_cast_fp16 = mul(x = tensor_7_cast_fp16, y = expanded_mask_7_cast_fp16)[name = string("input_11_cast_fp16")]; + string tensor_9_pad_type_0 = const()[name = string("tensor_9_pad_type_0"), val = string("valid")]; + tensor tensor_9_strides_0 = const()[name = string("tensor_9_strides_0"), val = tensor([1, 1])]; + tensor tensor_9_pad_0 = const()[name = string("tensor_9_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor tensor_9_dilations_0 = const()[name = string("tensor_9_dilations_0"), val = tensor([1, 1])]; + int32 tensor_9_groups_0 = const()[name = string("tensor_9_groups_0"), val = int32(1)]; + tensor encoder_pre_encode_conv_3_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8192))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(73792))))[name = string("encoder_pre_encode_conv_3_weight_to_fp16_quantized")]; + tensor encoder_pre_encode_conv_3_bias_to_fp16 = const()[name = string("encoder_pre_encode_conv_3_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(74368)))]; + tensor tensor_9_cast_fp16 = conv(bias = encoder_pre_encode_conv_3_bias_to_fp16, dilations = tensor_9_dilations_0, groups = tensor_9_groups_0, pad = tensor_9_pad_0, pad_type = tensor_9_pad_type_0, strides = tensor_9_strides_0, weight = encoder_pre_encode_conv_3_weight_to_fp16_quantized, x = input_11_cast_fp16)[name = string("tensor_9_cast_fp16")]; + tensor input_13_cast_fp16 = mul(x = tensor_9_cast_fp16, y = expanded_mask_7_cast_fp16)[name = string("input_13_cast_fp16")]; + tensor tensor_11_cast_fp16 = relu(x = input_13_cast_fp16)[name = string("tensor_11_cast_fp16")]; + tensor input_15_cast_fp16 = mul(x = tensor_11_cast_fp16, y = expanded_mask_7_cast_fp16)[name = string("input_15_cast_fp16")]; + tensor input_17_pad_0 = const()[name = string("input_17_pad_0"), val = tensor([0, 0, 0, 0, 2, 1, 2, 1])]; + string input_17_mode_0 = const()[name = string("input_17_mode_0"), val = string("constant")]; + fp16 const_41_to_fp16 = const()[name = string("const_41_to_fp16"), val = fp16(0x0p+0)]; + tensor input_17_cast_fp16 = pad(constant_val = const_41_to_fp16, mode = input_17_mode_0, pad = input_17_pad_0, x = input_15_cast_fp16)[name = string("input_17_cast_fp16")]; + string tensor_13_pad_type_0 = const()[name = string("tensor_13_pad_type_0"), val = string("valid")]; + tensor tensor_13_strides_0 = const()[name = string("tensor_13_strides_0"), val = tensor([2, 2])]; + int32 tensor_13_groups_0 = const()[name = string("tensor_13_groups_0"), val = int32(256)]; + tensor tensor_13_pad_0 = const()[name = string("tensor_13_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor tensor_13_dilations_0 = const()[name = string("tensor_13_dilations_0"), val = tensor([1, 1])]; + tensor encoder_pre_encode_conv_5_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(74944))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77312))))[name = string("encoder_pre_encode_conv_5_weight_to_fp16_quantized")]; + tensor encoder_pre_encode_conv_5_bias_to_fp16 = const()[name = string("encoder_pre_encode_conv_5_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77888)))]; + tensor tensor_13_cast_fp16 = conv(bias = encoder_pre_encode_conv_5_bias_to_fp16, dilations = tensor_13_dilations_0, groups = tensor_13_groups_0, pad = tensor_13_pad_0, pad_type = tensor_13_pad_type_0, strides = tensor_13_strides_0, weight = encoder_pre_encode_conv_5_weight_to_fp16_quantized, x = input_17_cast_fp16)[name = string("tensor_13_cast_fp16")]; + fp16 var_272_promoted_to_fp16 = const()[name = string("op_272_promoted_to_fp16"), val = fp16(0x1p+1)]; + tensor var_273_cast_fp16 = add(x = current_lengths_5_cast_fp16, y = var_272_promoted_to_fp16)[name = string("op_273_cast_fp16")]; + fp16 var_274_promoted_to_fp16 = const()[name = string("op_274_promoted_to_fp16"), val = fp16(0x1p+0)]; + tensor var_275_cast_fp16 = add(x = var_273_cast_fp16, y = var_274_promoted_to_fp16)[name = string("op_275_cast_fp16")]; + fp16 var_276_promoted_to_fp16 = const()[name = string("op_276_promoted_to_fp16"), val = fp16(0x1.8p+1)]; + tensor var_277_cast_fp16 = sub(x = var_275_cast_fp16, y = var_276_promoted_to_fp16)[name = string("op_277_cast_fp16")]; + fp16 var_57_promoted_2_to_fp16 = const()[name = string("op_57_promoted_2_to_fp16"), val = fp16(0x1p+1)]; + tensor floor_div_2_cast_fp16 = floor_div(x = var_277_cast_fp16, y = var_57_promoted_2_to_fp16)[name = string("floor_div_2_cast_fp16")]; + fp16 var_279_promoted_to_fp16 = const()[name = string("op_279_promoted_to_fp16"), val = fp16(0x1p+0)]; + tensor current_lengths_cast_fp16 = add(x = floor_div_2_cast_fp16, y = var_279_promoted_to_fp16)[name = string("current_lengths_cast_fp16")]; + string lengths_dtype_0 = const()[name = string("lengths_dtype_0"), val = string("int32")]; + tensor expand_dims_3 = const()[name = string("expand_dims_3"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78464)))]; + tensor var_288_axes_0 = const()[name = string("op_288_axes_0"), val = tensor([1])]; + tensor current_lengths_cast_fp16_to_int32 = cast(dtype = lengths_dtype_0, x = current_lengths_cast_fp16)[name = string("cast_15")]; + tensor var_288 = expand_dims(axes = var_288_axes_0, x = current_lengths_cast_fp16_to_int32)[name = string("op_288")]; + tensor time_mask = less(x = expand_dims_3, y = var_288)[name = string("time_mask")]; + tensor var_290_axes_0 = const()[name = string("op_290_axes_0"), val = tensor([-1])]; + tensor var_290 = expand_dims(axes = var_290_axes_0, x = time_mask)[name = string("op_290")]; + tensor var_292_reps_0 = const()[name = string("op_292_reps_0"), val = tensor([1, 1, 17])]; + tensor var_292 = tile(reps = var_292_reps_0, x = var_290)[name = string("op_292")]; + tensor var_298_axes_0 = const()[name = string("op_298_axes_0"), val = tensor([1])]; + string mask_7_to_fp16_dtype_0 = const()[name = string("mask_7_to_fp16_dtype_0"), val = string("fp16")]; + tensor var_292_to_fp16 = cast(dtype = mask_7_to_fp16_dtype_0, x = var_292)[name = string("cast_14")]; + tensor var_298_cast_fp16 = expand_dims(axes = var_298_axes_0, x = var_292_to_fp16)[name = string("op_298_cast_fp16")]; + tensor expanded_mask_13_reps_0 = const()[name = string("expanded_mask_13_reps_0"), val = tensor([1, 256, 1, 1])]; + tensor expanded_mask_13_cast_fp16 = tile(reps = expanded_mask_13_reps_0, x = var_298_cast_fp16)[name = string("expanded_mask_13_cast_fp16")]; + tensor input_19_cast_fp16 = mul(x = tensor_13_cast_fp16, y = expanded_mask_13_cast_fp16)[name = string("input_19_cast_fp16")]; + string tensor_15_pad_type_0 = const()[name = string("tensor_15_pad_type_0"), val = string("valid")]; + tensor tensor_15_strides_0 = const()[name = string("tensor_15_strides_0"), val = tensor([1, 1])]; + tensor tensor_15_pad_0 = const()[name = string("tensor_15_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor tensor_15_dilations_0 = const()[name = string("tensor_15_dilations_0"), val = tensor([1, 1])]; + int32 tensor_15_groups_0 = const()[name = string("tensor_15_groups_0"), val = int32(1)]; + tensor encoder_pre_encode_conv_6_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78592))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144192))))[name = string("encoder_pre_encode_conv_6_weight_to_fp16_quantized")]; + tensor encoder_pre_encode_conv_6_bias_to_fp16 = const()[name = string("encoder_pre_encode_conv_6_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144768)))]; + tensor tensor_15_cast_fp16 = conv(bias = encoder_pre_encode_conv_6_bias_to_fp16, dilations = tensor_15_dilations_0, groups = tensor_15_groups_0, pad = tensor_15_pad_0, pad_type = tensor_15_pad_type_0, strides = tensor_15_strides_0, weight = encoder_pre_encode_conv_6_weight_to_fp16_quantized, x = input_19_cast_fp16)[name = string("tensor_15_cast_fp16")]; + tensor input_21_cast_fp16 = mul(x = tensor_15_cast_fp16, y = expanded_mask_13_cast_fp16)[name = string("input_21_cast_fp16")]; + tensor tensor_cast_fp16 = relu(x = input_21_cast_fp16)[name = string("tensor_cast_fp16")]; + tensor x_3_cast_fp16 = mul(x = tensor_cast_fp16, y = expanded_mask_13_cast_fp16)[name = string("x_3_cast_fp16")]; + tensor var_332_perm_0 = const()[name = string("op_332_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_333 = const()[name = string("op_333"), val = tensor([1, 16, -1])]; + tensor var_332_cast_fp16 = transpose(perm = var_332_perm_0, x = x_3_cast_fp16)[name = string("transpose_366")]; + tensor input_23_cast_fp16 = reshape(shape = var_333, x = var_332_cast_fp16)[name = string("input_23_cast_fp16")]; + tensor encoder_pre_encode_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(145344))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4601856))))[name = string("encoder_pre_encode_out_weight_to_fp16_quantized")]; + tensor encoder_pre_encode_out_bias_to_fp16 = const()[name = string("encoder_pre_encode_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4603968)))]; + tensor linear_0_cast_fp16 = linear(bias = encoder_pre_encode_out_bias_to_fp16, weight = encoder_pre_encode_out_weight_to_fp16_quantized, x = input_23_cast_fp16)[name = string("linear_0_cast_fp16")]; + tensor var_343_begin_0 = const()[name = string("op_343_begin_0"), val = tensor([0, 2, 0])]; + tensor var_343_end_0 = const()[name = string("op_343_end_0"), val = tensor([1, 16, 1024])]; + tensor var_343_end_mask_0 = const()[name = string("op_343_end_mask_0"), val = tensor([true, true, true])]; + tensor var_343_cast_fp16 = slice_by_index(begin = var_343_begin_0, end = var_343_end_0, end_mask = var_343_end_mask_0, x = linear_0_cast_fp16)[name = string("op_343_cast_fp16")]; + int32 var_345 = const()[name = string("op_345"), val = int32(2)]; + tensor var_346 = sub(x = current_lengths_cast_fp16_to_int32, y = var_345)[name = string("op_346")]; + string var_346_promoted_to_fp16_dtype_0 = const()[name = string("op_346_promoted_to_fp16_dtype_0"), val = string("fp16")]; + fp16 var_63_promoted_to_fp16 = const()[name = string("op_63_promoted_to_fp16"), val = fp16(0x0p+0)]; + fp16 const_61_to_fp16 = const()[name = string("const_61_to_fp16"), val = fp16(inf)]; + tensor var_346_to_fp16 = cast(dtype = var_346_promoted_to_fp16_dtype_0, x = var_346)[name = string("cast_13")]; + tensor clip_0_cast_fp16 = clip(alpha = var_63_promoted_to_fp16, beta = const_61_to_fp16, x = var_346_to_fp16)[name = string("clip_0_cast_fp16")]; + tensor max_audio_length_1 = const()[name = string("max_audio_length_1"), val = tensor([14])]; + fp16 var_362_promoted_to_fp16 = const()[name = string("op_362_promoted_to_fp16"), val = fp16(0x1.5p+5)]; + tensor padding_length_cast_fp16 = add(x = clip_0_cast_fp16, y = var_362_promoted_to_fp16)[name = string("padding_length_cast_fp16")]; + int32 const_63 = const()[name = string("const_63"), val = int32(-1)]; + tensor var_364 = mul(x = cache_len, y = const_63)[name = string("op_364")]; + int32 var_365 = const()[name = string("op_365"), val = int32(42)]; + tensor offset = add(x = var_364, y = var_365)[name = string("offset")]; + tensor var_405_axes_0 = const()[name = string("op_405_axes_0"), val = tensor([-1])]; + tensor var_405_cast_fp16 = expand_dims(axes = var_405_axes_0, x = padding_length_cast_fp16)[name = string("op_405_cast_fp16")]; + tensor var_404_promoted_to_fp16 = const()[name = string("op_404_promoted_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4606080)))]; + tensor pad_mask_1_cast_fp16 = less(x = var_404_promoted_to_fp16, y = var_405_cast_fp16)[name = string("pad_mask_1_cast_fp16")]; + tensor expand_dims_5 = const()[name = string("expand_dims_5"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4606272)))]; + tensor var_411_axes_0 = const()[name = string("op_411_axes_0"), val = tensor([-1])]; + tensor var_411 = expand_dims(axes = var_411_axes_0, x = offset)[name = string("op_411")]; + tensor pad_mask_off = greater_equal(x = expand_dims_5, y = var_411)[name = string("pad_mask_off")]; + tensor pad_mask_3 = logical_and(x = pad_mask_off, y = pad_mask_1_cast_fp16)[name = string("pad_mask_3")]; + tensor var_414_axes_0 = const()[name = string("op_414_axes_0"), val = tensor([1])]; + tensor var_414 = expand_dims(axes = var_414_axes_0, x = pad_mask_3)[name = string("op_414")]; + tensor var_415 = const()[name = string("op_415"), val = tensor([1, 56, 1])]; + tensor pad_mask_for_att_mask_1 = tile(reps = var_415, x = var_414)[name = string("pad_mask_for_att_mask_1")]; + tensor var_417_perm_0 = const()[name = string("op_417_perm_0"), val = tensor([0, 2, 1])]; + tensor var_417 = transpose(perm = var_417_perm_0, x = pad_mask_for_att_mask_1)[name = string("transpose_365")]; + tensor pad_mask_for_att_mask = logical_and(x = pad_mask_for_att_mask_1, y = var_417)[name = string("pad_mask_for_att_mask")]; + tensor const_71 = const()[name = string("const_71"), val = tensor([[[true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true]]])]; + tensor att_mask_9 = logical_and(x = pad_mask_for_att_mask, y = const_71)[name = string("att_mask_9")]; + tensor att_mask = logical_not(x = att_mask_9)[name = string("att_mask")]; + tensor pad_mask_5 = logical_not(x = pad_mask_3)[name = string("pad_mask_5")]; + tensor pad_mask_begin_0 = const()[name = string("pad_mask_begin_0"), val = tensor([0, 42])]; + tensor pad_mask_end_0 = const()[name = string("pad_mask_end_0"), val = tensor([1, 56])]; + tensor pad_mask_end_mask_0 = const()[name = string("pad_mask_end_mask_0"), val = tensor([true, true])]; + tensor pad_mask = slice_by_index(begin = pad_mask_begin_0, end = pad_mask_end_0, end_mask = pad_mask_end_mask_0, x = pad_mask_5)[name = string("pad_mask")]; + tensor mask_9_begin_0 = const()[name = string("mask_9_begin_0"), val = tensor([0, 42, 0])]; + tensor mask_9_end_0 = const()[name = string("mask_9_end_0"), val = tensor([1, 56, 56])]; + tensor mask_9_end_mask_0 = const()[name = string("mask_9_end_mask_0"), val = tensor([true, true, true])]; + tensor mask_9 = slice_by_index(begin = mask_9_begin_0, end = mask_9_end_0, end_mask = mask_9_end_mask_0, x = att_mask)[name = string("mask_9")]; + tensor cache_1_begin_0 = const()[name = string("cache_1_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor cache_1_end_0 = const()[name = string("cache_1_end_0"), val = tensor([1, 1, 42, 1024])]; + tensor cache_1_end_mask_0 = const()[name = string("cache_1_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_1_squeeze_mask_0 = const()[name = string("cache_1_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_channel_to_fp16 = cast(dtype = cache_channel_to_fp16_dtype_0, x = cache_channel)[name = string("cast_12")]; + tensor value_3_cast_fp16 = transpose(perm = value_3_perm_0, x = cache_channel_to_fp16)[name = string("transpose_364")]; + tensor cache_1_cast_fp16 = slice_by_index(begin = cache_1_begin_0, end = cache_1_end_0, end_mask = cache_1_end_mask_0, squeeze_mask = cache_1_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_1_cast_fp16")]; + tensor cache_3_begin_0 = const()[name = string("cache_3_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor cache_3_end_0 = const()[name = string("cache_3_end_0"), val = tensor([1, 1, 1024, 8])]; + tensor cache_3_end_mask_0 = const()[name = string("cache_3_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_3_squeeze_mask_0 = const()[name = string("cache_3_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_time_to_fp16 = cast(dtype = cache_time_to_fp16_dtype_0, x = cache_time)[name = string("cast_11")]; + tensor value_5_cast_fp16 = transpose(perm = value_5_perm_0, x = cache_time_to_fp16)[name = string("transpose_363")]; + tensor cache_3_cast_fp16 = slice_by_index(begin = cache_3_begin_0, end = cache_3_end_0, end_mask = cache_3_end_mask_0, squeeze_mask = cache_3_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_3_cast_fp16")]; + tensor input_27_axes_0 = const()[name = string("input_27_axes_0"), val = tensor([-1])]; + tensor encoder_layers_0_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_0_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4606592)))]; + tensor encoder_layers_0_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_0_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4608704)))]; + fp16 var_43_to_fp16 = const()[name = string("op_43_to_fp16"), val = fp16(0x1.5p-17)]; + tensor input_27_cast_fp16 = layer_norm(axes = input_27_axes_0, beta = encoder_layers_0_norm_feed_forward1_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_0_norm_feed_forward1_weight_to_fp16, x = var_343_cast_fp16)[name = string("input_27_cast_fp16")]; + tensor encoder_layers_0_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4610816))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8805184))))[name = string("encoder_layers_0_feed_forward1_linear1_weight_to_fp16_quantized")]; + tensor encoder_layers_0_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_0_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8813440)))]; + tensor linear_1_cast_fp16 = linear(bias = encoder_layers_0_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_0_feed_forward1_linear1_weight_to_fp16_quantized, x = input_27_cast_fp16)[name = string("linear_1_cast_fp16")]; + tensor input_31_cast_fp16 = silu(x = linear_1_cast_fp16)[name = string("input_31_cast_fp16")]; + tensor encoder_layers_0_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8821696))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13016064))))[name = string("encoder_layers_0_feed_forward1_linear2_weight_to_fp16_quantized")]; + tensor encoder_layers_0_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_0_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13018176)))]; + tensor linear_2_cast_fp16 = linear(bias = encoder_layers_0_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_0_feed_forward1_linear2_weight_to_fp16_quantized, x = input_31_cast_fp16)[name = string("linear_2_cast_fp16")]; + fp16 var_456_to_fp16 = const()[name = string("op_456_to_fp16"), val = fp16(0x1p-1)]; + tensor var_457_cast_fp16 = mul(x = linear_2_cast_fp16, y = var_456_to_fp16)[name = string("op_457_cast_fp16")]; + tensor input_37_cast_fp16 = add(x = var_343_cast_fp16, y = var_457_cast_fp16)[name = string("input_37_cast_fp16")]; + tensor key_1_axes_0 = const()[name = string("key_1_axes_0"), val = tensor([-1])]; + tensor encoder_layers_0_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_0_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13020288)))]; + tensor encoder_layers_0_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_0_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13022400)))]; + tensor key_1_cast_fp16 = layer_norm(axes = key_1_axes_0, beta = encoder_layers_0_norm_self_att_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_0_norm_self_att_weight_to_fp16, x = input_37_cast_fp16)[name = string("key_1_cast_fp16")]; + bool input_39_interleave_0 = const()[name = string("input_39_interleave_0"), val = bool(false)]; + tensor input_39_cast_fp16 = concat(axis = var_69, interleave = input_39_interleave_0, values = (cache_1_cast_fp16, key_1_cast_fp16))[name = string("input_39_cast_fp16")]; + tensor var_479_begin_0 = const()[name = string("op_479_begin_0"), val = tensor([0, 14, 0])]; + tensor var_479_end_0 = const()[name = string("op_479_end_0"), val = tensor([1, 42, 1024])]; + tensor var_479_end_mask_0 = const()[name = string("op_479_end_mask_0"), val = tensor([true, true, true])]; + tensor var_479_cast_fp16 = slice_by_index(begin = var_479_begin_0, end = var_479_end_0, end_mask = var_479_end_mask_0, x = cache_1_cast_fp16)[name = string("op_479_cast_fp16")]; + bool var_485_interleave_0 = const()[name = string("op_485_interleave_0"), val = bool(false)]; + tensor var_485_cast_fp16 = concat(axis = var_69, interleave = var_485_interleave_0, values = (var_479_cast_fp16, key_1_cast_fp16))[name = string("op_485_cast_fp16")]; + tensor encoder_layers_0_self_attn_linear_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13024512))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14073152))))[name = string("encoder_layers_0_self_attn_linear_q_weight_to_fp16_quantized")]; + tensor encoder_layers_0_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_0_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14075264)))]; + tensor linear_3_cast_fp16 = linear(bias = encoder_layers_0_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_0_self_attn_linear_q_weight_to_fp16_quantized, x = key_1_cast_fp16)[name = string("linear_3_cast_fp16")]; + tensor var_490 = const()[name = string("op_490"), val = tensor([1, -1, 8, 128])]; + tensor q_1_cast_fp16 = reshape(shape = var_490, x = linear_3_cast_fp16)[name = string("q_1_cast_fp16")]; + tensor encoder_layers_0_self_attn_linear_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14077376))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15126016))))[name = string("encoder_layers_0_self_attn_linear_k_weight_to_fp16_quantized")]; + tensor encoder_layers_0_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_0_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15128128)))]; + tensor linear_4_cast_fp16 = linear(bias = encoder_layers_0_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_0_self_attn_linear_k_weight_to_fp16_quantized, x = input_39_cast_fp16)[name = string("linear_4_cast_fp16")]; + tensor var_495 = const()[name = string("op_495"), val = tensor([1, -1, 8, 128])]; + tensor k_1_cast_fp16 = reshape(shape = var_495, x = linear_4_cast_fp16)[name = string("k_1_cast_fp16")]; + tensor encoder_layers_0_self_attn_linear_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15130240))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16178880))))[name = string("encoder_layers_0_self_attn_linear_v_weight_to_fp16_quantized")]; + tensor encoder_layers_0_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_0_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16180992)))]; + tensor linear_5_cast_fp16 = linear(bias = encoder_layers_0_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_0_self_attn_linear_v_weight_to_fp16_quantized, x = input_39_cast_fp16)[name = string("linear_5_cast_fp16")]; + tensor var_500 = const()[name = string("op_500"), val = tensor([1, -1, 8, 128])]; + tensor v_1_cast_fp16 = reshape(shape = var_500, x = linear_5_cast_fp16)[name = string("v_1_cast_fp16")]; + tensor value_9_perm_0 = const()[name = string("value_9_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_0_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_0_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16183104)))]; + tensor var_513_cast_fp16 = add(x = q_1_cast_fp16, y = encoder_layers_0_self_attn_pos_bias_u_to_fp16)[name = string("op_513_cast_fp16")]; + tensor encoder_layers_0_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_0_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16185216)))]; + tensor var_515_cast_fp16 = add(x = q_1_cast_fp16, y = encoder_layers_0_self_attn_pos_bias_v_to_fp16)[name = string("op_515_cast_fp16")]; + tensor q_with_bias_v_1_perm_0 = const()[name = string("q_with_bias_v_1_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_7_transpose_x_0 = const()[name = string("x_7_transpose_x_0"), val = bool(false)]; + bool x_7_transpose_y_0 = const()[name = string("x_7_transpose_y_0"), val = bool(false)]; + tensor op_517_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16187328))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16301056))))[name = string("op_517_to_fp16_quantized")]; + tensor q_with_bias_v_1_cast_fp16 = transpose(perm = q_with_bias_v_1_perm_0, x = var_515_cast_fp16)[name = string("transpose_362")]; + tensor x_7_cast_fp16 = matmul(transpose_x = x_7_transpose_x_0, transpose_y = x_7_transpose_y_0, x = q_with_bias_v_1_cast_fp16, y = op_517_to_fp16_quantized)[name = string("x_7_cast_fp16")]; + tensor x_9_pad_0 = const()[name = string("x_9_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_9_mode_0 = const()[name = string("x_9_mode_0"), val = string("constant")]; + fp16 const_79_to_fp16 = const()[name = string("const_79_to_fp16"), val = fp16(0x0p+0)]; + tensor x_9_cast_fp16 = pad(constant_val = const_79_to_fp16, mode = x_9_mode_0, pad = x_9_pad_0, x = x_7_cast_fp16)[name = string("x_9_cast_fp16")]; + tensor var_525 = const()[name = string("op_525"), val = tensor([1, 8, -1, 14])]; + tensor x_11_cast_fp16 = reshape(shape = var_525, x = x_9_cast_fp16)[name = string("x_11_cast_fp16")]; + tensor var_529_begin_0 = const()[name = string("op_529_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_529_end_0 = const()[name = string("op_529_end_0"), val = tensor([1, 8, 112, 14])]; + tensor var_529_end_mask_0 = const()[name = string("op_529_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_529_cast_fp16 = slice_by_index(begin = var_529_begin_0, end = var_529_end_0, end_mask = var_529_end_mask_0, x = x_11_cast_fp16)[name = string("op_529_cast_fp16")]; + tensor var_530 = const()[name = string("op_530"), val = tensor([1, 8, 14, 111])]; + tensor matrix_bd_1_cast_fp16 = reshape(shape = var_530, x = var_529_cast_fp16)[name = string("matrix_bd_1_cast_fp16")]; + bool matrix_ac_1_transpose_x_0 = const()[name = string("matrix_ac_1_transpose_x_0"), val = bool(false)]; + bool matrix_ac_1_transpose_y_0 = const()[name = string("matrix_ac_1_transpose_y_0"), val = bool(false)]; + tensor transpose_96_perm_0 = const()[name = string("transpose_96_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_97_perm_0 = const()[name = string("transpose_97_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_97 = transpose(perm = transpose_97_perm_0, x = k_1_cast_fp16)[name = string("transpose_360")]; + tensor transpose_96 = transpose(perm = transpose_96_perm_0, x = var_513_cast_fp16)[name = string("transpose_361")]; + tensor matrix_ac_1_cast_fp16 = matmul(transpose_x = matrix_ac_1_transpose_x_0, transpose_y = matrix_ac_1_transpose_y_0, x = transpose_96, y = transpose_97)[name = string("matrix_ac_1_cast_fp16")]; + tensor matrix_bd_3_begin_0 = const()[name = string("matrix_bd_3_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_3_end_0 = const()[name = string("matrix_bd_3_end_0"), val = tensor([1, 8, 14, 56])]; + tensor matrix_bd_3_end_mask_0 = const()[name = string("matrix_bd_3_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_3_cast_fp16 = slice_by_index(begin = matrix_bd_3_begin_0, end = matrix_bd_3_end_0, end_mask = matrix_bd_3_end_mask_0, x = matrix_bd_1_cast_fp16)[name = string("matrix_bd_3_cast_fp16")]; + tensor var_539_cast_fp16 = add(x = matrix_ac_1_cast_fp16, y = matrix_bd_3_cast_fp16)[name = string("op_539_cast_fp16")]; + fp16 _inversed_scores_1_y_0_to_fp16 = const()[name = string("_inversed_scores_1_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_1_cast_fp16 = mul(x = var_539_cast_fp16, y = _inversed_scores_1_y_0_to_fp16)[name = string("_inversed_scores_1_cast_fp16")]; + tensor mask_11_axes_0 = const()[name = string("mask_11_axes_0"), val = tensor([1])]; + tensor mask_11 = expand_dims(axes = mask_11_axes_0, x = mask_9)[name = string("mask_11")]; + fp16 var_46_to_fp16 = const()[name = string("op_46_to_fp16"), val = fp16(-0x1.388p+13)]; + tensor scores_3_cast_fp16 = select(a = var_46_to_fp16, b = _inversed_scores_1_cast_fp16, cond = mask_11)[name = string("scores_3_cast_fp16")]; + tensor var_545_cast_fp16 = softmax(axis = var_60, x = scores_3_cast_fp16)[name = string("op_545_cast_fp16")]; + fp16 var_45_to_fp16 = const()[name = string("op_45_to_fp16"), val = fp16(0x0p+0)]; + tensor input_41_cast_fp16 = select(a = var_45_to_fp16, b = var_545_cast_fp16, cond = mask_11)[name = string("input_41_cast_fp16")]; + bool x_13_transpose_x_0 = const()[name = string("x_13_transpose_x_0"), val = bool(false)]; + bool x_13_transpose_y_0 = const()[name = string("x_13_transpose_y_0"), val = bool(false)]; + tensor value_9_cast_fp16 = transpose(perm = value_9_perm_0, x = v_1_cast_fp16)[name = string("transpose_359")]; + tensor x_13_cast_fp16 = matmul(transpose_x = x_13_transpose_x_0, transpose_y = x_13_transpose_y_0, x = input_41_cast_fp16, y = value_9_cast_fp16)[name = string("x_13_cast_fp16")]; + tensor var_549_perm_0 = const()[name = string("op_549_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_550 = const()[name = string("op_550"), val = tensor([1, -1, 1024])]; + tensor var_549_cast_fp16 = transpose(perm = var_549_perm_0, x = x_13_cast_fp16)[name = string("transpose_358")]; + tensor input_43_cast_fp16 = reshape(shape = var_550, x = var_549_cast_fp16)[name = string("input_43_cast_fp16")]; + tensor encoder_layers_0_self_attn_linear_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16301376))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17350016))))[name = string("encoder_layers_0_self_attn_linear_out_weight_to_fp16_quantized")]; + tensor encoder_layers_0_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_0_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17352128)))]; + tensor linear_7_cast_fp16 = linear(bias = encoder_layers_0_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_0_self_attn_linear_out_weight_to_fp16_quantized, x = input_43_cast_fp16)[name = string("linear_7_cast_fp16")]; + tensor input_47_cast_fp16 = add(x = input_37_cast_fp16, y = linear_7_cast_fp16)[name = string("input_47_cast_fp16")]; + tensor x_17_axes_0 = const()[name = string("x_17_axes_0"), val = tensor([-1])]; + tensor encoder_layers_0_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_0_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17354240)))]; + tensor encoder_layers_0_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_0_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17356352)))]; + tensor x_17_cast_fp16 = layer_norm(axes = x_17_axes_0, beta = encoder_layers_0_norm_conv_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_0_norm_conv_weight_to_fp16, x = input_47_cast_fp16)[name = string("x_17_cast_fp16")]; + tensor input_49_perm_0 = const()[name = string("input_49_perm_0"), val = tensor([0, 2, 1])]; + string input_51_pad_type_0 = const()[name = string("input_51_pad_type_0"), val = string("valid")]; + tensor input_51_strides_0 = const()[name = string("input_51_strides_0"), val = tensor([1])]; + tensor input_51_pad_0 = const()[name = string("input_51_pad_0"), val = tensor([0, 0])]; + tensor input_51_dilations_0 = const()[name = string("input_51_dilations_0"), val = tensor([1])]; + int32 input_51_groups_0 = const()[name = string("input_51_groups_0"), val = int32(1)]; + tensor encoder_layers_0_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17358464))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19455680))))[name = string("encoder_layers_0_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_49_cast_fp16 = transpose(perm = input_49_perm_0, x = x_17_cast_fp16)[name = string("transpose_357")]; + tensor input_51_cast_fp16 = conv(dilations = input_51_dilations_0, groups = input_51_groups_0, pad = input_51_pad_0, pad_type = input_51_pad_type_0, strides = input_51_strides_0, weight = encoder_layers_0_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_49_cast_fp16)[name = string("input_51_cast_fp16")]; + int32 x_19_split_num_splits_0 = const()[name = string("x_19_split_num_splits_0"), val = int32(2)]; + int32 x_19_split_axis_0 = const()[name = string("x_19_split_axis_0"), val = int32(1)]; + tensor x_19_split_cast_fp16_0, tensor x_19_split_cast_fp16_1 = split(axis = x_19_split_axis_0, num_splits = x_19_split_num_splits_0, x = input_51_cast_fp16)[name = string("x_19_split_cast_fp16")]; + tensor x_19_split_1_sigmoid_cast_fp16 = sigmoid(x = x_19_split_cast_fp16_1)[name = string("x_19_split_1_sigmoid_cast_fp16")]; + tensor x_19_cast_fp16 = mul(x = x_19_split_cast_fp16_0, y = x_19_split_1_sigmoid_cast_fp16)[name = string("x_19_cast_fp16")]; + tensor var_576_axes_0 = const()[name = string("op_576_axes_0"), val = tensor([1])]; + tensor var_576 = expand_dims(axes = var_576_axes_0, x = pad_mask)[name = string("op_576")]; + tensor input_53_cast_fp16 = select(a = var_45_to_fp16, b = x_19_cast_fp16, cond = var_576)[name = string("input_53_cast_fp16")]; + bool new_x_3_interleave_0 = const()[name = string("new_x_3_interleave_0"), val = bool(false)]; + tensor new_x_3_cast_fp16 = concat(axis = var_60, interleave = new_x_3_interleave_0, values = (cache_3_cast_fp16, input_53_cast_fp16))[name = string("new_x_3_cast_fp16")]; + tensor var_589_begin_0 = const()[name = string("op_589_begin_0"), val = tensor([0, 0, 14])]; + tensor var_589_end_0 = const()[name = string("op_589_end_0"), val = tensor([1, 1024, 22])]; + tensor var_589_end_mask_0 = const()[name = string("op_589_end_mask_0"), val = tensor([true, true, true])]; + tensor var_589_cast_fp16 = slice_by_index(begin = var_589_begin_0, end = var_589_end_0, end_mask = var_589_end_mask_0, x = new_x_3_cast_fp16)[name = string("op_589_cast_fp16")]; + string x_21_pad_type_0 = const()[name = string("x_21_pad_type_0"), val = string("valid")]; + int32 x_21_groups_0 = const()[name = string("x_21_groups_0"), val = int32(1024)]; + tensor x_21_strides_0 = const()[name = string("x_21_strides_0"), val = tensor([1])]; + tensor x_21_pad_0 = const()[name = string("x_21_pad_0"), val = tensor([0, 0])]; + tensor x_21_dilations_0 = const()[name = string("x_21_dilations_0"), val = tensor([1])]; + tensor encoder_layers_0_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19459840))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19469120))))[name = string("encoder_layers_0_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_21_cast_fp16 = conv(dilations = x_21_dilations_0, groups = x_21_groups_0, pad = x_21_pad_0, pad_type = x_21_pad_type_0, strides = x_21_strides_0, weight = encoder_layers_0_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_3_cast_fp16)[name = string("x_21_cast_fp16")]; + tensor input_55_perm_0 = const()[name = string("input_55_perm_0"), val = tensor([0, 2, 1])]; + tensor x_23_axes_0 = const()[name = string("x_23_axes_0"), val = tensor([-1])]; + tensor encoder_layers_0_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_0_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19471232)))]; + tensor encoder_layers_0_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_0_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19473344)))]; + tensor input_55_cast_fp16 = transpose(perm = input_55_perm_0, x = x_21_cast_fp16)[name = string("transpose_356")]; + tensor x_23_cast_fp16 = layer_norm(axes = x_23_axes_0, beta = encoder_layers_0_conv_batch_norm_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_0_conv_batch_norm_weight_to_fp16, x = input_55_cast_fp16)[name = string("x_23_cast_fp16")]; + tensor input_57_perm_0 = const()[name = string("input_57_perm_0"), val = tensor([0, 2, 1])]; + tensor input_57_cast_fp16 = transpose(perm = input_57_perm_0, x = x_23_cast_fp16)[name = string("transpose_355")]; + tensor input_59_cast_fp16 = silu(x = input_57_cast_fp16)[name = string("input_59_cast_fp16")]; + string x_25_pad_type_0 = const()[name = string("x_25_pad_type_0"), val = string("valid")]; + tensor x_25_strides_0 = const()[name = string("x_25_strides_0"), val = tensor([1])]; + tensor x_25_pad_0 = const()[name = string("x_25_pad_0"), val = tensor([0, 0])]; + tensor x_25_dilations_0 = const()[name = string("x_25_dilations_0"), val = tensor([1])]; + int32 x_25_groups_0 = const()[name = string("x_25_groups_0"), val = int32(1)]; + tensor encoder_layers_0_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19475456))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20524096))))[name = string("encoder_layers_0_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_25_cast_fp16 = conv(dilations = x_25_dilations_0, groups = x_25_groups_0, pad = x_25_pad_0, pad_type = x_25_pad_type_0, strides = x_25_strides_0, weight = encoder_layers_0_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_59_cast_fp16)[name = string("x_25_cast_fp16")]; + tensor input_61_perm_0 = const()[name = string("input_61_perm_0"), val = tensor([0, 2, 1])]; + tensor input_61_cast_fp16 = transpose(perm = input_61_perm_0, x = x_25_cast_fp16)[name = string("transpose_354")]; + tensor input_63_cast_fp16 = add(x = input_47_cast_fp16, y = input_61_cast_fp16)[name = string("input_63_cast_fp16")]; + tensor input_65_axes_0 = const()[name = string("input_65_axes_0"), val = tensor([-1])]; + tensor encoder_layers_0_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_0_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20526208)))]; + tensor encoder_layers_0_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_0_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20528320)))]; + tensor input_65_cast_fp16 = layer_norm(axes = input_65_axes_0, beta = encoder_layers_0_norm_feed_forward2_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_0_norm_feed_forward2_weight_to_fp16, x = input_63_cast_fp16)[name = string("input_65_cast_fp16")]; + tensor encoder_layers_0_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20530432))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24724800))))[name = string("encoder_layers_0_feed_forward2_linear1_weight_to_fp16_quantized")]; + tensor encoder_layers_0_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_0_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24733056)))]; + tensor linear_8_cast_fp16 = linear(bias = encoder_layers_0_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_0_feed_forward2_linear1_weight_to_fp16_quantized, x = input_65_cast_fp16)[name = string("linear_8_cast_fp16")]; + tensor input_69_cast_fp16 = silu(x = linear_8_cast_fp16)[name = string("input_69_cast_fp16")]; + tensor encoder_layers_0_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24741312))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28935680))))[name = string("encoder_layers_0_feed_forward2_linear2_weight_to_fp16_quantized")]; + tensor encoder_layers_0_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_0_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28937792)))]; + tensor linear_9_cast_fp16 = linear(bias = encoder_layers_0_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_0_feed_forward2_linear2_weight_to_fp16_quantized, x = input_69_cast_fp16)[name = string("linear_9_cast_fp16")]; + fp16 var_632_to_fp16 = const()[name = string("op_632_to_fp16"), val = fp16(0x1p-1)]; + tensor var_633_cast_fp16 = mul(x = linear_9_cast_fp16, y = var_632_to_fp16)[name = string("op_633_cast_fp16")]; + tensor input_75_cast_fp16 = add(x = input_63_cast_fp16, y = var_633_cast_fp16)[name = string("input_75_cast_fp16")]; + tensor input_77_axes_0 = const()[name = string("input_77_axes_0"), val = tensor([-1])]; + tensor encoder_layers_0_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_0_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28939904)))]; + tensor encoder_layers_0_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_0_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28942016)))]; + tensor input_77_cast_fp16 = layer_norm(axes = input_77_axes_0, beta = encoder_layers_0_norm_out_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_0_norm_out_weight_to_fp16, x = input_75_cast_fp16)[name = string("input_77_cast_fp16")]; + tensor cache_5_begin_0 = const()[name = string("cache_5_begin_0"), val = tensor([1, 0, 0, 0])]; + tensor cache_5_end_0 = const()[name = string("cache_5_end_0"), val = tensor([2, 1, 42, 1024])]; + tensor cache_5_end_mask_0 = const()[name = string("cache_5_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_5_squeeze_mask_0 = const()[name = string("cache_5_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_5_cast_fp16 = slice_by_index(begin = cache_5_begin_0, end = cache_5_end_0, end_mask = cache_5_end_mask_0, squeeze_mask = cache_5_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_5_cast_fp16")]; + tensor cache_7_begin_0 = const()[name = string("cache_7_begin_0"), val = tensor([1, 0, 0, 0])]; + tensor cache_7_end_0 = const()[name = string("cache_7_end_0"), val = tensor([2, 1, 1024, 8])]; + tensor cache_7_end_mask_0 = const()[name = string("cache_7_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_7_squeeze_mask_0 = const()[name = string("cache_7_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_7_cast_fp16 = slice_by_index(begin = cache_7_begin_0, end = cache_7_end_0, end_mask = cache_7_end_mask_0, squeeze_mask = cache_7_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_7_cast_fp16")]; + tensor input_79_axes_0 = const()[name = string("input_79_axes_0"), val = tensor([-1])]; + tensor encoder_layers_1_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_1_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28944128)))]; + tensor encoder_layers_1_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_1_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28946240)))]; + tensor input_79_cast_fp16 = layer_norm(axes = input_79_axes_0, beta = encoder_layers_1_norm_feed_forward1_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_1_norm_feed_forward1_weight_to_fp16, x = input_77_cast_fp16)[name = string("input_79_cast_fp16")]; + tensor encoder_layers_1_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28948352))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33142720))))[name = string("encoder_layers_1_feed_forward1_linear1_weight_to_fp16_quantized")]; + tensor encoder_layers_1_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_1_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33150976)))]; + tensor linear_10_cast_fp16 = linear(bias = encoder_layers_1_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_1_feed_forward1_linear1_weight_to_fp16_quantized, x = input_79_cast_fp16)[name = string("linear_10_cast_fp16")]; + tensor input_83_cast_fp16 = silu(x = linear_10_cast_fp16)[name = string("input_83_cast_fp16")]; + tensor encoder_layers_1_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33159232))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37353600))))[name = string("encoder_layers_1_feed_forward1_linear2_weight_to_fp16_quantized")]; + tensor encoder_layers_1_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_1_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37355712)))]; + tensor linear_11_cast_fp16 = linear(bias = encoder_layers_1_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_1_feed_forward1_linear2_weight_to_fp16_quantized, x = input_83_cast_fp16)[name = string("linear_11_cast_fp16")]; + fp16 var_669_to_fp16 = const()[name = string("op_669_to_fp16"), val = fp16(0x1p-1)]; + tensor var_670_cast_fp16 = mul(x = linear_11_cast_fp16, y = var_669_to_fp16)[name = string("op_670_cast_fp16")]; + tensor input_89_cast_fp16 = add(x = input_77_cast_fp16, y = var_670_cast_fp16)[name = string("input_89_cast_fp16")]; + tensor key_3_axes_0 = const()[name = string("key_3_axes_0"), val = tensor([-1])]; + tensor encoder_layers_1_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_1_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37357824)))]; + tensor encoder_layers_1_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_1_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37359936)))]; + tensor key_3_cast_fp16 = layer_norm(axes = key_3_axes_0, beta = encoder_layers_1_norm_self_att_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_1_norm_self_att_weight_to_fp16, x = input_89_cast_fp16)[name = string("key_3_cast_fp16")]; + bool input_91_interleave_0 = const()[name = string("input_91_interleave_0"), val = bool(false)]; + tensor input_91_cast_fp16 = concat(axis = var_69, interleave = input_91_interleave_0, values = (cache_5_cast_fp16, key_3_cast_fp16))[name = string("input_91_cast_fp16")]; + tensor var_692_begin_0 = const()[name = string("op_692_begin_0"), val = tensor([0, 14, 0])]; + tensor var_692_end_0 = const()[name = string("op_692_end_0"), val = tensor([1, 42, 1024])]; + tensor var_692_end_mask_0 = const()[name = string("op_692_end_mask_0"), val = tensor([true, true, true])]; + tensor var_692_cast_fp16 = slice_by_index(begin = var_692_begin_0, end = var_692_end_0, end_mask = var_692_end_mask_0, x = cache_5_cast_fp16)[name = string("op_692_cast_fp16")]; + bool var_698_interleave_0 = const()[name = string("op_698_interleave_0"), val = bool(false)]; + tensor var_698_cast_fp16 = concat(axis = var_69, interleave = var_698_interleave_0, values = (var_692_cast_fp16, key_3_cast_fp16))[name = string("op_698_cast_fp16")]; + tensor encoder_layers_1_self_attn_linear_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37362048))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38410688))))[name = string("encoder_layers_1_self_attn_linear_q_weight_to_fp16_quantized")]; + tensor encoder_layers_1_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_1_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38412800)))]; + tensor linear_12_cast_fp16 = linear(bias = encoder_layers_1_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_1_self_attn_linear_q_weight_to_fp16_quantized, x = key_3_cast_fp16)[name = string("linear_12_cast_fp16")]; + tensor var_703 = const()[name = string("op_703"), val = tensor([1, -1, 8, 128])]; + tensor q_7_cast_fp16 = reshape(shape = var_703, x = linear_12_cast_fp16)[name = string("q_7_cast_fp16")]; + tensor encoder_layers_1_self_attn_linear_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38414912))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39463552))))[name = string("encoder_layers_1_self_attn_linear_k_weight_to_fp16_quantized")]; + tensor encoder_layers_1_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_1_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39465664)))]; + tensor linear_13_cast_fp16 = linear(bias = encoder_layers_1_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_1_self_attn_linear_k_weight_to_fp16_quantized, x = input_91_cast_fp16)[name = string("linear_13_cast_fp16")]; + tensor var_708 = const()[name = string("op_708"), val = tensor([1, -1, 8, 128])]; + tensor k_5_cast_fp16 = reshape(shape = var_708, x = linear_13_cast_fp16)[name = string("k_5_cast_fp16")]; + tensor encoder_layers_1_self_attn_linear_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39467776))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40516416))))[name = string("encoder_layers_1_self_attn_linear_v_weight_to_fp16_quantized")]; + tensor encoder_layers_1_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_1_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40518528)))]; + tensor linear_14_cast_fp16 = linear(bias = encoder_layers_1_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_1_self_attn_linear_v_weight_to_fp16_quantized, x = input_91_cast_fp16)[name = string("linear_14_cast_fp16")]; + tensor var_713 = const()[name = string("op_713"), val = tensor([1, -1, 8, 128])]; + tensor v_3_cast_fp16 = reshape(shape = var_713, x = linear_14_cast_fp16)[name = string("v_3_cast_fp16")]; + tensor value_11_perm_0 = const()[name = string("value_11_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_1_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_1_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40520640)))]; + tensor var_726_cast_fp16 = add(x = q_7_cast_fp16, y = encoder_layers_1_self_attn_pos_bias_u_to_fp16)[name = string("op_726_cast_fp16")]; + tensor encoder_layers_1_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_1_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40522752)))]; + tensor var_728_cast_fp16 = add(x = q_7_cast_fp16, y = encoder_layers_1_self_attn_pos_bias_v_to_fp16)[name = string("op_728_cast_fp16")]; + tensor q_with_bias_v_3_perm_0 = const()[name = string("q_with_bias_v_3_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_33_transpose_x_0 = const()[name = string("x_33_transpose_x_0"), val = bool(false)]; + bool x_33_transpose_y_0 = const()[name = string("x_33_transpose_y_0"), val = bool(false)]; + tensor op_730_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40524864))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40638592))))[name = string("op_730_to_fp16_quantized")]; + tensor q_with_bias_v_3_cast_fp16 = transpose(perm = q_with_bias_v_3_perm_0, x = var_728_cast_fp16)[name = string("transpose_353")]; + tensor x_33_cast_fp16 = matmul(transpose_x = x_33_transpose_x_0, transpose_y = x_33_transpose_y_0, x = q_with_bias_v_3_cast_fp16, y = op_730_to_fp16_quantized)[name = string("x_33_cast_fp16")]; + tensor x_35_pad_0 = const()[name = string("x_35_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_35_mode_0 = const()[name = string("x_35_mode_0"), val = string("constant")]; + fp16 const_92_to_fp16 = const()[name = string("const_92_to_fp16"), val = fp16(0x0p+0)]; + tensor x_35_cast_fp16 = pad(constant_val = const_92_to_fp16, mode = x_35_mode_0, pad = x_35_pad_0, x = x_33_cast_fp16)[name = string("x_35_cast_fp16")]; + tensor var_738 = const()[name = string("op_738"), val = tensor([1, 8, -1, 14])]; + tensor x_37_cast_fp16 = reshape(shape = var_738, x = x_35_cast_fp16)[name = string("x_37_cast_fp16")]; + tensor var_742_begin_0 = const()[name = string("op_742_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_742_end_0 = const()[name = string("op_742_end_0"), val = tensor([1, 8, 112, 14])]; + tensor var_742_end_mask_0 = const()[name = string("op_742_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_742_cast_fp16 = slice_by_index(begin = var_742_begin_0, end = var_742_end_0, end_mask = var_742_end_mask_0, x = x_37_cast_fp16)[name = string("op_742_cast_fp16")]; + tensor var_743 = const()[name = string("op_743"), val = tensor([1, 8, 14, 111])]; + tensor matrix_bd_5_cast_fp16 = reshape(shape = var_743, x = var_742_cast_fp16)[name = string("matrix_bd_5_cast_fp16")]; + bool matrix_ac_3_transpose_x_0 = const()[name = string("matrix_ac_3_transpose_x_0"), val = bool(false)]; + bool matrix_ac_3_transpose_y_0 = const()[name = string("matrix_ac_3_transpose_y_0"), val = bool(false)]; + tensor transpose_98_perm_0 = const()[name = string("transpose_98_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_99_perm_0 = const()[name = string("transpose_99_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_99 = transpose(perm = transpose_99_perm_0, x = k_5_cast_fp16)[name = string("transpose_351")]; + tensor transpose_98 = transpose(perm = transpose_98_perm_0, x = var_726_cast_fp16)[name = string("transpose_352")]; + tensor matrix_ac_3_cast_fp16 = matmul(transpose_x = matrix_ac_3_transpose_x_0, transpose_y = matrix_ac_3_transpose_y_0, x = transpose_98, y = transpose_99)[name = string("matrix_ac_3_cast_fp16")]; + tensor matrix_bd_7_begin_0 = const()[name = string("matrix_bd_7_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_7_end_0 = const()[name = string("matrix_bd_7_end_0"), val = tensor([1, 8, 14, 56])]; + tensor matrix_bd_7_end_mask_0 = const()[name = string("matrix_bd_7_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_7_cast_fp16 = slice_by_index(begin = matrix_bd_7_begin_0, end = matrix_bd_7_end_0, end_mask = matrix_bd_7_end_mask_0, x = matrix_bd_5_cast_fp16)[name = string("matrix_bd_7_cast_fp16")]; + tensor var_752_cast_fp16 = add(x = matrix_ac_3_cast_fp16, y = matrix_bd_7_cast_fp16)[name = string("op_752_cast_fp16")]; + fp16 _inversed_scores_5_y_0_to_fp16 = const()[name = string("_inversed_scores_5_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_5_cast_fp16 = mul(x = var_752_cast_fp16, y = _inversed_scores_5_y_0_to_fp16)[name = string("_inversed_scores_5_cast_fp16")]; + tensor scores_7_cast_fp16 = select(a = var_46_to_fp16, b = _inversed_scores_5_cast_fp16, cond = mask_11)[name = string("scores_7_cast_fp16")]; + tensor var_758_cast_fp16 = softmax(axis = var_60, x = scores_7_cast_fp16)[name = string("op_758_cast_fp16")]; + tensor input_93_cast_fp16 = select(a = var_45_to_fp16, b = var_758_cast_fp16, cond = mask_11)[name = string("input_93_cast_fp16")]; + bool x_39_transpose_x_0 = const()[name = string("x_39_transpose_x_0"), val = bool(false)]; + bool x_39_transpose_y_0 = const()[name = string("x_39_transpose_y_0"), val = bool(false)]; + tensor value_11_cast_fp16 = transpose(perm = value_11_perm_0, x = v_3_cast_fp16)[name = string("transpose_350")]; + tensor x_39_cast_fp16 = matmul(transpose_x = x_39_transpose_x_0, transpose_y = x_39_transpose_y_0, x = input_93_cast_fp16, y = value_11_cast_fp16)[name = string("x_39_cast_fp16")]; + tensor var_762_perm_0 = const()[name = string("op_762_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_763 = const()[name = string("op_763"), val = tensor([1, -1, 1024])]; + tensor var_762_cast_fp16 = transpose(perm = var_762_perm_0, x = x_39_cast_fp16)[name = string("transpose_349")]; + tensor input_95_cast_fp16 = reshape(shape = var_763, x = var_762_cast_fp16)[name = string("input_95_cast_fp16")]; + tensor encoder_layers_1_self_attn_linear_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40638912))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41687552))))[name = string("encoder_layers_1_self_attn_linear_out_weight_to_fp16_quantized")]; + tensor encoder_layers_1_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_1_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41689664)))]; + tensor linear_16_cast_fp16 = linear(bias = encoder_layers_1_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_1_self_attn_linear_out_weight_to_fp16_quantized, x = input_95_cast_fp16)[name = string("linear_16_cast_fp16")]; + tensor input_99_cast_fp16 = add(x = input_89_cast_fp16, y = linear_16_cast_fp16)[name = string("input_99_cast_fp16")]; + tensor x_43_axes_0 = const()[name = string("x_43_axes_0"), val = tensor([-1])]; + tensor encoder_layers_1_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_1_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41691776)))]; + tensor encoder_layers_1_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_1_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41693888)))]; + tensor x_43_cast_fp16 = layer_norm(axes = x_43_axes_0, beta = encoder_layers_1_norm_conv_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_1_norm_conv_weight_to_fp16, x = input_99_cast_fp16)[name = string("x_43_cast_fp16")]; + tensor input_101_perm_0 = const()[name = string("input_101_perm_0"), val = tensor([0, 2, 1])]; + string input_103_pad_type_0 = const()[name = string("input_103_pad_type_0"), val = string("valid")]; + tensor input_103_strides_0 = const()[name = string("input_103_strides_0"), val = tensor([1])]; + tensor input_103_pad_0 = const()[name = string("input_103_pad_0"), val = tensor([0, 0])]; + tensor input_103_dilations_0 = const()[name = string("input_103_dilations_0"), val = tensor([1])]; + int32 input_103_groups_0 = const()[name = string("input_103_groups_0"), val = int32(1)]; + tensor encoder_layers_1_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41696000))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43793216))))[name = string("encoder_layers_1_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_101_cast_fp16 = transpose(perm = input_101_perm_0, x = x_43_cast_fp16)[name = string("transpose_348")]; + tensor input_103_cast_fp16 = conv(dilations = input_103_dilations_0, groups = input_103_groups_0, pad = input_103_pad_0, pad_type = input_103_pad_type_0, strides = input_103_strides_0, weight = encoder_layers_1_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_101_cast_fp16)[name = string("input_103_cast_fp16")]; + int32 x_45_split_num_splits_0 = const()[name = string("x_45_split_num_splits_0"), val = int32(2)]; + int32 x_45_split_axis_0 = const()[name = string("x_45_split_axis_0"), val = int32(1)]; + tensor x_45_split_cast_fp16_0, tensor x_45_split_cast_fp16_1 = split(axis = x_45_split_axis_0, num_splits = x_45_split_num_splits_0, x = input_103_cast_fp16)[name = string("x_45_split_cast_fp16")]; + tensor x_45_split_1_sigmoid_cast_fp16 = sigmoid(x = x_45_split_cast_fp16_1)[name = string("x_45_split_1_sigmoid_cast_fp16")]; + tensor x_45_cast_fp16 = mul(x = x_45_split_cast_fp16_0, y = x_45_split_1_sigmoid_cast_fp16)[name = string("x_45_cast_fp16")]; + tensor input_105_cast_fp16 = select(a = var_45_to_fp16, b = x_45_cast_fp16, cond = var_576)[name = string("input_105_cast_fp16")]; + bool new_x_7_interleave_0 = const()[name = string("new_x_7_interleave_0"), val = bool(false)]; + tensor new_x_7_cast_fp16 = concat(axis = var_60, interleave = new_x_7_interleave_0, values = (cache_7_cast_fp16, input_105_cast_fp16))[name = string("new_x_7_cast_fp16")]; + tensor var_802_begin_0 = const()[name = string("op_802_begin_0"), val = tensor([0, 0, 14])]; + tensor var_802_end_0 = const()[name = string("op_802_end_0"), val = tensor([1, 1024, 22])]; + tensor var_802_end_mask_0 = const()[name = string("op_802_end_mask_0"), val = tensor([true, true, true])]; + tensor var_802_cast_fp16 = slice_by_index(begin = var_802_begin_0, end = var_802_end_0, end_mask = var_802_end_mask_0, x = new_x_7_cast_fp16)[name = string("op_802_cast_fp16")]; + string x_47_pad_type_0 = const()[name = string("x_47_pad_type_0"), val = string("valid")]; + int32 x_47_groups_0 = const()[name = string("x_47_groups_0"), val = int32(1024)]; + tensor x_47_strides_0 = const()[name = string("x_47_strides_0"), val = tensor([1])]; + tensor x_47_pad_0 = const()[name = string("x_47_pad_0"), val = tensor([0, 0])]; + tensor x_47_dilations_0 = const()[name = string("x_47_dilations_0"), val = tensor([1])]; + tensor encoder_layers_1_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43797376))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43806656))))[name = string("encoder_layers_1_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_47_cast_fp16 = conv(dilations = x_47_dilations_0, groups = x_47_groups_0, pad = x_47_pad_0, pad_type = x_47_pad_type_0, strides = x_47_strides_0, weight = encoder_layers_1_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_7_cast_fp16)[name = string("x_47_cast_fp16")]; + tensor input_107_perm_0 = const()[name = string("input_107_perm_0"), val = tensor([0, 2, 1])]; + tensor x_49_axes_0 = const()[name = string("x_49_axes_0"), val = tensor([-1])]; + tensor encoder_layers_1_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_1_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43808768)))]; + tensor encoder_layers_1_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_1_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43810880)))]; + tensor input_107_cast_fp16 = transpose(perm = input_107_perm_0, x = x_47_cast_fp16)[name = string("transpose_347")]; + tensor x_49_cast_fp16 = layer_norm(axes = x_49_axes_0, beta = encoder_layers_1_conv_batch_norm_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_1_conv_batch_norm_weight_to_fp16, x = input_107_cast_fp16)[name = string("x_49_cast_fp16")]; + tensor input_109_perm_0 = const()[name = string("input_109_perm_0"), val = tensor([0, 2, 1])]; + tensor input_109_cast_fp16 = transpose(perm = input_109_perm_0, x = x_49_cast_fp16)[name = string("transpose_346")]; + tensor input_111_cast_fp16 = silu(x = input_109_cast_fp16)[name = string("input_111_cast_fp16")]; + string x_51_pad_type_0 = const()[name = string("x_51_pad_type_0"), val = string("valid")]; + tensor x_51_strides_0 = const()[name = string("x_51_strides_0"), val = tensor([1])]; + tensor x_51_pad_0 = const()[name = string("x_51_pad_0"), val = tensor([0, 0])]; + tensor x_51_dilations_0 = const()[name = string("x_51_dilations_0"), val = tensor([1])]; + int32 x_51_groups_0 = const()[name = string("x_51_groups_0"), val = int32(1)]; + tensor encoder_layers_1_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43812992))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44861632))))[name = string("encoder_layers_1_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_51_cast_fp16 = conv(dilations = x_51_dilations_0, groups = x_51_groups_0, pad = x_51_pad_0, pad_type = x_51_pad_type_0, strides = x_51_strides_0, weight = encoder_layers_1_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_111_cast_fp16)[name = string("x_51_cast_fp16")]; + tensor input_113_perm_0 = const()[name = string("input_113_perm_0"), val = tensor([0, 2, 1])]; + tensor input_113_cast_fp16 = transpose(perm = input_113_perm_0, x = x_51_cast_fp16)[name = string("transpose_345")]; + tensor input_115_cast_fp16 = add(x = input_99_cast_fp16, y = input_113_cast_fp16)[name = string("input_115_cast_fp16")]; + tensor input_117_axes_0 = const()[name = string("input_117_axes_0"), val = tensor([-1])]; + tensor encoder_layers_1_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_1_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44863744)))]; + tensor encoder_layers_1_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_1_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44865856)))]; + tensor input_117_cast_fp16 = layer_norm(axes = input_117_axes_0, beta = encoder_layers_1_norm_feed_forward2_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_1_norm_feed_forward2_weight_to_fp16, x = input_115_cast_fp16)[name = string("input_117_cast_fp16")]; + tensor encoder_layers_1_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44867968))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49062336))))[name = string("encoder_layers_1_feed_forward2_linear1_weight_to_fp16_quantized")]; + tensor encoder_layers_1_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_1_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49070592)))]; + tensor linear_17_cast_fp16 = linear(bias = encoder_layers_1_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_1_feed_forward2_linear1_weight_to_fp16_quantized, x = input_117_cast_fp16)[name = string("linear_17_cast_fp16")]; + tensor input_121_cast_fp16 = silu(x = linear_17_cast_fp16)[name = string("input_121_cast_fp16")]; + tensor encoder_layers_1_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49078848))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53273216))))[name = string("encoder_layers_1_feed_forward2_linear2_weight_to_fp16_quantized")]; + tensor encoder_layers_1_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_1_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53275328)))]; + tensor linear_18_cast_fp16 = linear(bias = encoder_layers_1_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_1_feed_forward2_linear2_weight_to_fp16_quantized, x = input_121_cast_fp16)[name = string("linear_18_cast_fp16")]; + fp16 var_845_to_fp16 = const()[name = string("op_845_to_fp16"), val = fp16(0x1p-1)]; + tensor var_846_cast_fp16 = mul(x = linear_18_cast_fp16, y = var_845_to_fp16)[name = string("op_846_cast_fp16")]; + tensor input_127_cast_fp16 = add(x = input_115_cast_fp16, y = var_846_cast_fp16)[name = string("input_127_cast_fp16")]; + tensor input_129_axes_0 = const()[name = string("input_129_axes_0"), val = tensor([-1])]; + tensor encoder_layers_1_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_1_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53277440)))]; + tensor encoder_layers_1_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_1_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53279552)))]; + tensor input_129_cast_fp16 = layer_norm(axes = input_129_axes_0, beta = encoder_layers_1_norm_out_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_1_norm_out_weight_to_fp16, x = input_127_cast_fp16)[name = string("input_129_cast_fp16")]; + tensor cache_9_begin_0 = const()[name = string("cache_9_begin_0"), val = tensor([2, 0, 0, 0])]; + tensor cache_9_end_0 = const()[name = string("cache_9_end_0"), val = tensor([3, 1, 42, 1024])]; + tensor cache_9_end_mask_0 = const()[name = string("cache_9_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_9_squeeze_mask_0 = const()[name = string("cache_9_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_9_cast_fp16 = slice_by_index(begin = cache_9_begin_0, end = cache_9_end_0, end_mask = cache_9_end_mask_0, squeeze_mask = cache_9_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_9_cast_fp16")]; + tensor cache_11_begin_0 = const()[name = string("cache_11_begin_0"), val = tensor([2, 0, 0, 0])]; + tensor cache_11_end_0 = const()[name = string("cache_11_end_0"), val = tensor([3, 1, 1024, 8])]; + tensor cache_11_end_mask_0 = const()[name = string("cache_11_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_11_squeeze_mask_0 = const()[name = string("cache_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_11_cast_fp16 = slice_by_index(begin = cache_11_begin_0, end = cache_11_end_0, end_mask = cache_11_end_mask_0, squeeze_mask = cache_11_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_11_cast_fp16")]; + tensor input_131_axes_0 = const()[name = string("input_131_axes_0"), val = tensor([-1])]; + tensor encoder_layers_2_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_2_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53281664)))]; + tensor encoder_layers_2_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_2_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53283776)))]; + tensor input_131_cast_fp16 = layer_norm(axes = input_131_axes_0, beta = encoder_layers_2_norm_feed_forward1_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_2_norm_feed_forward1_weight_to_fp16, x = input_129_cast_fp16)[name = string("input_131_cast_fp16")]; + tensor encoder_layers_2_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53285888))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57480256))))[name = string("encoder_layers_2_feed_forward1_linear1_weight_to_fp16_quantized")]; + tensor encoder_layers_2_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_2_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57488512)))]; + tensor linear_19_cast_fp16 = linear(bias = encoder_layers_2_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_2_feed_forward1_linear1_weight_to_fp16_quantized, x = input_131_cast_fp16)[name = string("linear_19_cast_fp16")]; + tensor input_135_cast_fp16 = silu(x = linear_19_cast_fp16)[name = string("input_135_cast_fp16")]; + tensor encoder_layers_2_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57496768))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61691136))))[name = string("encoder_layers_2_feed_forward1_linear2_weight_to_fp16_quantized")]; + tensor encoder_layers_2_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_2_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61693248)))]; + tensor linear_20_cast_fp16 = linear(bias = encoder_layers_2_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_2_feed_forward1_linear2_weight_to_fp16_quantized, x = input_135_cast_fp16)[name = string("linear_20_cast_fp16")]; + fp16 var_882_to_fp16 = const()[name = string("op_882_to_fp16"), val = fp16(0x1p-1)]; + tensor var_883_cast_fp16 = mul(x = linear_20_cast_fp16, y = var_882_to_fp16)[name = string("op_883_cast_fp16")]; + tensor input_141_cast_fp16 = add(x = input_129_cast_fp16, y = var_883_cast_fp16)[name = string("input_141_cast_fp16")]; + tensor key_5_axes_0 = const()[name = string("key_5_axes_0"), val = tensor([-1])]; + tensor encoder_layers_2_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_2_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61695360)))]; + tensor encoder_layers_2_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_2_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61697472)))]; + tensor key_5_cast_fp16 = layer_norm(axes = key_5_axes_0, beta = encoder_layers_2_norm_self_att_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_2_norm_self_att_weight_to_fp16, x = input_141_cast_fp16)[name = string("key_5_cast_fp16")]; + bool input_143_interleave_0 = const()[name = string("input_143_interleave_0"), val = bool(false)]; + tensor input_143_cast_fp16 = concat(axis = var_69, interleave = input_143_interleave_0, values = (cache_9_cast_fp16, key_5_cast_fp16))[name = string("input_143_cast_fp16")]; + tensor var_905_begin_0 = const()[name = string("op_905_begin_0"), val = tensor([0, 14, 0])]; + tensor var_905_end_0 = const()[name = string("op_905_end_0"), val = tensor([1, 42, 1024])]; + tensor var_905_end_mask_0 = const()[name = string("op_905_end_mask_0"), val = tensor([true, true, true])]; + tensor var_905_cast_fp16 = slice_by_index(begin = var_905_begin_0, end = var_905_end_0, end_mask = var_905_end_mask_0, x = cache_9_cast_fp16)[name = string("op_905_cast_fp16")]; + bool var_911_interleave_0 = const()[name = string("op_911_interleave_0"), val = bool(false)]; + tensor var_911_cast_fp16 = concat(axis = var_69, interleave = var_911_interleave_0, values = (var_905_cast_fp16, key_5_cast_fp16))[name = string("op_911_cast_fp16")]; + tensor encoder_layers_2_self_attn_linear_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61699584))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62748224))))[name = string("encoder_layers_2_self_attn_linear_q_weight_to_fp16_quantized")]; + tensor encoder_layers_2_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_2_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62750336)))]; + tensor linear_21_cast_fp16 = linear(bias = encoder_layers_2_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_2_self_attn_linear_q_weight_to_fp16_quantized, x = key_5_cast_fp16)[name = string("linear_21_cast_fp16")]; + tensor var_916 = const()[name = string("op_916"), val = tensor([1, -1, 8, 128])]; + tensor q_13_cast_fp16 = reshape(shape = var_916, x = linear_21_cast_fp16)[name = string("q_13_cast_fp16")]; + tensor encoder_layers_2_self_attn_linear_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62752448))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(63801088))))[name = string("encoder_layers_2_self_attn_linear_k_weight_to_fp16_quantized")]; + tensor encoder_layers_2_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_2_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(63803200)))]; + tensor linear_22_cast_fp16 = linear(bias = encoder_layers_2_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_2_self_attn_linear_k_weight_to_fp16_quantized, x = input_143_cast_fp16)[name = string("linear_22_cast_fp16")]; + tensor var_921 = const()[name = string("op_921"), val = tensor([1, -1, 8, 128])]; + tensor k_9_cast_fp16 = reshape(shape = var_921, x = linear_22_cast_fp16)[name = string("k_9_cast_fp16")]; + tensor encoder_layers_2_self_attn_linear_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(63805312))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64853952))))[name = string("encoder_layers_2_self_attn_linear_v_weight_to_fp16_quantized")]; + tensor encoder_layers_2_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_2_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64856064)))]; + tensor linear_23_cast_fp16 = linear(bias = encoder_layers_2_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_2_self_attn_linear_v_weight_to_fp16_quantized, x = input_143_cast_fp16)[name = string("linear_23_cast_fp16")]; + tensor var_926 = const()[name = string("op_926"), val = tensor([1, -1, 8, 128])]; + tensor v_5_cast_fp16 = reshape(shape = var_926, x = linear_23_cast_fp16)[name = string("v_5_cast_fp16")]; + tensor value_13_perm_0 = const()[name = string("value_13_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_2_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_2_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64858176)))]; + tensor var_939_cast_fp16 = add(x = q_13_cast_fp16, y = encoder_layers_2_self_attn_pos_bias_u_to_fp16)[name = string("op_939_cast_fp16")]; + tensor encoder_layers_2_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_2_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64860288)))]; + tensor var_941_cast_fp16 = add(x = q_13_cast_fp16, y = encoder_layers_2_self_attn_pos_bias_v_to_fp16)[name = string("op_941_cast_fp16")]; + tensor q_with_bias_v_5_perm_0 = const()[name = string("q_with_bias_v_5_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_59_transpose_x_0 = const()[name = string("x_59_transpose_x_0"), val = bool(false)]; + bool x_59_transpose_y_0 = const()[name = string("x_59_transpose_y_0"), val = bool(false)]; + tensor op_943_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64862400))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64976128))))[name = string("op_943_to_fp16_quantized")]; + tensor q_with_bias_v_5_cast_fp16 = transpose(perm = q_with_bias_v_5_perm_0, x = var_941_cast_fp16)[name = string("transpose_344")]; + tensor x_59_cast_fp16 = matmul(transpose_x = x_59_transpose_x_0, transpose_y = x_59_transpose_y_0, x = q_with_bias_v_5_cast_fp16, y = op_943_to_fp16_quantized)[name = string("x_59_cast_fp16")]; + tensor x_61_pad_0 = const()[name = string("x_61_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_61_mode_0 = const()[name = string("x_61_mode_0"), val = string("constant")]; + fp16 const_105_to_fp16 = const()[name = string("const_105_to_fp16"), val = fp16(0x0p+0)]; + tensor x_61_cast_fp16 = pad(constant_val = const_105_to_fp16, mode = x_61_mode_0, pad = x_61_pad_0, x = x_59_cast_fp16)[name = string("x_61_cast_fp16")]; + tensor var_951 = const()[name = string("op_951"), val = tensor([1, 8, -1, 14])]; + tensor x_63_cast_fp16 = reshape(shape = var_951, x = x_61_cast_fp16)[name = string("x_63_cast_fp16")]; + tensor var_955_begin_0 = const()[name = string("op_955_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_955_end_0 = const()[name = string("op_955_end_0"), val = tensor([1, 8, 112, 14])]; + tensor var_955_end_mask_0 = const()[name = string("op_955_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_955_cast_fp16 = slice_by_index(begin = var_955_begin_0, end = var_955_end_0, end_mask = var_955_end_mask_0, x = x_63_cast_fp16)[name = string("op_955_cast_fp16")]; + tensor var_956 = const()[name = string("op_956"), val = tensor([1, 8, 14, 111])]; + tensor matrix_bd_9_cast_fp16 = reshape(shape = var_956, x = var_955_cast_fp16)[name = string("matrix_bd_9_cast_fp16")]; + bool matrix_ac_5_transpose_x_0 = const()[name = string("matrix_ac_5_transpose_x_0"), val = bool(false)]; + bool matrix_ac_5_transpose_y_0 = const()[name = string("matrix_ac_5_transpose_y_0"), val = bool(false)]; + tensor transpose_100_perm_0 = const()[name = string("transpose_100_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_101_perm_0 = const()[name = string("transpose_101_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_101 = transpose(perm = transpose_101_perm_0, x = k_9_cast_fp16)[name = string("transpose_342")]; + tensor transpose_100 = transpose(perm = transpose_100_perm_0, x = var_939_cast_fp16)[name = string("transpose_343")]; + tensor matrix_ac_5_cast_fp16 = matmul(transpose_x = matrix_ac_5_transpose_x_0, transpose_y = matrix_ac_5_transpose_y_0, x = transpose_100, y = transpose_101)[name = string("matrix_ac_5_cast_fp16")]; + tensor matrix_bd_11_begin_0 = const()[name = string("matrix_bd_11_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_11_end_0 = const()[name = string("matrix_bd_11_end_0"), val = tensor([1, 8, 14, 56])]; + tensor matrix_bd_11_end_mask_0 = const()[name = string("matrix_bd_11_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_11_cast_fp16 = slice_by_index(begin = matrix_bd_11_begin_0, end = matrix_bd_11_end_0, end_mask = matrix_bd_11_end_mask_0, x = matrix_bd_9_cast_fp16)[name = string("matrix_bd_11_cast_fp16")]; + tensor var_965_cast_fp16 = add(x = matrix_ac_5_cast_fp16, y = matrix_bd_11_cast_fp16)[name = string("op_965_cast_fp16")]; + fp16 _inversed_scores_9_y_0_to_fp16 = const()[name = string("_inversed_scores_9_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_9_cast_fp16 = mul(x = var_965_cast_fp16, y = _inversed_scores_9_y_0_to_fp16)[name = string("_inversed_scores_9_cast_fp16")]; + tensor scores_11_cast_fp16 = select(a = var_46_to_fp16, b = _inversed_scores_9_cast_fp16, cond = mask_11)[name = string("scores_11_cast_fp16")]; + tensor var_971_cast_fp16 = softmax(axis = var_60, x = scores_11_cast_fp16)[name = string("op_971_cast_fp16")]; + tensor input_145_cast_fp16 = select(a = var_45_to_fp16, b = var_971_cast_fp16, cond = mask_11)[name = string("input_145_cast_fp16")]; + bool x_65_transpose_x_0 = const()[name = string("x_65_transpose_x_0"), val = bool(false)]; + bool x_65_transpose_y_0 = const()[name = string("x_65_transpose_y_0"), val = bool(false)]; + tensor value_13_cast_fp16 = transpose(perm = value_13_perm_0, x = v_5_cast_fp16)[name = string("transpose_341")]; + tensor x_65_cast_fp16 = matmul(transpose_x = x_65_transpose_x_0, transpose_y = x_65_transpose_y_0, x = input_145_cast_fp16, y = value_13_cast_fp16)[name = string("x_65_cast_fp16")]; + tensor var_975_perm_0 = const()[name = string("op_975_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_976 = const()[name = string("op_976"), val = tensor([1, -1, 1024])]; + tensor var_975_cast_fp16 = transpose(perm = var_975_perm_0, x = x_65_cast_fp16)[name = string("transpose_340")]; + tensor input_147_cast_fp16 = reshape(shape = var_976, x = var_975_cast_fp16)[name = string("input_147_cast_fp16")]; + tensor encoder_layers_2_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64976448))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65762944))))[name = string("encoder_layers_2_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_2_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_2_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65763136)))]; + tensor linear_25_cast_fp16 = linear(bias = encoder_layers_2_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_2_self_attn_linear_out_weight_to_fp16_palettized, x = input_147_cast_fp16)[name = string("linear_25_cast_fp16")]; + tensor input_151_cast_fp16 = add(x = input_141_cast_fp16, y = linear_25_cast_fp16)[name = string("input_151_cast_fp16")]; + tensor x_69_axes_0 = const()[name = string("x_69_axes_0"), val = tensor([-1])]; + tensor encoder_layers_2_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_2_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65765248)))]; + tensor encoder_layers_2_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_2_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65767360)))]; + tensor x_69_cast_fp16 = layer_norm(axes = x_69_axes_0, beta = encoder_layers_2_norm_conv_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_2_norm_conv_weight_to_fp16, x = input_151_cast_fp16)[name = string("x_69_cast_fp16")]; + tensor input_153_perm_0 = const()[name = string("input_153_perm_0"), val = tensor([0, 2, 1])]; + string input_155_pad_type_0 = const()[name = string("input_155_pad_type_0"), val = string("valid")]; + tensor input_155_strides_0 = const()[name = string("input_155_strides_0"), val = tensor([1])]; + tensor input_155_pad_0 = const()[name = string("input_155_pad_0"), val = tensor([0, 0])]; + tensor input_155_dilations_0 = const()[name = string("input_155_dilations_0"), val = tensor([1])]; + int32 input_155_groups_0 = const()[name = string("input_155_groups_0"), val = int32(1)]; + tensor encoder_layers_2_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65769472))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67866688))))[name = string("encoder_layers_2_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_153_cast_fp16 = transpose(perm = input_153_perm_0, x = x_69_cast_fp16)[name = string("transpose_339")]; + tensor input_155_cast_fp16 = conv(dilations = input_155_dilations_0, groups = input_155_groups_0, pad = input_155_pad_0, pad_type = input_155_pad_type_0, strides = input_155_strides_0, weight = encoder_layers_2_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_153_cast_fp16)[name = string("input_155_cast_fp16")]; + int32 x_71_split_num_splits_0 = const()[name = string("x_71_split_num_splits_0"), val = int32(2)]; + int32 x_71_split_axis_0 = const()[name = string("x_71_split_axis_0"), val = int32(1)]; + tensor x_71_split_cast_fp16_0, tensor x_71_split_cast_fp16_1 = split(axis = x_71_split_axis_0, num_splits = x_71_split_num_splits_0, x = input_155_cast_fp16)[name = string("x_71_split_cast_fp16")]; + tensor x_71_split_1_sigmoid_cast_fp16 = sigmoid(x = x_71_split_cast_fp16_1)[name = string("x_71_split_1_sigmoid_cast_fp16")]; + tensor x_71_cast_fp16 = mul(x = x_71_split_cast_fp16_0, y = x_71_split_1_sigmoid_cast_fp16)[name = string("x_71_cast_fp16")]; + tensor input_157_cast_fp16 = select(a = var_45_to_fp16, b = x_71_cast_fp16, cond = var_576)[name = string("input_157_cast_fp16")]; + bool new_x_11_interleave_0 = const()[name = string("new_x_11_interleave_0"), val = bool(false)]; + tensor new_x_11_cast_fp16 = concat(axis = var_60, interleave = new_x_11_interleave_0, values = (cache_11_cast_fp16, input_157_cast_fp16))[name = string("new_x_11_cast_fp16")]; + tensor var_1015_begin_0 = const()[name = string("op_1015_begin_0"), val = tensor([0, 0, 14])]; + tensor var_1015_end_0 = const()[name = string("op_1015_end_0"), val = tensor([1, 1024, 22])]; + tensor var_1015_end_mask_0 = const()[name = string("op_1015_end_mask_0"), val = tensor([true, true, true])]; + tensor var_1015_cast_fp16 = slice_by_index(begin = var_1015_begin_0, end = var_1015_end_0, end_mask = var_1015_end_mask_0, x = new_x_11_cast_fp16)[name = string("op_1015_cast_fp16")]; + string x_73_pad_type_0 = const()[name = string("x_73_pad_type_0"), val = string("valid")]; + int32 x_73_groups_0 = const()[name = string("x_73_groups_0"), val = int32(1024)]; + tensor x_73_strides_0 = const()[name = string("x_73_strides_0"), val = tensor([1])]; + tensor x_73_pad_0 = const()[name = string("x_73_pad_0"), val = tensor([0, 0])]; + tensor x_73_dilations_0 = const()[name = string("x_73_dilations_0"), val = tensor([1])]; + tensor encoder_layers_2_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67870848))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67880128))))[name = string("encoder_layers_2_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_73_cast_fp16 = conv(dilations = x_73_dilations_0, groups = x_73_groups_0, pad = x_73_pad_0, pad_type = x_73_pad_type_0, strides = x_73_strides_0, weight = encoder_layers_2_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_11_cast_fp16)[name = string("x_73_cast_fp16")]; + tensor input_159_perm_0 = const()[name = string("input_159_perm_0"), val = tensor([0, 2, 1])]; + tensor x_75_axes_0 = const()[name = string("x_75_axes_0"), val = tensor([-1])]; + tensor encoder_layers_2_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_2_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67882240)))]; + tensor encoder_layers_2_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_2_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67884352)))]; + tensor input_159_cast_fp16 = transpose(perm = input_159_perm_0, x = x_73_cast_fp16)[name = string("transpose_338")]; + tensor x_75_cast_fp16 = layer_norm(axes = x_75_axes_0, beta = encoder_layers_2_conv_batch_norm_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_2_conv_batch_norm_weight_to_fp16, x = input_159_cast_fp16)[name = string("x_75_cast_fp16")]; + tensor input_161_perm_0 = const()[name = string("input_161_perm_0"), val = tensor([0, 2, 1])]; + tensor input_161_cast_fp16 = transpose(perm = input_161_perm_0, x = x_75_cast_fp16)[name = string("transpose_337")]; + tensor input_163_cast_fp16 = silu(x = input_161_cast_fp16)[name = string("input_163_cast_fp16")]; + string x_77_pad_type_0 = const()[name = string("x_77_pad_type_0"), val = string("valid")]; + tensor x_77_strides_0 = const()[name = string("x_77_strides_0"), val = tensor([1])]; + tensor x_77_pad_0 = const()[name = string("x_77_pad_0"), val = tensor([0, 0])]; + tensor x_77_dilations_0 = const()[name = string("x_77_dilations_0"), val = tensor([1])]; + int32 x_77_groups_0 = const()[name = string("x_77_groups_0"), val = int32(1)]; + tensor encoder_layers_2_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67886464))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68935104))))[name = string("encoder_layers_2_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_77_cast_fp16 = conv(dilations = x_77_dilations_0, groups = x_77_groups_0, pad = x_77_pad_0, pad_type = x_77_pad_type_0, strides = x_77_strides_0, weight = encoder_layers_2_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_163_cast_fp16)[name = string("x_77_cast_fp16")]; + tensor input_165_perm_0 = const()[name = string("input_165_perm_0"), val = tensor([0, 2, 1])]; + tensor input_165_cast_fp16 = transpose(perm = input_165_perm_0, x = x_77_cast_fp16)[name = string("transpose_336")]; + tensor input_167_cast_fp16 = add(x = input_151_cast_fp16, y = input_165_cast_fp16)[name = string("input_167_cast_fp16")]; + tensor input_169_axes_0 = const()[name = string("input_169_axes_0"), val = tensor([-1])]; + tensor encoder_layers_2_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_2_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68937216)))]; + tensor encoder_layers_2_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_2_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68939328)))]; + tensor input_169_cast_fp16 = layer_norm(axes = input_169_axes_0, beta = encoder_layers_2_norm_feed_forward2_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_2_norm_feed_forward2_weight_to_fp16, x = input_167_cast_fp16)[name = string("input_169_cast_fp16")]; + tensor encoder_layers_2_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68941440))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72087232))))[name = string("encoder_layers_2_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_2_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_2_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72087424)))]; + tensor linear_26_cast_fp16 = linear(bias = encoder_layers_2_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_2_feed_forward2_linear1_weight_to_fp16_palettized, x = input_169_cast_fp16)[name = string("linear_26_cast_fp16")]; + tensor input_173_cast_fp16 = silu(x = linear_26_cast_fp16)[name = string("input_173_cast_fp16")]; + tensor encoder_layers_2_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72095680))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75241472))))[name = string("encoder_layers_2_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_2_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_2_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75241664)))]; + tensor linear_27_cast_fp16 = linear(bias = encoder_layers_2_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_2_feed_forward2_linear2_weight_to_fp16_palettized, x = input_173_cast_fp16)[name = string("linear_27_cast_fp16")]; + fp16 var_1058_to_fp16 = const()[name = string("op_1058_to_fp16"), val = fp16(0x1p-1)]; + tensor var_1059_cast_fp16 = mul(x = linear_27_cast_fp16, y = var_1058_to_fp16)[name = string("op_1059_cast_fp16")]; + tensor input_179_cast_fp16 = add(x = input_167_cast_fp16, y = var_1059_cast_fp16)[name = string("input_179_cast_fp16")]; + tensor input_181_axes_0 = const()[name = string("input_181_axes_0"), val = tensor([-1])]; + tensor encoder_layers_2_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_2_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75243776)))]; + tensor encoder_layers_2_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_2_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75245888)))]; + tensor input_181_cast_fp16 = layer_norm(axes = input_181_axes_0, beta = encoder_layers_2_norm_out_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_2_norm_out_weight_to_fp16, x = input_179_cast_fp16)[name = string("input_181_cast_fp16")]; + tensor cache_13_begin_0 = const()[name = string("cache_13_begin_0"), val = tensor([3, 0, 0, 0])]; + tensor cache_13_end_0 = const()[name = string("cache_13_end_0"), val = tensor([4, 1, 42, 1024])]; + tensor cache_13_end_mask_0 = const()[name = string("cache_13_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_13_squeeze_mask_0 = const()[name = string("cache_13_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_13_cast_fp16 = slice_by_index(begin = cache_13_begin_0, end = cache_13_end_0, end_mask = cache_13_end_mask_0, squeeze_mask = cache_13_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_13_cast_fp16")]; + tensor cache_15_begin_0 = const()[name = string("cache_15_begin_0"), val = tensor([3, 0, 0, 0])]; + tensor cache_15_end_0 = const()[name = string("cache_15_end_0"), val = tensor([4, 1, 1024, 8])]; + tensor cache_15_end_mask_0 = const()[name = string("cache_15_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_15_squeeze_mask_0 = const()[name = string("cache_15_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_15_cast_fp16 = slice_by_index(begin = cache_15_begin_0, end = cache_15_end_0, end_mask = cache_15_end_mask_0, squeeze_mask = cache_15_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_15_cast_fp16")]; + tensor input_183_axes_0 = const()[name = string("input_183_axes_0"), val = tensor([-1])]; + tensor encoder_layers_3_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_3_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75248000)))]; + tensor encoder_layers_3_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_3_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75250112)))]; + tensor input_183_cast_fp16 = layer_norm(axes = input_183_axes_0, beta = encoder_layers_3_norm_feed_forward1_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_3_norm_feed_forward1_weight_to_fp16, x = input_181_cast_fp16)[name = string("input_183_cast_fp16")]; + tensor encoder_layers_3_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75252224))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78398016))))[name = string("encoder_layers_3_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_3_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_3_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78398208)))]; + tensor linear_28_cast_fp16 = linear(bias = encoder_layers_3_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_3_feed_forward1_linear1_weight_to_fp16_palettized, x = input_183_cast_fp16)[name = string("linear_28_cast_fp16")]; + tensor input_187_cast_fp16 = silu(x = linear_28_cast_fp16)[name = string("input_187_cast_fp16")]; + tensor encoder_layers_3_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78406464))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81552256))))[name = string("encoder_layers_3_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_3_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_3_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81552448)))]; + tensor linear_29_cast_fp16 = linear(bias = encoder_layers_3_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_3_feed_forward1_linear2_weight_to_fp16_palettized, x = input_187_cast_fp16)[name = string("linear_29_cast_fp16")]; + fp16 var_1095_to_fp16 = const()[name = string("op_1095_to_fp16"), val = fp16(0x1p-1)]; + tensor var_1096_cast_fp16 = mul(x = linear_29_cast_fp16, y = var_1095_to_fp16)[name = string("op_1096_cast_fp16")]; + tensor input_193_cast_fp16 = add(x = input_181_cast_fp16, y = var_1096_cast_fp16)[name = string("input_193_cast_fp16")]; + tensor key_7_axes_0 = const()[name = string("key_7_axes_0"), val = tensor([-1])]; + tensor encoder_layers_3_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_3_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81554560)))]; + tensor encoder_layers_3_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_3_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81556672)))]; + tensor key_7_cast_fp16 = layer_norm(axes = key_7_axes_0, beta = encoder_layers_3_norm_self_att_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_3_norm_self_att_weight_to_fp16, x = input_193_cast_fp16)[name = string("key_7_cast_fp16")]; + bool input_195_interleave_0 = const()[name = string("input_195_interleave_0"), val = bool(false)]; + tensor input_195_cast_fp16 = concat(axis = var_69, interleave = input_195_interleave_0, values = (cache_13_cast_fp16, key_7_cast_fp16))[name = string("input_195_cast_fp16")]; + tensor var_1118_begin_0 = const()[name = string("op_1118_begin_0"), val = tensor([0, 14, 0])]; + tensor var_1118_end_0 = const()[name = string("op_1118_end_0"), val = tensor([1, 42, 1024])]; + tensor var_1118_end_mask_0 = const()[name = string("op_1118_end_mask_0"), val = tensor([true, true, true])]; + tensor var_1118_cast_fp16 = slice_by_index(begin = var_1118_begin_0, end = var_1118_end_0, end_mask = var_1118_end_mask_0, x = cache_13_cast_fp16)[name = string("op_1118_cast_fp16")]; + bool var_1124_interleave_0 = const()[name = string("op_1124_interleave_0"), val = bool(false)]; + tensor var_1124_cast_fp16 = concat(axis = var_69, interleave = var_1124_interleave_0, values = (var_1118_cast_fp16, key_7_cast_fp16))[name = string("op_1124_cast_fp16")]; + tensor encoder_layers_3_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81558784))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82345280))))[name = string("encoder_layers_3_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_3_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_3_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82345472)))]; + tensor linear_30_cast_fp16 = linear(bias = encoder_layers_3_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_3_self_attn_linear_q_weight_to_fp16_palettized, x = key_7_cast_fp16)[name = string("linear_30_cast_fp16")]; + tensor var_1129 = const()[name = string("op_1129"), val = tensor([1, -1, 8, 128])]; + tensor q_19_cast_fp16 = reshape(shape = var_1129, x = linear_30_cast_fp16)[name = string("q_19_cast_fp16")]; + tensor encoder_layers_3_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82347584))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83134080))))[name = string("encoder_layers_3_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_3_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_3_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83134272)))]; + tensor linear_31_cast_fp16 = linear(bias = encoder_layers_3_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_3_self_attn_linear_k_weight_to_fp16_palettized, x = input_195_cast_fp16)[name = string("linear_31_cast_fp16")]; + tensor var_1134 = const()[name = string("op_1134"), val = tensor([1, -1, 8, 128])]; + tensor k_13_cast_fp16 = reshape(shape = var_1134, x = linear_31_cast_fp16)[name = string("k_13_cast_fp16")]; + tensor encoder_layers_3_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83136384))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83922880))))[name = string("encoder_layers_3_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_3_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_3_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83923072)))]; + tensor linear_32_cast_fp16 = linear(bias = encoder_layers_3_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_3_self_attn_linear_v_weight_to_fp16_palettized, x = input_195_cast_fp16)[name = string("linear_32_cast_fp16")]; + tensor var_1139 = const()[name = string("op_1139"), val = tensor([1, -1, 8, 128])]; + tensor v_7_cast_fp16 = reshape(shape = var_1139, x = linear_32_cast_fp16)[name = string("v_7_cast_fp16")]; + tensor value_15_perm_0 = const()[name = string("value_15_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_3_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_3_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83925184)))]; + tensor var_1152_cast_fp16 = add(x = q_19_cast_fp16, y = encoder_layers_3_self_attn_pos_bias_u_to_fp16)[name = string("op_1152_cast_fp16")]; + tensor encoder_layers_3_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_3_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83927296)))]; + tensor var_1154_cast_fp16 = add(x = q_19_cast_fp16, y = encoder_layers_3_self_attn_pos_bias_v_to_fp16)[name = string("op_1154_cast_fp16")]; + tensor q_with_bias_v_7_perm_0 = const()[name = string("q_with_bias_v_7_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_85_transpose_x_0 = const()[name = string("x_85_transpose_x_0"), val = bool(false)]; + bool x_85_transpose_y_0 = const()[name = string("x_85_transpose_y_0"), val = bool(false)]; + tensor op_1156_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83929408))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84043136))))[name = string("op_1156_to_fp16_quantized")]; + tensor q_with_bias_v_7_cast_fp16 = transpose(perm = q_with_bias_v_7_perm_0, x = var_1154_cast_fp16)[name = string("transpose_335")]; + tensor x_85_cast_fp16 = matmul(transpose_x = x_85_transpose_x_0, transpose_y = x_85_transpose_y_0, x = q_with_bias_v_7_cast_fp16, y = op_1156_to_fp16_quantized)[name = string("x_85_cast_fp16")]; + tensor x_87_pad_0 = const()[name = string("x_87_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_87_mode_0 = const()[name = string("x_87_mode_0"), val = string("constant")]; + fp16 const_118_to_fp16 = const()[name = string("const_118_to_fp16"), val = fp16(0x0p+0)]; + tensor x_87_cast_fp16 = pad(constant_val = const_118_to_fp16, mode = x_87_mode_0, pad = x_87_pad_0, x = x_85_cast_fp16)[name = string("x_87_cast_fp16")]; + tensor var_1164 = const()[name = string("op_1164"), val = tensor([1, 8, -1, 14])]; + tensor x_89_cast_fp16 = reshape(shape = var_1164, x = x_87_cast_fp16)[name = string("x_89_cast_fp16")]; + tensor var_1168_begin_0 = const()[name = string("op_1168_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_1168_end_0 = const()[name = string("op_1168_end_0"), val = tensor([1, 8, 112, 14])]; + tensor var_1168_end_mask_0 = const()[name = string("op_1168_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_1168_cast_fp16 = slice_by_index(begin = var_1168_begin_0, end = var_1168_end_0, end_mask = var_1168_end_mask_0, x = x_89_cast_fp16)[name = string("op_1168_cast_fp16")]; + tensor var_1169 = const()[name = string("op_1169"), val = tensor([1, 8, 14, 111])]; + tensor matrix_bd_13_cast_fp16 = reshape(shape = var_1169, x = var_1168_cast_fp16)[name = string("matrix_bd_13_cast_fp16")]; + bool matrix_ac_7_transpose_x_0 = const()[name = string("matrix_ac_7_transpose_x_0"), val = bool(false)]; + bool matrix_ac_7_transpose_y_0 = const()[name = string("matrix_ac_7_transpose_y_0"), val = bool(false)]; + tensor transpose_102_perm_0 = const()[name = string("transpose_102_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_103_perm_0 = const()[name = string("transpose_103_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_103 = transpose(perm = transpose_103_perm_0, x = k_13_cast_fp16)[name = string("transpose_333")]; + tensor transpose_102 = transpose(perm = transpose_102_perm_0, x = var_1152_cast_fp16)[name = string("transpose_334")]; + tensor matrix_ac_7_cast_fp16 = matmul(transpose_x = matrix_ac_7_transpose_x_0, transpose_y = matrix_ac_7_transpose_y_0, x = transpose_102, y = transpose_103)[name = string("matrix_ac_7_cast_fp16")]; + tensor matrix_bd_15_begin_0 = const()[name = string("matrix_bd_15_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_15_end_0 = const()[name = string("matrix_bd_15_end_0"), val = tensor([1, 8, 14, 56])]; + tensor matrix_bd_15_end_mask_0 = const()[name = string("matrix_bd_15_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_15_cast_fp16 = slice_by_index(begin = matrix_bd_15_begin_0, end = matrix_bd_15_end_0, end_mask = matrix_bd_15_end_mask_0, x = matrix_bd_13_cast_fp16)[name = string("matrix_bd_15_cast_fp16")]; + tensor var_1178_cast_fp16 = add(x = matrix_ac_7_cast_fp16, y = matrix_bd_15_cast_fp16)[name = string("op_1178_cast_fp16")]; + fp16 _inversed_scores_13_y_0_to_fp16 = const()[name = string("_inversed_scores_13_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_13_cast_fp16 = mul(x = var_1178_cast_fp16, y = _inversed_scores_13_y_0_to_fp16)[name = string("_inversed_scores_13_cast_fp16")]; + tensor scores_15_cast_fp16 = select(a = var_46_to_fp16, b = _inversed_scores_13_cast_fp16, cond = mask_11)[name = string("scores_15_cast_fp16")]; + tensor var_1184_cast_fp16 = softmax(axis = var_60, x = scores_15_cast_fp16)[name = string("op_1184_cast_fp16")]; + tensor input_197_cast_fp16 = select(a = var_45_to_fp16, b = var_1184_cast_fp16, cond = mask_11)[name = string("input_197_cast_fp16")]; + bool x_91_transpose_x_0 = const()[name = string("x_91_transpose_x_0"), val = bool(false)]; + bool x_91_transpose_y_0 = const()[name = string("x_91_transpose_y_0"), val = bool(false)]; + tensor value_15_cast_fp16 = transpose(perm = value_15_perm_0, x = v_7_cast_fp16)[name = string("transpose_332")]; + tensor x_91_cast_fp16 = matmul(transpose_x = x_91_transpose_x_0, transpose_y = x_91_transpose_y_0, x = input_197_cast_fp16, y = value_15_cast_fp16)[name = string("x_91_cast_fp16")]; + tensor var_1188_perm_0 = const()[name = string("op_1188_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1189 = const()[name = string("op_1189"), val = tensor([1, -1, 1024])]; + tensor var_1188_cast_fp16 = transpose(perm = var_1188_perm_0, x = x_91_cast_fp16)[name = string("transpose_331")]; + tensor input_199_cast_fp16 = reshape(shape = var_1189, x = var_1188_cast_fp16)[name = string("input_199_cast_fp16")]; + tensor encoder_layers_3_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84043456))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84829952))))[name = string("encoder_layers_3_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_3_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_3_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84830144)))]; + tensor linear_34_cast_fp16 = linear(bias = encoder_layers_3_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_3_self_attn_linear_out_weight_to_fp16_palettized, x = input_199_cast_fp16)[name = string("linear_34_cast_fp16")]; + tensor input_203_cast_fp16 = add(x = input_193_cast_fp16, y = linear_34_cast_fp16)[name = string("input_203_cast_fp16")]; + tensor x_95_axes_0 = const()[name = string("x_95_axes_0"), val = tensor([-1])]; + tensor encoder_layers_3_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_3_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84832256)))]; + tensor encoder_layers_3_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_3_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84834368)))]; + tensor x_95_cast_fp16 = layer_norm(axes = x_95_axes_0, beta = encoder_layers_3_norm_conv_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_3_norm_conv_weight_to_fp16, x = input_203_cast_fp16)[name = string("x_95_cast_fp16")]; + tensor input_205_perm_0 = const()[name = string("input_205_perm_0"), val = tensor([0, 2, 1])]; + string input_207_pad_type_0 = const()[name = string("input_207_pad_type_0"), val = string("valid")]; + tensor input_207_strides_0 = const()[name = string("input_207_strides_0"), val = tensor([1])]; + tensor input_207_pad_0 = const()[name = string("input_207_pad_0"), val = tensor([0, 0])]; + tensor input_207_dilations_0 = const()[name = string("input_207_dilations_0"), val = tensor([1])]; + int32 input_207_groups_0 = const()[name = string("input_207_groups_0"), val = int32(1)]; + tensor encoder_layers_3_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84836480))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86933696))))[name = string("encoder_layers_3_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_205_cast_fp16 = transpose(perm = input_205_perm_0, x = x_95_cast_fp16)[name = string("transpose_330")]; + tensor input_207_cast_fp16 = conv(dilations = input_207_dilations_0, groups = input_207_groups_0, pad = input_207_pad_0, pad_type = input_207_pad_type_0, strides = input_207_strides_0, weight = encoder_layers_3_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_205_cast_fp16)[name = string("input_207_cast_fp16")]; + int32 x_97_split_num_splits_0 = const()[name = string("x_97_split_num_splits_0"), val = int32(2)]; + int32 x_97_split_axis_0 = const()[name = string("x_97_split_axis_0"), val = int32(1)]; + tensor x_97_split_cast_fp16_0, tensor x_97_split_cast_fp16_1 = split(axis = x_97_split_axis_0, num_splits = x_97_split_num_splits_0, x = input_207_cast_fp16)[name = string("x_97_split_cast_fp16")]; + tensor x_97_split_1_sigmoid_cast_fp16 = sigmoid(x = x_97_split_cast_fp16_1)[name = string("x_97_split_1_sigmoid_cast_fp16")]; + tensor x_97_cast_fp16 = mul(x = x_97_split_cast_fp16_0, y = x_97_split_1_sigmoid_cast_fp16)[name = string("x_97_cast_fp16")]; + tensor input_209_cast_fp16 = select(a = var_45_to_fp16, b = x_97_cast_fp16, cond = var_576)[name = string("input_209_cast_fp16")]; + bool new_x_15_interleave_0 = const()[name = string("new_x_15_interleave_0"), val = bool(false)]; + tensor new_x_15_cast_fp16 = concat(axis = var_60, interleave = new_x_15_interleave_0, values = (cache_15_cast_fp16, input_209_cast_fp16))[name = string("new_x_15_cast_fp16")]; + tensor var_1228_begin_0 = const()[name = string("op_1228_begin_0"), val = tensor([0, 0, 14])]; + tensor var_1228_end_0 = const()[name = string("op_1228_end_0"), val = tensor([1, 1024, 22])]; + tensor var_1228_end_mask_0 = const()[name = string("op_1228_end_mask_0"), val = tensor([true, true, true])]; + tensor var_1228_cast_fp16 = slice_by_index(begin = var_1228_begin_0, end = var_1228_end_0, end_mask = var_1228_end_mask_0, x = new_x_15_cast_fp16)[name = string("op_1228_cast_fp16")]; + string x_99_pad_type_0 = const()[name = string("x_99_pad_type_0"), val = string("valid")]; + int32 x_99_groups_0 = const()[name = string("x_99_groups_0"), val = int32(1024)]; + tensor x_99_strides_0 = const()[name = string("x_99_strides_0"), val = tensor([1])]; + tensor x_99_pad_0 = const()[name = string("x_99_pad_0"), val = tensor([0, 0])]; + tensor x_99_dilations_0 = const()[name = string("x_99_dilations_0"), val = tensor([1])]; + tensor encoder_layers_3_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86937856))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86947136))))[name = string("encoder_layers_3_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_99_cast_fp16 = conv(dilations = x_99_dilations_0, groups = x_99_groups_0, pad = x_99_pad_0, pad_type = x_99_pad_type_0, strides = x_99_strides_0, weight = encoder_layers_3_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_15_cast_fp16)[name = string("x_99_cast_fp16")]; + tensor input_211_perm_0 = const()[name = string("input_211_perm_0"), val = tensor([0, 2, 1])]; + tensor x_101_axes_0 = const()[name = string("x_101_axes_0"), val = tensor([-1])]; + tensor encoder_layers_3_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_3_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86949248)))]; + tensor encoder_layers_3_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_3_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86951360)))]; + tensor input_211_cast_fp16 = transpose(perm = input_211_perm_0, x = x_99_cast_fp16)[name = string("transpose_329")]; + tensor x_101_cast_fp16 = layer_norm(axes = x_101_axes_0, beta = encoder_layers_3_conv_batch_norm_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_3_conv_batch_norm_weight_to_fp16, x = input_211_cast_fp16)[name = string("x_101_cast_fp16")]; + tensor input_213_perm_0 = const()[name = string("input_213_perm_0"), val = tensor([0, 2, 1])]; + tensor input_213_cast_fp16 = transpose(perm = input_213_perm_0, x = x_101_cast_fp16)[name = string("transpose_328")]; + tensor input_215_cast_fp16 = silu(x = input_213_cast_fp16)[name = string("input_215_cast_fp16")]; + string x_103_pad_type_0 = const()[name = string("x_103_pad_type_0"), val = string("valid")]; + tensor x_103_strides_0 = const()[name = string("x_103_strides_0"), val = tensor([1])]; + tensor x_103_pad_0 = const()[name = string("x_103_pad_0"), val = tensor([0, 0])]; + tensor x_103_dilations_0 = const()[name = string("x_103_dilations_0"), val = tensor([1])]; + int32 x_103_groups_0 = const()[name = string("x_103_groups_0"), val = int32(1)]; + tensor encoder_layers_3_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86953472))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88002112))))[name = string("encoder_layers_3_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_103_cast_fp16 = conv(dilations = x_103_dilations_0, groups = x_103_groups_0, pad = x_103_pad_0, pad_type = x_103_pad_type_0, strides = x_103_strides_0, weight = encoder_layers_3_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_215_cast_fp16)[name = string("x_103_cast_fp16")]; + tensor input_217_perm_0 = const()[name = string("input_217_perm_0"), val = tensor([0, 2, 1])]; + tensor input_217_cast_fp16 = transpose(perm = input_217_perm_0, x = x_103_cast_fp16)[name = string("transpose_327")]; + tensor input_219_cast_fp16 = add(x = input_203_cast_fp16, y = input_217_cast_fp16)[name = string("input_219_cast_fp16")]; + tensor input_221_axes_0 = const()[name = string("input_221_axes_0"), val = tensor([-1])]; + tensor encoder_layers_3_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_3_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88004224)))]; + tensor encoder_layers_3_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_3_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88006336)))]; + tensor input_221_cast_fp16 = layer_norm(axes = input_221_axes_0, beta = encoder_layers_3_norm_feed_forward2_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_3_norm_feed_forward2_weight_to_fp16, x = input_219_cast_fp16)[name = string("input_221_cast_fp16")]; + tensor encoder_layers_3_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88008448))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(91154240))))[name = string("encoder_layers_3_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_3_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_3_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(91154432)))]; + tensor linear_35_cast_fp16 = linear(bias = encoder_layers_3_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_3_feed_forward2_linear1_weight_to_fp16_palettized, x = input_221_cast_fp16)[name = string("linear_35_cast_fp16")]; + tensor input_225_cast_fp16 = silu(x = linear_35_cast_fp16)[name = string("input_225_cast_fp16")]; + tensor encoder_layers_3_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(91162688))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94308480))))[name = string("encoder_layers_3_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_3_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_3_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94308672)))]; + tensor linear_36_cast_fp16 = linear(bias = encoder_layers_3_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_3_feed_forward2_linear2_weight_to_fp16_palettized, x = input_225_cast_fp16)[name = string("linear_36_cast_fp16")]; + fp16 var_1271_to_fp16 = const()[name = string("op_1271_to_fp16"), val = fp16(0x1p-1)]; + tensor var_1272_cast_fp16 = mul(x = linear_36_cast_fp16, y = var_1271_to_fp16)[name = string("op_1272_cast_fp16")]; + tensor input_231_cast_fp16 = add(x = input_219_cast_fp16, y = var_1272_cast_fp16)[name = string("input_231_cast_fp16")]; + tensor input_233_axes_0 = const()[name = string("input_233_axes_0"), val = tensor([-1])]; + tensor encoder_layers_3_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_3_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94310784)))]; + tensor encoder_layers_3_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_3_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94312896)))]; + tensor input_233_cast_fp16 = layer_norm(axes = input_233_axes_0, beta = encoder_layers_3_norm_out_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_3_norm_out_weight_to_fp16, x = input_231_cast_fp16)[name = string("input_233_cast_fp16")]; + tensor cache_17_begin_0 = const()[name = string("cache_17_begin_0"), val = tensor([4, 0, 0, 0])]; + tensor cache_17_end_0 = const()[name = string("cache_17_end_0"), val = tensor([5, 1, 42, 1024])]; + tensor cache_17_end_mask_0 = const()[name = string("cache_17_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_17_squeeze_mask_0 = const()[name = string("cache_17_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_17_cast_fp16 = slice_by_index(begin = cache_17_begin_0, end = cache_17_end_0, end_mask = cache_17_end_mask_0, squeeze_mask = cache_17_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_17_cast_fp16")]; + tensor cache_19_begin_0 = const()[name = string("cache_19_begin_0"), val = tensor([4, 0, 0, 0])]; + tensor cache_19_end_0 = const()[name = string("cache_19_end_0"), val = tensor([5, 1, 1024, 8])]; + tensor cache_19_end_mask_0 = const()[name = string("cache_19_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_19_squeeze_mask_0 = const()[name = string("cache_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_19_cast_fp16 = slice_by_index(begin = cache_19_begin_0, end = cache_19_end_0, end_mask = cache_19_end_mask_0, squeeze_mask = cache_19_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_19_cast_fp16")]; + tensor input_235_axes_0 = const()[name = string("input_235_axes_0"), val = tensor([-1])]; + tensor encoder_layers_4_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_4_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94315008)))]; + tensor encoder_layers_4_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_4_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94317120)))]; + tensor input_235_cast_fp16 = layer_norm(axes = input_235_axes_0, beta = encoder_layers_4_norm_feed_forward1_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_4_norm_feed_forward1_weight_to_fp16, x = input_233_cast_fp16)[name = string("input_235_cast_fp16")]; + tensor encoder_layers_4_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94319232))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97465024))))[name = string("encoder_layers_4_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_4_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_4_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97465216)))]; + tensor linear_37_cast_fp16 = linear(bias = encoder_layers_4_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_4_feed_forward1_linear1_weight_to_fp16_palettized, x = input_235_cast_fp16)[name = string("linear_37_cast_fp16")]; + tensor input_239_cast_fp16 = silu(x = linear_37_cast_fp16)[name = string("input_239_cast_fp16")]; + tensor encoder_layers_4_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97473472))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100619264))))[name = string("encoder_layers_4_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_4_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_4_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100619456)))]; + tensor linear_38_cast_fp16 = linear(bias = encoder_layers_4_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_4_feed_forward1_linear2_weight_to_fp16_palettized, x = input_239_cast_fp16)[name = string("linear_38_cast_fp16")]; + fp16 var_1308_to_fp16 = const()[name = string("op_1308_to_fp16"), val = fp16(0x1p-1)]; + tensor var_1309_cast_fp16 = mul(x = linear_38_cast_fp16, y = var_1308_to_fp16)[name = string("op_1309_cast_fp16")]; + tensor input_245_cast_fp16 = add(x = input_233_cast_fp16, y = var_1309_cast_fp16)[name = string("input_245_cast_fp16")]; + tensor key_9_axes_0 = const()[name = string("key_9_axes_0"), val = tensor([-1])]; + tensor encoder_layers_4_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_4_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100621568)))]; + tensor encoder_layers_4_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_4_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100623680)))]; + tensor key_9_cast_fp16 = layer_norm(axes = key_9_axes_0, beta = encoder_layers_4_norm_self_att_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_4_norm_self_att_weight_to_fp16, x = input_245_cast_fp16)[name = string("key_9_cast_fp16")]; + bool input_247_interleave_0 = const()[name = string("input_247_interleave_0"), val = bool(false)]; + tensor input_247_cast_fp16 = concat(axis = var_69, interleave = input_247_interleave_0, values = (cache_17_cast_fp16, key_9_cast_fp16))[name = string("input_247_cast_fp16")]; + tensor var_1331_begin_0 = const()[name = string("op_1331_begin_0"), val = tensor([0, 14, 0])]; + tensor var_1331_end_0 = const()[name = string("op_1331_end_0"), val = tensor([1, 42, 1024])]; + tensor var_1331_end_mask_0 = const()[name = string("op_1331_end_mask_0"), val = tensor([true, true, true])]; + tensor var_1331_cast_fp16 = slice_by_index(begin = var_1331_begin_0, end = var_1331_end_0, end_mask = var_1331_end_mask_0, x = cache_17_cast_fp16)[name = string("op_1331_cast_fp16")]; + bool var_1337_interleave_0 = const()[name = string("op_1337_interleave_0"), val = bool(false)]; + tensor var_1337_cast_fp16 = concat(axis = var_69, interleave = var_1337_interleave_0, values = (var_1331_cast_fp16, key_9_cast_fp16))[name = string("op_1337_cast_fp16")]; + tensor encoder_layers_4_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100625792))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(101412288))))[name = string("encoder_layers_4_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_4_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_4_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(101412480)))]; + tensor linear_39_cast_fp16 = linear(bias = encoder_layers_4_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_4_self_attn_linear_q_weight_to_fp16_palettized, x = key_9_cast_fp16)[name = string("linear_39_cast_fp16")]; + tensor var_1342 = const()[name = string("op_1342"), val = tensor([1, -1, 8, 128])]; + tensor q_25_cast_fp16 = reshape(shape = var_1342, x = linear_39_cast_fp16)[name = string("q_25_cast_fp16")]; + tensor encoder_layers_4_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(101414592))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102201088))))[name = string("encoder_layers_4_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_4_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_4_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102201280)))]; + tensor linear_40_cast_fp16 = linear(bias = encoder_layers_4_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_4_self_attn_linear_k_weight_to_fp16_palettized, x = input_247_cast_fp16)[name = string("linear_40_cast_fp16")]; + tensor var_1347 = const()[name = string("op_1347"), val = tensor([1, -1, 8, 128])]; + tensor k_17_cast_fp16 = reshape(shape = var_1347, x = linear_40_cast_fp16)[name = string("k_17_cast_fp16")]; + tensor encoder_layers_4_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102203392))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102989888))))[name = string("encoder_layers_4_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_4_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_4_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102990080)))]; + tensor linear_41_cast_fp16 = linear(bias = encoder_layers_4_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_4_self_attn_linear_v_weight_to_fp16_palettized, x = input_247_cast_fp16)[name = string("linear_41_cast_fp16")]; + tensor var_1352 = const()[name = string("op_1352"), val = tensor([1, -1, 8, 128])]; + tensor v_9_cast_fp16 = reshape(shape = var_1352, x = linear_41_cast_fp16)[name = string("v_9_cast_fp16")]; + tensor value_17_perm_0 = const()[name = string("value_17_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_4_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_4_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102992192)))]; + tensor var_1365_cast_fp16 = add(x = q_25_cast_fp16, y = encoder_layers_4_self_attn_pos_bias_u_to_fp16)[name = string("op_1365_cast_fp16")]; + tensor encoder_layers_4_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_4_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102994304)))]; + tensor var_1367_cast_fp16 = add(x = q_25_cast_fp16, y = encoder_layers_4_self_attn_pos_bias_v_to_fp16)[name = string("op_1367_cast_fp16")]; + tensor q_with_bias_v_9_perm_0 = const()[name = string("q_with_bias_v_9_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_111_transpose_x_0 = const()[name = string("x_111_transpose_x_0"), val = bool(false)]; + bool x_111_transpose_y_0 = const()[name = string("x_111_transpose_y_0"), val = bool(false)]; + tensor op_1369_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102996416))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103110144))))[name = string("op_1369_to_fp16_quantized")]; + tensor q_with_bias_v_9_cast_fp16 = transpose(perm = q_with_bias_v_9_perm_0, x = var_1367_cast_fp16)[name = string("transpose_326")]; + tensor x_111_cast_fp16 = matmul(transpose_x = x_111_transpose_x_0, transpose_y = x_111_transpose_y_0, x = q_with_bias_v_9_cast_fp16, y = op_1369_to_fp16_quantized)[name = string("x_111_cast_fp16")]; + tensor x_113_pad_0 = const()[name = string("x_113_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_113_mode_0 = const()[name = string("x_113_mode_0"), val = string("constant")]; + fp16 const_131_to_fp16 = const()[name = string("const_131_to_fp16"), val = fp16(0x0p+0)]; + tensor x_113_cast_fp16 = pad(constant_val = const_131_to_fp16, mode = x_113_mode_0, pad = x_113_pad_0, x = x_111_cast_fp16)[name = string("x_113_cast_fp16")]; + tensor var_1377 = const()[name = string("op_1377"), val = tensor([1, 8, -1, 14])]; + tensor x_115_cast_fp16 = reshape(shape = var_1377, x = x_113_cast_fp16)[name = string("x_115_cast_fp16")]; + tensor var_1381_begin_0 = const()[name = string("op_1381_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_1381_end_0 = const()[name = string("op_1381_end_0"), val = tensor([1, 8, 112, 14])]; + tensor var_1381_end_mask_0 = const()[name = string("op_1381_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_1381_cast_fp16 = slice_by_index(begin = var_1381_begin_0, end = var_1381_end_0, end_mask = var_1381_end_mask_0, x = x_115_cast_fp16)[name = string("op_1381_cast_fp16")]; + tensor var_1382 = const()[name = string("op_1382"), val = tensor([1, 8, 14, 111])]; + tensor matrix_bd_17_cast_fp16 = reshape(shape = var_1382, x = var_1381_cast_fp16)[name = string("matrix_bd_17_cast_fp16")]; + bool matrix_ac_9_transpose_x_0 = const()[name = string("matrix_ac_9_transpose_x_0"), val = bool(false)]; + bool matrix_ac_9_transpose_y_0 = const()[name = string("matrix_ac_9_transpose_y_0"), val = bool(false)]; + tensor transpose_104_perm_0 = const()[name = string("transpose_104_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_105_perm_0 = const()[name = string("transpose_105_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_105 = transpose(perm = transpose_105_perm_0, x = k_17_cast_fp16)[name = string("transpose_324")]; + tensor transpose_104 = transpose(perm = transpose_104_perm_0, x = var_1365_cast_fp16)[name = string("transpose_325")]; + tensor matrix_ac_9_cast_fp16 = matmul(transpose_x = matrix_ac_9_transpose_x_0, transpose_y = matrix_ac_9_transpose_y_0, x = transpose_104, y = transpose_105)[name = string("matrix_ac_9_cast_fp16")]; + tensor matrix_bd_19_begin_0 = const()[name = string("matrix_bd_19_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_19_end_0 = const()[name = string("matrix_bd_19_end_0"), val = tensor([1, 8, 14, 56])]; + tensor matrix_bd_19_end_mask_0 = const()[name = string("matrix_bd_19_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_19_cast_fp16 = slice_by_index(begin = matrix_bd_19_begin_0, end = matrix_bd_19_end_0, end_mask = matrix_bd_19_end_mask_0, x = matrix_bd_17_cast_fp16)[name = string("matrix_bd_19_cast_fp16")]; + tensor var_1391_cast_fp16 = add(x = matrix_ac_9_cast_fp16, y = matrix_bd_19_cast_fp16)[name = string("op_1391_cast_fp16")]; + fp16 _inversed_scores_17_y_0_to_fp16 = const()[name = string("_inversed_scores_17_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_17_cast_fp16 = mul(x = var_1391_cast_fp16, y = _inversed_scores_17_y_0_to_fp16)[name = string("_inversed_scores_17_cast_fp16")]; + tensor scores_19_cast_fp16 = select(a = var_46_to_fp16, b = _inversed_scores_17_cast_fp16, cond = mask_11)[name = string("scores_19_cast_fp16")]; + tensor var_1397_cast_fp16 = softmax(axis = var_60, x = scores_19_cast_fp16)[name = string("op_1397_cast_fp16")]; + tensor input_249_cast_fp16 = select(a = var_45_to_fp16, b = var_1397_cast_fp16, cond = mask_11)[name = string("input_249_cast_fp16")]; + bool x_117_transpose_x_0 = const()[name = string("x_117_transpose_x_0"), val = bool(false)]; + bool x_117_transpose_y_0 = const()[name = string("x_117_transpose_y_0"), val = bool(false)]; + tensor value_17_cast_fp16 = transpose(perm = value_17_perm_0, x = v_9_cast_fp16)[name = string("transpose_323")]; + tensor x_117_cast_fp16 = matmul(transpose_x = x_117_transpose_x_0, transpose_y = x_117_transpose_y_0, x = input_249_cast_fp16, y = value_17_cast_fp16)[name = string("x_117_cast_fp16")]; + tensor var_1401_perm_0 = const()[name = string("op_1401_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1402 = const()[name = string("op_1402"), val = tensor([1, -1, 1024])]; + tensor var_1401_cast_fp16 = transpose(perm = var_1401_perm_0, x = x_117_cast_fp16)[name = string("transpose_322")]; + tensor input_251_cast_fp16 = reshape(shape = var_1402, x = var_1401_cast_fp16)[name = string("input_251_cast_fp16")]; + tensor encoder_layers_4_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103110464))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103896960))))[name = string("encoder_layers_4_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_4_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_4_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103897152)))]; + tensor linear_43_cast_fp16 = linear(bias = encoder_layers_4_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_4_self_attn_linear_out_weight_to_fp16_palettized, x = input_251_cast_fp16)[name = string("linear_43_cast_fp16")]; + tensor input_255_cast_fp16 = add(x = input_245_cast_fp16, y = linear_43_cast_fp16)[name = string("input_255_cast_fp16")]; + tensor x_121_axes_0 = const()[name = string("x_121_axes_0"), val = tensor([-1])]; + tensor encoder_layers_4_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_4_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103899264)))]; + tensor encoder_layers_4_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_4_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103901376)))]; + tensor x_121_cast_fp16 = layer_norm(axes = x_121_axes_0, beta = encoder_layers_4_norm_conv_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_4_norm_conv_weight_to_fp16, x = input_255_cast_fp16)[name = string("x_121_cast_fp16")]; + tensor input_257_perm_0 = const()[name = string("input_257_perm_0"), val = tensor([0, 2, 1])]; + string input_259_pad_type_0 = const()[name = string("input_259_pad_type_0"), val = string("valid")]; + tensor input_259_strides_0 = const()[name = string("input_259_strides_0"), val = tensor([1])]; + tensor input_259_pad_0 = const()[name = string("input_259_pad_0"), val = tensor([0, 0])]; + tensor input_259_dilations_0 = const()[name = string("input_259_dilations_0"), val = tensor([1])]; + int32 input_259_groups_0 = const()[name = string("input_259_groups_0"), val = int32(1)]; + tensor encoder_layers_4_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103903488))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106000704))))[name = string("encoder_layers_4_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_257_cast_fp16 = transpose(perm = input_257_perm_0, x = x_121_cast_fp16)[name = string("transpose_321")]; + tensor input_259_cast_fp16 = conv(dilations = input_259_dilations_0, groups = input_259_groups_0, pad = input_259_pad_0, pad_type = input_259_pad_type_0, strides = input_259_strides_0, weight = encoder_layers_4_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_257_cast_fp16)[name = string("input_259_cast_fp16")]; + int32 x_123_split_num_splits_0 = const()[name = string("x_123_split_num_splits_0"), val = int32(2)]; + int32 x_123_split_axis_0 = const()[name = string("x_123_split_axis_0"), val = int32(1)]; + tensor x_123_split_cast_fp16_0, tensor x_123_split_cast_fp16_1 = split(axis = x_123_split_axis_0, num_splits = x_123_split_num_splits_0, x = input_259_cast_fp16)[name = string("x_123_split_cast_fp16")]; + tensor x_123_split_1_sigmoid_cast_fp16 = sigmoid(x = x_123_split_cast_fp16_1)[name = string("x_123_split_1_sigmoid_cast_fp16")]; + tensor x_123_cast_fp16 = mul(x = x_123_split_cast_fp16_0, y = x_123_split_1_sigmoid_cast_fp16)[name = string("x_123_cast_fp16")]; + tensor input_261_cast_fp16 = select(a = var_45_to_fp16, b = x_123_cast_fp16, cond = var_576)[name = string("input_261_cast_fp16")]; + bool new_x_19_interleave_0 = const()[name = string("new_x_19_interleave_0"), val = bool(false)]; + tensor new_x_19_cast_fp16 = concat(axis = var_60, interleave = new_x_19_interleave_0, values = (cache_19_cast_fp16, input_261_cast_fp16))[name = string("new_x_19_cast_fp16")]; + tensor var_1441_begin_0 = const()[name = string("op_1441_begin_0"), val = tensor([0, 0, 14])]; + tensor var_1441_end_0 = const()[name = string("op_1441_end_0"), val = tensor([1, 1024, 22])]; + tensor var_1441_end_mask_0 = const()[name = string("op_1441_end_mask_0"), val = tensor([true, true, true])]; + tensor var_1441_cast_fp16 = slice_by_index(begin = var_1441_begin_0, end = var_1441_end_0, end_mask = var_1441_end_mask_0, x = new_x_19_cast_fp16)[name = string("op_1441_cast_fp16")]; + string x_125_pad_type_0 = const()[name = string("x_125_pad_type_0"), val = string("valid")]; + int32 x_125_groups_0 = const()[name = string("x_125_groups_0"), val = int32(1024)]; + tensor x_125_strides_0 = const()[name = string("x_125_strides_0"), val = tensor([1])]; + tensor x_125_pad_0 = const()[name = string("x_125_pad_0"), val = tensor([0, 0])]; + tensor x_125_dilations_0 = const()[name = string("x_125_dilations_0"), val = tensor([1])]; + tensor encoder_layers_4_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106004864))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106014144))))[name = string("encoder_layers_4_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_125_cast_fp16 = conv(dilations = x_125_dilations_0, groups = x_125_groups_0, pad = x_125_pad_0, pad_type = x_125_pad_type_0, strides = x_125_strides_0, weight = encoder_layers_4_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_19_cast_fp16)[name = string("x_125_cast_fp16")]; + tensor input_263_perm_0 = const()[name = string("input_263_perm_0"), val = tensor([0, 2, 1])]; + tensor x_127_axes_0 = const()[name = string("x_127_axes_0"), val = tensor([-1])]; + tensor encoder_layers_4_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_4_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106016256)))]; + tensor encoder_layers_4_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_4_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106018368)))]; + tensor input_263_cast_fp16 = transpose(perm = input_263_perm_0, x = x_125_cast_fp16)[name = string("transpose_320")]; + tensor x_127_cast_fp16 = layer_norm(axes = x_127_axes_0, beta = encoder_layers_4_conv_batch_norm_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_4_conv_batch_norm_weight_to_fp16, x = input_263_cast_fp16)[name = string("x_127_cast_fp16")]; + tensor input_265_perm_0 = const()[name = string("input_265_perm_0"), val = tensor([0, 2, 1])]; + tensor input_265_cast_fp16 = transpose(perm = input_265_perm_0, x = x_127_cast_fp16)[name = string("transpose_319")]; + tensor input_267_cast_fp16 = silu(x = input_265_cast_fp16)[name = string("input_267_cast_fp16")]; + string x_129_pad_type_0 = const()[name = string("x_129_pad_type_0"), val = string("valid")]; + tensor x_129_strides_0 = const()[name = string("x_129_strides_0"), val = tensor([1])]; + tensor x_129_pad_0 = const()[name = string("x_129_pad_0"), val = tensor([0, 0])]; + tensor x_129_dilations_0 = const()[name = string("x_129_dilations_0"), val = tensor([1])]; + int32 x_129_groups_0 = const()[name = string("x_129_groups_0"), val = int32(1)]; + tensor encoder_layers_4_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106020480))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107069120))))[name = string("encoder_layers_4_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_129_cast_fp16 = conv(dilations = x_129_dilations_0, groups = x_129_groups_0, pad = x_129_pad_0, pad_type = x_129_pad_type_0, strides = x_129_strides_0, weight = encoder_layers_4_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_267_cast_fp16)[name = string("x_129_cast_fp16")]; + tensor input_269_perm_0 = const()[name = string("input_269_perm_0"), val = tensor([0, 2, 1])]; + tensor input_269_cast_fp16 = transpose(perm = input_269_perm_0, x = x_129_cast_fp16)[name = string("transpose_318")]; + tensor input_271_cast_fp16 = add(x = input_255_cast_fp16, y = input_269_cast_fp16)[name = string("input_271_cast_fp16")]; + tensor input_273_axes_0 = const()[name = string("input_273_axes_0"), val = tensor([-1])]; + tensor encoder_layers_4_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_4_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107071232)))]; + tensor encoder_layers_4_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_4_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107073344)))]; + tensor input_273_cast_fp16 = layer_norm(axes = input_273_axes_0, beta = encoder_layers_4_norm_feed_forward2_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_4_norm_feed_forward2_weight_to_fp16, x = input_271_cast_fp16)[name = string("input_273_cast_fp16")]; + tensor encoder_layers_4_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107075456))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110221248))))[name = string("encoder_layers_4_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_4_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_4_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110221440)))]; + tensor linear_44_cast_fp16 = linear(bias = encoder_layers_4_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_4_feed_forward2_linear1_weight_to_fp16_palettized, x = input_273_cast_fp16)[name = string("linear_44_cast_fp16")]; + tensor input_277_cast_fp16 = silu(x = linear_44_cast_fp16)[name = string("input_277_cast_fp16")]; + tensor encoder_layers_4_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110229696))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113375488))))[name = string("encoder_layers_4_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_4_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_4_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113375680)))]; + tensor linear_45_cast_fp16 = linear(bias = encoder_layers_4_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_4_feed_forward2_linear2_weight_to_fp16_palettized, x = input_277_cast_fp16)[name = string("linear_45_cast_fp16")]; + fp16 var_1484_to_fp16 = const()[name = string("op_1484_to_fp16"), val = fp16(0x1p-1)]; + tensor var_1485_cast_fp16 = mul(x = linear_45_cast_fp16, y = var_1484_to_fp16)[name = string("op_1485_cast_fp16")]; + tensor input_283_cast_fp16 = add(x = input_271_cast_fp16, y = var_1485_cast_fp16)[name = string("input_283_cast_fp16")]; + tensor input_285_axes_0 = const()[name = string("input_285_axes_0"), val = tensor([-1])]; + tensor encoder_layers_4_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_4_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113377792)))]; + tensor encoder_layers_4_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_4_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113379904)))]; + tensor input_285_cast_fp16 = layer_norm(axes = input_285_axes_0, beta = encoder_layers_4_norm_out_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_4_norm_out_weight_to_fp16, x = input_283_cast_fp16)[name = string("input_285_cast_fp16")]; + tensor cache_21_begin_0 = const()[name = string("cache_21_begin_0"), val = tensor([5, 0, 0, 0])]; + tensor cache_21_end_0 = const()[name = string("cache_21_end_0"), val = tensor([6, 1, 42, 1024])]; + tensor cache_21_end_mask_0 = const()[name = string("cache_21_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_21_squeeze_mask_0 = const()[name = string("cache_21_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_21_cast_fp16 = slice_by_index(begin = cache_21_begin_0, end = cache_21_end_0, end_mask = cache_21_end_mask_0, squeeze_mask = cache_21_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_21_cast_fp16")]; + tensor cache_23_begin_0 = const()[name = string("cache_23_begin_0"), val = tensor([5, 0, 0, 0])]; + tensor cache_23_end_0 = const()[name = string("cache_23_end_0"), val = tensor([6, 1, 1024, 8])]; + tensor cache_23_end_mask_0 = const()[name = string("cache_23_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_23_squeeze_mask_0 = const()[name = string("cache_23_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_23_cast_fp16 = slice_by_index(begin = cache_23_begin_0, end = cache_23_end_0, end_mask = cache_23_end_mask_0, squeeze_mask = cache_23_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_23_cast_fp16")]; + tensor input_287_axes_0 = const()[name = string("input_287_axes_0"), val = tensor([-1])]; + tensor encoder_layers_5_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_5_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113382016)))]; + tensor encoder_layers_5_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_5_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113384128)))]; + tensor input_287_cast_fp16 = layer_norm(axes = input_287_axes_0, beta = encoder_layers_5_norm_feed_forward1_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_5_norm_feed_forward1_weight_to_fp16, x = input_285_cast_fp16)[name = string("input_287_cast_fp16")]; + tensor encoder_layers_5_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113386240))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116532032))))[name = string("encoder_layers_5_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_5_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_5_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116532224)))]; + tensor linear_46_cast_fp16 = linear(bias = encoder_layers_5_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_5_feed_forward1_linear1_weight_to_fp16_palettized, x = input_287_cast_fp16)[name = string("linear_46_cast_fp16")]; + tensor input_291_cast_fp16 = silu(x = linear_46_cast_fp16)[name = string("input_291_cast_fp16")]; + tensor encoder_layers_5_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116540480))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119686272))))[name = string("encoder_layers_5_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_5_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_5_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119686464)))]; + tensor linear_47_cast_fp16 = linear(bias = encoder_layers_5_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_5_feed_forward1_linear2_weight_to_fp16_palettized, x = input_291_cast_fp16)[name = string("linear_47_cast_fp16")]; + fp16 var_1521_to_fp16 = const()[name = string("op_1521_to_fp16"), val = fp16(0x1p-1)]; + tensor var_1522_cast_fp16 = mul(x = linear_47_cast_fp16, y = var_1521_to_fp16)[name = string("op_1522_cast_fp16")]; + tensor input_297_cast_fp16 = add(x = input_285_cast_fp16, y = var_1522_cast_fp16)[name = string("input_297_cast_fp16")]; + tensor key_11_axes_0 = const()[name = string("key_11_axes_0"), val = tensor([-1])]; + tensor encoder_layers_5_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_5_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119688576)))]; + tensor encoder_layers_5_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_5_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119690688)))]; + tensor key_11_cast_fp16 = layer_norm(axes = key_11_axes_0, beta = encoder_layers_5_norm_self_att_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_5_norm_self_att_weight_to_fp16, x = input_297_cast_fp16)[name = string("key_11_cast_fp16")]; + bool input_299_interleave_0 = const()[name = string("input_299_interleave_0"), val = bool(false)]; + tensor input_299_cast_fp16 = concat(axis = var_69, interleave = input_299_interleave_0, values = (cache_21_cast_fp16, key_11_cast_fp16))[name = string("input_299_cast_fp16")]; + tensor var_1544_begin_0 = const()[name = string("op_1544_begin_0"), val = tensor([0, 14, 0])]; + tensor var_1544_end_0 = const()[name = string("op_1544_end_0"), val = tensor([1, 42, 1024])]; + tensor var_1544_end_mask_0 = const()[name = string("op_1544_end_mask_0"), val = tensor([true, true, true])]; + tensor var_1544_cast_fp16 = slice_by_index(begin = var_1544_begin_0, end = var_1544_end_0, end_mask = var_1544_end_mask_0, x = cache_21_cast_fp16)[name = string("op_1544_cast_fp16")]; + bool var_1550_interleave_0 = const()[name = string("op_1550_interleave_0"), val = bool(false)]; + tensor var_1550_cast_fp16 = concat(axis = var_69, interleave = var_1550_interleave_0, values = (var_1544_cast_fp16, key_11_cast_fp16))[name = string("op_1550_cast_fp16")]; + tensor encoder_layers_5_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119692800))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120479296))))[name = string("encoder_layers_5_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_5_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_5_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120479488)))]; + tensor linear_48_cast_fp16 = linear(bias = encoder_layers_5_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_5_self_attn_linear_q_weight_to_fp16_palettized, x = key_11_cast_fp16)[name = string("linear_48_cast_fp16")]; + tensor var_1555 = const()[name = string("op_1555"), val = tensor([1, -1, 8, 128])]; + tensor q_31_cast_fp16 = reshape(shape = var_1555, x = linear_48_cast_fp16)[name = string("q_31_cast_fp16")]; + tensor encoder_layers_5_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120481600))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121268096))))[name = string("encoder_layers_5_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_5_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_5_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121268288)))]; + tensor linear_49_cast_fp16 = linear(bias = encoder_layers_5_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_5_self_attn_linear_k_weight_to_fp16_palettized, x = input_299_cast_fp16)[name = string("linear_49_cast_fp16")]; + tensor var_1560 = const()[name = string("op_1560"), val = tensor([1, -1, 8, 128])]; + tensor k_21_cast_fp16 = reshape(shape = var_1560, x = linear_49_cast_fp16)[name = string("k_21_cast_fp16")]; + tensor encoder_layers_5_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121270400))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122056896))))[name = string("encoder_layers_5_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_5_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_5_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122057088)))]; + tensor linear_50_cast_fp16 = linear(bias = encoder_layers_5_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_5_self_attn_linear_v_weight_to_fp16_palettized, x = input_299_cast_fp16)[name = string("linear_50_cast_fp16")]; + tensor var_1565 = const()[name = string("op_1565"), val = tensor([1, -1, 8, 128])]; + tensor v_11_cast_fp16 = reshape(shape = var_1565, x = linear_50_cast_fp16)[name = string("v_11_cast_fp16")]; + tensor value_19_perm_0 = const()[name = string("value_19_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_5_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_5_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122059200)))]; + tensor var_1578_cast_fp16 = add(x = q_31_cast_fp16, y = encoder_layers_5_self_attn_pos_bias_u_to_fp16)[name = string("op_1578_cast_fp16")]; + tensor encoder_layers_5_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_5_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122061312)))]; + tensor var_1580_cast_fp16 = add(x = q_31_cast_fp16, y = encoder_layers_5_self_attn_pos_bias_v_to_fp16)[name = string("op_1580_cast_fp16")]; + tensor q_with_bias_v_11_perm_0 = const()[name = string("q_with_bias_v_11_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_137_transpose_x_0 = const()[name = string("x_137_transpose_x_0"), val = bool(false)]; + bool x_137_transpose_y_0 = const()[name = string("x_137_transpose_y_0"), val = bool(false)]; + tensor op_1582_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122063424))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122177152))))[name = string("op_1582_to_fp16_quantized")]; + tensor q_with_bias_v_11_cast_fp16 = transpose(perm = q_with_bias_v_11_perm_0, x = var_1580_cast_fp16)[name = string("transpose_317")]; + tensor x_137_cast_fp16 = matmul(transpose_x = x_137_transpose_x_0, transpose_y = x_137_transpose_y_0, x = q_with_bias_v_11_cast_fp16, y = op_1582_to_fp16_quantized)[name = string("x_137_cast_fp16")]; + tensor x_139_pad_0 = const()[name = string("x_139_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_139_mode_0 = const()[name = string("x_139_mode_0"), val = string("constant")]; + fp16 const_144_to_fp16 = const()[name = string("const_144_to_fp16"), val = fp16(0x0p+0)]; + tensor x_139_cast_fp16 = pad(constant_val = const_144_to_fp16, mode = x_139_mode_0, pad = x_139_pad_0, x = x_137_cast_fp16)[name = string("x_139_cast_fp16")]; + tensor var_1590 = const()[name = string("op_1590"), val = tensor([1, 8, -1, 14])]; + tensor x_141_cast_fp16 = reshape(shape = var_1590, x = x_139_cast_fp16)[name = string("x_141_cast_fp16")]; + tensor var_1594_begin_0 = const()[name = string("op_1594_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_1594_end_0 = const()[name = string("op_1594_end_0"), val = tensor([1, 8, 112, 14])]; + tensor var_1594_end_mask_0 = const()[name = string("op_1594_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_1594_cast_fp16 = slice_by_index(begin = var_1594_begin_0, end = var_1594_end_0, end_mask = var_1594_end_mask_0, x = x_141_cast_fp16)[name = string("op_1594_cast_fp16")]; + tensor var_1595 = const()[name = string("op_1595"), val = tensor([1, 8, 14, 111])]; + tensor matrix_bd_21_cast_fp16 = reshape(shape = var_1595, x = var_1594_cast_fp16)[name = string("matrix_bd_21_cast_fp16")]; + bool matrix_ac_11_transpose_x_0 = const()[name = string("matrix_ac_11_transpose_x_0"), val = bool(false)]; + bool matrix_ac_11_transpose_y_0 = const()[name = string("matrix_ac_11_transpose_y_0"), val = bool(false)]; + tensor transpose_106_perm_0 = const()[name = string("transpose_106_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_107_perm_0 = const()[name = string("transpose_107_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_107 = transpose(perm = transpose_107_perm_0, x = k_21_cast_fp16)[name = string("transpose_315")]; + tensor transpose_106 = transpose(perm = transpose_106_perm_0, x = var_1578_cast_fp16)[name = string("transpose_316")]; + tensor matrix_ac_11_cast_fp16 = matmul(transpose_x = matrix_ac_11_transpose_x_0, transpose_y = matrix_ac_11_transpose_y_0, x = transpose_106, y = transpose_107)[name = string("matrix_ac_11_cast_fp16")]; + tensor matrix_bd_23_begin_0 = const()[name = string("matrix_bd_23_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_23_end_0 = const()[name = string("matrix_bd_23_end_0"), val = tensor([1, 8, 14, 56])]; + tensor matrix_bd_23_end_mask_0 = const()[name = string("matrix_bd_23_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_23_cast_fp16 = slice_by_index(begin = matrix_bd_23_begin_0, end = matrix_bd_23_end_0, end_mask = matrix_bd_23_end_mask_0, x = matrix_bd_21_cast_fp16)[name = string("matrix_bd_23_cast_fp16")]; + tensor var_1604_cast_fp16 = add(x = matrix_ac_11_cast_fp16, y = matrix_bd_23_cast_fp16)[name = string("op_1604_cast_fp16")]; + fp16 _inversed_scores_21_y_0_to_fp16 = const()[name = string("_inversed_scores_21_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_21_cast_fp16 = mul(x = var_1604_cast_fp16, y = _inversed_scores_21_y_0_to_fp16)[name = string("_inversed_scores_21_cast_fp16")]; + tensor scores_23_cast_fp16 = select(a = var_46_to_fp16, b = _inversed_scores_21_cast_fp16, cond = mask_11)[name = string("scores_23_cast_fp16")]; + tensor var_1610_cast_fp16 = softmax(axis = var_60, x = scores_23_cast_fp16)[name = string("op_1610_cast_fp16")]; + tensor input_301_cast_fp16 = select(a = var_45_to_fp16, b = var_1610_cast_fp16, cond = mask_11)[name = string("input_301_cast_fp16")]; + bool x_143_transpose_x_0 = const()[name = string("x_143_transpose_x_0"), val = bool(false)]; + bool x_143_transpose_y_0 = const()[name = string("x_143_transpose_y_0"), val = bool(false)]; + tensor value_19_cast_fp16 = transpose(perm = value_19_perm_0, x = v_11_cast_fp16)[name = string("transpose_314")]; + tensor x_143_cast_fp16 = matmul(transpose_x = x_143_transpose_x_0, transpose_y = x_143_transpose_y_0, x = input_301_cast_fp16, y = value_19_cast_fp16)[name = string("x_143_cast_fp16")]; + tensor var_1614_perm_0 = const()[name = string("op_1614_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1615 = const()[name = string("op_1615"), val = tensor([1, -1, 1024])]; + tensor var_1614_cast_fp16 = transpose(perm = var_1614_perm_0, x = x_143_cast_fp16)[name = string("transpose_313")]; + tensor input_303_cast_fp16 = reshape(shape = var_1615, x = var_1614_cast_fp16)[name = string("input_303_cast_fp16")]; + tensor encoder_layers_5_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122177472))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122963968))))[name = string("encoder_layers_5_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_5_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_5_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122964160)))]; + tensor linear_52_cast_fp16 = linear(bias = encoder_layers_5_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_5_self_attn_linear_out_weight_to_fp16_palettized, x = input_303_cast_fp16)[name = string("linear_52_cast_fp16")]; + tensor input_307_cast_fp16 = add(x = input_297_cast_fp16, y = linear_52_cast_fp16)[name = string("input_307_cast_fp16")]; + tensor x_147_axes_0 = const()[name = string("x_147_axes_0"), val = tensor([-1])]; + tensor encoder_layers_5_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_5_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122966272)))]; + tensor encoder_layers_5_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_5_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122968384)))]; + tensor x_147_cast_fp16 = layer_norm(axes = x_147_axes_0, beta = encoder_layers_5_norm_conv_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_5_norm_conv_weight_to_fp16, x = input_307_cast_fp16)[name = string("x_147_cast_fp16")]; + tensor input_309_perm_0 = const()[name = string("input_309_perm_0"), val = tensor([0, 2, 1])]; + string input_311_pad_type_0 = const()[name = string("input_311_pad_type_0"), val = string("valid")]; + tensor input_311_strides_0 = const()[name = string("input_311_strides_0"), val = tensor([1])]; + tensor input_311_pad_0 = const()[name = string("input_311_pad_0"), val = tensor([0, 0])]; + tensor input_311_dilations_0 = const()[name = string("input_311_dilations_0"), val = tensor([1])]; + int32 input_311_groups_0 = const()[name = string("input_311_groups_0"), val = int32(1)]; + tensor encoder_layers_5_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122970496))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125067712))))[name = string("encoder_layers_5_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_309_cast_fp16 = transpose(perm = input_309_perm_0, x = x_147_cast_fp16)[name = string("transpose_312")]; + tensor input_311_cast_fp16 = conv(dilations = input_311_dilations_0, groups = input_311_groups_0, pad = input_311_pad_0, pad_type = input_311_pad_type_0, strides = input_311_strides_0, weight = encoder_layers_5_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_309_cast_fp16)[name = string("input_311_cast_fp16")]; + int32 x_149_split_num_splits_0 = const()[name = string("x_149_split_num_splits_0"), val = int32(2)]; + int32 x_149_split_axis_0 = const()[name = string("x_149_split_axis_0"), val = int32(1)]; + tensor x_149_split_cast_fp16_0, tensor x_149_split_cast_fp16_1 = split(axis = x_149_split_axis_0, num_splits = x_149_split_num_splits_0, x = input_311_cast_fp16)[name = string("x_149_split_cast_fp16")]; + tensor x_149_split_1_sigmoid_cast_fp16 = sigmoid(x = x_149_split_cast_fp16_1)[name = string("x_149_split_1_sigmoid_cast_fp16")]; + tensor x_149_cast_fp16 = mul(x = x_149_split_cast_fp16_0, y = x_149_split_1_sigmoid_cast_fp16)[name = string("x_149_cast_fp16")]; + tensor input_313_cast_fp16 = select(a = var_45_to_fp16, b = x_149_cast_fp16, cond = var_576)[name = string("input_313_cast_fp16")]; + bool new_x_23_interleave_0 = const()[name = string("new_x_23_interleave_0"), val = bool(false)]; + tensor new_x_23_cast_fp16 = concat(axis = var_60, interleave = new_x_23_interleave_0, values = (cache_23_cast_fp16, input_313_cast_fp16))[name = string("new_x_23_cast_fp16")]; + tensor var_1654_begin_0 = const()[name = string("op_1654_begin_0"), val = tensor([0, 0, 14])]; + tensor var_1654_end_0 = const()[name = string("op_1654_end_0"), val = tensor([1, 1024, 22])]; + tensor var_1654_end_mask_0 = const()[name = string("op_1654_end_mask_0"), val = tensor([true, true, true])]; + tensor var_1654_cast_fp16 = slice_by_index(begin = var_1654_begin_0, end = var_1654_end_0, end_mask = var_1654_end_mask_0, x = new_x_23_cast_fp16)[name = string("op_1654_cast_fp16")]; + string x_151_pad_type_0 = const()[name = string("x_151_pad_type_0"), val = string("valid")]; + int32 x_151_groups_0 = const()[name = string("x_151_groups_0"), val = int32(1024)]; + tensor x_151_strides_0 = const()[name = string("x_151_strides_0"), val = tensor([1])]; + tensor x_151_pad_0 = const()[name = string("x_151_pad_0"), val = tensor([0, 0])]; + tensor x_151_dilations_0 = const()[name = string("x_151_dilations_0"), val = tensor([1])]; + tensor encoder_layers_5_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125071872))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125081152))))[name = string("encoder_layers_5_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_151_cast_fp16 = conv(dilations = x_151_dilations_0, groups = x_151_groups_0, pad = x_151_pad_0, pad_type = x_151_pad_type_0, strides = x_151_strides_0, weight = encoder_layers_5_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_23_cast_fp16)[name = string("x_151_cast_fp16")]; + tensor input_315_perm_0 = const()[name = string("input_315_perm_0"), val = tensor([0, 2, 1])]; + tensor x_153_axes_0 = const()[name = string("x_153_axes_0"), val = tensor([-1])]; + tensor encoder_layers_5_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_5_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125083264)))]; + tensor encoder_layers_5_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_5_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125085376)))]; + tensor input_315_cast_fp16 = transpose(perm = input_315_perm_0, x = x_151_cast_fp16)[name = string("transpose_311")]; + tensor x_153_cast_fp16 = layer_norm(axes = x_153_axes_0, beta = encoder_layers_5_conv_batch_norm_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_5_conv_batch_norm_weight_to_fp16, x = input_315_cast_fp16)[name = string("x_153_cast_fp16")]; + tensor input_317_perm_0 = const()[name = string("input_317_perm_0"), val = tensor([0, 2, 1])]; + tensor input_317_cast_fp16 = transpose(perm = input_317_perm_0, x = x_153_cast_fp16)[name = string("transpose_310")]; + tensor input_319_cast_fp16 = silu(x = input_317_cast_fp16)[name = string("input_319_cast_fp16")]; + string x_155_pad_type_0 = const()[name = string("x_155_pad_type_0"), val = string("valid")]; + tensor x_155_strides_0 = const()[name = string("x_155_strides_0"), val = tensor([1])]; + tensor x_155_pad_0 = const()[name = string("x_155_pad_0"), val = tensor([0, 0])]; + tensor x_155_dilations_0 = const()[name = string("x_155_dilations_0"), val = tensor([1])]; + int32 x_155_groups_0 = const()[name = string("x_155_groups_0"), val = int32(1)]; + tensor encoder_layers_5_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125087488))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126136128))))[name = string("encoder_layers_5_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_155_cast_fp16 = conv(dilations = x_155_dilations_0, groups = x_155_groups_0, pad = x_155_pad_0, pad_type = x_155_pad_type_0, strides = x_155_strides_0, weight = encoder_layers_5_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_319_cast_fp16)[name = string("x_155_cast_fp16")]; + tensor input_321_perm_0 = const()[name = string("input_321_perm_0"), val = tensor([0, 2, 1])]; + tensor input_321_cast_fp16 = transpose(perm = input_321_perm_0, x = x_155_cast_fp16)[name = string("transpose_309")]; + tensor input_323_cast_fp16 = add(x = input_307_cast_fp16, y = input_321_cast_fp16)[name = string("input_323_cast_fp16")]; + tensor input_325_axes_0 = const()[name = string("input_325_axes_0"), val = tensor([-1])]; + tensor encoder_layers_5_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_5_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126138240)))]; + tensor encoder_layers_5_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_5_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126140352)))]; + tensor input_325_cast_fp16 = layer_norm(axes = input_325_axes_0, beta = encoder_layers_5_norm_feed_forward2_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_5_norm_feed_forward2_weight_to_fp16, x = input_323_cast_fp16)[name = string("input_325_cast_fp16")]; + tensor encoder_layers_5_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126142464))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(129288256))))[name = string("encoder_layers_5_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_5_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_5_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(129288448)))]; + tensor linear_53_cast_fp16 = linear(bias = encoder_layers_5_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_5_feed_forward2_linear1_weight_to_fp16_palettized, x = input_325_cast_fp16)[name = string("linear_53_cast_fp16")]; + tensor input_329_cast_fp16 = silu(x = linear_53_cast_fp16)[name = string("input_329_cast_fp16")]; + tensor encoder_layers_5_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(129296704))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132442496))))[name = string("encoder_layers_5_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_5_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_5_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132442688)))]; + tensor linear_54_cast_fp16 = linear(bias = encoder_layers_5_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_5_feed_forward2_linear2_weight_to_fp16_palettized, x = input_329_cast_fp16)[name = string("linear_54_cast_fp16")]; + fp16 var_1697_to_fp16 = const()[name = string("op_1697_to_fp16"), val = fp16(0x1p-1)]; + tensor var_1698_cast_fp16 = mul(x = linear_54_cast_fp16, y = var_1697_to_fp16)[name = string("op_1698_cast_fp16")]; + tensor input_335_cast_fp16 = add(x = input_323_cast_fp16, y = var_1698_cast_fp16)[name = string("input_335_cast_fp16")]; + tensor input_337_axes_0 = const()[name = string("input_337_axes_0"), val = tensor([-1])]; + tensor encoder_layers_5_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_5_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132444800)))]; + tensor encoder_layers_5_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_5_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132446912)))]; + tensor input_337_cast_fp16 = layer_norm(axes = input_337_axes_0, beta = encoder_layers_5_norm_out_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_5_norm_out_weight_to_fp16, x = input_335_cast_fp16)[name = string("input_337_cast_fp16")]; + tensor cache_25_begin_0 = const()[name = string("cache_25_begin_0"), val = tensor([6, 0, 0, 0])]; + tensor cache_25_end_0 = const()[name = string("cache_25_end_0"), val = tensor([7, 1, 42, 1024])]; + tensor cache_25_end_mask_0 = const()[name = string("cache_25_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_25_squeeze_mask_0 = const()[name = string("cache_25_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_25_cast_fp16 = slice_by_index(begin = cache_25_begin_0, end = cache_25_end_0, end_mask = cache_25_end_mask_0, squeeze_mask = cache_25_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_25_cast_fp16")]; + tensor cache_27_begin_0 = const()[name = string("cache_27_begin_0"), val = tensor([6, 0, 0, 0])]; + tensor cache_27_end_0 = const()[name = string("cache_27_end_0"), val = tensor([7, 1, 1024, 8])]; + tensor cache_27_end_mask_0 = const()[name = string("cache_27_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_27_squeeze_mask_0 = const()[name = string("cache_27_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_27_cast_fp16 = slice_by_index(begin = cache_27_begin_0, end = cache_27_end_0, end_mask = cache_27_end_mask_0, squeeze_mask = cache_27_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_27_cast_fp16")]; + tensor input_339_axes_0 = const()[name = string("input_339_axes_0"), val = tensor([-1])]; + tensor encoder_layers_6_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_6_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132449024)))]; + tensor encoder_layers_6_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_6_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132451136)))]; + tensor input_339_cast_fp16 = layer_norm(axes = input_339_axes_0, beta = encoder_layers_6_norm_feed_forward1_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_6_norm_feed_forward1_weight_to_fp16, x = input_337_cast_fp16)[name = string("input_339_cast_fp16")]; + tensor encoder_layers_6_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132453248))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135599040))))[name = string("encoder_layers_6_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_6_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_6_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135599232)))]; + tensor linear_55_cast_fp16 = linear(bias = encoder_layers_6_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_6_feed_forward1_linear1_weight_to_fp16_palettized, x = input_339_cast_fp16)[name = string("linear_55_cast_fp16")]; + tensor input_343_cast_fp16 = silu(x = linear_55_cast_fp16)[name = string("input_343_cast_fp16")]; + tensor encoder_layers_6_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135607488))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138753280))))[name = string("encoder_layers_6_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_6_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_6_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138753472)))]; + tensor linear_56_cast_fp16 = linear(bias = encoder_layers_6_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_6_feed_forward1_linear2_weight_to_fp16_palettized, x = input_343_cast_fp16)[name = string("linear_56_cast_fp16")]; + fp16 var_1734_to_fp16 = const()[name = string("op_1734_to_fp16"), val = fp16(0x1p-1)]; + tensor var_1735_cast_fp16 = mul(x = linear_56_cast_fp16, y = var_1734_to_fp16)[name = string("op_1735_cast_fp16")]; + tensor input_349_cast_fp16 = add(x = input_337_cast_fp16, y = var_1735_cast_fp16)[name = string("input_349_cast_fp16")]; + tensor key_13_axes_0 = const()[name = string("key_13_axes_0"), val = tensor([-1])]; + tensor encoder_layers_6_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_6_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138755584)))]; + tensor encoder_layers_6_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_6_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138757696)))]; + tensor key_13_cast_fp16 = layer_norm(axes = key_13_axes_0, beta = encoder_layers_6_norm_self_att_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_6_norm_self_att_weight_to_fp16, x = input_349_cast_fp16)[name = string("key_13_cast_fp16")]; + bool input_351_interleave_0 = const()[name = string("input_351_interleave_0"), val = bool(false)]; + tensor input_351_cast_fp16 = concat(axis = var_69, interleave = input_351_interleave_0, values = (cache_25_cast_fp16, key_13_cast_fp16))[name = string("input_351_cast_fp16")]; + tensor var_1757_begin_0 = const()[name = string("op_1757_begin_0"), val = tensor([0, 14, 0])]; + tensor var_1757_end_0 = const()[name = string("op_1757_end_0"), val = tensor([1, 42, 1024])]; + tensor var_1757_end_mask_0 = const()[name = string("op_1757_end_mask_0"), val = tensor([true, true, true])]; + tensor var_1757_cast_fp16 = slice_by_index(begin = var_1757_begin_0, end = var_1757_end_0, end_mask = var_1757_end_mask_0, x = cache_25_cast_fp16)[name = string("op_1757_cast_fp16")]; + bool var_1763_interleave_0 = const()[name = string("op_1763_interleave_0"), val = bool(false)]; + tensor var_1763_cast_fp16 = concat(axis = var_69, interleave = var_1763_interleave_0, values = (var_1757_cast_fp16, key_13_cast_fp16))[name = string("op_1763_cast_fp16")]; + tensor encoder_layers_6_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138759808))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(139546304))))[name = string("encoder_layers_6_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_6_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_6_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(139546496)))]; + tensor linear_57_cast_fp16 = linear(bias = encoder_layers_6_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_6_self_attn_linear_q_weight_to_fp16_palettized, x = key_13_cast_fp16)[name = string("linear_57_cast_fp16")]; + tensor var_1768 = const()[name = string("op_1768"), val = tensor([1, -1, 8, 128])]; + tensor q_37_cast_fp16 = reshape(shape = var_1768, x = linear_57_cast_fp16)[name = string("q_37_cast_fp16")]; + tensor encoder_layers_6_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(139548608))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140335104))))[name = string("encoder_layers_6_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_6_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_6_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140335296)))]; + tensor linear_58_cast_fp16 = linear(bias = encoder_layers_6_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_6_self_attn_linear_k_weight_to_fp16_palettized, x = input_351_cast_fp16)[name = string("linear_58_cast_fp16")]; + tensor var_1773 = const()[name = string("op_1773"), val = tensor([1, -1, 8, 128])]; + tensor k_25_cast_fp16 = reshape(shape = var_1773, x = linear_58_cast_fp16)[name = string("k_25_cast_fp16")]; + tensor encoder_layers_6_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140337408))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141123904))))[name = string("encoder_layers_6_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_6_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_6_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141124096)))]; + tensor linear_59_cast_fp16 = linear(bias = encoder_layers_6_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_6_self_attn_linear_v_weight_to_fp16_palettized, x = input_351_cast_fp16)[name = string("linear_59_cast_fp16")]; + tensor var_1778 = const()[name = string("op_1778"), val = tensor([1, -1, 8, 128])]; + tensor v_13_cast_fp16 = reshape(shape = var_1778, x = linear_59_cast_fp16)[name = string("v_13_cast_fp16")]; + tensor value_21_perm_0 = const()[name = string("value_21_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_6_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_6_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141126208)))]; + tensor var_1791_cast_fp16 = add(x = q_37_cast_fp16, y = encoder_layers_6_self_attn_pos_bias_u_to_fp16)[name = string("op_1791_cast_fp16")]; + tensor encoder_layers_6_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_6_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141128320)))]; + tensor var_1793_cast_fp16 = add(x = q_37_cast_fp16, y = encoder_layers_6_self_attn_pos_bias_v_to_fp16)[name = string("op_1793_cast_fp16")]; + tensor q_with_bias_v_13_perm_0 = const()[name = string("q_with_bias_v_13_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_163_transpose_x_0 = const()[name = string("x_163_transpose_x_0"), val = bool(false)]; + bool x_163_transpose_y_0 = const()[name = string("x_163_transpose_y_0"), val = bool(false)]; + tensor op_1795_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141130432))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141244160))))[name = string("op_1795_to_fp16_quantized")]; + tensor q_with_bias_v_13_cast_fp16 = transpose(perm = q_with_bias_v_13_perm_0, x = var_1793_cast_fp16)[name = string("transpose_308")]; + tensor x_163_cast_fp16 = matmul(transpose_x = x_163_transpose_x_0, transpose_y = x_163_transpose_y_0, x = q_with_bias_v_13_cast_fp16, y = op_1795_to_fp16_quantized)[name = string("x_163_cast_fp16")]; + tensor x_165_pad_0 = const()[name = string("x_165_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_165_mode_0 = const()[name = string("x_165_mode_0"), val = string("constant")]; + fp16 const_157_to_fp16 = const()[name = string("const_157_to_fp16"), val = fp16(0x0p+0)]; + tensor x_165_cast_fp16 = pad(constant_val = const_157_to_fp16, mode = x_165_mode_0, pad = x_165_pad_0, x = x_163_cast_fp16)[name = string("x_165_cast_fp16")]; + tensor var_1803 = const()[name = string("op_1803"), val = tensor([1, 8, -1, 14])]; + tensor x_167_cast_fp16 = reshape(shape = var_1803, x = x_165_cast_fp16)[name = string("x_167_cast_fp16")]; + tensor var_1807_begin_0 = const()[name = string("op_1807_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_1807_end_0 = const()[name = string("op_1807_end_0"), val = tensor([1, 8, 112, 14])]; + tensor var_1807_end_mask_0 = const()[name = string("op_1807_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_1807_cast_fp16 = slice_by_index(begin = var_1807_begin_0, end = var_1807_end_0, end_mask = var_1807_end_mask_0, x = x_167_cast_fp16)[name = string("op_1807_cast_fp16")]; + tensor var_1808 = const()[name = string("op_1808"), val = tensor([1, 8, 14, 111])]; + tensor matrix_bd_25_cast_fp16 = reshape(shape = var_1808, x = var_1807_cast_fp16)[name = string("matrix_bd_25_cast_fp16")]; + bool matrix_ac_13_transpose_x_0 = const()[name = string("matrix_ac_13_transpose_x_0"), val = bool(false)]; + bool matrix_ac_13_transpose_y_0 = const()[name = string("matrix_ac_13_transpose_y_0"), val = bool(false)]; + tensor transpose_108_perm_0 = const()[name = string("transpose_108_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_109_perm_0 = const()[name = string("transpose_109_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_109 = transpose(perm = transpose_109_perm_0, x = k_25_cast_fp16)[name = string("transpose_306")]; + tensor transpose_108 = transpose(perm = transpose_108_perm_0, x = var_1791_cast_fp16)[name = string("transpose_307")]; + tensor matrix_ac_13_cast_fp16 = matmul(transpose_x = matrix_ac_13_transpose_x_0, transpose_y = matrix_ac_13_transpose_y_0, x = transpose_108, y = transpose_109)[name = string("matrix_ac_13_cast_fp16")]; + tensor matrix_bd_27_begin_0 = const()[name = string("matrix_bd_27_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_27_end_0 = const()[name = string("matrix_bd_27_end_0"), val = tensor([1, 8, 14, 56])]; + tensor matrix_bd_27_end_mask_0 = const()[name = string("matrix_bd_27_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_27_cast_fp16 = slice_by_index(begin = matrix_bd_27_begin_0, end = matrix_bd_27_end_0, end_mask = matrix_bd_27_end_mask_0, x = matrix_bd_25_cast_fp16)[name = string("matrix_bd_27_cast_fp16")]; + tensor var_1817_cast_fp16 = add(x = matrix_ac_13_cast_fp16, y = matrix_bd_27_cast_fp16)[name = string("op_1817_cast_fp16")]; + fp16 _inversed_scores_25_y_0_to_fp16 = const()[name = string("_inversed_scores_25_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_25_cast_fp16 = mul(x = var_1817_cast_fp16, y = _inversed_scores_25_y_0_to_fp16)[name = string("_inversed_scores_25_cast_fp16")]; + tensor scores_27_cast_fp16 = select(a = var_46_to_fp16, b = _inversed_scores_25_cast_fp16, cond = mask_11)[name = string("scores_27_cast_fp16")]; + tensor var_1823_cast_fp16 = softmax(axis = var_60, x = scores_27_cast_fp16)[name = string("op_1823_cast_fp16")]; + tensor input_353_cast_fp16 = select(a = var_45_to_fp16, b = var_1823_cast_fp16, cond = mask_11)[name = string("input_353_cast_fp16")]; + bool x_169_transpose_x_0 = const()[name = string("x_169_transpose_x_0"), val = bool(false)]; + bool x_169_transpose_y_0 = const()[name = string("x_169_transpose_y_0"), val = bool(false)]; + tensor value_21_cast_fp16 = transpose(perm = value_21_perm_0, x = v_13_cast_fp16)[name = string("transpose_305")]; + tensor x_169_cast_fp16 = matmul(transpose_x = x_169_transpose_x_0, transpose_y = x_169_transpose_y_0, x = input_353_cast_fp16, y = value_21_cast_fp16)[name = string("x_169_cast_fp16")]; + tensor var_1827_perm_0 = const()[name = string("op_1827_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1828 = const()[name = string("op_1828"), val = tensor([1, -1, 1024])]; + tensor var_1827_cast_fp16 = transpose(perm = var_1827_perm_0, x = x_169_cast_fp16)[name = string("transpose_304")]; + tensor input_355_cast_fp16 = reshape(shape = var_1828, x = var_1827_cast_fp16)[name = string("input_355_cast_fp16")]; + tensor encoder_layers_6_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141244480))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142030976))))[name = string("encoder_layers_6_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_6_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_6_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142031168)))]; + tensor linear_61_cast_fp16 = linear(bias = encoder_layers_6_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_6_self_attn_linear_out_weight_to_fp16_palettized, x = input_355_cast_fp16)[name = string("linear_61_cast_fp16")]; + tensor input_359_cast_fp16 = add(x = input_349_cast_fp16, y = linear_61_cast_fp16)[name = string("input_359_cast_fp16")]; + tensor x_173_axes_0 = const()[name = string("x_173_axes_0"), val = tensor([-1])]; + tensor encoder_layers_6_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_6_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142033280)))]; + tensor encoder_layers_6_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_6_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142035392)))]; + tensor x_173_cast_fp16 = layer_norm(axes = x_173_axes_0, beta = encoder_layers_6_norm_conv_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_6_norm_conv_weight_to_fp16, x = input_359_cast_fp16)[name = string("x_173_cast_fp16")]; + tensor input_361_perm_0 = const()[name = string("input_361_perm_0"), val = tensor([0, 2, 1])]; + string input_363_pad_type_0 = const()[name = string("input_363_pad_type_0"), val = string("valid")]; + tensor input_363_strides_0 = const()[name = string("input_363_strides_0"), val = tensor([1])]; + tensor input_363_pad_0 = const()[name = string("input_363_pad_0"), val = tensor([0, 0])]; + tensor input_363_dilations_0 = const()[name = string("input_363_dilations_0"), val = tensor([1])]; + int32 input_363_groups_0 = const()[name = string("input_363_groups_0"), val = int32(1)]; + tensor encoder_layers_6_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142037504))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144134720))))[name = string("encoder_layers_6_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_361_cast_fp16 = transpose(perm = input_361_perm_0, x = x_173_cast_fp16)[name = string("transpose_303")]; + tensor input_363_cast_fp16 = conv(dilations = input_363_dilations_0, groups = input_363_groups_0, pad = input_363_pad_0, pad_type = input_363_pad_type_0, strides = input_363_strides_0, weight = encoder_layers_6_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_361_cast_fp16)[name = string("input_363_cast_fp16")]; + int32 x_175_split_num_splits_0 = const()[name = string("x_175_split_num_splits_0"), val = int32(2)]; + int32 x_175_split_axis_0 = const()[name = string("x_175_split_axis_0"), val = int32(1)]; + tensor x_175_split_cast_fp16_0, tensor x_175_split_cast_fp16_1 = split(axis = x_175_split_axis_0, num_splits = x_175_split_num_splits_0, x = input_363_cast_fp16)[name = string("x_175_split_cast_fp16")]; + tensor x_175_split_1_sigmoid_cast_fp16 = sigmoid(x = x_175_split_cast_fp16_1)[name = string("x_175_split_1_sigmoid_cast_fp16")]; + tensor x_175_cast_fp16 = mul(x = x_175_split_cast_fp16_0, y = x_175_split_1_sigmoid_cast_fp16)[name = string("x_175_cast_fp16")]; + tensor input_365_cast_fp16 = select(a = var_45_to_fp16, b = x_175_cast_fp16, cond = var_576)[name = string("input_365_cast_fp16")]; + bool new_x_27_interleave_0 = const()[name = string("new_x_27_interleave_0"), val = bool(false)]; + tensor new_x_27_cast_fp16 = concat(axis = var_60, interleave = new_x_27_interleave_0, values = (cache_27_cast_fp16, input_365_cast_fp16))[name = string("new_x_27_cast_fp16")]; + tensor var_1867_begin_0 = const()[name = string("op_1867_begin_0"), val = tensor([0, 0, 14])]; + tensor var_1867_end_0 = const()[name = string("op_1867_end_0"), val = tensor([1, 1024, 22])]; + tensor var_1867_end_mask_0 = const()[name = string("op_1867_end_mask_0"), val = tensor([true, true, true])]; + tensor var_1867_cast_fp16 = slice_by_index(begin = var_1867_begin_0, end = var_1867_end_0, end_mask = var_1867_end_mask_0, x = new_x_27_cast_fp16)[name = string("op_1867_cast_fp16")]; + string x_177_pad_type_0 = const()[name = string("x_177_pad_type_0"), val = string("valid")]; + int32 x_177_groups_0 = const()[name = string("x_177_groups_0"), val = int32(1024)]; + tensor x_177_strides_0 = const()[name = string("x_177_strides_0"), val = tensor([1])]; + tensor x_177_pad_0 = const()[name = string("x_177_pad_0"), val = tensor([0, 0])]; + tensor x_177_dilations_0 = const()[name = string("x_177_dilations_0"), val = tensor([1])]; + tensor encoder_layers_6_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144138880))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144148160))))[name = string("encoder_layers_6_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_177_cast_fp16 = conv(dilations = x_177_dilations_0, groups = x_177_groups_0, pad = x_177_pad_0, pad_type = x_177_pad_type_0, strides = x_177_strides_0, weight = encoder_layers_6_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_27_cast_fp16)[name = string("x_177_cast_fp16")]; + tensor input_367_perm_0 = const()[name = string("input_367_perm_0"), val = tensor([0, 2, 1])]; + tensor x_179_axes_0 = const()[name = string("x_179_axes_0"), val = tensor([-1])]; + tensor encoder_layers_6_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_6_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144150272)))]; + tensor encoder_layers_6_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_6_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144152384)))]; + tensor input_367_cast_fp16 = transpose(perm = input_367_perm_0, x = x_177_cast_fp16)[name = string("transpose_302")]; + tensor x_179_cast_fp16 = layer_norm(axes = x_179_axes_0, beta = encoder_layers_6_conv_batch_norm_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_6_conv_batch_norm_weight_to_fp16, x = input_367_cast_fp16)[name = string("x_179_cast_fp16")]; + tensor input_369_perm_0 = const()[name = string("input_369_perm_0"), val = tensor([0, 2, 1])]; + tensor input_369_cast_fp16 = transpose(perm = input_369_perm_0, x = x_179_cast_fp16)[name = string("transpose_301")]; + tensor input_371_cast_fp16 = silu(x = input_369_cast_fp16)[name = string("input_371_cast_fp16")]; + string x_181_pad_type_0 = const()[name = string("x_181_pad_type_0"), val = string("valid")]; + tensor x_181_strides_0 = const()[name = string("x_181_strides_0"), val = tensor([1])]; + tensor x_181_pad_0 = const()[name = string("x_181_pad_0"), val = tensor([0, 0])]; + tensor x_181_dilations_0 = const()[name = string("x_181_dilations_0"), val = tensor([1])]; + int32 x_181_groups_0 = const()[name = string("x_181_groups_0"), val = int32(1)]; + tensor encoder_layers_6_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144154496))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(145203136))))[name = string("encoder_layers_6_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_181_cast_fp16 = conv(dilations = x_181_dilations_0, groups = x_181_groups_0, pad = x_181_pad_0, pad_type = x_181_pad_type_0, strides = x_181_strides_0, weight = encoder_layers_6_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_371_cast_fp16)[name = string("x_181_cast_fp16")]; + tensor input_373_perm_0 = const()[name = string("input_373_perm_0"), val = tensor([0, 2, 1])]; + tensor input_373_cast_fp16 = transpose(perm = input_373_perm_0, x = x_181_cast_fp16)[name = string("transpose_300")]; + tensor input_375_cast_fp16 = add(x = input_359_cast_fp16, y = input_373_cast_fp16)[name = string("input_375_cast_fp16")]; + tensor input_377_axes_0 = const()[name = string("input_377_axes_0"), val = tensor([-1])]; + tensor encoder_layers_6_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_6_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(145205248)))]; + tensor encoder_layers_6_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_6_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(145207360)))]; + tensor input_377_cast_fp16 = layer_norm(axes = input_377_axes_0, beta = encoder_layers_6_norm_feed_forward2_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_6_norm_feed_forward2_weight_to_fp16, x = input_375_cast_fp16)[name = string("input_377_cast_fp16")]; + tensor encoder_layers_6_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(145209472))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(148355264))))[name = string("encoder_layers_6_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_6_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_6_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(148355456)))]; + tensor linear_62_cast_fp16 = linear(bias = encoder_layers_6_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_6_feed_forward2_linear1_weight_to_fp16_palettized, x = input_377_cast_fp16)[name = string("linear_62_cast_fp16")]; + tensor input_381_cast_fp16 = silu(x = linear_62_cast_fp16)[name = string("input_381_cast_fp16")]; + tensor encoder_layers_6_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(148363712))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151509504))))[name = string("encoder_layers_6_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_6_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_6_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151509696)))]; + tensor linear_63_cast_fp16 = linear(bias = encoder_layers_6_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_6_feed_forward2_linear2_weight_to_fp16_palettized, x = input_381_cast_fp16)[name = string("linear_63_cast_fp16")]; + fp16 var_1910_to_fp16 = const()[name = string("op_1910_to_fp16"), val = fp16(0x1p-1)]; + tensor var_1911_cast_fp16 = mul(x = linear_63_cast_fp16, y = var_1910_to_fp16)[name = string("op_1911_cast_fp16")]; + tensor input_387_cast_fp16 = add(x = input_375_cast_fp16, y = var_1911_cast_fp16)[name = string("input_387_cast_fp16")]; + tensor input_389_axes_0 = const()[name = string("input_389_axes_0"), val = tensor([-1])]; + tensor encoder_layers_6_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_6_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151511808)))]; + tensor encoder_layers_6_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_6_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151513920)))]; + tensor input_389_cast_fp16 = layer_norm(axes = input_389_axes_0, beta = encoder_layers_6_norm_out_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_6_norm_out_weight_to_fp16, x = input_387_cast_fp16)[name = string("input_389_cast_fp16")]; + tensor cache_29_begin_0 = const()[name = string("cache_29_begin_0"), val = tensor([7, 0, 0, 0])]; + tensor cache_29_end_0 = const()[name = string("cache_29_end_0"), val = tensor([8, 1, 42, 1024])]; + tensor cache_29_end_mask_0 = const()[name = string("cache_29_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_29_squeeze_mask_0 = const()[name = string("cache_29_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_29_cast_fp16 = slice_by_index(begin = cache_29_begin_0, end = cache_29_end_0, end_mask = cache_29_end_mask_0, squeeze_mask = cache_29_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_29_cast_fp16")]; + tensor cache_31_begin_0 = const()[name = string("cache_31_begin_0"), val = tensor([7, 0, 0, 0])]; + tensor cache_31_end_0 = const()[name = string("cache_31_end_0"), val = tensor([8, 1, 1024, 8])]; + tensor cache_31_end_mask_0 = const()[name = string("cache_31_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_31_squeeze_mask_0 = const()[name = string("cache_31_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_31_cast_fp16 = slice_by_index(begin = cache_31_begin_0, end = cache_31_end_0, end_mask = cache_31_end_mask_0, squeeze_mask = cache_31_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_31_cast_fp16")]; + tensor input_391_axes_0 = const()[name = string("input_391_axes_0"), val = tensor([-1])]; + tensor encoder_layers_7_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_7_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151516032)))]; + tensor encoder_layers_7_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_7_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151518144)))]; + tensor input_391_cast_fp16 = layer_norm(axes = input_391_axes_0, beta = encoder_layers_7_norm_feed_forward1_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_7_norm_feed_forward1_weight_to_fp16, x = input_389_cast_fp16)[name = string("input_391_cast_fp16")]; + tensor encoder_layers_7_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151520256))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(154666048))))[name = string("encoder_layers_7_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_7_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_7_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(154666240)))]; + tensor linear_64_cast_fp16 = linear(bias = encoder_layers_7_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_7_feed_forward1_linear1_weight_to_fp16_palettized, x = input_391_cast_fp16)[name = string("linear_64_cast_fp16")]; + tensor input_395_cast_fp16 = silu(x = linear_64_cast_fp16)[name = string("input_395_cast_fp16")]; + tensor encoder_layers_7_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(154674496))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(157820288))))[name = string("encoder_layers_7_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_7_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_7_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(157820480)))]; + tensor linear_65_cast_fp16 = linear(bias = encoder_layers_7_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_7_feed_forward1_linear2_weight_to_fp16_palettized, x = input_395_cast_fp16)[name = string("linear_65_cast_fp16")]; + fp16 var_1947_to_fp16 = const()[name = string("op_1947_to_fp16"), val = fp16(0x1p-1)]; + tensor var_1948_cast_fp16 = mul(x = linear_65_cast_fp16, y = var_1947_to_fp16)[name = string("op_1948_cast_fp16")]; + tensor input_401_cast_fp16 = add(x = input_389_cast_fp16, y = var_1948_cast_fp16)[name = string("input_401_cast_fp16")]; + tensor key_15_axes_0 = const()[name = string("key_15_axes_0"), val = tensor([-1])]; + tensor encoder_layers_7_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_7_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(157822592)))]; + tensor encoder_layers_7_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_7_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(157824704)))]; + tensor key_15_cast_fp16 = layer_norm(axes = key_15_axes_0, beta = encoder_layers_7_norm_self_att_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_7_norm_self_att_weight_to_fp16, x = input_401_cast_fp16)[name = string("key_15_cast_fp16")]; + bool input_403_interleave_0 = const()[name = string("input_403_interleave_0"), val = bool(false)]; + tensor input_403_cast_fp16 = concat(axis = var_69, interleave = input_403_interleave_0, values = (cache_29_cast_fp16, key_15_cast_fp16))[name = string("input_403_cast_fp16")]; + tensor var_1970_begin_0 = const()[name = string("op_1970_begin_0"), val = tensor([0, 14, 0])]; + tensor var_1970_end_0 = const()[name = string("op_1970_end_0"), val = tensor([1, 42, 1024])]; + tensor var_1970_end_mask_0 = const()[name = string("op_1970_end_mask_0"), val = tensor([true, true, true])]; + tensor var_1970_cast_fp16 = slice_by_index(begin = var_1970_begin_0, end = var_1970_end_0, end_mask = var_1970_end_mask_0, x = cache_29_cast_fp16)[name = string("op_1970_cast_fp16")]; + bool var_1976_interleave_0 = const()[name = string("op_1976_interleave_0"), val = bool(false)]; + tensor var_1976_cast_fp16 = concat(axis = var_69, interleave = var_1976_interleave_0, values = (var_1970_cast_fp16, key_15_cast_fp16))[name = string("op_1976_cast_fp16")]; + tensor encoder_layers_7_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(157826816))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158613312))))[name = string("encoder_layers_7_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_7_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_7_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158613504)))]; + tensor linear_66_cast_fp16 = linear(bias = encoder_layers_7_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_7_self_attn_linear_q_weight_to_fp16_palettized, x = key_15_cast_fp16)[name = string("linear_66_cast_fp16")]; + tensor var_1981 = const()[name = string("op_1981"), val = tensor([1, -1, 8, 128])]; + tensor q_43_cast_fp16 = reshape(shape = var_1981, x = linear_66_cast_fp16)[name = string("q_43_cast_fp16")]; + tensor encoder_layers_7_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158615616))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(159402112))))[name = string("encoder_layers_7_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_7_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_7_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(159402304)))]; + tensor linear_67_cast_fp16 = linear(bias = encoder_layers_7_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_7_self_attn_linear_k_weight_to_fp16_palettized, x = input_403_cast_fp16)[name = string("linear_67_cast_fp16")]; + tensor var_1986 = const()[name = string("op_1986"), val = tensor([1, -1, 8, 128])]; + tensor k_29_cast_fp16 = reshape(shape = var_1986, x = linear_67_cast_fp16)[name = string("k_29_cast_fp16")]; + tensor encoder_layers_7_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(159404416))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160190912))))[name = string("encoder_layers_7_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_7_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_7_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160191104)))]; + tensor linear_68_cast_fp16 = linear(bias = encoder_layers_7_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_7_self_attn_linear_v_weight_to_fp16_palettized, x = input_403_cast_fp16)[name = string("linear_68_cast_fp16")]; + tensor var_1991 = const()[name = string("op_1991"), val = tensor([1, -1, 8, 128])]; + tensor v_15_cast_fp16 = reshape(shape = var_1991, x = linear_68_cast_fp16)[name = string("v_15_cast_fp16")]; + tensor value_23_perm_0 = const()[name = string("value_23_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_7_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_7_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160193216)))]; + tensor var_2004_cast_fp16 = add(x = q_43_cast_fp16, y = encoder_layers_7_self_attn_pos_bias_u_to_fp16)[name = string("op_2004_cast_fp16")]; + tensor encoder_layers_7_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_7_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160195328)))]; + tensor var_2006_cast_fp16 = add(x = q_43_cast_fp16, y = encoder_layers_7_self_attn_pos_bias_v_to_fp16)[name = string("op_2006_cast_fp16")]; + tensor q_with_bias_v_15_perm_0 = const()[name = string("q_with_bias_v_15_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_189_transpose_x_0 = const()[name = string("x_189_transpose_x_0"), val = bool(false)]; + bool x_189_transpose_y_0 = const()[name = string("x_189_transpose_y_0"), val = bool(false)]; + tensor op_2008_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160197440))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160311168))))[name = string("op_2008_to_fp16_quantized")]; + tensor q_with_bias_v_15_cast_fp16 = transpose(perm = q_with_bias_v_15_perm_0, x = var_2006_cast_fp16)[name = string("transpose_299")]; + tensor x_189_cast_fp16 = matmul(transpose_x = x_189_transpose_x_0, transpose_y = x_189_transpose_y_0, x = q_with_bias_v_15_cast_fp16, y = op_2008_to_fp16_quantized)[name = string("x_189_cast_fp16")]; + tensor x_191_pad_0 = const()[name = string("x_191_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_191_mode_0 = const()[name = string("x_191_mode_0"), val = string("constant")]; + fp16 const_170_to_fp16 = const()[name = string("const_170_to_fp16"), val = fp16(0x0p+0)]; + tensor x_191_cast_fp16 = pad(constant_val = const_170_to_fp16, mode = x_191_mode_0, pad = x_191_pad_0, x = x_189_cast_fp16)[name = string("x_191_cast_fp16")]; + tensor var_2016 = const()[name = string("op_2016"), val = tensor([1, 8, -1, 14])]; + tensor x_193_cast_fp16 = reshape(shape = var_2016, x = x_191_cast_fp16)[name = string("x_193_cast_fp16")]; + tensor var_2020_begin_0 = const()[name = string("op_2020_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_2020_end_0 = const()[name = string("op_2020_end_0"), val = tensor([1, 8, 112, 14])]; + tensor var_2020_end_mask_0 = const()[name = string("op_2020_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_2020_cast_fp16 = slice_by_index(begin = var_2020_begin_0, end = var_2020_end_0, end_mask = var_2020_end_mask_0, x = x_193_cast_fp16)[name = string("op_2020_cast_fp16")]; + tensor var_2021 = const()[name = string("op_2021"), val = tensor([1, 8, 14, 111])]; + tensor matrix_bd_29_cast_fp16 = reshape(shape = var_2021, x = var_2020_cast_fp16)[name = string("matrix_bd_29_cast_fp16")]; + bool matrix_ac_15_transpose_x_0 = const()[name = string("matrix_ac_15_transpose_x_0"), val = bool(false)]; + bool matrix_ac_15_transpose_y_0 = const()[name = string("matrix_ac_15_transpose_y_0"), val = bool(false)]; + tensor transpose_110_perm_0 = const()[name = string("transpose_110_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_111_perm_0 = const()[name = string("transpose_111_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_111 = transpose(perm = transpose_111_perm_0, x = k_29_cast_fp16)[name = string("transpose_297")]; + tensor transpose_110 = transpose(perm = transpose_110_perm_0, x = var_2004_cast_fp16)[name = string("transpose_298")]; + tensor matrix_ac_15_cast_fp16 = matmul(transpose_x = matrix_ac_15_transpose_x_0, transpose_y = matrix_ac_15_transpose_y_0, x = transpose_110, y = transpose_111)[name = string("matrix_ac_15_cast_fp16")]; + tensor matrix_bd_31_begin_0 = const()[name = string("matrix_bd_31_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_31_end_0 = const()[name = string("matrix_bd_31_end_0"), val = tensor([1, 8, 14, 56])]; + tensor matrix_bd_31_end_mask_0 = const()[name = string("matrix_bd_31_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_31_cast_fp16 = slice_by_index(begin = matrix_bd_31_begin_0, end = matrix_bd_31_end_0, end_mask = matrix_bd_31_end_mask_0, x = matrix_bd_29_cast_fp16)[name = string("matrix_bd_31_cast_fp16")]; + tensor var_2030_cast_fp16 = add(x = matrix_ac_15_cast_fp16, y = matrix_bd_31_cast_fp16)[name = string("op_2030_cast_fp16")]; + fp16 _inversed_scores_29_y_0_to_fp16 = const()[name = string("_inversed_scores_29_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_29_cast_fp16 = mul(x = var_2030_cast_fp16, y = _inversed_scores_29_y_0_to_fp16)[name = string("_inversed_scores_29_cast_fp16")]; + tensor scores_31_cast_fp16 = select(a = var_46_to_fp16, b = _inversed_scores_29_cast_fp16, cond = mask_11)[name = string("scores_31_cast_fp16")]; + tensor var_2036_cast_fp16 = softmax(axis = var_60, x = scores_31_cast_fp16)[name = string("op_2036_cast_fp16")]; + tensor input_405_cast_fp16 = select(a = var_45_to_fp16, b = var_2036_cast_fp16, cond = mask_11)[name = string("input_405_cast_fp16")]; + bool x_195_transpose_x_0 = const()[name = string("x_195_transpose_x_0"), val = bool(false)]; + bool x_195_transpose_y_0 = const()[name = string("x_195_transpose_y_0"), val = bool(false)]; + tensor value_23_cast_fp16 = transpose(perm = value_23_perm_0, x = v_15_cast_fp16)[name = string("transpose_296")]; + tensor x_195_cast_fp16 = matmul(transpose_x = x_195_transpose_x_0, transpose_y = x_195_transpose_y_0, x = input_405_cast_fp16, y = value_23_cast_fp16)[name = string("x_195_cast_fp16")]; + tensor var_2040_perm_0 = const()[name = string("op_2040_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_2041 = const()[name = string("op_2041"), val = tensor([1, -1, 1024])]; + tensor var_2040_cast_fp16 = transpose(perm = var_2040_perm_0, x = x_195_cast_fp16)[name = string("transpose_295")]; + tensor input_407_cast_fp16 = reshape(shape = var_2041, x = var_2040_cast_fp16)[name = string("input_407_cast_fp16")]; + tensor encoder_layers_7_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160311488))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161097984))))[name = string("encoder_layers_7_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_7_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_7_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161098176)))]; + tensor linear_70_cast_fp16 = linear(bias = encoder_layers_7_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_7_self_attn_linear_out_weight_to_fp16_palettized, x = input_407_cast_fp16)[name = string("linear_70_cast_fp16")]; + tensor input_411_cast_fp16 = add(x = input_401_cast_fp16, y = linear_70_cast_fp16)[name = string("input_411_cast_fp16")]; + tensor x_199_axes_0 = const()[name = string("x_199_axes_0"), val = tensor([-1])]; + tensor encoder_layers_7_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_7_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161100288)))]; + tensor encoder_layers_7_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_7_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161102400)))]; + tensor x_199_cast_fp16 = layer_norm(axes = x_199_axes_0, beta = encoder_layers_7_norm_conv_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_7_norm_conv_weight_to_fp16, x = input_411_cast_fp16)[name = string("x_199_cast_fp16")]; + tensor input_413_perm_0 = const()[name = string("input_413_perm_0"), val = tensor([0, 2, 1])]; + string input_415_pad_type_0 = const()[name = string("input_415_pad_type_0"), val = string("valid")]; + tensor input_415_strides_0 = const()[name = string("input_415_strides_0"), val = tensor([1])]; + tensor input_415_pad_0 = const()[name = string("input_415_pad_0"), val = tensor([0, 0])]; + tensor input_415_dilations_0 = const()[name = string("input_415_dilations_0"), val = tensor([1])]; + int32 input_415_groups_0 = const()[name = string("input_415_groups_0"), val = int32(1)]; + tensor encoder_layers_7_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161104512))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163201728))))[name = string("encoder_layers_7_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_413_cast_fp16 = transpose(perm = input_413_perm_0, x = x_199_cast_fp16)[name = string("transpose_294")]; + tensor input_415_cast_fp16 = conv(dilations = input_415_dilations_0, groups = input_415_groups_0, pad = input_415_pad_0, pad_type = input_415_pad_type_0, strides = input_415_strides_0, weight = encoder_layers_7_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_413_cast_fp16)[name = string("input_415_cast_fp16")]; + int32 x_201_split_num_splits_0 = const()[name = string("x_201_split_num_splits_0"), val = int32(2)]; + int32 x_201_split_axis_0 = const()[name = string("x_201_split_axis_0"), val = int32(1)]; + tensor x_201_split_cast_fp16_0, tensor x_201_split_cast_fp16_1 = split(axis = x_201_split_axis_0, num_splits = x_201_split_num_splits_0, x = input_415_cast_fp16)[name = string("x_201_split_cast_fp16")]; + tensor x_201_split_1_sigmoid_cast_fp16 = sigmoid(x = x_201_split_cast_fp16_1)[name = string("x_201_split_1_sigmoid_cast_fp16")]; + tensor x_201_cast_fp16 = mul(x = x_201_split_cast_fp16_0, y = x_201_split_1_sigmoid_cast_fp16)[name = string("x_201_cast_fp16")]; + tensor input_417_cast_fp16 = select(a = var_45_to_fp16, b = x_201_cast_fp16, cond = var_576)[name = string("input_417_cast_fp16")]; + bool new_x_31_interleave_0 = const()[name = string("new_x_31_interleave_0"), val = bool(false)]; + tensor new_x_31_cast_fp16 = concat(axis = var_60, interleave = new_x_31_interleave_0, values = (cache_31_cast_fp16, input_417_cast_fp16))[name = string("new_x_31_cast_fp16")]; + tensor var_2080_begin_0 = const()[name = string("op_2080_begin_0"), val = tensor([0, 0, 14])]; + tensor var_2080_end_0 = const()[name = string("op_2080_end_0"), val = tensor([1, 1024, 22])]; + tensor var_2080_end_mask_0 = const()[name = string("op_2080_end_mask_0"), val = tensor([true, true, true])]; + tensor var_2080_cast_fp16 = slice_by_index(begin = var_2080_begin_0, end = var_2080_end_0, end_mask = var_2080_end_mask_0, x = new_x_31_cast_fp16)[name = string("op_2080_cast_fp16")]; + string x_203_pad_type_0 = const()[name = string("x_203_pad_type_0"), val = string("valid")]; + int32 x_203_groups_0 = const()[name = string("x_203_groups_0"), val = int32(1024)]; + tensor x_203_strides_0 = const()[name = string("x_203_strides_0"), val = tensor([1])]; + tensor x_203_pad_0 = const()[name = string("x_203_pad_0"), val = tensor([0, 0])]; + tensor x_203_dilations_0 = const()[name = string("x_203_dilations_0"), val = tensor([1])]; + tensor encoder_layers_7_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163205888))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163215168))))[name = string("encoder_layers_7_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_203_cast_fp16 = conv(dilations = x_203_dilations_0, groups = x_203_groups_0, pad = x_203_pad_0, pad_type = x_203_pad_type_0, strides = x_203_strides_0, weight = encoder_layers_7_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_31_cast_fp16)[name = string("x_203_cast_fp16")]; + tensor input_419_perm_0 = const()[name = string("input_419_perm_0"), val = tensor([0, 2, 1])]; + tensor x_205_axes_0 = const()[name = string("x_205_axes_0"), val = tensor([-1])]; + tensor encoder_layers_7_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_7_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163217280)))]; + tensor encoder_layers_7_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_7_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163219392)))]; + tensor input_419_cast_fp16 = transpose(perm = input_419_perm_0, x = x_203_cast_fp16)[name = string("transpose_293")]; + tensor x_205_cast_fp16 = layer_norm(axes = x_205_axes_0, beta = encoder_layers_7_conv_batch_norm_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_7_conv_batch_norm_weight_to_fp16, x = input_419_cast_fp16)[name = string("x_205_cast_fp16")]; + tensor input_421_perm_0 = const()[name = string("input_421_perm_0"), val = tensor([0, 2, 1])]; + tensor input_421_cast_fp16 = transpose(perm = input_421_perm_0, x = x_205_cast_fp16)[name = string("transpose_292")]; + tensor input_423_cast_fp16 = silu(x = input_421_cast_fp16)[name = string("input_423_cast_fp16")]; + string x_207_pad_type_0 = const()[name = string("x_207_pad_type_0"), val = string("valid")]; + tensor x_207_strides_0 = const()[name = string("x_207_strides_0"), val = tensor([1])]; + tensor x_207_pad_0 = const()[name = string("x_207_pad_0"), val = tensor([0, 0])]; + tensor x_207_dilations_0 = const()[name = string("x_207_dilations_0"), val = tensor([1])]; + int32 x_207_groups_0 = const()[name = string("x_207_groups_0"), val = int32(1)]; + tensor encoder_layers_7_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163221504))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(164270144))))[name = string("encoder_layers_7_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_207_cast_fp16 = conv(dilations = x_207_dilations_0, groups = x_207_groups_0, pad = x_207_pad_0, pad_type = x_207_pad_type_0, strides = x_207_strides_0, weight = encoder_layers_7_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_423_cast_fp16)[name = string("x_207_cast_fp16")]; + tensor input_425_perm_0 = const()[name = string("input_425_perm_0"), val = tensor([0, 2, 1])]; + tensor input_425_cast_fp16 = transpose(perm = input_425_perm_0, x = x_207_cast_fp16)[name = string("transpose_291")]; + tensor input_427_cast_fp16 = add(x = input_411_cast_fp16, y = input_425_cast_fp16)[name = string("input_427_cast_fp16")]; + tensor input_429_axes_0 = const()[name = string("input_429_axes_0"), val = tensor([-1])]; + tensor encoder_layers_7_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_7_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(164272256)))]; + tensor encoder_layers_7_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_7_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(164274368)))]; + tensor input_429_cast_fp16 = layer_norm(axes = input_429_axes_0, beta = encoder_layers_7_norm_feed_forward2_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_7_norm_feed_forward2_weight_to_fp16, x = input_427_cast_fp16)[name = string("input_429_cast_fp16")]; + tensor encoder_layers_7_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(164276480))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167422272))))[name = string("encoder_layers_7_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_7_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_7_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167422464)))]; + tensor linear_71_cast_fp16 = linear(bias = encoder_layers_7_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_7_feed_forward2_linear1_weight_to_fp16_palettized, x = input_429_cast_fp16)[name = string("linear_71_cast_fp16")]; + tensor input_433_cast_fp16 = silu(x = linear_71_cast_fp16)[name = string("input_433_cast_fp16")]; + tensor encoder_layers_7_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167430720))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170576512))))[name = string("encoder_layers_7_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_7_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_7_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170576704)))]; + tensor linear_72_cast_fp16 = linear(bias = encoder_layers_7_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_7_feed_forward2_linear2_weight_to_fp16_palettized, x = input_433_cast_fp16)[name = string("linear_72_cast_fp16")]; + fp16 var_2123_to_fp16 = const()[name = string("op_2123_to_fp16"), val = fp16(0x1p-1)]; + tensor var_2124_cast_fp16 = mul(x = linear_72_cast_fp16, y = var_2123_to_fp16)[name = string("op_2124_cast_fp16")]; + tensor input_439_cast_fp16 = add(x = input_427_cast_fp16, y = var_2124_cast_fp16)[name = string("input_439_cast_fp16")]; + tensor input_441_axes_0 = const()[name = string("input_441_axes_0"), val = tensor([-1])]; + tensor encoder_layers_7_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_7_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170578816)))]; + tensor encoder_layers_7_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_7_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170580928)))]; + tensor input_441_cast_fp16 = layer_norm(axes = input_441_axes_0, beta = encoder_layers_7_norm_out_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_7_norm_out_weight_to_fp16, x = input_439_cast_fp16)[name = string("input_441_cast_fp16")]; + tensor cache_33_begin_0 = const()[name = string("cache_33_begin_0"), val = tensor([8, 0, 0, 0])]; + tensor cache_33_end_0 = const()[name = string("cache_33_end_0"), val = tensor([9, 1, 42, 1024])]; + tensor cache_33_end_mask_0 = const()[name = string("cache_33_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_33_squeeze_mask_0 = const()[name = string("cache_33_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_33_cast_fp16 = slice_by_index(begin = cache_33_begin_0, end = cache_33_end_0, end_mask = cache_33_end_mask_0, squeeze_mask = cache_33_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_33_cast_fp16")]; + tensor cache_35_begin_0 = const()[name = string("cache_35_begin_0"), val = tensor([8, 0, 0, 0])]; + tensor cache_35_end_0 = const()[name = string("cache_35_end_0"), val = tensor([9, 1, 1024, 8])]; + tensor cache_35_end_mask_0 = const()[name = string("cache_35_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_35_squeeze_mask_0 = const()[name = string("cache_35_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_35_cast_fp16 = slice_by_index(begin = cache_35_begin_0, end = cache_35_end_0, end_mask = cache_35_end_mask_0, squeeze_mask = cache_35_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_35_cast_fp16")]; + tensor input_443_axes_0 = const()[name = string("input_443_axes_0"), val = tensor([-1])]; + tensor encoder_layers_8_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_8_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170583040)))]; + tensor encoder_layers_8_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_8_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170585152)))]; + tensor input_443_cast_fp16 = layer_norm(axes = input_443_axes_0, beta = encoder_layers_8_norm_feed_forward1_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_8_norm_feed_forward1_weight_to_fp16, x = input_441_cast_fp16)[name = string("input_443_cast_fp16")]; + tensor encoder_layers_8_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170587264))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(173733056))))[name = string("encoder_layers_8_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_8_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_8_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(173733248)))]; + tensor linear_73_cast_fp16 = linear(bias = encoder_layers_8_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_8_feed_forward1_linear1_weight_to_fp16_palettized, x = input_443_cast_fp16)[name = string("linear_73_cast_fp16")]; + tensor input_447_cast_fp16 = silu(x = linear_73_cast_fp16)[name = string("input_447_cast_fp16")]; + tensor encoder_layers_8_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(173741504))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(176887296))))[name = string("encoder_layers_8_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_8_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_8_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(176887488)))]; + tensor linear_74_cast_fp16 = linear(bias = encoder_layers_8_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_8_feed_forward1_linear2_weight_to_fp16_palettized, x = input_447_cast_fp16)[name = string("linear_74_cast_fp16")]; + fp16 var_2160_to_fp16 = const()[name = string("op_2160_to_fp16"), val = fp16(0x1p-1)]; + tensor var_2161_cast_fp16 = mul(x = linear_74_cast_fp16, y = var_2160_to_fp16)[name = string("op_2161_cast_fp16")]; + tensor input_453_cast_fp16 = add(x = input_441_cast_fp16, y = var_2161_cast_fp16)[name = string("input_453_cast_fp16")]; + tensor key_17_axes_0 = const()[name = string("key_17_axes_0"), val = tensor([-1])]; + tensor encoder_layers_8_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_8_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(176889600)))]; + tensor encoder_layers_8_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_8_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(176891712)))]; + tensor key_17_cast_fp16 = layer_norm(axes = key_17_axes_0, beta = encoder_layers_8_norm_self_att_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_8_norm_self_att_weight_to_fp16, x = input_453_cast_fp16)[name = string("key_17_cast_fp16")]; + bool input_455_interleave_0 = const()[name = string("input_455_interleave_0"), val = bool(false)]; + tensor input_455_cast_fp16 = concat(axis = var_69, interleave = input_455_interleave_0, values = (cache_33_cast_fp16, key_17_cast_fp16))[name = string("input_455_cast_fp16")]; + tensor var_2183_begin_0 = const()[name = string("op_2183_begin_0"), val = tensor([0, 14, 0])]; + tensor var_2183_end_0 = const()[name = string("op_2183_end_0"), val = tensor([1, 42, 1024])]; + tensor var_2183_end_mask_0 = const()[name = string("op_2183_end_mask_0"), val = tensor([true, true, true])]; + tensor var_2183_cast_fp16 = slice_by_index(begin = var_2183_begin_0, end = var_2183_end_0, end_mask = var_2183_end_mask_0, x = cache_33_cast_fp16)[name = string("op_2183_cast_fp16")]; + bool var_2189_interleave_0 = const()[name = string("op_2189_interleave_0"), val = bool(false)]; + tensor var_2189_cast_fp16 = concat(axis = var_69, interleave = var_2189_interleave_0, values = (var_2183_cast_fp16, key_17_cast_fp16))[name = string("op_2189_cast_fp16")]; + tensor encoder_layers_8_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(176893824))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(177680320))))[name = string("encoder_layers_8_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_8_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_8_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(177680512)))]; + tensor linear_75_cast_fp16 = linear(bias = encoder_layers_8_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_8_self_attn_linear_q_weight_to_fp16_palettized, x = key_17_cast_fp16)[name = string("linear_75_cast_fp16")]; + tensor var_2194 = const()[name = string("op_2194"), val = tensor([1, -1, 8, 128])]; + tensor q_49_cast_fp16 = reshape(shape = var_2194, x = linear_75_cast_fp16)[name = string("q_49_cast_fp16")]; + tensor encoder_layers_8_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(177682624))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(178469120))))[name = string("encoder_layers_8_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_8_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_8_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(178469312)))]; + tensor linear_76_cast_fp16 = linear(bias = encoder_layers_8_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_8_self_attn_linear_k_weight_to_fp16_palettized, x = input_455_cast_fp16)[name = string("linear_76_cast_fp16")]; + tensor var_2199 = const()[name = string("op_2199"), val = tensor([1, -1, 8, 128])]; + tensor k_33_cast_fp16 = reshape(shape = var_2199, x = linear_76_cast_fp16)[name = string("k_33_cast_fp16")]; + tensor encoder_layers_8_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(178471424))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179257920))))[name = string("encoder_layers_8_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_8_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_8_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179258112)))]; + tensor linear_77_cast_fp16 = linear(bias = encoder_layers_8_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_8_self_attn_linear_v_weight_to_fp16_palettized, x = input_455_cast_fp16)[name = string("linear_77_cast_fp16")]; + tensor var_2204 = const()[name = string("op_2204"), val = tensor([1, -1, 8, 128])]; + tensor v_17_cast_fp16 = reshape(shape = var_2204, x = linear_77_cast_fp16)[name = string("v_17_cast_fp16")]; + tensor value_25_perm_0 = const()[name = string("value_25_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_8_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_8_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179260224)))]; + tensor var_2217_cast_fp16 = add(x = q_49_cast_fp16, y = encoder_layers_8_self_attn_pos_bias_u_to_fp16)[name = string("op_2217_cast_fp16")]; + tensor encoder_layers_8_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_8_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179262336)))]; + tensor var_2219_cast_fp16 = add(x = q_49_cast_fp16, y = encoder_layers_8_self_attn_pos_bias_v_to_fp16)[name = string("op_2219_cast_fp16")]; + tensor q_with_bias_v_17_perm_0 = const()[name = string("q_with_bias_v_17_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_215_transpose_x_0 = const()[name = string("x_215_transpose_x_0"), val = bool(false)]; + bool x_215_transpose_y_0 = const()[name = string("x_215_transpose_y_0"), val = bool(false)]; + tensor op_2221_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179264448))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179378176))))[name = string("op_2221_to_fp16_quantized")]; + tensor q_with_bias_v_17_cast_fp16 = transpose(perm = q_with_bias_v_17_perm_0, x = var_2219_cast_fp16)[name = string("transpose_290")]; + tensor x_215_cast_fp16 = matmul(transpose_x = x_215_transpose_x_0, transpose_y = x_215_transpose_y_0, x = q_with_bias_v_17_cast_fp16, y = op_2221_to_fp16_quantized)[name = string("x_215_cast_fp16")]; + tensor x_217_pad_0 = const()[name = string("x_217_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_217_mode_0 = const()[name = string("x_217_mode_0"), val = string("constant")]; + fp16 const_183_to_fp16 = const()[name = string("const_183_to_fp16"), val = fp16(0x0p+0)]; + tensor x_217_cast_fp16 = pad(constant_val = const_183_to_fp16, mode = x_217_mode_0, pad = x_217_pad_0, x = x_215_cast_fp16)[name = string("x_217_cast_fp16")]; + tensor var_2229 = const()[name = string("op_2229"), val = tensor([1, 8, -1, 14])]; + tensor x_219_cast_fp16 = reshape(shape = var_2229, x = x_217_cast_fp16)[name = string("x_219_cast_fp16")]; + tensor var_2233_begin_0 = const()[name = string("op_2233_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_2233_end_0 = const()[name = string("op_2233_end_0"), val = tensor([1, 8, 112, 14])]; + tensor var_2233_end_mask_0 = const()[name = string("op_2233_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_2233_cast_fp16 = slice_by_index(begin = var_2233_begin_0, end = var_2233_end_0, end_mask = var_2233_end_mask_0, x = x_219_cast_fp16)[name = string("op_2233_cast_fp16")]; + tensor var_2234 = const()[name = string("op_2234"), val = tensor([1, 8, 14, 111])]; + tensor matrix_bd_33_cast_fp16 = reshape(shape = var_2234, x = var_2233_cast_fp16)[name = string("matrix_bd_33_cast_fp16")]; + bool matrix_ac_17_transpose_x_0 = const()[name = string("matrix_ac_17_transpose_x_0"), val = bool(false)]; + bool matrix_ac_17_transpose_y_0 = const()[name = string("matrix_ac_17_transpose_y_0"), val = bool(false)]; + tensor transpose_112_perm_0 = const()[name = string("transpose_112_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_113_perm_0 = const()[name = string("transpose_113_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_113 = transpose(perm = transpose_113_perm_0, x = k_33_cast_fp16)[name = string("transpose_288")]; + tensor transpose_112 = transpose(perm = transpose_112_perm_0, x = var_2217_cast_fp16)[name = string("transpose_289")]; + tensor matrix_ac_17_cast_fp16 = matmul(transpose_x = matrix_ac_17_transpose_x_0, transpose_y = matrix_ac_17_transpose_y_0, x = transpose_112, y = transpose_113)[name = string("matrix_ac_17_cast_fp16")]; + tensor matrix_bd_35_begin_0 = const()[name = string("matrix_bd_35_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_35_end_0 = const()[name = string("matrix_bd_35_end_0"), val = tensor([1, 8, 14, 56])]; + tensor matrix_bd_35_end_mask_0 = const()[name = string("matrix_bd_35_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_35_cast_fp16 = slice_by_index(begin = matrix_bd_35_begin_0, end = matrix_bd_35_end_0, end_mask = matrix_bd_35_end_mask_0, x = matrix_bd_33_cast_fp16)[name = string("matrix_bd_35_cast_fp16")]; + tensor var_2243_cast_fp16 = add(x = matrix_ac_17_cast_fp16, y = matrix_bd_35_cast_fp16)[name = string("op_2243_cast_fp16")]; + fp16 _inversed_scores_33_y_0_to_fp16 = const()[name = string("_inversed_scores_33_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_33_cast_fp16 = mul(x = var_2243_cast_fp16, y = _inversed_scores_33_y_0_to_fp16)[name = string("_inversed_scores_33_cast_fp16")]; + tensor scores_35_cast_fp16 = select(a = var_46_to_fp16, b = _inversed_scores_33_cast_fp16, cond = mask_11)[name = string("scores_35_cast_fp16")]; + tensor var_2249_cast_fp16 = softmax(axis = var_60, x = scores_35_cast_fp16)[name = string("op_2249_cast_fp16")]; + tensor input_457_cast_fp16 = select(a = var_45_to_fp16, b = var_2249_cast_fp16, cond = mask_11)[name = string("input_457_cast_fp16")]; + bool x_221_transpose_x_0 = const()[name = string("x_221_transpose_x_0"), val = bool(false)]; + bool x_221_transpose_y_0 = const()[name = string("x_221_transpose_y_0"), val = bool(false)]; + tensor value_25_cast_fp16 = transpose(perm = value_25_perm_0, x = v_17_cast_fp16)[name = string("transpose_287")]; + tensor x_221_cast_fp16 = matmul(transpose_x = x_221_transpose_x_0, transpose_y = x_221_transpose_y_0, x = input_457_cast_fp16, y = value_25_cast_fp16)[name = string("x_221_cast_fp16")]; + tensor var_2253_perm_0 = const()[name = string("op_2253_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_2254 = const()[name = string("op_2254"), val = tensor([1, -1, 1024])]; + tensor var_2253_cast_fp16 = transpose(perm = var_2253_perm_0, x = x_221_cast_fp16)[name = string("transpose_286")]; + tensor input_459_cast_fp16 = reshape(shape = var_2254, x = var_2253_cast_fp16)[name = string("input_459_cast_fp16")]; + tensor encoder_layers_8_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179378496))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(180164992))))[name = string("encoder_layers_8_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_8_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_8_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(180165184)))]; + tensor linear_79_cast_fp16 = linear(bias = encoder_layers_8_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_8_self_attn_linear_out_weight_to_fp16_palettized, x = input_459_cast_fp16)[name = string("linear_79_cast_fp16")]; + tensor input_463_cast_fp16 = add(x = input_453_cast_fp16, y = linear_79_cast_fp16)[name = string("input_463_cast_fp16")]; + tensor x_225_axes_0 = const()[name = string("x_225_axes_0"), val = tensor([-1])]; + tensor encoder_layers_8_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_8_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(180167296)))]; + tensor encoder_layers_8_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_8_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(180169408)))]; + tensor x_225_cast_fp16 = layer_norm(axes = x_225_axes_0, beta = encoder_layers_8_norm_conv_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_8_norm_conv_weight_to_fp16, x = input_463_cast_fp16)[name = string("x_225_cast_fp16")]; + tensor input_465_perm_0 = const()[name = string("input_465_perm_0"), val = tensor([0, 2, 1])]; + string input_467_pad_type_0 = const()[name = string("input_467_pad_type_0"), val = string("valid")]; + tensor input_467_strides_0 = const()[name = string("input_467_strides_0"), val = tensor([1])]; + tensor input_467_pad_0 = const()[name = string("input_467_pad_0"), val = tensor([0, 0])]; + tensor input_467_dilations_0 = const()[name = string("input_467_dilations_0"), val = tensor([1])]; + int32 input_467_groups_0 = const()[name = string("input_467_groups_0"), val = int32(1)]; + tensor encoder_layers_8_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(180171520))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182268736))))[name = string("encoder_layers_8_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_465_cast_fp16 = transpose(perm = input_465_perm_0, x = x_225_cast_fp16)[name = string("transpose_285")]; + tensor input_467_cast_fp16 = conv(dilations = input_467_dilations_0, groups = input_467_groups_0, pad = input_467_pad_0, pad_type = input_467_pad_type_0, strides = input_467_strides_0, weight = encoder_layers_8_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_465_cast_fp16)[name = string("input_467_cast_fp16")]; + int32 x_227_split_num_splits_0 = const()[name = string("x_227_split_num_splits_0"), val = int32(2)]; + int32 x_227_split_axis_0 = const()[name = string("x_227_split_axis_0"), val = int32(1)]; + tensor x_227_split_cast_fp16_0, tensor x_227_split_cast_fp16_1 = split(axis = x_227_split_axis_0, num_splits = x_227_split_num_splits_0, x = input_467_cast_fp16)[name = string("x_227_split_cast_fp16")]; + tensor x_227_split_1_sigmoid_cast_fp16 = sigmoid(x = x_227_split_cast_fp16_1)[name = string("x_227_split_1_sigmoid_cast_fp16")]; + tensor x_227_cast_fp16 = mul(x = x_227_split_cast_fp16_0, y = x_227_split_1_sigmoid_cast_fp16)[name = string("x_227_cast_fp16")]; + tensor input_469_cast_fp16 = select(a = var_45_to_fp16, b = x_227_cast_fp16, cond = var_576)[name = string("input_469_cast_fp16")]; + bool new_x_35_interleave_0 = const()[name = string("new_x_35_interleave_0"), val = bool(false)]; + tensor new_x_35_cast_fp16 = concat(axis = var_60, interleave = new_x_35_interleave_0, values = (cache_35_cast_fp16, input_469_cast_fp16))[name = string("new_x_35_cast_fp16")]; + tensor var_2293_begin_0 = const()[name = string("op_2293_begin_0"), val = tensor([0, 0, 14])]; + tensor var_2293_end_0 = const()[name = string("op_2293_end_0"), val = tensor([1, 1024, 22])]; + tensor var_2293_end_mask_0 = const()[name = string("op_2293_end_mask_0"), val = tensor([true, true, true])]; + tensor var_2293_cast_fp16 = slice_by_index(begin = var_2293_begin_0, end = var_2293_end_0, end_mask = var_2293_end_mask_0, x = new_x_35_cast_fp16)[name = string("op_2293_cast_fp16")]; + string x_229_pad_type_0 = const()[name = string("x_229_pad_type_0"), val = string("valid")]; + int32 x_229_groups_0 = const()[name = string("x_229_groups_0"), val = int32(1024)]; + tensor x_229_strides_0 = const()[name = string("x_229_strides_0"), val = tensor([1])]; + tensor x_229_pad_0 = const()[name = string("x_229_pad_0"), val = tensor([0, 0])]; + tensor x_229_dilations_0 = const()[name = string("x_229_dilations_0"), val = tensor([1])]; + tensor encoder_layers_8_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182272896))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182282176))))[name = string("encoder_layers_8_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_229_cast_fp16 = conv(dilations = x_229_dilations_0, groups = x_229_groups_0, pad = x_229_pad_0, pad_type = x_229_pad_type_0, strides = x_229_strides_0, weight = encoder_layers_8_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_35_cast_fp16)[name = string("x_229_cast_fp16")]; + tensor input_471_perm_0 = const()[name = string("input_471_perm_0"), val = tensor([0, 2, 1])]; + tensor x_231_axes_0 = const()[name = string("x_231_axes_0"), val = tensor([-1])]; + tensor encoder_layers_8_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_8_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182284288)))]; + tensor encoder_layers_8_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_8_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182286400)))]; + tensor input_471_cast_fp16 = transpose(perm = input_471_perm_0, x = x_229_cast_fp16)[name = string("transpose_284")]; + tensor x_231_cast_fp16 = layer_norm(axes = x_231_axes_0, beta = encoder_layers_8_conv_batch_norm_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_8_conv_batch_norm_weight_to_fp16, x = input_471_cast_fp16)[name = string("x_231_cast_fp16")]; + tensor input_473_perm_0 = const()[name = string("input_473_perm_0"), val = tensor([0, 2, 1])]; + tensor input_473_cast_fp16 = transpose(perm = input_473_perm_0, x = x_231_cast_fp16)[name = string("transpose_283")]; + tensor input_475_cast_fp16 = silu(x = input_473_cast_fp16)[name = string("input_475_cast_fp16")]; + string x_233_pad_type_0 = const()[name = string("x_233_pad_type_0"), val = string("valid")]; + tensor x_233_strides_0 = const()[name = string("x_233_strides_0"), val = tensor([1])]; + tensor x_233_pad_0 = const()[name = string("x_233_pad_0"), val = tensor([0, 0])]; + tensor x_233_dilations_0 = const()[name = string("x_233_dilations_0"), val = tensor([1])]; + int32 x_233_groups_0 = const()[name = string("x_233_groups_0"), val = int32(1)]; + tensor encoder_layers_8_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182288512))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183337152))))[name = string("encoder_layers_8_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_233_cast_fp16 = conv(dilations = x_233_dilations_0, groups = x_233_groups_0, pad = x_233_pad_0, pad_type = x_233_pad_type_0, strides = x_233_strides_0, weight = encoder_layers_8_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_475_cast_fp16)[name = string("x_233_cast_fp16")]; + tensor input_477_perm_0 = const()[name = string("input_477_perm_0"), val = tensor([0, 2, 1])]; + tensor input_477_cast_fp16 = transpose(perm = input_477_perm_0, x = x_233_cast_fp16)[name = string("transpose_282")]; + tensor input_479_cast_fp16 = add(x = input_463_cast_fp16, y = input_477_cast_fp16)[name = string("input_479_cast_fp16")]; + tensor input_481_axes_0 = const()[name = string("input_481_axes_0"), val = tensor([-1])]; + tensor encoder_layers_8_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_8_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183339264)))]; + tensor encoder_layers_8_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_8_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183341376)))]; + tensor input_481_cast_fp16 = layer_norm(axes = input_481_axes_0, beta = encoder_layers_8_norm_feed_forward2_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_8_norm_feed_forward2_weight_to_fp16, x = input_479_cast_fp16)[name = string("input_481_cast_fp16")]; + tensor encoder_layers_8_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183343488))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186489280))))[name = string("encoder_layers_8_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_8_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_8_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186489472)))]; + tensor linear_80_cast_fp16 = linear(bias = encoder_layers_8_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_8_feed_forward2_linear1_weight_to_fp16_palettized, x = input_481_cast_fp16)[name = string("linear_80_cast_fp16")]; + tensor input_485_cast_fp16 = silu(x = linear_80_cast_fp16)[name = string("input_485_cast_fp16")]; + tensor encoder_layers_8_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186497728))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(189643520))))[name = string("encoder_layers_8_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_8_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_8_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(189643712)))]; + tensor linear_81_cast_fp16 = linear(bias = encoder_layers_8_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_8_feed_forward2_linear2_weight_to_fp16_palettized, x = input_485_cast_fp16)[name = string("linear_81_cast_fp16")]; + fp16 var_2336_to_fp16 = const()[name = string("op_2336_to_fp16"), val = fp16(0x1p-1)]; + tensor var_2337_cast_fp16 = mul(x = linear_81_cast_fp16, y = var_2336_to_fp16)[name = string("op_2337_cast_fp16")]; + tensor input_491_cast_fp16 = add(x = input_479_cast_fp16, y = var_2337_cast_fp16)[name = string("input_491_cast_fp16")]; + tensor input_493_axes_0 = const()[name = string("input_493_axes_0"), val = tensor([-1])]; + tensor encoder_layers_8_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_8_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(189645824)))]; + tensor encoder_layers_8_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_8_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(189647936)))]; + tensor input_493_cast_fp16 = layer_norm(axes = input_493_axes_0, beta = encoder_layers_8_norm_out_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_8_norm_out_weight_to_fp16, x = input_491_cast_fp16)[name = string("input_493_cast_fp16")]; + tensor cache_37_begin_0 = const()[name = string("cache_37_begin_0"), val = tensor([9, 0, 0, 0])]; + tensor cache_37_end_0 = const()[name = string("cache_37_end_0"), val = tensor([10, 1, 42, 1024])]; + tensor cache_37_end_mask_0 = const()[name = string("cache_37_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_37_squeeze_mask_0 = const()[name = string("cache_37_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_37_cast_fp16 = slice_by_index(begin = cache_37_begin_0, end = cache_37_end_0, end_mask = cache_37_end_mask_0, squeeze_mask = cache_37_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_37_cast_fp16")]; + tensor cache_39_begin_0 = const()[name = string("cache_39_begin_0"), val = tensor([9, 0, 0, 0])]; + tensor cache_39_end_0 = const()[name = string("cache_39_end_0"), val = tensor([10, 1, 1024, 8])]; + tensor cache_39_end_mask_0 = const()[name = string("cache_39_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_39_squeeze_mask_0 = const()[name = string("cache_39_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_39_cast_fp16 = slice_by_index(begin = cache_39_begin_0, end = cache_39_end_0, end_mask = cache_39_end_mask_0, squeeze_mask = cache_39_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_39_cast_fp16")]; + tensor input_495_axes_0 = const()[name = string("input_495_axes_0"), val = tensor([-1])]; + tensor encoder_layers_9_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_9_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(189650048)))]; + tensor encoder_layers_9_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_9_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(189652160)))]; + tensor input_495_cast_fp16 = layer_norm(axes = input_495_axes_0, beta = encoder_layers_9_norm_feed_forward1_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_9_norm_feed_forward1_weight_to_fp16, x = input_493_cast_fp16)[name = string("input_495_cast_fp16")]; + tensor encoder_layers_9_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(189654272))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(192800064))))[name = string("encoder_layers_9_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_9_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_9_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(192800256)))]; + tensor linear_82_cast_fp16 = linear(bias = encoder_layers_9_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_9_feed_forward1_linear1_weight_to_fp16_palettized, x = input_495_cast_fp16)[name = string("linear_82_cast_fp16")]; + tensor input_499_cast_fp16 = silu(x = linear_82_cast_fp16)[name = string("input_499_cast_fp16")]; + tensor encoder_layers_9_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(192808512))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(195954304))))[name = string("encoder_layers_9_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_9_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_9_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(195954496)))]; + tensor linear_83_cast_fp16 = linear(bias = encoder_layers_9_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_9_feed_forward1_linear2_weight_to_fp16_palettized, x = input_499_cast_fp16)[name = string("linear_83_cast_fp16")]; + fp16 var_2373_to_fp16 = const()[name = string("op_2373_to_fp16"), val = fp16(0x1p-1)]; + tensor var_2374_cast_fp16 = mul(x = linear_83_cast_fp16, y = var_2373_to_fp16)[name = string("op_2374_cast_fp16")]; + tensor input_505_cast_fp16 = add(x = input_493_cast_fp16, y = var_2374_cast_fp16)[name = string("input_505_cast_fp16")]; + tensor key_19_axes_0 = const()[name = string("key_19_axes_0"), val = tensor([-1])]; + tensor encoder_layers_9_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_9_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(195956608)))]; + tensor encoder_layers_9_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_9_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(195958720)))]; + tensor key_19_cast_fp16 = layer_norm(axes = key_19_axes_0, beta = encoder_layers_9_norm_self_att_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_9_norm_self_att_weight_to_fp16, x = input_505_cast_fp16)[name = string("key_19_cast_fp16")]; + bool input_507_interleave_0 = const()[name = string("input_507_interleave_0"), val = bool(false)]; + tensor input_507_cast_fp16 = concat(axis = var_69, interleave = input_507_interleave_0, values = (cache_37_cast_fp16, key_19_cast_fp16))[name = string("input_507_cast_fp16")]; + tensor var_2396_begin_0 = const()[name = string("op_2396_begin_0"), val = tensor([0, 14, 0])]; + tensor var_2396_end_0 = const()[name = string("op_2396_end_0"), val = tensor([1, 42, 1024])]; + tensor var_2396_end_mask_0 = const()[name = string("op_2396_end_mask_0"), val = tensor([true, true, true])]; + tensor var_2396_cast_fp16 = slice_by_index(begin = var_2396_begin_0, end = var_2396_end_0, end_mask = var_2396_end_mask_0, x = cache_37_cast_fp16)[name = string("op_2396_cast_fp16")]; + bool var_2402_interleave_0 = const()[name = string("op_2402_interleave_0"), val = bool(false)]; + tensor var_2402_cast_fp16 = concat(axis = var_69, interleave = var_2402_interleave_0, values = (var_2396_cast_fp16, key_19_cast_fp16))[name = string("op_2402_cast_fp16")]; + tensor encoder_layers_9_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(195960832))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196747328))))[name = string("encoder_layers_9_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_9_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_9_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196747520)))]; + tensor linear_84_cast_fp16 = linear(bias = encoder_layers_9_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_9_self_attn_linear_q_weight_to_fp16_palettized, x = key_19_cast_fp16)[name = string("linear_84_cast_fp16")]; + tensor var_2407 = const()[name = string("op_2407"), val = tensor([1, -1, 8, 128])]; + tensor q_55_cast_fp16 = reshape(shape = var_2407, x = linear_84_cast_fp16)[name = string("q_55_cast_fp16")]; + tensor encoder_layers_9_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196749632))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197536128))))[name = string("encoder_layers_9_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_9_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_9_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197536320)))]; + tensor linear_85_cast_fp16 = linear(bias = encoder_layers_9_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_9_self_attn_linear_k_weight_to_fp16_palettized, x = input_507_cast_fp16)[name = string("linear_85_cast_fp16")]; + tensor var_2412 = const()[name = string("op_2412"), val = tensor([1, -1, 8, 128])]; + tensor k_37_cast_fp16 = reshape(shape = var_2412, x = linear_85_cast_fp16)[name = string("k_37_cast_fp16")]; + tensor encoder_layers_9_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197538432))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198324928))))[name = string("encoder_layers_9_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_9_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_9_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198325120)))]; + tensor linear_86_cast_fp16 = linear(bias = encoder_layers_9_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_9_self_attn_linear_v_weight_to_fp16_palettized, x = input_507_cast_fp16)[name = string("linear_86_cast_fp16")]; + tensor var_2417 = const()[name = string("op_2417"), val = tensor([1, -1, 8, 128])]; + tensor v_19_cast_fp16 = reshape(shape = var_2417, x = linear_86_cast_fp16)[name = string("v_19_cast_fp16")]; + tensor value_27_perm_0 = const()[name = string("value_27_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_9_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_9_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198327232)))]; + tensor var_2430_cast_fp16 = add(x = q_55_cast_fp16, y = encoder_layers_9_self_attn_pos_bias_u_to_fp16)[name = string("op_2430_cast_fp16")]; + tensor encoder_layers_9_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_9_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198329344)))]; + tensor var_2432_cast_fp16 = add(x = q_55_cast_fp16, y = encoder_layers_9_self_attn_pos_bias_v_to_fp16)[name = string("op_2432_cast_fp16")]; + tensor q_with_bias_v_19_perm_0 = const()[name = string("q_with_bias_v_19_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_241_transpose_x_0 = const()[name = string("x_241_transpose_x_0"), val = bool(false)]; + bool x_241_transpose_y_0 = const()[name = string("x_241_transpose_y_0"), val = bool(false)]; + tensor op_2434_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198331456))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198445184))))[name = string("op_2434_to_fp16_quantized")]; + tensor q_with_bias_v_19_cast_fp16 = transpose(perm = q_with_bias_v_19_perm_0, x = var_2432_cast_fp16)[name = string("transpose_281")]; + tensor x_241_cast_fp16 = matmul(transpose_x = x_241_transpose_x_0, transpose_y = x_241_transpose_y_0, x = q_with_bias_v_19_cast_fp16, y = op_2434_to_fp16_quantized)[name = string("x_241_cast_fp16")]; + tensor x_243_pad_0 = const()[name = string("x_243_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_243_mode_0 = const()[name = string("x_243_mode_0"), val = string("constant")]; + fp16 const_196_to_fp16 = const()[name = string("const_196_to_fp16"), val = fp16(0x0p+0)]; + tensor x_243_cast_fp16 = pad(constant_val = const_196_to_fp16, mode = x_243_mode_0, pad = x_243_pad_0, x = x_241_cast_fp16)[name = string("x_243_cast_fp16")]; + tensor var_2442 = const()[name = string("op_2442"), val = tensor([1, 8, -1, 14])]; + tensor x_245_cast_fp16 = reshape(shape = var_2442, x = x_243_cast_fp16)[name = string("x_245_cast_fp16")]; + tensor var_2446_begin_0 = const()[name = string("op_2446_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_2446_end_0 = const()[name = string("op_2446_end_0"), val = tensor([1, 8, 112, 14])]; + tensor var_2446_end_mask_0 = const()[name = string("op_2446_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_2446_cast_fp16 = slice_by_index(begin = var_2446_begin_0, end = var_2446_end_0, end_mask = var_2446_end_mask_0, x = x_245_cast_fp16)[name = string("op_2446_cast_fp16")]; + tensor var_2447 = const()[name = string("op_2447"), val = tensor([1, 8, 14, 111])]; + tensor matrix_bd_37_cast_fp16 = reshape(shape = var_2447, x = var_2446_cast_fp16)[name = string("matrix_bd_37_cast_fp16")]; + bool matrix_ac_19_transpose_x_0 = const()[name = string("matrix_ac_19_transpose_x_0"), val = bool(false)]; + bool matrix_ac_19_transpose_y_0 = const()[name = string("matrix_ac_19_transpose_y_0"), val = bool(false)]; + tensor transpose_114_perm_0 = const()[name = string("transpose_114_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_115_perm_0 = const()[name = string("transpose_115_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_115 = transpose(perm = transpose_115_perm_0, x = k_37_cast_fp16)[name = string("transpose_279")]; + tensor transpose_114 = transpose(perm = transpose_114_perm_0, x = var_2430_cast_fp16)[name = string("transpose_280")]; + tensor matrix_ac_19_cast_fp16 = matmul(transpose_x = matrix_ac_19_transpose_x_0, transpose_y = matrix_ac_19_transpose_y_0, x = transpose_114, y = transpose_115)[name = string("matrix_ac_19_cast_fp16")]; + tensor matrix_bd_39_begin_0 = const()[name = string("matrix_bd_39_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_39_end_0 = const()[name = string("matrix_bd_39_end_0"), val = tensor([1, 8, 14, 56])]; + tensor matrix_bd_39_end_mask_0 = const()[name = string("matrix_bd_39_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_39_cast_fp16 = slice_by_index(begin = matrix_bd_39_begin_0, end = matrix_bd_39_end_0, end_mask = matrix_bd_39_end_mask_0, x = matrix_bd_37_cast_fp16)[name = string("matrix_bd_39_cast_fp16")]; + tensor var_2456_cast_fp16 = add(x = matrix_ac_19_cast_fp16, y = matrix_bd_39_cast_fp16)[name = string("op_2456_cast_fp16")]; + fp16 _inversed_scores_37_y_0_to_fp16 = const()[name = string("_inversed_scores_37_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_37_cast_fp16 = mul(x = var_2456_cast_fp16, y = _inversed_scores_37_y_0_to_fp16)[name = string("_inversed_scores_37_cast_fp16")]; + tensor scores_39_cast_fp16 = select(a = var_46_to_fp16, b = _inversed_scores_37_cast_fp16, cond = mask_11)[name = string("scores_39_cast_fp16")]; + tensor var_2462_cast_fp16 = softmax(axis = var_60, x = scores_39_cast_fp16)[name = string("op_2462_cast_fp16")]; + tensor input_509_cast_fp16 = select(a = var_45_to_fp16, b = var_2462_cast_fp16, cond = mask_11)[name = string("input_509_cast_fp16")]; + bool x_247_transpose_x_0 = const()[name = string("x_247_transpose_x_0"), val = bool(false)]; + bool x_247_transpose_y_0 = const()[name = string("x_247_transpose_y_0"), val = bool(false)]; + tensor value_27_cast_fp16 = transpose(perm = value_27_perm_0, x = v_19_cast_fp16)[name = string("transpose_278")]; + tensor x_247_cast_fp16 = matmul(transpose_x = x_247_transpose_x_0, transpose_y = x_247_transpose_y_0, x = input_509_cast_fp16, y = value_27_cast_fp16)[name = string("x_247_cast_fp16")]; + tensor var_2466_perm_0 = const()[name = string("op_2466_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_2467 = const()[name = string("op_2467"), val = tensor([1, -1, 1024])]; + tensor var_2466_cast_fp16 = transpose(perm = var_2466_perm_0, x = x_247_cast_fp16)[name = string("transpose_277")]; + tensor input_511_cast_fp16 = reshape(shape = var_2467, x = var_2466_cast_fp16)[name = string("input_511_cast_fp16")]; + tensor encoder_layers_9_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198445504))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199232000))))[name = string("encoder_layers_9_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_9_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_9_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199232192)))]; + tensor linear_88_cast_fp16 = linear(bias = encoder_layers_9_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_9_self_attn_linear_out_weight_to_fp16_palettized, x = input_511_cast_fp16)[name = string("linear_88_cast_fp16")]; + tensor input_515_cast_fp16 = add(x = input_505_cast_fp16, y = linear_88_cast_fp16)[name = string("input_515_cast_fp16")]; + tensor x_251_axes_0 = const()[name = string("x_251_axes_0"), val = tensor([-1])]; + tensor encoder_layers_9_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_9_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199234304)))]; + tensor encoder_layers_9_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_9_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199236416)))]; + tensor x_251_cast_fp16 = layer_norm(axes = x_251_axes_0, beta = encoder_layers_9_norm_conv_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_9_norm_conv_weight_to_fp16, x = input_515_cast_fp16)[name = string("x_251_cast_fp16")]; + tensor input_517_perm_0 = const()[name = string("input_517_perm_0"), val = tensor([0, 2, 1])]; + string input_519_pad_type_0 = const()[name = string("input_519_pad_type_0"), val = string("valid")]; + tensor input_519_strides_0 = const()[name = string("input_519_strides_0"), val = tensor([1])]; + tensor input_519_pad_0 = const()[name = string("input_519_pad_0"), val = tensor([0, 0])]; + tensor input_519_dilations_0 = const()[name = string("input_519_dilations_0"), val = tensor([1])]; + int32 input_519_groups_0 = const()[name = string("input_519_groups_0"), val = int32(1)]; + tensor encoder_layers_9_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199238528))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(201335744))))[name = string("encoder_layers_9_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_517_cast_fp16 = transpose(perm = input_517_perm_0, x = x_251_cast_fp16)[name = string("transpose_276")]; + tensor input_519_cast_fp16 = conv(dilations = input_519_dilations_0, groups = input_519_groups_0, pad = input_519_pad_0, pad_type = input_519_pad_type_0, strides = input_519_strides_0, weight = encoder_layers_9_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_517_cast_fp16)[name = string("input_519_cast_fp16")]; + int32 x_253_split_num_splits_0 = const()[name = string("x_253_split_num_splits_0"), val = int32(2)]; + int32 x_253_split_axis_0 = const()[name = string("x_253_split_axis_0"), val = int32(1)]; + tensor x_253_split_cast_fp16_0, tensor x_253_split_cast_fp16_1 = split(axis = x_253_split_axis_0, num_splits = x_253_split_num_splits_0, x = input_519_cast_fp16)[name = string("x_253_split_cast_fp16")]; + tensor x_253_split_1_sigmoid_cast_fp16 = sigmoid(x = x_253_split_cast_fp16_1)[name = string("x_253_split_1_sigmoid_cast_fp16")]; + tensor x_253_cast_fp16 = mul(x = x_253_split_cast_fp16_0, y = x_253_split_1_sigmoid_cast_fp16)[name = string("x_253_cast_fp16")]; + tensor input_521_cast_fp16 = select(a = var_45_to_fp16, b = x_253_cast_fp16, cond = var_576)[name = string("input_521_cast_fp16")]; + bool new_x_39_interleave_0 = const()[name = string("new_x_39_interleave_0"), val = bool(false)]; + tensor new_x_39_cast_fp16 = concat(axis = var_60, interleave = new_x_39_interleave_0, values = (cache_39_cast_fp16, input_521_cast_fp16))[name = string("new_x_39_cast_fp16")]; + tensor var_2506_begin_0 = const()[name = string("op_2506_begin_0"), val = tensor([0, 0, 14])]; + tensor var_2506_end_0 = const()[name = string("op_2506_end_0"), val = tensor([1, 1024, 22])]; + tensor var_2506_end_mask_0 = const()[name = string("op_2506_end_mask_0"), val = tensor([true, true, true])]; + tensor var_2506_cast_fp16 = slice_by_index(begin = var_2506_begin_0, end = var_2506_end_0, end_mask = var_2506_end_mask_0, x = new_x_39_cast_fp16)[name = string("op_2506_cast_fp16")]; + string x_255_pad_type_0 = const()[name = string("x_255_pad_type_0"), val = string("valid")]; + int32 x_255_groups_0 = const()[name = string("x_255_groups_0"), val = int32(1024)]; + tensor x_255_strides_0 = const()[name = string("x_255_strides_0"), val = tensor([1])]; + tensor x_255_pad_0 = const()[name = string("x_255_pad_0"), val = tensor([0, 0])]; + tensor x_255_dilations_0 = const()[name = string("x_255_dilations_0"), val = tensor([1])]; + tensor encoder_layers_9_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(201339904))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(201349184))))[name = string("encoder_layers_9_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_255_cast_fp16 = conv(dilations = x_255_dilations_0, groups = x_255_groups_0, pad = x_255_pad_0, pad_type = x_255_pad_type_0, strides = x_255_strides_0, weight = encoder_layers_9_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_39_cast_fp16)[name = string("x_255_cast_fp16")]; + tensor input_523_perm_0 = const()[name = string("input_523_perm_0"), val = tensor([0, 2, 1])]; + tensor x_257_axes_0 = const()[name = string("x_257_axes_0"), val = tensor([-1])]; + tensor encoder_layers_9_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_9_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(201351296)))]; + tensor encoder_layers_9_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_9_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(201353408)))]; + tensor input_523_cast_fp16 = transpose(perm = input_523_perm_0, x = x_255_cast_fp16)[name = string("transpose_275")]; + tensor x_257_cast_fp16 = layer_norm(axes = x_257_axes_0, beta = encoder_layers_9_conv_batch_norm_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_9_conv_batch_norm_weight_to_fp16, x = input_523_cast_fp16)[name = string("x_257_cast_fp16")]; + tensor input_525_perm_0 = const()[name = string("input_525_perm_0"), val = tensor([0, 2, 1])]; + tensor input_525_cast_fp16 = transpose(perm = input_525_perm_0, x = x_257_cast_fp16)[name = string("transpose_274")]; + tensor input_527_cast_fp16 = silu(x = input_525_cast_fp16)[name = string("input_527_cast_fp16")]; + string x_259_pad_type_0 = const()[name = string("x_259_pad_type_0"), val = string("valid")]; + tensor x_259_strides_0 = const()[name = string("x_259_strides_0"), val = tensor([1])]; + tensor x_259_pad_0 = const()[name = string("x_259_pad_0"), val = tensor([0, 0])]; + tensor x_259_dilations_0 = const()[name = string("x_259_dilations_0"), val = tensor([1])]; + int32 x_259_groups_0 = const()[name = string("x_259_groups_0"), val = int32(1)]; + tensor encoder_layers_9_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(201355520))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202404160))))[name = string("encoder_layers_9_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_259_cast_fp16 = conv(dilations = x_259_dilations_0, groups = x_259_groups_0, pad = x_259_pad_0, pad_type = x_259_pad_type_0, strides = x_259_strides_0, weight = encoder_layers_9_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_527_cast_fp16)[name = string("x_259_cast_fp16")]; + tensor input_529_perm_0 = const()[name = string("input_529_perm_0"), val = tensor([0, 2, 1])]; + tensor input_529_cast_fp16 = transpose(perm = input_529_perm_0, x = x_259_cast_fp16)[name = string("transpose_273")]; + tensor input_531_cast_fp16 = add(x = input_515_cast_fp16, y = input_529_cast_fp16)[name = string("input_531_cast_fp16")]; + tensor input_533_axes_0 = const()[name = string("input_533_axes_0"), val = tensor([-1])]; + tensor encoder_layers_9_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_9_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202406272)))]; + tensor encoder_layers_9_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_9_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202408384)))]; + tensor input_533_cast_fp16 = layer_norm(axes = input_533_axes_0, beta = encoder_layers_9_norm_feed_forward2_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_9_norm_feed_forward2_weight_to_fp16, x = input_531_cast_fp16)[name = string("input_533_cast_fp16")]; + tensor encoder_layers_9_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202410496))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(205556288))))[name = string("encoder_layers_9_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_9_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_9_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(205556480)))]; + tensor linear_89_cast_fp16 = linear(bias = encoder_layers_9_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_9_feed_forward2_linear1_weight_to_fp16_palettized, x = input_533_cast_fp16)[name = string("linear_89_cast_fp16")]; + tensor input_537_cast_fp16 = silu(x = linear_89_cast_fp16)[name = string("input_537_cast_fp16")]; + tensor encoder_layers_9_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(205564736))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(208710528))))[name = string("encoder_layers_9_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_9_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_9_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(208710720)))]; + tensor linear_90_cast_fp16 = linear(bias = encoder_layers_9_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_9_feed_forward2_linear2_weight_to_fp16_palettized, x = input_537_cast_fp16)[name = string("linear_90_cast_fp16")]; + fp16 var_2549_to_fp16 = const()[name = string("op_2549_to_fp16"), val = fp16(0x1p-1)]; + tensor var_2550_cast_fp16 = mul(x = linear_90_cast_fp16, y = var_2549_to_fp16)[name = string("op_2550_cast_fp16")]; + tensor input_543_cast_fp16 = add(x = input_531_cast_fp16, y = var_2550_cast_fp16)[name = string("input_543_cast_fp16")]; + tensor input_545_axes_0 = const()[name = string("input_545_axes_0"), val = tensor([-1])]; + tensor encoder_layers_9_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_9_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(208712832)))]; + tensor encoder_layers_9_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_9_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(208714944)))]; + tensor input_545_cast_fp16 = layer_norm(axes = input_545_axes_0, beta = encoder_layers_9_norm_out_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_9_norm_out_weight_to_fp16, x = input_543_cast_fp16)[name = string("input_545_cast_fp16")]; + tensor cache_41_begin_0 = const()[name = string("cache_41_begin_0"), val = tensor([10, 0, 0, 0])]; + tensor cache_41_end_0 = const()[name = string("cache_41_end_0"), val = tensor([11, 1, 42, 1024])]; + tensor cache_41_end_mask_0 = const()[name = string("cache_41_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_41_squeeze_mask_0 = const()[name = string("cache_41_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_41_cast_fp16 = slice_by_index(begin = cache_41_begin_0, end = cache_41_end_0, end_mask = cache_41_end_mask_0, squeeze_mask = cache_41_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_41_cast_fp16")]; + tensor cache_43_begin_0 = const()[name = string("cache_43_begin_0"), val = tensor([10, 0, 0, 0])]; + tensor cache_43_end_0 = const()[name = string("cache_43_end_0"), val = tensor([11, 1, 1024, 8])]; + tensor cache_43_end_mask_0 = const()[name = string("cache_43_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_43_squeeze_mask_0 = const()[name = string("cache_43_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_43_cast_fp16 = slice_by_index(begin = cache_43_begin_0, end = cache_43_end_0, end_mask = cache_43_end_mask_0, squeeze_mask = cache_43_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_43_cast_fp16")]; + tensor input_547_axes_0 = const()[name = string("input_547_axes_0"), val = tensor([-1])]; + tensor encoder_layers_10_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_10_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(208717056)))]; + tensor encoder_layers_10_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_10_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(208719168)))]; + tensor input_547_cast_fp16 = layer_norm(axes = input_547_axes_0, beta = encoder_layers_10_norm_feed_forward1_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_10_norm_feed_forward1_weight_to_fp16, x = input_545_cast_fp16)[name = string("input_547_cast_fp16")]; + tensor encoder_layers_10_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(208721280))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211867072))))[name = string("encoder_layers_10_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_10_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_10_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211867264)))]; + tensor linear_91_cast_fp16 = linear(bias = encoder_layers_10_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_10_feed_forward1_linear1_weight_to_fp16_palettized, x = input_547_cast_fp16)[name = string("linear_91_cast_fp16")]; + tensor input_551_cast_fp16 = silu(x = linear_91_cast_fp16)[name = string("input_551_cast_fp16")]; + tensor encoder_layers_10_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211875520))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215021312))))[name = string("encoder_layers_10_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_10_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_10_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215021504)))]; + tensor linear_92_cast_fp16 = linear(bias = encoder_layers_10_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_10_feed_forward1_linear2_weight_to_fp16_palettized, x = input_551_cast_fp16)[name = string("linear_92_cast_fp16")]; + fp16 var_2586_to_fp16 = const()[name = string("op_2586_to_fp16"), val = fp16(0x1p-1)]; + tensor var_2587_cast_fp16 = mul(x = linear_92_cast_fp16, y = var_2586_to_fp16)[name = string("op_2587_cast_fp16")]; + tensor input_557_cast_fp16 = add(x = input_545_cast_fp16, y = var_2587_cast_fp16)[name = string("input_557_cast_fp16")]; + tensor key_21_axes_0 = const()[name = string("key_21_axes_0"), val = tensor([-1])]; + tensor encoder_layers_10_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_10_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215023616)))]; + tensor encoder_layers_10_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_10_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215025728)))]; + tensor key_21_cast_fp16 = layer_norm(axes = key_21_axes_0, beta = encoder_layers_10_norm_self_att_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_10_norm_self_att_weight_to_fp16, x = input_557_cast_fp16)[name = string("key_21_cast_fp16")]; + bool input_559_interleave_0 = const()[name = string("input_559_interleave_0"), val = bool(false)]; + tensor input_559_cast_fp16 = concat(axis = var_69, interleave = input_559_interleave_0, values = (cache_41_cast_fp16, key_21_cast_fp16))[name = string("input_559_cast_fp16")]; + tensor var_2609_begin_0 = const()[name = string("op_2609_begin_0"), val = tensor([0, 14, 0])]; + tensor var_2609_end_0 = const()[name = string("op_2609_end_0"), val = tensor([1, 42, 1024])]; + tensor var_2609_end_mask_0 = const()[name = string("op_2609_end_mask_0"), val = tensor([true, true, true])]; + tensor var_2609_cast_fp16 = slice_by_index(begin = var_2609_begin_0, end = var_2609_end_0, end_mask = var_2609_end_mask_0, x = cache_41_cast_fp16)[name = string("op_2609_cast_fp16")]; + bool var_2615_interleave_0 = const()[name = string("op_2615_interleave_0"), val = bool(false)]; + tensor var_2615_cast_fp16 = concat(axis = var_69, interleave = var_2615_interleave_0, values = (var_2609_cast_fp16, key_21_cast_fp16))[name = string("op_2615_cast_fp16")]; + tensor encoder_layers_10_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215027840))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215814336))))[name = string("encoder_layers_10_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_10_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_10_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215814528)))]; + tensor linear_93_cast_fp16 = linear(bias = encoder_layers_10_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_10_self_attn_linear_q_weight_to_fp16_palettized, x = key_21_cast_fp16)[name = string("linear_93_cast_fp16")]; + tensor var_2620 = const()[name = string("op_2620"), val = tensor([1, -1, 8, 128])]; + tensor q_61_cast_fp16 = reshape(shape = var_2620, x = linear_93_cast_fp16)[name = string("q_61_cast_fp16")]; + tensor encoder_layers_10_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215816640))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216603136))))[name = string("encoder_layers_10_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_10_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_10_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216603328)))]; + tensor linear_94_cast_fp16 = linear(bias = encoder_layers_10_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_10_self_attn_linear_k_weight_to_fp16_palettized, x = input_559_cast_fp16)[name = string("linear_94_cast_fp16")]; + tensor var_2625 = const()[name = string("op_2625"), val = tensor([1, -1, 8, 128])]; + tensor k_41_cast_fp16 = reshape(shape = var_2625, x = linear_94_cast_fp16)[name = string("k_41_cast_fp16")]; + tensor encoder_layers_10_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216605440))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217391936))))[name = string("encoder_layers_10_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_10_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_10_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217392128)))]; + tensor linear_95_cast_fp16 = linear(bias = encoder_layers_10_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_10_self_attn_linear_v_weight_to_fp16_palettized, x = input_559_cast_fp16)[name = string("linear_95_cast_fp16")]; + tensor var_2630 = const()[name = string("op_2630"), val = tensor([1, -1, 8, 128])]; + tensor v_21_cast_fp16 = reshape(shape = var_2630, x = linear_95_cast_fp16)[name = string("v_21_cast_fp16")]; + tensor value_29_perm_0 = const()[name = string("value_29_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_10_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_10_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217394240)))]; + tensor var_2643_cast_fp16 = add(x = q_61_cast_fp16, y = encoder_layers_10_self_attn_pos_bias_u_to_fp16)[name = string("op_2643_cast_fp16")]; + tensor encoder_layers_10_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_10_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217396352)))]; + tensor var_2645_cast_fp16 = add(x = q_61_cast_fp16, y = encoder_layers_10_self_attn_pos_bias_v_to_fp16)[name = string("op_2645_cast_fp16")]; + tensor q_with_bias_v_21_perm_0 = const()[name = string("q_with_bias_v_21_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_267_transpose_x_0 = const()[name = string("x_267_transpose_x_0"), val = bool(false)]; + bool x_267_transpose_y_0 = const()[name = string("x_267_transpose_y_0"), val = bool(false)]; + tensor op_2647_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217398464))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217512192))))[name = string("op_2647_to_fp16_quantized")]; + tensor q_with_bias_v_21_cast_fp16 = transpose(perm = q_with_bias_v_21_perm_0, x = var_2645_cast_fp16)[name = string("transpose_272")]; + tensor x_267_cast_fp16 = matmul(transpose_x = x_267_transpose_x_0, transpose_y = x_267_transpose_y_0, x = q_with_bias_v_21_cast_fp16, y = op_2647_to_fp16_quantized)[name = string("x_267_cast_fp16")]; + tensor x_269_pad_0 = const()[name = string("x_269_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_269_mode_0 = const()[name = string("x_269_mode_0"), val = string("constant")]; + fp16 const_209_to_fp16 = const()[name = string("const_209_to_fp16"), val = fp16(0x0p+0)]; + tensor x_269_cast_fp16 = pad(constant_val = const_209_to_fp16, mode = x_269_mode_0, pad = x_269_pad_0, x = x_267_cast_fp16)[name = string("x_269_cast_fp16")]; + tensor var_2655 = const()[name = string("op_2655"), val = tensor([1, 8, -1, 14])]; + tensor x_271_cast_fp16 = reshape(shape = var_2655, x = x_269_cast_fp16)[name = string("x_271_cast_fp16")]; + tensor var_2659_begin_0 = const()[name = string("op_2659_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_2659_end_0 = const()[name = string("op_2659_end_0"), val = tensor([1, 8, 112, 14])]; + tensor var_2659_end_mask_0 = const()[name = string("op_2659_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_2659_cast_fp16 = slice_by_index(begin = var_2659_begin_0, end = var_2659_end_0, end_mask = var_2659_end_mask_0, x = x_271_cast_fp16)[name = string("op_2659_cast_fp16")]; + tensor var_2660 = const()[name = string("op_2660"), val = tensor([1, 8, 14, 111])]; + tensor matrix_bd_41_cast_fp16 = reshape(shape = var_2660, x = var_2659_cast_fp16)[name = string("matrix_bd_41_cast_fp16")]; + bool matrix_ac_21_transpose_x_0 = const()[name = string("matrix_ac_21_transpose_x_0"), val = bool(false)]; + bool matrix_ac_21_transpose_y_0 = const()[name = string("matrix_ac_21_transpose_y_0"), val = bool(false)]; + tensor transpose_116_perm_0 = const()[name = string("transpose_116_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_117_perm_0 = const()[name = string("transpose_117_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_117 = transpose(perm = transpose_117_perm_0, x = k_41_cast_fp16)[name = string("transpose_270")]; + tensor transpose_116 = transpose(perm = transpose_116_perm_0, x = var_2643_cast_fp16)[name = string("transpose_271")]; + tensor matrix_ac_21_cast_fp16 = matmul(transpose_x = matrix_ac_21_transpose_x_0, transpose_y = matrix_ac_21_transpose_y_0, x = transpose_116, y = transpose_117)[name = string("matrix_ac_21_cast_fp16")]; + tensor matrix_bd_43_begin_0 = const()[name = string("matrix_bd_43_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_43_end_0 = const()[name = string("matrix_bd_43_end_0"), val = tensor([1, 8, 14, 56])]; + tensor matrix_bd_43_end_mask_0 = const()[name = string("matrix_bd_43_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_43_cast_fp16 = slice_by_index(begin = matrix_bd_43_begin_0, end = matrix_bd_43_end_0, end_mask = matrix_bd_43_end_mask_0, x = matrix_bd_41_cast_fp16)[name = string("matrix_bd_43_cast_fp16")]; + tensor var_2669_cast_fp16 = add(x = matrix_ac_21_cast_fp16, y = matrix_bd_43_cast_fp16)[name = string("op_2669_cast_fp16")]; + fp16 _inversed_scores_41_y_0_to_fp16 = const()[name = string("_inversed_scores_41_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_41_cast_fp16 = mul(x = var_2669_cast_fp16, y = _inversed_scores_41_y_0_to_fp16)[name = string("_inversed_scores_41_cast_fp16")]; + tensor scores_43_cast_fp16 = select(a = var_46_to_fp16, b = _inversed_scores_41_cast_fp16, cond = mask_11)[name = string("scores_43_cast_fp16")]; + tensor var_2675_cast_fp16 = softmax(axis = var_60, x = scores_43_cast_fp16)[name = string("op_2675_cast_fp16")]; + tensor input_561_cast_fp16 = select(a = var_45_to_fp16, b = var_2675_cast_fp16, cond = mask_11)[name = string("input_561_cast_fp16")]; + bool x_273_transpose_x_0 = const()[name = string("x_273_transpose_x_0"), val = bool(false)]; + bool x_273_transpose_y_0 = const()[name = string("x_273_transpose_y_0"), val = bool(false)]; + tensor value_29_cast_fp16 = transpose(perm = value_29_perm_0, x = v_21_cast_fp16)[name = string("transpose_269")]; + tensor x_273_cast_fp16 = matmul(transpose_x = x_273_transpose_x_0, transpose_y = x_273_transpose_y_0, x = input_561_cast_fp16, y = value_29_cast_fp16)[name = string("x_273_cast_fp16")]; + tensor var_2679_perm_0 = const()[name = string("op_2679_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_2680 = const()[name = string("op_2680"), val = tensor([1, -1, 1024])]; + tensor var_2679_cast_fp16 = transpose(perm = var_2679_perm_0, x = x_273_cast_fp16)[name = string("transpose_268")]; + tensor input_563_cast_fp16 = reshape(shape = var_2680, x = var_2679_cast_fp16)[name = string("input_563_cast_fp16")]; + tensor encoder_layers_10_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217512512))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218299008))))[name = string("encoder_layers_10_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_10_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_10_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218299200)))]; + tensor linear_97_cast_fp16 = linear(bias = encoder_layers_10_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_10_self_attn_linear_out_weight_to_fp16_palettized, x = input_563_cast_fp16)[name = string("linear_97_cast_fp16")]; + tensor input_567_cast_fp16 = add(x = input_557_cast_fp16, y = linear_97_cast_fp16)[name = string("input_567_cast_fp16")]; + tensor x_277_axes_0 = const()[name = string("x_277_axes_0"), val = tensor([-1])]; + tensor encoder_layers_10_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_10_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218301312)))]; + tensor encoder_layers_10_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_10_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218303424)))]; + tensor x_277_cast_fp16 = layer_norm(axes = x_277_axes_0, beta = encoder_layers_10_norm_conv_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_10_norm_conv_weight_to_fp16, x = input_567_cast_fp16)[name = string("x_277_cast_fp16")]; + tensor input_569_perm_0 = const()[name = string("input_569_perm_0"), val = tensor([0, 2, 1])]; + string input_571_pad_type_0 = const()[name = string("input_571_pad_type_0"), val = string("valid")]; + tensor input_571_strides_0 = const()[name = string("input_571_strides_0"), val = tensor([1])]; + tensor input_571_pad_0 = const()[name = string("input_571_pad_0"), val = tensor([0, 0])]; + tensor input_571_dilations_0 = const()[name = string("input_571_dilations_0"), val = tensor([1])]; + int32 input_571_groups_0 = const()[name = string("input_571_groups_0"), val = int32(1)]; + tensor encoder_layers_10_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218305536))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220402752))))[name = string("encoder_layers_10_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_569_cast_fp16 = transpose(perm = input_569_perm_0, x = x_277_cast_fp16)[name = string("transpose_267")]; + tensor input_571_cast_fp16 = conv(dilations = input_571_dilations_0, groups = input_571_groups_0, pad = input_571_pad_0, pad_type = input_571_pad_type_0, strides = input_571_strides_0, weight = encoder_layers_10_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_569_cast_fp16)[name = string("input_571_cast_fp16")]; + int32 x_279_split_num_splits_0 = const()[name = string("x_279_split_num_splits_0"), val = int32(2)]; + int32 x_279_split_axis_0 = const()[name = string("x_279_split_axis_0"), val = int32(1)]; + tensor x_279_split_cast_fp16_0, tensor x_279_split_cast_fp16_1 = split(axis = x_279_split_axis_0, num_splits = x_279_split_num_splits_0, x = input_571_cast_fp16)[name = string("x_279_split_cast_fp16")]; + tensor x_279_split_1_sigmoid_cast_fp16 = sigmoid(x = x_279_split_cast_fp16_1)[name = string("x_279_split_1_sigmoid_cast_fp16")]; + tensor x_279_cast_fp16 = mul(x = x_279_split_cast_fp16_0, y = x_279_split_1_sigmoid_cast_fp16)[name = string("x_279_cast_fp16")]; + tensor input_573_cast_fp16 = select(a = var_45_to_fp16, b = x_279_cast_fp16, cond = var_576)[name = string("input_573_cast_fp16")]; + bool new_x_43_interleave_0 = const()[name = string("new_x_43_interleave_0"), val = bool(false)]; + tensor new_x_43_cast_fp16 = concat(axis = var_60, interleave = new_x_43_interleave_0, values = (cache_43_cast_fp16, input_573_cast_fp16))[name = string("new_x_43_cast_fp16")]; + tensor var_2719_begin_0 = const()[name = string("op_2719_begin_0"), val = tensor([0, 0, 14])]; + tensor var_2719_end_0 = const()[name = string("op_2719_end_0"), val = tensor([1, 1024, 22])]; + tensor var_2719_end_mask_0 = const()[name = string("op_2719_end_mask_0"), val = tensor([true, true, true])]; + tensor var_2719_cast_fp16 = slice_by_index(begin = var_2719_begin_0, end = var_2719_end_0, end_mask = var_2719_end_mask_0, x = new_x_43_cast_fp16)[name = string("op_2719_cast_fp16")]; + string x_281_pad_type_0 = const()[name = string("x_281_pad_type_0"), val = string("valid")]; + int32 x_281_groups_0 = const()[name = string("x_281_groups_0"), val = int32(1024)]; + tensor x_281_strides_0 = const()[name = string("x_281_strides_0"), val = tensor([1])]; + tensor x_281_pad_0 = const()[name = string("x_281_pad_0"), val = tensor([0, 0])]; + tensor x_281_dilations_0 = const()[name = string("x_281_dilations_0"), val = tensor([1])]; + tensor encoder_layers_10_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220406912))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220416192))))[name = string("encoder_layers_10_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_281_cast_fp16 = conv(dilations = x_281_dilations_0, groups = x_281_groups_0, pad = x_281_pad_0, pad_type = x_281_pad_type_0, strides = x_281_strides_0, weight = encoder_layers_10_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_43_cast_fp16)[name = string("x_281_cast_fp16")]; + tensor input_575_perm_0 = const()[name = string("input_575_perm_0"), val = tensor([0, 2, 1])]; + tensor x_283_axes_0 = const()[name = string("x_283_axes_0"), val = tensor([-1])]; + tensor encoder_layers_10_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_10_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220418304)))]; + tensor encoder_layers_10_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_10_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220420416)))]; + tensor input_575_cast_fp16 = transpose(perm = input_575_perm_0, x = x_281_cast_fp16)[name = string("transpose_266")]; + tensor x_283_cast_fp16 = layer_norm(axes = x_283_axes_0, beta = encoder_layers_10_conv_batch_norm_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_10_conv_batch_norm_weight_to_fp16, x = input_575_cast_fp16)[name = string("x_283_cast_fp16")]; + tensor input_577_perm_0 = const()[name = string("input_577_perm_0"), val = tensor([0, 2, 1])]; + tensor input_577_cast_fp16 = transpose(perm = input_577_perm_0, x = x_283_cast_fp16)[name = string("transpose_265")]; + tensor input_579_cast_fp16 = silu(x = input_577_cast_fp16)[name = string("input_579_cast_fp16")]; + string x_285_pad_type_0 = const()[name = string("x_285_pad_type_0"), val = string("valid")]; + tensor x_285_strides_0 = const()[name = string("x_285_strides_0"), val = tensor([1])]; + tensor x_285_pad_0 = const()[name = string("x_285_pad_0"), val = tensor([0, 0])]; + tensor x_285_dilations_0 = const()[name = string("x_285_dilations_0"), val = tensor([1])]; + int32 x_285_groups_0 = const()[name = string("x_285_groups_0"), val = int32(1)]; + tensor encoder_layers_10_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220422528))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221471168))))[name = string("encoder_layers_10_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_285_cast_fp16 = conv(dilations = x_285_dilations_0, groups = x_285_groups_0, pad = x_285_pad_0, pad_type = x_285_pad_type_0, strides = x_285_strides_0, weight = encoder_layers_10_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_579_cast_fp16)[name = string("x_285_cast_fp16")]; + tensor input_581_perm_0 = const()[name = string("input_581_perm_0"), val = tensor([0, 2, 1])]; + tensor input_581_cast_fp16 = transpose(perm = input_581_perm_0, x = x_285_cast_fp16)[name = string("transpose_264")]; + tensor input_583_cast_fp16 = add(x = input_567_cast_fp16, y = input_581_cast_fp16)[name = string("input_583_cast_fp16")]; + tensor input_585_axes_0 = const()[name = string("input_585_axes_0"), val = tensor([-1])]; + tensor encoder_layers_10_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_10_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221473280)))]; + tensor encoder_layers_10_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_10_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221475392)))]; + tensor input_585_cast_fp16 = layer_norm(axes = input_585_axes_0, beta = encoder_layers_10_norm_feed_forward2_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_10_norm_feed_forward2_weight_to_fp16, x = input_583_cast_fp16)[name = string("input_585_cast_fp16")]; + tensor encoder_layers_10_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221477504))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(224623296))))[name = string("encoder_layers_10_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_10_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_10_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(224623488)))]; + tensor linear_98_cast_fp16 = linear(bias = encoder_layers_10_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_10_feed_forward2_linear1_weight_to_fp16_palettized, x = input_585_cast_fp16)[name = string("linear_98_cast_fp16")]; + tensor input_589_cast_fp16 = silu(x = linear_98_cast_fp16)[name = string("input_589_cast_fp16")]; + tensor encoder_layers_10_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(224631744))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227777536))))[name = string("encoder_layers_10_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_10_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_10_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227777728)))]; + tensor linear_99_cast_fp16 = linear(bias = encoder_layers_10_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_10_feed_forward2_linear2_weight_to_fp16_palettized, x = input_589_cast_fp16)[name = string("linear_99_cast_fp16")]; + fp16 var_2762_to_fp16 = const()[name = string("op_2762_to_fp16"), val = fp16(0x1p-1)]; + tensor var_2763_cast_fp16 = mul(x = linear_99_cast_fp16, y = var_2762_to_fp16)[name = string("op_2763_cast_fp16")]; + tensor input_595_cast_fp16 = add(x = input_583_cast_fp16, y = var_2763_cast_fp16)[name = string("input_595_cast_fp16")]; + tensor input_597_axes_0 = const()[name = string("input_597_axes_0"), val = tensor([-1])]; + tensor encoder_layers_10_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_10_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227779840)))]; + tensor encoder_layers_10_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_10_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227781952)))]; + tensor input_597_cast_fp16 = layer_norm(axes = input_597_axes_0, beta = encoder_layers_10_norm_out_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_10_norm_out_weight_to_fp16, x = input_595_cast_fp16)[name = string("input_597_cast_fp16")]; + tensor cache_45_begin_0 = const()[name = string("cache_45_begin_0"), val = tensor([11, 0, 0, 0])]; + tensor cache_45_end_0 = const()[name = string("cache_45_end_0"), val = tensor([12, 1, 42, 1024])]; + tensor cache_45_end_mask_0 = const()[name = string("cache_45_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_45_squeeze_mask_0 = const()[name = string("cache_45_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_45_cast_fp16 = slice_by_index(begin = cache_45_begin_0, end = cache_45_end_0, end_mask = cache_45_end_mask_0, squeeze_mask = cache_45_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_45_cast_fp16")]; + tensor cache_47_begin_0 = const()[name = string("cache_47_begin_0"), val = tensor([11, 0, 0, 0])]; + tensor cache_47_end_0 = const()[name = string("cache_47_end_0"), val = tensor([12, 1, 1024, 8])]; + tensor cache_47_end_mask_0 = const()[name = string("cache_47_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_47_squeeze_mask_0 = const()[name = string("cache_47_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_47_cast_fp16 = slice_by_index(begin = cache_47_begin_0, end = cache_47_end_0, end_mask = cache_47_end_mask_0, squeeze_mask = cache_47_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_47_cast_fp16")]; + tensor input_599_axes_0 = const()[name = string("input_599_axes_0"), val = tensor([-1])]; + tensor encoder_layers_11_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_11_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227784064)))]; + tensor encoder_layers_11_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_11_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227786176)))]; + tensor input_599_cast_fp16 = layer_norm(axes = input_599_axes_0, beta = encoder_layers_11_norm_feed_forward1_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_11_norm_feed_forward1_weight_to_fp16, x = input_597_cast_fp16)[name = string("input_599_cast_fp16")]; + tensor encoder_layers_11_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227788288))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(230934080))))[name = string("encoder_layers_11_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_11_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_11_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(230934272)))]; + tensor linear_100_cast_fp16 = linear(bias = encoder_layers_11_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_11_feed_forward1_linear1_weight_to_fp16_palettized, x = input_599_cast_fp16)[name = string("linear_100_cast_fp16")]; + tensor input_603_cast_fp16 = silu(x = linear_100_cast_fp16)[name = string("input_603_cast_fp16")]; + tensor encoder_layers_11_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(230942528))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234088320))))[name = string("encoder_layers_11_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_11_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_11_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234088512)))]; + tensor linear_101_cast_fp16 = linear(bias = encoder_layers_11_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_11_feed_forward1_linear2_weight_to_fp16_palettized, x = input_603_cast_fp16)[name = string("linear_101_cast_fp16")]; + fp16 var_2799_to_fp16 = const()[name = string("op_2799_to_fp16"), val = fp16(0x1p-1)]; + tensor var_2800_cast_fp16 = mul(x = linear_101_cast_fp16, y = var_2799_to_fp16)[name = string("op_2800_cast_fp16")]; + tensor input_609_cast_fp16 = add(x = input_597_cast_fp16, y = var_2800_cast_fp16)[name = string("input_609_cast_fp16")]; + tensor key_23_axes_0 = const()[name = string("key_23_axes_0"), val = tensor([-1])]; + tensor encoder_layers_11_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_11_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234090624)))]; + tensor encoder_layers_11_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_11_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234092736)))]; + tensor key_23_cast_fp16 = layer_norm(axes = key_23_axes_0, beta = encoder_layers_11_norm_self_att_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_11_norm_self_att_weight_to_fp16, x = input_609_cast_fp16)[name = string("key_23_cast_fp16")]; + bool input_611_interleave_0 = const()[name = string("input_611_interleave_0"), val = bool(false)]; + tensor input_611_cast_fp16 = concat(axis = var_69, interleave = input_611_interleave_0, values = (cache_45_cast_fp16, key_23_cast_fp16))[name = string("input_611_cast_fp16")]; + tensor var_2822_begin_0 = const()[name = string("op_2822_begin_0"), val = tensor([0, 14, 0])]; + tensor var_2822_end_0 = const()[name = string("op_2822_end_0"), val = tensor([1, 42, 1024])]; + tensor var_2822_end_mask_0 = const()[name = string("op_2822_end_mask_0"), val = tensor([true, true, true])]; + tensor var_2822_cast_fp16 = slice_by_index(begin = var_2822_begin_0, end = var_2822_end_0, end_mask = var_2822_end_mask_0, x = cache_45_cast_fp16)[name = string("op_2822_cast_fp16")]; + bool var_2828_interleave_0 = const()[name = string("op_2828_interleave_0"), val = bool(false)]; + tensor var_2828_cast_fp16 = concat(axis = var_69, interleave = var_2828_interleave_0, values = (var_2822_cast_fp16, key_23_cast_fp16))[name = string("op_2828_cast_fp16")]; + tensor encoder_layers_11_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234094848))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234881344))))[name = string("encoder_layers_11_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_11_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_11_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234881536)))]; + tensor linear_102_cast_fp16 = linear(bias = encoder_layers_11_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_11_self_attn_linear_q_weight_to_fp16_palettized, x = key_23_cast_fp16)[name = string("linear_102_cast_fp16")]; + tensor var_2833 = const()[name = string("op_2833"), val = tensor([1, -1, 8, 128])]; + tensor q_67_cast_fp16 = reshape(shape = var_2833, x = linear_102_cast_fp16)[name = string("q_67_cast_fp16")]; + tensor encoder_layers_11_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234883648))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(235670144))))[name = string("encoder_layers_11_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_11_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_11_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(235670336)))]; + tensor linear_103_cast_fp16 = linear(bias = encoder_layers_11_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_11_self_attn_linear_k_weight_to_fp16_palettized, x = input_611_cast_fp16)[name = string("linear_103_cast_fp16")]; + tensor var_2838 = const()[name = string("op_2838"), val = tensor([1, -1, 8, 128])]; + tensor k_45_cast_fp16 = reshape(shape = var_2838, x = linear_103_cast_fp16)[name = string("k_45_cast_fp16")]; + tensor encoder_layers_11_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(235672448))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236458944))))[name = string("encoder_layers_11_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_11_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_11_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236459136)))]; + tensor linear_104_cast_fp16 = linear(bias = encoder_layers_11_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_11_self_attn_linear_v_weight_to_fp16_palettized, x = input_611_cast_fp16)[name = string("linear_104_cast_fp16")]; + tensor var_2843 = const()[name = string("op_2843"), val = tensor([1, -1, 8, 128])]; + tensor v_23_cast_fp16 = reshape(shape = var_2843, x = linear_104_cast_fp16)[name = string("v_23_cast_fp16")]; + tensor value_31_perm_0 = const()[name = string("value_31_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_11_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_11_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236461248)))]; + tensor var_2856_cast_fp16 = add(x = q_67_cast_fp16, y = encoder_layers_11_self_attn_pos_bias_u_to_fp16)[name = string("op_2856_cast_fp16")]; + tensor encoder_layers_11_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_11_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236463360)))]; + tensor var_2858_cast_fp16 = add(x = q_67_cast_fp16, y = encoder_layers_11_self_attn_pos_bias_v_to_fp16)[name = string("op_2858_cast_fp16")]; + tensor q_with_bias_v_23_perm_0 = const()[name = string("q_with_bias_v_23_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_293_transpose_x_0 = const()[name = string("x_293_transpose_x_0"), val = bool(false)]; + bool x_293_transpose_y_0 = const()[name = string("x_293_transpose_y_0"), val = bool(false)]; + tensor op_2860_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236465472))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236579200))))[name = string("op_2860_to_fp16_quantized")]; + tensor q_with_bias_v_23_cast_fp16 = transpose(perm = q_with_bias_v_23_perm_0, x = var_2858_cast_fp16)[name = string("transpose_263")]; + tensor x_293_cast_fp16 = matmul(transpose_x = x_293_transpose_x_0, transpose_y = x_293_transpose_y_0, x = q_with_bias_v_23_cast_fp16, y = op_2860_to_fp16_quantized)[name = string("x_293_cast_fp16")]; + tensor x_295_pad_0 = const()[name = string("x_295_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_295_mode_0 = const()[name = string("x_295_mode_0"), val = string("constant")]; + fp16 const_222_to_fp16 = const()[name = string("const_222_to_fp16"), val = fp16(0x0p+0)]; + tensor x_295_cast_fp16 = pad(constant_val = const_222_to_fp16, mode = x_295_mode_0, pad = x_295_pad_0, x = x_293_cast_fp16)[name = string("x_295_cast_fp16")]; + tensor var_2868 = const()[name = string("op_2868"), val = tensor([1, 8, -1, 14])]; + tensor x_297_cast_fp16 = reshape(shape = var_2868, x = x_295_cast_fp16)[name = string("x_297_cast_fp16")]; + tensor var_2872_begin_0 = const()[name = string("op_2872_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_2872_end_0 = const()[name = string("op_2872_end_0"), val = tensor([1, 8, 112, 14])]; + tensor var_2872_end_mask_0 = const()[name = string("op_2872_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_2872_cast_fp16 = slice_by_index(begin = var_2872_begin_0, end = var_2872_end_0, end_mask = var_2872_end_mask_0, x = x_297_cast_fp16)[name = string("op_2872_cast_fp16")]; + tensor var_2873 = const()[name = string("op_2873"), val = tensor([1, 8, 14, 111])]; + tensor matrix_bd_45_cast_fp16 = reshape(shape = var_2873, x = var_2872_cast_fp16)[name = string("matrix_bd_45_cast_fp16")]; + bool matrix_ac_23_transpose_x_0 = const()[name = string("matrix_ac_23_transpose_x_0"), val = bool(false)]; + bool matrix_ac_23_transpose_y_0 = const()[name = string("matrix_ac_23_transpose_y_0"), val = bool(false)]; + tensor transpose_118_perm_0 = const()[name = string("transpose_118_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_119_perm_0 = const()[name = string("transpose_119_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_119 = transpose(perm = transpose_119_perm_0, x = k_45_cast_fp16)[name = string("transpose_261")]; + tensor transpose_118 = transpose(perm = transpose_118_perm_0, x = var_2856_cast_fp16)[name = string("transpose_262")]; + tensor matrix_ac_23_cast_fp16 = matmul(transpose_x = matrix_ac_23_transpose_x_0, transpose_y = matrix_ac_23_transpose_y_0, x = transpose_118, y = transpose_119)[name = string("matrix_ac_23_cast_fp16")]; + tensor matrix_bd_47_begin_0 = const()[name = string("matrix_bd_47_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_47_end_0 = const()[name = string("matrix_bd_47_end_0"), val = tensor([1, 8, 14, 56])]; + tensor matrix_bd_47_end_mask_0 = const()[name = string("matrix_bd_47_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_47_cast_fp16 = slice_by_index(begin = matrix_bd_47_begin_0, end = matrix_bd_47_end_0, end_mask = matrix_bd_47_end_mask_0, x = matrix_bd_45_cast_fp16)[name = string("matrix_bd_47_cast_fp16")]; + tensor var_2882_cast_fp16 = add(x = matrix_ac_23_cast_fp16, y = matrix_bd_47_cast_fp16)[name = string("op_2882_cast_fp16")]; + fp16 _inversed_scores_45_y_0_to_fp16 = const()[name = string("_inversed_scores_45_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_45_cast_fp16 = mul(x = var_2882_cast_fp16, y = _inversed_scores_45_y_0_to_fp16)[name = string("_inversed_scores_45_cast_fp16")]; + tensor scores_47_cast_fp16 = select(a = var_46_to_fp16, b = _inversed_scores_45_cast_fp16, cond = mask_11)[name = string("scores_47_cast_fp16")]; + tensor var_2888_cast_fp16 = softmax(axis = var_60, x = scores_47_cast_fp16)[name = string("op_2888_cast_fp16")]; + tensor input_613_cast_fp16 = select(a = var_45_to_fp16, b = var_2888_cast_fp16, cond = mask_11)[name = string("input_613_cast_fp16")]; + bool x_299_transpose_x_0 = const()[name = string("x_299_transpose_x_0"), val = bool(false)]; + bool x_299_transpose_y_0 = const()[name = string("x_299_transpose_y_0"), val = bool(false)]; + tensor value_31_cast_fp16 = transpose(perm = value_31_perm_0, x = v_23_cast_fp16)[name = string("transpose_260")]; + tensor x_299_cast_fp16 = matmul(transpose_x = x_299_transpose_x_0, transpose_y = x_299_transpose_y_0, x = input_613_cast_fp16, y = value_31_cast_fp16)[name = string("x_299_cast_fp16")]; + tensor var_2892_perm_0 = const()[name = string("op_2892_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_2893 = const()[name = string("op_2893"), val = tensor([1, -1, 1024])]; + tensor var_2892_cast_fp16 = transpose(perm = var_2892_perm_0, x = x_299_cast_fp16)[name = string("transpose_259")]; + tensor input_615_cast_fp16 = reshape(shape = var_2893, x = var_2892_cast_fp16)[name = string("input_615_cast_fp16")]; + tensor encoder_layers_11_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236579520))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(237366016))))[name = string("encoder_layers_11_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_11_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_11_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(237366208)))]; + tensor linear_106_cast_fp16 = linear(bias = encoder_layers_11_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_11_self_attn_linear_out_weight_to_fp16_palettized, x = input_615_cast_fp16)[name = string("linear_106_cast_fp16")]; + tensor input_619_cast_fp16 = add(x = input_609_cast_fp16, y = linear_106_cast_fp16)[name = string("input_619_cast_fp16")]; + tensor x_303_axes_0 = const()[name = string("x_303_axes_0"), val = tensor([-1])]; + tensor encoder_layers_11_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_11_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(237368320)))]; + tensor encoder_layers_11_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_11_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(237370432)))]; + tensor x_303_cast_fp16 = layer_norm(axes = x_303_axes_0, beta = encoder_layers_11_norm_conv_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_11_norm_conv_weight_to_fp16, x = input_619_cast_fp16)[name = string("x_303_cast_fp16")]; + tensor input_621_perm_0 = const()[name = string("input_621_perm_0"), val = tensor([0, 2, 1])]; + string input_623_pad_type_0 = const()[name = string("input_623_pad_type_0"), val = string("valid")]; + tensor input_623_strides_0 = const()[name = string("input_623_strides_0"), val = tensor([1])]; + tensor input_623_pad_0 = const()[name = string("input_623_pad_0"), val = tensor([0, 0])]; + tensor input_623_dilations_0 = const()[name = string("input_623_dilations_0"), val = tensor([1])]; + int32 input_623_groups_0 = const()[name = string("input_623_groups_0"), val = int32(1)]; + tensor encoder_layers_11_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(237372544))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(239469760))))[name = string("encoder_layers_11_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_621_cast_fp16 = transpose(perm = input_621_perm_0, x = x_303_cast_fp16)[name = string("transpose_258")]; + tensor input_623_cast_fp16 = conv(dilations = input_623_dilations_0, groups = input_623_groups_0, pad = input_623_pad_0, pad_type = input_623_pad_type_0, strides = input_623_strides_0, weight = encoder_layers_11_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_621_cast_fp16)[name = string("input_623_cast_fp16")]; + int32 x_305_split_num_splits_0 = const()[name = string("x_305_split_num_splits_0"), val = int32(2)]; + int32 x_305_split_axis_0 = const()[name = string("x_305_split_axis_0"), val = int32(1)]; + tensor x_305_split_cast_fp16_0, tensor x_305_split_cast_fp16_1 = split(axis = x_305_split_axis_0, num_splits = x_305_split_num_splits_0, x = input_623_cast_fp16)[name = string("x_305_split_cast_fp16")]; + tensor x_305_split_1_sigmoid_cast_fp16 = sigmoid(x = x_305_split_cast_fp16_1)[name = string("x_305_split_1_sigmoid_cast_fp16")]; + tensor x_305_cast_fp16 = mul(x = x_305_split_cast_fp16_0, y = x_305_split_1_sigmoid_cast_fp16)[name = string("x_305_cast_fp16")]; + tensor input_625_cast_fp16 = select(a = var_45_to_fp16, b = x_305_cast_fp16, cond = var_576)[name = string("input_625_cast_fp16")]; + bool new_x_47_interleave_0 = const()[name = string("new_x_47_interleave_0"), val = bool(false)]; + tensor new_x_47_cast_fp16 = concat(axis = var_60, interleave = new_x_47_interleave_0, values = (cache_47_cast_fp16, input_625_cast_fp16))[name = string("new_x_47_cast_fp16")]; + tensor var_2932_begin_0 = const()[name = string("op_2932_begin_0"), val = tensor([0, 0, 14])]; + tensor var_2932_end_0 = const()[name = string("op_2932_end_0"), val = tensor([1, 1024, 22])]; + tensor var_2932_end_mask_0 = const()[name = string("op_2932_end_mask_0"), val = tensor([true, true, true])]; + tensor var_2932_cast_fp16 = slice_by_index(begin = var_2932_begin_0, end = var_2932_end_0, end_mask = var_2932_end_mask_0, x = new_x_47_cast_fp16)[name = string("op_2932_cast_fp16")]; + string x_307_pad_type_0 = const()[name = string("x_307_pad_type_0"), val = string("valid")]; + int32 x_307_groups_0 = const()[name = string("x_307_groups_0"), val = int32(1024)]; + tensor x_307_strides_0 = const()[name = string("x_307_strides_0"), val = tensor([1])]; + tensor x_307_pad_0 = const()[name = string("x_307_pad_0"), val = tensor([0, 0])]; + tensor x_307_dilations_0 = const()[name = string("x_307_dilations_0"), val = tensor([1])]; + tensor encoder_layers_11_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(239473920))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(239483200))))[name = string("encoder_layers_11_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_307_cast_fp16 = conv(dilations = x_307_dilations_0, groups = x_307_groups_0, pad = x_307_pad_0, pad_type = x_307_pad_type_0, strides = x_307_strides_0, weight = encoder_layers_11_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_47_cast_fp16)[name = string("x_307_cast_fp16")]; + tensor input_627_perm_0 = const()[name = string("input_627_perm_0"), val = tensor([0, 2, 1])]; + tensor x_309_axes_0 = const()[name = string("x_309_axes_0"), val = tensor([-1])]; + tensor encoder_layers_11_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_11_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(239485312)))]; + tensor encoder_layers_11_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_11_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(239487424)))]; + tensor input_627_cast_fp16 = transpose(perm = input_627_perm_0, x = x_307_cast_fp16)[name = string("transpose_257")]; + tensor x_309_cast_fp16 = layer_norm(axes = x_309_axes_0, beta = encoder_layers_11_conv_batch_norm_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_11_conv_batch_norm_weight_to_fp16, x = input_627_cast_fp16)[name = string("x_309_cast_fp16")]; + tensor input_629_perm_0 = const()[name = string("input_629_perm_0"), val = tensor([0, 2, 1])]; + tensor input_629_cast_fp16 = transpose(perm = input_629_perm_0, x = x_309_cast_fp16)[name = string("transpose_256")]; + tensor input_631_cast_fp16 = silu(x = input_629_cast_fp16)[name = string("input_631_cast_fp16")]; + string x_311_pad_type_0 = const()[name = string("x_311_pad_type_0"), val = string("valid")]; + tensor x_311_strides_0 = const()[name = string("x_311_strides_0"), val = tensor([1])]; + tensor x_311_pad_0 = const()[name = string("x_311_pad_0"), val = tensor([0, 0])]; + tensor x_311_dilations_0 = const()[name = string("x_311_dilations_0"), val = tensor([1])]; + int32 x_311_groups_0 = const()[name = string("x_311_groups_0"), val = int32(1)]; + tensor encoder_layers_11_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(239489536))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240538176))))[name = string("encoder_layers_11_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_311_cast_fp16 = conv(dilations = x_311_dilations_0, groups = x_311_groups_0, pad = x_311_pad_0, pad_type = x_311_pad_type_0, strides = x_311_strides_0, weight = encoder_layers_11_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_631_cast_fp16)[name = string("x_311_cast_fp16")]; + tensor input_633_perm_0 = const()[name = string("input_633_perm_0"), val = tensor([0, 2, 1])]; + tensor input_633_cast_fp16 = transpose(perm = input_633_perm_0, x = x_311_cast_fp16)[name = string("transpose_255")]; + tensor input_635_cast_fp16 = add(x = input_619_cast_fp16, y = input_633_cast_fp16)[name = string("input_635_cast_fp16")]; + tensor input_637_axes_0 = const()[name = string("input_637_axes_0"), val = tensor([-1])]; + tensor encoder_layers_11_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_11_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240540288)))]; + tensor encoder_layers_11_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_11_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240542400)))]; + tensor input_637_cast_fp16 = layer_norm(axes = input_637_axes_0, beta = encoder_layers_11_norm_feed_forward2_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_11_norm_feed_forward2_weight_to_fp16, x = input_635_cast_fp16)[name = string("input_637_cast_fp16")]; + tensor encoder_layers_11_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240544512))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(243690304))))[name = string("encoder_layers_11_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_11_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_11_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(243690496)))]; + tensor linear_107_cast_fp16 = linear(bias = encoder_layers_11_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_11_feed_forward2_linear1_weight_to_fp16_palettized, x = input_637_cast_fp16)[name = string("linear_107_cast_fp16")]; + tensor input_641_cast_fp16 = silu(x = linear_107_cast_fp16)[name = string("input_641_cast_fp16")]; + tensor encoder_layers_11_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(243698752))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(246844544))))[name = string("encoder_layers_11_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_11_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_11_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(246844736)))]; + tensor linear_108_cast_fp16 = linear(bias = encoder_layers_11_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_11_feed_forward2_linear2_weight_to_fp16_palettized, x = input_641_cast_fp16)[name = string("linear_108_cast_fp16")]; + fp16 var_2975_to_fp16 = const()[name = string("op_2975_to_fp16"), val = fp16(0x1p-1)]; + tensor var_2976_cast_fp16 = mul(x = linear_108_cast_fp16, y = var_2975_to_fp16)[name = string("op_2976_cast_fp16")]; + tensor input_647_cast_fp16 = add(x = input_635_cast_fp16, y = var_2976_cast_fp16)[name = string("input_647_cast_fp16")]; + tensor input_649_axes_0 = const()[name = string("input_649_axes_0"), val = tensor([-1])]; + tensor encoder_layers_11_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_11_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(246846848)))]; + tensor encoder_layers_11_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_11_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(246848960)))]; + tensor input_649_cast_fp16 = layer_norm(axes = input_649_axes_0, beta = encoder_layers_11_norm_out_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_11_norm_out_weight_to_fp16, x = input_647_cast_fp16)[name = string("input_649_cast_fp16")]; + tensor cache_49_begin_0 = const()[name = string("cache_49_begin_0"), val = tensor([12, 0, 0, 0])]; + tensor cache_49_end_0 = const()[name = string("cache_49_end_0"), val = tensor([13, 1, 42, 1024])]; + tensor cache_49_end_mask_0 = const()[name = string("cache_49_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_49_squeeze_mask_0 = const()[name = string("cache_49_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_49_cast_fp16 = slice_by_index(begin = cache_49_begin_0, end = cache_49_end_0, end_mask = cache_49_end_mask_0, squeeze_mask = cache_49_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_49_cast_fp16")]; + tensor cache_51_begin_0 = const()[name = string("cache_51_begin_0"), val = tensor([12, 0, 0, 0])]; + tensor cache_51_end_0 = const()[name = string("cache_51_end_0"), val = tensor([13, 1, 1024, 8])]; + tensor cache_51_end_mask_0 = const()[name = string("cache_51_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_51_squeeze_mask_0 = const()[name = string("cache_51_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_51_cast_fp16 = slice_by_index(begin = cache_51_begin_0, end = cache_51_end_0, end_mask = cache_51_end_mask_0, squeeze_mask = cache_51_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_51_cast_fp16")]; + tensor input_651_axes_0 = const()[name = string("input_651_axes_0"), val = tensor([-1])]; + tensor encoder_layers_12_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_12_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(246851072)))]; + tensor encoder_layers_12_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_12_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(246853184)))]; + tensor input_651_cast_fp16 = layer_norm(axes = input_651_axes_0, beta = encoder_layers_12_norm_feed_forward1_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_12_norm_feed_forward1_weight_to_fp16, x = input_649_cast_fp16)[name = string("input_651_cast_fp16")]; + tensor encoder_layers_12_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(246855296))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(250001088))))[name = string("encoder_layers_12_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_12_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_12_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(250001280)))]; + tensor linear_109_cast_fp16 = linear(bias = encoder_layers_12_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_12_feed_forward1_linear1_weight_to_fp16_palettized, x = input_651_cast_fp16)[name = string("linear_109_cast_fp16")]; + tensor input_655_cast_fp16 = silu(x = linear_109_cast_fp16)[name = string("input_655_cast_fp16")]; + tensor encoder_layers_12_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(250009536))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253155328))))[name = string("encoder_layers_12_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_12_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_12_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253155520)))]; + tensor linear_110_cast_fp16 = linear(bias = encoder_layers_12_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_12_feed_forward1_linear2_weight_to_fp16_palettized, x = input_655_cast_fp16)[name = string("linear_110_cast_fp16")]; + fp16 var_3012_to_fp16 = const()[name = string("op_3012_to_fp16"), val = fp16(0x1p-1)]; + tensor var_3013_cast_fp16 = mul(x = linear_110_cast_fp16, y = var_3012_to_fp16)[name = string("op_3013_cast_fp16")]; + tensor input_661_cast_fp16 = add(x = input_649_cast_fp16, y = var_3013_cast_fp16)[name = string("input_661_cast_fp16")]; + tensor key_25_axes_0 = const()[name = string("key_25_axes_0"), val = tensor([-1])]; + tensor encoder_layers_12_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_12_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253157632)))]; + tensor encoder_layers_12_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_12_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253159744)))]; + tensor key_25_cast_fp16 = layer_norm(axes = key_25_axes_0, beta = encoder_layers_12_norm_self_att_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_12_norm_self_att_weight_to_fp16, x = input_661_cast_fp16)[name = string("key_25_cast_fp16")]; + bool input_663_interleave_0 = const()[name = string("input_663_interleave_0"), val = bool(false)]; + tensor input_663_cast_fp16 = concat(axis = var_69, interleave = input_663_interleave_0, values = (cache_49_cast_fp16, key_25_cast_fp16))[name = string("input_663_cast_fp16")]; + tensor var_3035_begin_0 = const()[name = string("op_3035_begin_0"), val = tensor([0, 14, 0])]; + tensor var_3035_end_0 = const()[name = string("op_3035_end_0"), val = tensor([1, 42, 1024])]; + tensor var_3035_end_mask_0 = const()[name = string("op_3035_end_mask_0"), val = tensor([true, true, true])]; + tensor var_3035_cast_fp16 = slice_by_index(begin = var_3035_begin_0, end = var_3035_end_0, end_mask = var_3035_end_mask_0, x = cache_49_cast_fp16)[name = string("op_3035_cast_fp16")]; + bool var_3041_interleave_0 = const()[name = string("op_3041_interleave_0"), val = bool(false)]; + tensor var_3041_cast_fp16 = concat(axis = var_69, interleave = var_3041_interleave_0, values = (var_3035_cast_fp16, key_25_cast_fp16))[name = string("op_3041_cast_fp16")]; + tensor encoder_layers_12_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253161856))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253948352))))[name = string("encoder_layers_12_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_12_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_12_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253948544)))]; + tensor linear_111_cast_fp16 = linear(bias = encoder_layers_12_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_12_self_attn_linear_q_weight_to_fp16_palettized, x = key_25_cast_fp16)[name = string("linear_111_cast_fp16")]; + tensor var_3046 = const()[name = string("op_3046"), val = tensor([1, -1, 8, 128])]; + tensor q_73_cast_fp16 = reshape(shape = var_3046, x = linear_111_cast_fp16)[name = string("q_73_cast_fp16")]; + tensor encoder_layers_12_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253950656))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(254737152))))[name = string("encoder_layers_12_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_12_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_12_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(254737344)))]; + tensor linear_112_cast_fp16 = linear(bias = encoder_layers_12_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_12_self_attn_linear_k_weight_to_fp16_palettized, x = input_663_cast_fp16)[name = string("linear_112_cast_fp16")]; + tensor var_3051 = const()[name = string("op_3051"), val = tensor([1, -1, 8, 128])]; + tensor k_49_cast_fp16 = reshape(shape = var_3051, x = linear_112_cast_fp16)[name = string("k_49_cast_fp16")]; + tensor encoder_layers_12_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(254739456))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255525952))))[name = string("encoder_layers_12_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_12_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_12_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255526144)))]; + tensor linear_113_cast_fp16 = linear(bias = encoder_layers_12_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_12_self_attn_linear_v_weight_to_fp16_palettized, x = input_663_cast_fp16)[name = string("linear_113_cast_fp16")]; + tensor var_3056 = const()[name = string("op_3056"), val = tensor([1, -1, 8, 128])]; + tensor v_25_cast_fp16 = reshape(shape = var_3056, x = linear_113_cast_fp16)[name = string("v_25_cast_fp16")]; + tensor value_33_perm_0 = const()[name = string("value_33_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_12_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_12_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255528256)))]; + tensor var_3069_cast_fp16 = add(x = q_73_cast_fp16, y = encoder_layers_12_self_attn_pos_bias_u_to_fp16)[name = string("op_3069_cast_fp16")]; + tensor encoder_layers_12_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_12_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255530368)))]; + tensor var_3071_cast_fp16 = add(x = q_73_cast_fp16, y = encoder_layers_12_self_attn_pos_bias_v_to_fp16)[name = string("op_3071_cast_fp16")]; + tensor q_with_bias_v_25_perm_0 = const()[name = string("q_with_bias_v_25_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_319_transpose_x_0 = const()[name = string("x_319_transpose_x_0"), val = bool(false)]; + bool x_319_transpose_y_0 = const()[name = string("x_319_transpose_y_0"), val = bool(false)]; + tensor op_3073_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255532480))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255646208))))[name = string("op_3073_to_fp16_quantized")]; + tensor q_with_bias_v_25_cast_fp16 = transpose(perm = q_with_bias_v_25_perm_0, x = var_3071_cast_fp16)[name = string("transpose_254")]; + tensor x_319_cast_fp16 = matmul(transpose_x = x_319_transpose_x_0, transpose_y = x_319_transpose_y_0, x = q_with_bias_v_25_cast_fp16, y = op_3073_to_fp16_quantized)[name = string("x_319_cast_fp16")]; + tensor x_321_pad_0 = const()[name = string("x_321_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_321_mode_0 = const()[name = string("x_321_mode_0"), val = string("constant")]; + fp16 const_235_to_fp16 = const()[name = string("const_235_to_fp16"), val = fp16(0x0p+0)]; + tensor x_321_cast_fp16 = pad(constant_val = const_235_to_fp16, mode = x_321_mode_0, pad = x_321_pad_0, x = x_319_cast_fp16)[name = string("x_321_cast_fp16")]; + tensor var_3081 = const()[name = string("op_3081"), val = tensor([1, 8, -1, 14])]; + tensor x_323_cast_fp16 = reshape(shape = var_3081, x = x_321_cast_fp16)[name = string("x_323_cast_fp16")]; + tensor var_3085_begin_0 = const()[name = string("op_3085_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_3085_end_0 = const()[name = string("op_3085_end_0"), val = tensor([1, 8, 112, 14])]; + tensor var_3085_end_mask_0 = const()[name = string("op_3085_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_3085_cast_fp16 = slice_by_index(begin = var_3085_begin_0, end = var_3085_end_0, end_mask = var_3085_end_mask_0, x = x_323_cast_fp16)[name = string("op_3085_cast_fp16")]; + tensor var_3086 = const()[name = string("op_3086"), val = tensor([1, 8, 14, 111])]; + tensor matrix_bd_49_cast_fp16 = reshape(shape = var_3086, x = var_3085_cast_fp16)[name = string("matrix_bd_49_cast_fp16")]; + bool matrix_ac_25_transpose_x_0 = const()[name = string("matrix_ac_25_transpose_x_0"), val = bool(false)]; + bool matrix_ac_25_transpose_y_0 = const()[name = string("matrix_ac_25_transpose_y_0"), val = bool(false)]; + tensor transpose_120_perm_0 = const()[name = string("transpose_120_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_121_perm_0 = const()[name = string("transpose_121_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_121 = transpose(perm = transpose_121_perm_0, x = k_49_cast_fp16)[name = string("transpose_252")]; + tensor transpose_120 = transpose(perm = transpose_120_perm_0, x = var_3069_cast_fp16)[name = string("transpose_253")]; + tensor matrix_ac_25_cast_fp16 = matmul(transpose_x = matrix_ac_25_transpose_x_0, transpose_y = matrix_ac_25_transpose_y_0, x = transpose_120, y = transpose_121)[name = string("matrix_ac_25_cast_fp16")]; + tensor matrix_bd_51_begin_0 = const()[name = string("matrix_bd_51_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_51_end_0 = const()[name = string("matrix_bd_51_end_0"), val = tensor([1, 8, 14, 56])]; + tensor matrix_bd_51_end_mask_0 = const()[name = string("matrix_bd_51_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_51_cast_fp16 = slice_by_index(begin = matrix_bd_51_begin_0, end = matrix_bd_51_end_0, end_mask = matrix_bd_51_end_mask_0, x = matrix_bd_49_cast_fp16)[name = string("matrix_bd_51_cast_fp16")]; + tensor var_3095_cast_fp16 = add(x = matrix_ac_25_cast_fp16, y = matrix_bd_51_cast_fp16)[name = string("op_3095_cast_fp16")]; + fp16 _inversed_scores_49_y_0_to_fp16 = const()[name = string("_inversed_scores_49_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_49_cast_fp16 = mul(x = var_3095_cast_fp16, y = _inversed_scores_49_y_0_to_fp16)[name = string("_inversed_scores_49_cast_fp16")]; + tensor scores_51_cast_fp16 = select(a = var_46_to_fp16, b = _inversed_scores_49_cast_fp16, cond = mask_11)[name = string("scores_51_cast_fp16")]; + tensor var_3101_cast_fp16 = softmax(axis = var_60, x = scores_51_cast_fp16)[name = string("op_3101_cast_fp16")]; + tensor input_665_cast_fp16 = select(a = var_45_to_fp16, b = var_3101_cast_fp16, cond = mask_11)[name = string("input_665_cast_fp16")]; + bool x_325_transpose_x_0 = const()[name = string("x_325_transpose_x_0"), val = bool(false)]; + bool x_325_transpose_y_0 = const()[name = string("x_325_transpose_y_0"), val = bool(false)]; + tensor value_33_cast_fp16 = transpose(perm = value_33_perm_0, x = v_25_cast_fp16)[name = string("transpose_251")]; + tensor x_325_cast_fp16 = matmul(transpose_x = x_325_transpose_x_0, transpose_y = x_325_transpose_y_0, x = input_665_cast_fp16, y = value_33_cast_fp16)[name = string("x_325_cast_fp16")]; + tensor var_3105_perm_0 = const()[name = string("op_3105_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_3106 = const()[name = string("op_3106"), val = tensor([1, -1, 1024])]; + tensor var_3105_cast_fp16 = transpose(perm = var_3105_perm_0, x = x_325_cast_fp16)[name = string("transpose_250")]; + tensor input_667_cast_fp16 = reshape(shape = var_3106, x = var_3105_cast_fp16)[name = string("input_667_cast_fp16")]; + tensor encoder_layers_12_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255646528))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256433024))))[name = string("encoder_layers_12_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_12_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_12_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256433216)))]; + tensor linear_115_cast_fp16 = linear(bias = encoder_layers_12_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_12_self_attn_linear_out_weight_to_fp16_palettized, x = input_667_cast_fp16)[name = string("linear_115_cast_fp16")]; + tensor input_671_cast_fp16 = add(x = input_661_cast_fp16, y = linear_115_cast_fp16)[name = string("input_671_cast_fp16")]; + tensor x_329_axes_0 = const()[name = string("x_329_axes_0"), val = tensor([-1])]; + tensor encoder_layers_12_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_12_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256435328)))]; + tensor encoder_layers_12_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_12_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256437440)))]; + tensor x_329_cast_fp16 = layer_norm(axes = x_329_axes_0, beta = encoder_layers_12_norm_conv_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_12_norm_conv_weight_to_fp16, x = input_671_cast_fp16)[name = string("x_329_cast_fp16")]; + tensor input_673_perm_0 = const()[name = string("input_673_perm_0"), val = tensor([0, 2, 1])]; + string input_675_pad_type_0 = const()[name = string("input_675_pad_type_0"), val = string("valid")]; + tensor input_675_strides_0 = const()[name = string("input_675_strides_0"), val = tensor([1])]; + tensor input_675_pad_0 = const()[name = string("input_675_pad_0"), val = tensor([0, 0])]; + tensor input_675_dilations_0 = const()[name = string("input_675_dilations_0"), val = tensor([1])]; + int32 input_675_groups_0 = const()[name = string("input_675_groups_0"), val = int32(1)]; + tensor encoder_layers_12_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256439552))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(258536768))))[name = string("encoder_layers_12_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_673_cast_fp16 = transpose(perm = input_673_perm_0, x = x_329_cast_fp16)[name = string("transpose_249")]; + tensor input_675_cast_fp16 = conv(dilations = input_675_dilations_0, groups = input_675_groups_0, pad = input_675_pad_0, pad_type = input_675_pad_type_0, strides = input_675_strides_0, weight = encoder_layers_12_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_673_cast_fp16)[name = string("input_675_cast_fp16")]; + int32 x_331_split_num_splits_0 = const()[name = string("x_331_split_num_splits_0"), val = int32(2)]; + int32 x_331_split_axis_0 = const()[name = string("x_331_split_axis_0"), val = int32(1)]; + tensor x_331_split_cast_fp16_0, tensor x_331_split_cast_fp16_1 = split(axis = x_331_split_axis_0, num_splits = x_331_split_num_splits_0, x = input_675_cast_fp16)[name = string("x_331_split_cast_fp16")]; + tensor x_331_split_1_sigmoid_cast_fp16 = sigmoid(x = x_331_split_cast_fp16_1)[name = string("x_331_split_1_sigmoid_cast_fp16")]; + tensor x_331_cast_fp16 = mul(x = x_331_split_cast_fp16_0, y = x_331_split_1_sigmoid_cast_fp16)[name = string("x_331_cast_fp16")]; + tensor input_677_cast_fp16 = select(a = var_45_to_fp16, b = x_331_cast_fp16, cond = var_576)[name = string("input_677_cast_fp16")]; + bool new_x_51_interleave_0 = const()[name = string("new_x_51_interleave_0"), val = bool(false)]; + tensor new_x_51_cast_fp16 = concat(axis = var_60, interleave = new_x_51_interleave_0, values = (cache_51_cast_fp16, input_677_cast_fp16))[name = string("new_x_51_cast_fp16")]; + tensor var_3145_begin_0 = const()[name = string("op_3145_begin_0"), val = tensor([0, 0, 14])]; + tensor var_3145_end_0 = const()[name = string("op_3145_end_0"), val = tensor([1, 1024, 22])]; + tensor var_3145_end_mask_0 = const()[name = string("op_3145_end_mask_0"), val = tensor([true, true, true])]; + tensor var_3145_cast_fp16 = slice_by_index(begin = var_3145_begin_0, end = var_3145_end_0, end_mask = var_3145_end_mask_0, x = new_x_51_cast_fp16)[name = string("op_3145_cast_fp16")]; + string x_333_pad_type_0 = const()[name = string("x_333_pad_type_0"), val = string("valid")]; + int32 x_333_groups_0 = const()[name = string("x_333_groups_0"), val = int32(1024)]; + tensor x_333_strides_0 = const()[name = string("x_333_strides_0"), val = tensor([1])]; + tensor x_333_pad_0 = const()[name = string("x_333_pad_0"), val = tensor([0, 0])]; + tensor x_333_dilations_0 = const()[name = string("x_333_dilations_0"), val = tensor([1])]; + tensor encoder_layers_12_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(258540928))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(258550208))))[name = string("encoder_layers_12_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_333_cast_fp16 = conv(dilations = x_333_dilations_0, groups = x_333_groups_0, pad = x_333_pad_0, pad_type = x_333_pad_type_0, strides = x_333_strides_0, weight = encoder_layers_12_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_51_cast_fp16)[name = string("x_333_cast_fp16")]; + tensor input_679_perm_0 = const()[name = string("input_679_perm_0"), val = tensor([0, 2, 1])]; + tensor x_335_axes_0 = const()[name = string("x_335_axes_0"), val = tensor([-1])]; + tensor encoder_layers_12_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_12_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(258552320)))]; + tensor encoder_layers_12_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_12_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(258554432)))]; + tensor input_679_cast_fp16 = transpose(perm = input_679_perm_0, x = x_333_cast_fp16)[name = string("transpose_248")]; + tensor x_335_cast_fp16 = layer_norm(axes = x_335_axes_0, beta = encoder_layers_12_conv_batch_norm_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_12_conv_batch_norm_weight_to_fp16, x = input_679_cast_fp16)[name = string("x_335_cast_fp16")]; + tensor input_681_perm_0 = const()[name = string("input_681_perm_0"), val = tensor([0, 2, 1])]; + tensor input_681_cast_fp16 = transpose(perm = input_681_perm_0, x = x_335_cast_fp16)[name = string("transpose_247")]; + tensor input_683_cast_fp16 = silu(x = input_681_cast_fp16)[name = string("input_683_cast_fp16")]; + string x_337_pad_type_0 = const()[name = string("x_337_pad_type_0"), val = string("valid")]; + tensor x_337_strides_0 = const()[name = string("x_337_strides_0"), val = tensor([1])]; + tensor x_337_pad_0 = const()[name = string("x_337_pad_0"), val = tensor([0, 0])]; + tensor x_337_dilations_0 = const()[name = string("x_337_dilations_0"), val = tensor([1])]; + int32 x_337_groups_0 = const()[name = string("x_337_groups_0"), val = int32(1)]; + tensor encoder_layers_12_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(258556544))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259605184))))[name = string("encoder_layers_12_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_337_cast_fp16 = conv(dilations = x_337_dilations_0, groups = x_337_groups_0, pad = x_337_pad_0, pad_type = x_337_pad_type_0, strides = x_337_strides_0, weight = encoder_layers_12_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_683_cast_fp16)[name = string("x_337_cast_fp16")]; + tensor input_685_perm_0 = const()[name = string("input_685_perm_0"), val = tensor([0, 2, 1])]; + tensor input_685_cast_fp16 = transpose(perm = input_685_perm_0, x = x_337_cast_fp16)[name = string("transpose_246")]; + tensor input_687_cast_fp16 = add(x = input_671_cast_fp16, y = input_685_cast_fp16)[name = string("input_687_cast_fp16")]; + tensor input_689_axes_0 = const()[name = string("input_689_axes_0"), val = tensor([-1])]; + tensor encoder_layers_12_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_12_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259607296)))]; + tensor encoder_layers_12_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_12_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259609408)))]; + tensor input_689_cast_fp16 = layer_norm(axes = input_689_axes_0, beta = encoder_layers_12_norm_feed_forward2_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_12_norm_feed_forward2_weight_to_fp16, x = input_687_cast_fp16)[name = string("input_689_cast_fp16")]; + tensor encoder_layers_12_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259611520))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(262757312))))[name = string("encoder_layers_12_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_12_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_12_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(262757504)))]; + tensor linear_116_cast_fp16 = linear(bias = encoder_layers_12_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_12_feed_forward2_linear1_weight_to_fp16_palettized, x = input_689_cast_fp16)[name = string("linear_116_cast_fp16")]; + tensor input_693_cast_fp16 = silu(x = linear_116_cast_fp16)[name = string("input_693_cast_fp16")]; + tensor encoder_layers_12_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(262765760))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(265911552))))[name = string("encoder_layers_12_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_12_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_12_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(265911744)))]; + tensor linear_117_cast_fp16 = linear(bias = encoder_layers_12_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_12_feed_forward2_linear2_weight_to_fp16_palettized, x = input_693_cast_fp16)[name = string("linear_117_cast_fp16")]; + fp16 var_3188_to_fp16 = const()[name = string("op_3188_to_fp16"), val = fp16(0x1p-1)]; + tensor var_3189_cast_fp16 = mul(x = linear_117_cast_fp16, y = var_3188_to_fp16)[name = string("op_3189_cast_fp16")]; + tensor input_699_cast_fp16 = add(x = input_687_cast_fp16, y = var_3189_cast_fp16)[name = string("input_699_cast_fp16")]; + tensor input_701_axes_0 = const()[name = string("input_701_axes_0"), val = tensor([-1])]; + tensor encoder_layers_12_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_12_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(265913856)))]; + tensor encoder_layers_12_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_12_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(265915968)))]; + tensor input_701_cast_fp16 = layer_norm(axes = input_701_axes_0, beta = encoder_layers_12_norm_out_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_12_norm_out_weight_to_fp16, x = input_699_cast_fp16)[name = string("input_701_cast_fp16")]; + tensor cache_53_begin_0 = const()[name = string("cache_53_begin_0"), val = tensor([13, 0, 0, 0])]; + tensor cache_53_end_0 = const()[name = string("cache_53_end_0"), val = tensor([14, 1, 42, 1024])]; + tensor cache_53_end_mask_0 = const()[name = string("cache_53_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_53_squeeze_mask_0 = const()[name = string("cache_53_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_53_cast_fp16 = slice_by_index(begin = cache_53_begin_0, end = cache_53_end_0, end_mask = cache_53_end_mask_0, squeeze_mask = cache_53_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_53_cast_fp16")]; + tensor cache_55_begin_0 = const()[name = string("cache_55_begin_0"), val = tensor([13, 0, 0, 0])]; + tensor cache_55_end_0 = const()[name = string("cache_55_end_0"), val = tensor([14, 1, 1024, 8])]; + tensor cache_55_end_mask_0 = const()[name = string("cache_55_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_55_squeeze_mask_0 = const()[name = string("cache_55_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_55_cast_fp16 = slice_by_index(begin = cache_55_begin_0, end = cache_55_end_0, end_mask = cache_55_end_mask_0, squeeze_mask = cache_55_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_55_cast_fp16")]; + tensor input_703_axes_0 = const()[name = string("input_703_axes_0"), val = tensor([-1])]; + tensor encoder_layers_13_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_13_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(265918080)))]; + tensor encoder_layers_13_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_13_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(265920192)))]; + tensor input_703_cast_fp16 = layer_norm(axes = input_703_axes_0, beta = encoder_layers_13_norm_feed_forward1_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_13_norm_feed_forward1_weight_to_fp16, x = input_701_cast_fp16)[name = string("input_703_cast_fp16")]; + tensor encoder_layers_13_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(265922304))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(269068096))))[name = string("encoder_layers_13_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_13_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_13_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(269068288)))]; + tensor linear_118_cast_fp16 = linear(bias = encoder_layers_13_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_13_feed_forward1_linear1_weight_to_fp16_palettized, x = input_703_cast_fp16)[name = string("linear_118_cast_fp16")]; + tensor input_707_cast_fp16 = silu(x = linear_118_cast_fp16)[name = string("input_707_cast_fp16")]; + tensor encoder_layers_13_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(269076544))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272222336))))[name = string("encoder_layers_13_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_13_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_13_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272222528)))]; + tensor linear_119_cast_fp16 = linear(bias = encoder_layers_13_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_13_feed_forward1_linear2_weight_to_fp16_palettized, x = input_707_cast_fp16)[name = string("linear_119_cast_fp16")]; + fp16 var_3225_to_fp16 = const()[name = string("op_3225_to_fp16"), val = fp16(0x1p-1)]; + tensor var_3226_cast_fp16 = mul(x = linear_119_cast_fp16, y = var_3225_to_fp16)[name = string("op_3226_cast_fp16")]; + tensor input_713_cast_fp16 = add(x = input_701_cast_fp16, y = var_3226_cast_fp16)[name = string("input_713_cast_fp16")]; + tensor key_27_axes_0 = const()[name = string("key_27_axes_0"), val = tensor([-1])]; + tensor encoder_layers_13_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_13_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272224640)))]; + tensor encoder_layers_13_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_13_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272226752)))]; + tensor key_27_cast_fp16 = layer_norm(axes = key_27_axes_0, beta = encoder_layers_13_norm_self_att_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_13_norm_self_att_weight_to_fp16, x = input_713_cast_fp16)[name = string("key_27_cast_fp16")]; + bool input_715_interleave_0 = const()[name = string("input_715_interleave_0"), val = bool(false)]; + tensor input_715_cast_fp16 = concat(axis = var_69, interleave = input_715_interleave_0, values = (cache_53_cast_fp16, key_27_cast_fp16))[name = string("input_715_cast_fp16")]; + tensor var_3248_begin_0 = const()[name = string("op_3248_begin_0"), val = tensor([0, 14, 0])]; + tensor var_3248_end_0 = const()[name = string("op_3248_end_0"), val = tensor([1, 42, 1024])]; + tensor var_3248_end_mask_0 = const()[name = string("op_3248_end_mask_0"), val = tensor([true, true, true])]; + tensor var_3248_cast_fp16 = slice_by_index(begin = var_3248_begin_0, end = var_3248_end_0, end_mask = var_3248_end_mask_0, x = cache_53_cast_fp16)[name = string("op_3248_cast_fp16")]; + bool var_3254_interleave_0 = const()[name = string("op_3254_interleave_0"), val = bool(false)]; + tensor var_3254_cast_fp16 = concat(axis = var_69, interleave = var_3254_interleave_0, values = (var_3248_cast_fp16, key_27_cast_fp16))[name = string("op_3254_cast_fp16")]; + tensor encoder_layers_13_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272228864))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(273015360))))[name = string("encoder_layers_13_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_13_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_13_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(273015552)))]; + tensor linear_120_cast_fp16 = linear(bias = encoder_layers_13_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_13_self_attn_linear_q_weight_to_fp16_palettized, x = key_27_cast_fp16)[name = string("linear_120_cast_fp16")]; + tensor var_3259 = const()[name = string("op_3259"), val = tensor([1, -1, 8, 128])]; + tensor q_79_cast_fp16 = reshape(shape = var_3259, x = linear_120_cast_fp16)[name = string("q_79_cast_fp16")]; + tensor encoder_layers_13_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(273017664))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(273804160))))[name = string("encoder_layers_13_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_13_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_13_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(273804352)))]; + tensor linear_121_cast_fp16 = linear(bias = encoder_layers_13_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_13_self_attn_linear_k_weight_to_fp16_palettized, x = input_715_cast_fp16)[name = string("linear_121_cast_fp16")]; + tensor var_3264 = const()[name = string("op_3264"), val = tensor([1, -1, 8, 128])]; + tensor k_53_cast_fp16 = reshape(shape = var_3264, x = linear_121_cast_fp16)[name = string("k_53_cast_fp16")]; + tensor encoder_layers_13_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(273806464))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274592960))))[name = string("encoder_layers_13_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_13_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_13_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274593152)))]; + tensor linear_122_cast_fp16 = linear(bias = encoder_layers_13_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_13_self_attn_linear_v_weight_to_fp16_palettized, x = input_715_cast_fp16)[name = string("linear_122_cast_fp16")]; + tensor var_3269 = const()[name = string("op_3269"), val = tensor([1, -1, 8, 128])]; + tensor v_27_cast_fp16 = reshape(shape = var_3269, x = linear_122_cast_fp16)[name = string("v_27_cast_fp16")]; + tensor value_35_perm_0 = const()[name = string("value_35_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_13_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_13_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274595264)))]; + tensor var_3282_cast_fp16 = add(x = q_79_cast_fp16, y = encoder_layers_13_self_attn_pos_bias_u_to_fp16)[name = string("op_3282_cast_fp16")]; + tensor encoder_layers_13_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_13_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274597376)))]; + tensor var_3284_cast_fp16 = add(x = q_79_cast_fp16, y = encoder_layers_13_self_attn_pos_bias_v_to_fp16)[name = string("op_3284_cast_fp16")]; + tensor q_with_bias_v_27_perm_0 = const()[name = string("q_with_bias_v_27_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_345_transpose_x_0 = const()[name = string("x_345_transpose_x_0"), val = bool(false)]; + bool x_345_transpose_y_0 = const()[name = string("x_345_transpose_y_0"), val = bool(false)]; + tensor op_3286_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274599488))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274713216))))[name = string("op_3286_to_fp16_quantized")]; + tensor q_with_bias_v_27_cast_fp16 = transpose(perm = q_with_bias_v_27_perm_0, x = var_3284_cast_fp16)[name = string("transpose_245")]; + tensor x_345_cast_fp16 = matmul(transpose_x = x_345_transpose_x_0, transpose_y = x_345_transpose_y_0, x = q_with_bias_v_27_cast_fp16, y = op_3286_to_fp16_quantized)[name = string("x_345_cast_fp16")]; + tensor x_347_pad_0 = const()[name = string("x_347_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_347_mode_0 = const()[name = string("x_347_mode_0"), val = string("constant")]; + fp16 const_248_to_fp16 = const()[name = string("const_248_to_fp16"), val = fp16(0x0p+0)]; + tensor x_347_cast_fp16 = pad(constant_val = const_248_to_fp16, mode = x_347_mode_0, pad = x_347_pad_0, x = x_345_cast_fp16)[name = string("x_347_cast_fp16")]; + tensor var_3294 = const()[name = string("op_3294"), val = tensor([1, 8, -1, 14])]; + tensor x_349_cast_fp16 = reshape(shape = var_3294, x = x_347_cast_fp16)[name = string("x_349_cast_fp16")]; + tensor var_3298_begin_0 = const()[name = string("op_3298_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_3298_end_0 = const()[name = string("op_3298_end_0"), val = tensor([1, 8, 112, 14])]; + tensor var_3298_end_mask_0 = const()[name = string("op_3298_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_3298_cast_fp16 = slice_by_index(begin = var_3298_begin_0, end = var_3298_end_0, end_mask = var_3298_end_mask_0, x = x_349_cast_fp16)[name = string("op_3298_cast_fp16")]; + tensor var_3299 = const()[name = string("op_3299"), val = tensor([1, 8, 14, 111])]; + tensor matrix_bd_53_cast_fp16 = reshape(shape = var_3299, x = var_3298_cast_fp16)[name = string("matrix_bd_53_cast_fp16")]; + bool matrix_ac_27_transpose_x_0 = const()[name = string("matrix_ac_27_transpose_x_0"), val = bool(false)]; + bool matrix_ac_27_transpose_y_0 = const()[name = string("matrix_ac_27_transpose_y_0"), val = bool(false)]; + tensor transpose_122_perm_0 = const()[name = string("transpose_122_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_123_perm_0 = const()[name = string("transpose_123_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_123 = transpose(perm = transpose_123_perm_0, x = k_53_cast_fp16)[name = string("transpose_243")]; + tensor transpose_122 = transpose(perm = transpose_122_perm_0, x = var_3282_cast_fp16)[name = string("transpose_244")]; + tensor matrix_ac_27_cast_fp16 = matmul(transpose_x = matrix_ac_27_transpose_x_0, transpose_y = matrix_ac_27_transpose_y_0, x = transpose_122, y = transpose_123)[name = string("matrix_ac_27_cast_fp16")]; + tensor matrix_bd_55_begin_0 = const()[name = string("matrix_bd_55_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_55_end_0 = const()[name = string("matrix_bd_55_end_0"), val = tensor([1, 8, 14, 56])]; + tensor matrix_bd_55_end_mask_0 = const()[name = string("matrix_bd_55_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_55_cast_fp16 = slice_by_index(begin = matrix_bd_55_begin_0, end = matrix_bd_55_end_0, end_mask = matrix_bd_55_end_mask_0, x = matrix_bd_53_cast_fp16)[name = string("matrix_bd_55_cast_fp16")]; + tensor var_3308_cast_fp16 = add(x = matrix_ac_27_cast_fp16, y = matrix_bd_55_cast_fp16)[name = string("op_3308_cast_fp16")]; + fp16 _inversed_scores_53_y_0_to_fp16 = const()[name = string("_inversed_scores_53_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_53_cast_fp16 = mul(x = var_3308_cast_fp16, y = _inversed_scores_53_y_0_to_fp16)[name = string("_inversed_scores_53_cast_fp16")]; + tensor scores_55_cast_fp16 = select(a = var_46_to_fp16, b = _inversed_scores_53_cast_fp16, cond = mask_11)[name = string("scores_55_cast_fp16")]; + tensor var_3314_cast_fp16 = softmax(axis = var_60, x = scores_55_cast_fp16)[name = string("op_3314_cast_fp16")]; + tensor input_717_cast_fp16 = select(a = var_45_to_fp16, b = var_3314_cast_fp16, cond = mask_11)[name = string("input_717_cast_fp16")]; + bool x_351_transpose_x_0 = const()[name = string("x_351_transpose_x_0"), val = bool(false)]; + bool x_351_transpose_y_0 = const()[name = string("x_351_transpose_y_0"), val = bool(false)]; + tensor value_35_cast_fp16 = transpose(perm = value_35_perm_0, x = v_27_cast_fp16)[name = string("transpose_242")]; + tensor x_351_cast_fp16 = matmul(transpose_x = x_351_transpose_x_0, transpose_y = x_351_transpose_y_0, x = input_717_cast_fp16, y = value_35_cast_fp16)[name = string("x_351_cast_fp16")]; + tensor var_3318_perm_0 = const()[name = string("op_3318_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_3319 = const()[name = string("op_3319"), val = tensor([1, -1, 1024])]; + tensor var_3318_cast_fp16 = transpose(perm = var_3318_perm_0, x = x_351_cast_fp16)[name = string("transpose_241")]; + tensor input_719_cast_fp16 = reshape(shape = var_3319, x = var_3318_cast_fp16)[name = string("input_719_cast_fp16")]; + tensor encoder_layers_13_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274713536))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275500032))))[name = string("encoder_layers_13_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_13_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_13_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275500224)))]; + tensor linear_124_cast_fp16 = linear(bias = encoder_layers_13_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_13_self_attn_linear_out_weight_to_fp16_palettized, x = input_719_cast_fp16)[name = string("linear_124_cast_fp16")]; + tensor input_723_cast_fp16 = add(x = input_713_cast_fp16, y = linear_124_cast_fp16)[name = string("input_723_cast_fp16")]; + tensor x_355_axes_0 = const()[name = string("x_355_axes_0"), val = tensor([-1])]; + tensor encoder_layers_13_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_13_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275502336)))]; + tensor encoder_layers_13_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_13_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275504448)))]; + tensor x_355_cast_fp16 = layer_norm(axes = x_355_axes_0, beta = encoder_layers_13_norm_conv_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_13_norm_conv_weight_to_fp16, x = input_723_cast_fp16)[name = string("x_355_cast_fp16")]; + tensor input_725_perm_0 = const()[name = string("input_725_perm_0"), val = tensor([0, 2, 1])]; + string input_727_pad_type_0 = const()[name = string("input_727_pad_type_0"), val = string("valid")]; + tensor input_727_strides_0 = const()[name = string("input_727_strides_0"), val = tensor([1])]; + tensor input_727_pad_0 = const()[name = string("input_727_pad_0"), val = tensor([0, 0])]; + tensor input_727_dilations_0 = const()[name = string("input_727_dilations_0"), val = tensor([1])]; + int32 input_727_groups_0 = const()[name = string("input_727_groups_0"), val = int32(1)]; + tensor encoder_layers_13_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275506560))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(277603776))))[name = string("encoder_layers_13_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_725_cast_fp16 = transpose(perm = input_725_perm_0, x = x_355_cast_fp16)[name = string("transpose_240")]; + tensor input_727_cast_fp16 = conv(dilations = input_727_dilations_0, groups = input_727_groups_0, pad = input_727_pad_0, pad_type = input_727_pad_type_0, strides = input_727_strides_0, weight = encoder_layers_13_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_725_cast_fp16)[name = string("input_727_cast_fp16")]; + int32 x_357_split_num_splits_0 = const()[name = string("x_357_split_num_splits_0"), val = int32(2)]; + int32 x_357_split_axis_0 = const()[name = string("x_357_split_axis_0"), val = int32(1)]; + tensor x_357_split_cast_fp16_0, tensor x_357_split_cast_fp16_1 = split(axis = x_357_split_axis_0, num_splits = x_357_split_num_splits_0, x = input_727_cast_fp16)[name = string("x_357_split_cast_fp16")]; + tensor x_357_split_1_sigmoid_cast_fp16 = sigmoid(x = x_357_split_cast_fp16_1)[name = string("x_357_split_1_sigmoid_cast_fp16")]; + tensor x_357_cast_fp16 = mul(x = x_357_split_cast_fp16_0, y = x_357_split_1_sigmoid_cast_fp16)[name = string("x_357_cast_fp16")]; + tensor input_729_cast_fp16 = select(a = var_45_to_fp16, b = x_357_cast_fp16, cond = var_576)[name = string("input_729_cast_fp16")]; + bool new_x_55_interleave_0 = const()[name = string("new_x_55_interleave_0"), val = bool(false)]; + tensor new_x_55_cast_fp16 = concat(axis = var_60, interleave = new_x_55_interleave_0, values = (cache_55_cast_fp16, input_729_cast_fp16))[name = string("new_x_55_cast_fp16")]; + tensor var_3358_begin_0 = const()[name = string("op_3358_begin_0"), val = tensor([0, 0, 14])]; + tensor var_3358_end_0 = const()[name = string("op_3358_end_0"), val = tensor([1, 1024, 22])]; + tensor var_3358_end_mask_0 = const()[name = string("op_3358_end_mask_0"), val = tensor([true, true, true])]; + tensor var_3358_cast_fp16 = slice_by_index(begin = var_3358_begin_0, end = var_3358_end_0, end_mask = var_3358_end_mask_0, x = new_x_55_cast_fp16)[name = string("op_3358_cast_fp16")]; + string x_359_pad_type_0 = const()[name = string("x_359_pad_type_0"), val = string("valid")]; + int32 x_359_groups_0 = const()[name = string("x_359_groups_0"), val = int32(1024)]; + tensor x_359_strides_0 = const()[name = string("x_359_strides_0"), val = tensor([1])]; + tensor x_359_pad_0 = const()[name = string("x_359_pad_0"), val = tensor([0, 0])]; + tensor x_359_dilations_0 = const()[name = string("x_359_dilations_0"), val = tensor([1])]; + tensor encoder_layers_13_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(277607936))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(277617216))))[name = string("encoder_layers_13_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_359_cast_fp16 = conv(dilations = x_359_dilations_0, groups = x_359_groups_0, pad = x_359_pad_0, pad_type = x_359_pad_type_0, strides = x_359_strides_0, weight = encoder_layers_13_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_55_cast_fp16)[name = string("x_359_cast_fp16")]; + tensor input_731_perm_0 = const()[name = string("input_731_perm_0"), val = tensor([0, 2, 1])]; + tensor x_361_axes_0 = const()[name = string("x_361_axes_0"), val = tensor([-1])]; + tensor encoder_layers_13_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_13_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(277619328)))]; + tensor encoder_layers_13_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_13_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(277621440)))]; + tensor input_731_cast_fp16 = transpose(perm = input_731_perm_0, x = x_359_cast_fp16)[name = string("transpose_239")]; + tensor x_361_cast_fp16 = layer_norm(axes = x_361_axes_0, beta = encoder_layers_13_conv_batch_norm_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_13_conv_batch_norm_weight_to_fp16, x = input_731_cast_fp16)[name = string("x_361_cast_fp16")]; + tensor input_733_perm_0 = const()[name = string("input_733_perm_0"), val = tensor([0, 2, 1])]; + tensor input_733_cast_fp16 = transpose(perm = input_733_perm_0, x = x_361_cast_fp16)[name = string("transpose_238")]; + tensor input_735_cast_fp16 = silu(x = input_733_cast_fp16)[name = string("input_735_cast_fp16")]; + string x_363_pad_type_0 = const()[name = string("x_363_pad_type_0"), val = string("valid")]; + tensor x_363_strides_0 = const()[name = string("x_363_strides_0"), val = tensor([1])]; + tensor x_363_pad_0 = const()[name = string("x_363_pad_0"), val = tensor([0, 0])]; + tensor x_363_dilations_0 = const()[name = string("x_363_dilations_0"), val = tensor([1])]; + int32 x_363_groups_0 = const()[name = string("x_363_groups_0"), val = int32(1)]; + tensor encoder_layers_13_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(277623552))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278672192))))[name = string("encoder_layers_13_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_363_cast_fp16 = conv(dilations = x_363_dilations_0, groups = x_363_groups_0, pad = x_363_pad_0, pad_type = x_363_pad_type_0, strides = x_363_strides_0, weight = encoder_layers_13_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_735_cast_fp16)[name = string("x_363_cast_fp16")]; + tensor input_737_perm_0 = const()[name = string("input_737_perm_0"), val = tensor([0, 2, 1])]; + tensor input_737_cast_fp16 = transpose(perm = input_737_perm_0, x = x_363_cast_fp16)[name = string("transpose_237")]; + tensor input_739_cast_fp16 = add(x = input_723_cast_fp16, y = input_737_cast_fp16)[name = string("input_739_cast_fp16")]; + tensor input_741_axes_0 = const()[name = string("input_741_axes_0"), val = tensor([-1])]; + tensor encoder_layers_13_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_13_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278674304)))]; + tensor encoder_layers_13_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_13_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278676416)))]; + tensor input_741_cast_fp16 = layer_norm(axes = input_741_axes_0, beta = encoder_layers_13_norm_feed_forward2_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_13_norm_feed_forward2_weight_to_fp16, x = input_739_cast_fp16)[name = string("input_741_cast_fp16")]; + tensor encoder_layers_13_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278678528))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(281824320))))[name = string("encoder_layers_13_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_13_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_13_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(281824512)))]; + tensor linear_125_cast_fp16 = linear(bias = encoder_layers_13_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_13_feed_forward2_linear1_weight_to_fp16_palettized, x = input_741_cast_fp16)[name = string("linear_125_cast_fp16")]; + tensor input_745_cast_fp16 = silu(x = linear_125_cast_fp16)[name = string("input_745_cast_fp16")]; + tensor encoder_layers_13_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(281832768))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(284978560))))[name = string("encoder_layers_13_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_13_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_13_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(284978752)))]; + tensor linear_126_cast_fp16 = linear(bias = encoder_layers_13_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_13_feed_forward2_linear2_weight_to_fp16_palettized, x = input_745_cast_fp16)[name = string("linear_126_cast_fp16")]; + fp16 var_3401_to_fp16 = const()[name = string("op_3401_to_fp16"), val = fp16(0x1p-1)]; + tensor var_3402_cast_fp16 = mul(x = linear_126_cast_fp16, y = var_3401_to_fp16)[name = string("op_3402_cast_fp16")]; + tensor input_751_cast_fp16 = add(x = input_739_cast_fp16, y = var_3402_cast_fp16)[name = string("input_751_cast_fp16")]; + tensor input_753_axes_0 = const()[name = string("input_753_axes_0"), val = tensor([-1])]; + tensor encoder_layers_13_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_13_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(284980864)))]; + tensor encoder_layers_13_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_13_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(284982976)))]; + tensor input_753_cast_fp16 = layer_norm(axes = input_753_axes_0, beta = encoder_layers_13_norm_out_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_13_norm_out_weight_to_fp16, x = input_751_cast_fp16)[name = string("input_753_cast_fp16")]; + tensor cache_57_begin_0 = const()[name = string("cache_57_begin_0"), val = tensor([14, 0, 0, 0])]; + tensor cache_57_end_0 = const()[name = string("cache_57_end_0"), val = tensor([15, 1, 42, 1024])]; + tensor cache_57_end_mask_0 = const()[name = string("cache_57_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_57_squeeze_mask_0 = const()[name = string("cache_57_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_57_cast_fp16 = slice_by_index(begin = cache_57_begin_0, end = cache_57_end_0, end_mask = cache_57_end_mask_0, squeeze_mask = cache_57_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_57_cast_fp16")]; + tensor cache_59_begin_0 = const()[name = string("cache_59_begin_0"), val = tensor([14, 0, 0, 0])]; + tensor cache_59_end_0 = const()[name = string("cache_59_end_0"), val = tensor([15, 1, 1024, 8])]; + tensor cache_59_end_mask_0 = const()[name = string("cache_59_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_59_squeeze_mask_0 = const()[name = string("cache_59_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_59_cast_fp16 = slice_by_index(begin = cache_59_begin_0, end = cache_59_end_0, end_mask = cache_59_end_mask_0, squeeze_mask = cache_59_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_59_cast_fp16")]; + tensor input_755_axes_0 = const()[name = string("input_755_axes_0"), val = tensor([-1])]; + tensor encoder_layers_14_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_14_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(284985088)))]; + tensor encoder_layers_14_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_14_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(284987200)))]; + tensor input_755_cast_fp16 = layer_norm(axes = input_755_axes_0, beta = encoder_layers_14_norm_feed_forward1_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_14_norm_feed_forward1_weight_to_fp16, x = input_753_cast_fp16)[name = string("input_755_cast_fp16")]; + tensor encoder_layers_14_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(284989312))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(288135104))))[name = string("encoder_layers_14_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_14_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_14_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(288135296)))]; + tensor linear_127_cast_fp16 = linear(bias = encoder_layers_14_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_14_feed_forward1_linear1_weight_to_fp16_palettized, x = input_755_cast_fp16)[name = string("linear_127_cast_fp16")]; + tensor input_759_cast_fp16 = silu(x = linear_127_cast_fp16)[name = string("input_759_cast_fp16")]; + tensor encoder_layers_14_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(288143552))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291289344))))[name = string("encoder_layers_14_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_14_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_14_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291289536)))]; + tensor linear_128_cast_fp16 = linear(bias = encoder_layers_14_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_14_feed_forward1_linear2_weight_to_fp16_palettized, x = input_759_cast_fp16)[name = string("linear_128_cast_fp16")]; + fp16 var_3438_to_fp16 = const()[name = string("op_3438_to_fp16"), val = fp16(0x1p-1)]; + tensor var_3439_cast_fp16 = mul(x = linear_128_cast_fp16, y = var_3438_to_fp16)[name = string("op_3439_cast_fp16")]; + tensor input_765_cast_fp16 = add(x = input_753_cast_fp16, y = var_3439_cast_fp16)[name = string("input_765_cast_fp16")]; + tensor key_29_axes_0 = const()[name = string("key_29_axes_0"), val = tensor([-1])]; + tensor encoder_layers_14_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_14_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291291648)))]; + tensor encoder_layers_14_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_14_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291293760)))]; + tensor key_29_cast_fp16 = layer_norm(axes = key_29_axes_0, beta = encoder_layers_14_norm_self_att_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_14_norm_self_att_weight_to_fp16, x = input_765_cast_fp16)[name = string("key_29_cast_fp16")]; + bool input_767_interleave_0 = const()[name = string("input_767_interleave_0"), val = bool(false)]; + tensor input_767_cast_fp16 = concat(axis = var_69, interleave = input_767_interleave_0, values = (cache_57_cast_fp16, key_29_cast_fp16))[name = string("input_767_cast_fp16")]; + tensor var_3461_begin_0 = const()[name = string("op_3461_begin_0"), val = tensor([0, 14, 0])]; + tensor var_3461_end_0 = const()[name = string("op_3461_end_0"), val = tensor([1, 42, 1024])]; + tensor var_3461_end_mask_0 = const()[name = string("op_3461_end_mask_0"), val = tensor([true, true, true])]; + tensor var_3461_cast_fp16 = slice_by_index(begin = var_3461_begin_0, end = var_3461_end_0, end_mask = var_3461_end_mask_0, x = cache_57_cast_fp16)[name = string("op_3461_cast_fp16")]; + bool var_3467_interleave_0 = const()[name = string("op_3467_interleave_0"), val = bool(false)]; + tensor var_3467_cast_fp16 = concat(axis = var_69, interleave = var_3467_interleave_0, values = (var_3461_cast_fp16, key_29_cast_fp16))[name = string("op_3467_cast_fp16")]; + tensor encoder_layers_14_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291295872))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(292082368))))[name = string("encoder_layers_14_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_14_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_14_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(292082560)))]; + tensor linear_129_cast_fp16 = linear(bias = encoder_layers_14_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_14_self_attn_linear_q_weight_to_fp16_palettized, x = key_29_cast_fp16)[name = string("linear_129_cast_fp16")]; + tensor var_3472 = const()[name = string("op_3472"), val = tensor([1, -1, 8, 128])]; + tensor q_85_cast_fp16 = reshape(shape = var_3472, x = linear_129_cast_fp16)[name = string("q_85_cast_fp16")]; + tensor encoder_layers_14_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(292084672))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(292871168))))[name = string("encoder_layers_14_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_14_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_14_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(292871360)))]; + tensor linear_130_cast_fp16 = linear(bias = encoder_layers_14_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_14_self_attn_linear_k_weight_to_fp16_palettized, x = input_767_cast_fp16)[name = string("linear_130_cast_fp16")]; + tensor var_3477 = const()[name = string("op_3477"), val = tensor([1, -1, 8, 128])]; + tensor k_57_cast_fp16 = reshape(shape = var_3477, x = linear_130_cast_fp16)[name = string("k_57_cast_fp16")]; + tensor encoder_layers_14_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(292873472))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293659968))))[name = string("encoder_layers_14_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_14_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_14_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293660160)))]; + tensor linear_131_cast_fp16 = linear(bias = encoder_layers_14_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_14_self_attn_linear_v_weight_to_fp16_palettized, x = input_767_cast_fp16)[name = string("linear_131_cast_fp16")]; + tensor var_3482 = const()[name = string("op_3482"), val = tensor([1, -1, 8, 128])]; + tensor v_29_cast_fp16 = reshape(shape = var_3482, x = linear_131_cast_fp16)[name = string("v_29_cast_fp16")]; + tensor value_37_perm_0 = const()[name = string("value_37_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_14_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_14_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293662272)))]; + tensor var_3495_cast_fp16 = add(x = q_85_cast_fp16, y = encoder_layers_14_self_attn_pos_bias_u_to_fp16)[name = string("op_3495_cast_fp16")]; + tensor encoder_layers_14_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_14_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293664384)))]; + tensor var_3497_cast_fp16 = add(x = q_85_cast_fp16, y = encoder_layers_14_self_attn_pos_bias_v_to_fp16)[name = string("op_3497_cast_fp16")]; + tensor q_with_bias_v_29_perm_0 = const()[name = string("q_with_bias_v_29_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_371_transpose_x_0 = const()[name = string("x_371_transpose_x_0"), val = bool(false)]; + bool x_371_transpose_y_0 = const()[name = string("x_371_transpose_y_0"), val = bool(false)]; + tensor op_3499_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293666496))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293780224))))[name = string("op_3499_to_fp16_quantized")]; + tensor q_with_bias_v_29_cast_fp16 = transpose(perm = q_with_bias_v_29_perm_0, x = var_3497_cast_fp16)[name = string("transpose_236")]; + tensor x_371_cast_fp16 = matmul(transpose_x = x_371_transpose_x_0, transpose_y = x_371_transpose_y_0, x = q_with_bias_v_29_cast_fp16, y = op_3499_to_fp16_quantized)[name = string("x_371_cast_fp16")]; + tensor x_373_pad_0 = const()[name = string("x_373_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_373_mode_0 = const()[name = string("x_373_mode_0"), val = string("constant")]; + fp16 const_261_to_fp16 = const()[name = string("const_261_to_fp16"), val = fp16(0x0p+0)]; + tensor x_373_cast_fp16 = pad(constant_val = const_261_to_fp16, mode = x_373_mode_0, pad = x_373_pad_0, x = x_371_cast_fp16)[name = string("x_373_cast_fp16")]; + tensor var_3507 = const()[name = string("op_3507"), val = tensor([1, 8, -1, 14])]; + tensor x_375_cast_fp16 = reshape(shape = var_3507, x = x_373_cast_fp16)[name = string("x_375_cast_fp16")]; + tensor var_3511_begin_0 = const()[name = string("op_3511_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_3511_end_0 = const()[name = string("op_3511_end_0"), val = tensor([1, 8, 112, 14])]; + tensor var_3511_end_mask_0 = const()[name = string("op_3511_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_3511_cast_fp16 = slice_by_index(begin = var_3511_begin_0, end = var_3511_end_0, end_mask = var_3511_end_mask_0, x = x_375_cast_fp16)[name = string("op_3511_cast_fp16")]; + tensor var_3512 = const()[name = string("op_3512"), val = tensor([1, 8, 14, 111])]; + tensor matrix_bd_57_cast_fp16 = reshape(shape = var_3512, x = var_3511_cast_fp16)[name = string("matrix_bd_57_cast_fp16")]; + bool matrix_ac_29_transpose_x_0 = const()[name = string("matrix_ac_29_transpose_x_0"), val = bool(false)]; + bool matrix_ac_29_transpose_y_0 = const()[name = string("matrix_ac_29_transpose_y_0"), val = bool(false)]; + tensor transpose_124_perm_0 = const()[name = string("transpose_124_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_125_perm_0 = const()[name = string("transpose_125_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_125 = transpose(perm = transpose_125_perm_0, x = k_57_cast_fp16)[name = string("transpose_234")]; + tensor transpose_124 = transpose(perm = transpose_124_perm_0, x = var_3495_cast_fp16)[name = string("transpose_235")]; + tensor matrix_ac_29_cast_fp16 = matmul(transpose_x = matrix_ac_29_transpose_x_0, transpose_y = matrix_ac_29_transpose_y_0, x = transpose_124, y = transpose_125)[name = string("matrix_ac_29_cast_fp16")]; + tensor matrix_bd_59_begin_0 = const()[name = string("matrix_bd_59_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_59_end_0 = const()[name = string("matrix_bd_59_end_0"), val = tensor([1, 8, 14, 56])]; + tensor matrix_bd_59_end_mask_0 = const()[name = string("matrix_bd_59_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_59_cast_fp16 = slice_by_index(begin = matrix_bd_59_begin_0, end = matrix_bd_59_end_0, end_mask = matrix_bd_59_end_mask_0, x = matrix_bd_57_cast_fp16)[name = string("matrix_bd_59_cast_fp16")]; + tensor var_3521_cast_fp16 = add(x = matrix_ac_29_cast_fp16, y = matrix_bd_59_cast_fp16)[name = string("op_3521_cast_fp16")]; + fp16 _inversed_scores_57_y_0_to_fp16 = const()[name = string("_inversed_scores_57_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_57_cast_fp16 = mul(x = var_3521_cast_fp16, y = _inversed_scores_57_y_0_to_fp16)[name = string("_inversed_scores_57_cast_fp16")]; + tensor scores_59_cast_fp16 = select(a = var_46_to_fp16, b = _inversed_scores_57_cast_fp16, cond = mask_11)[name = string("scores_59_cast_fp16")]; + tensor var_3527_cast_fp16 = softmax(axis = var_60, x = scores_59_cast_fp16)[name = string("op_3527_cast_fp16")]; + tensor input_769_cast_fp16 = select(a = var_45_to_fp16, b = var_3527_cast_fp16, cond = mask_11)[name = string("input_769_cast_fp16")]; + bool x_377_transpose_x_0 = const()[name = string("x_377_transpose_x_0"), val = bool(false)]; + bool x_377_transpose_y_0 = const()[name = string("x_377_transpose_y_0"), val = bool(false)]; + tensor value_37_cast_fp16 = transpose(perm = value_37_perm_0, x = v_29_cast_fp16)[name = string("transpose_233")]; + tensor x_377_cast_fp16 = matmul(transpose_x = x_377_transpose_x_0, transpose_y = x_377_transpose_y_0, x = input_769_cast_fp16, y = value_37_cast_fp16)[name = string("x_377_cast_fp16")]; + tensor var_3531_perm_0 = const()[name = string("op_3531_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_3532 = const()[name = string("op_3532"), val = tensor([1, -1, 1024])]; + tensor var_3531_cast_fp16 = transpose(perm = var_3531_perm_0, x = x_377_cast_fp16)[name = string("transpose_232")]; + tensor input_771_cast_fp16 = reshape(shape = var_3532, x = var_3531_cast_fp16)[name = string("input_771_cast_fp16")]; + tensor encoder_layers_14_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293780544))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294567040))))[name = string("encoder_layers_14_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_14_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_14_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294567232)))]; + tensor linear_133_cast_fp16 = linear(bias = encoder_layers_14_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_14_self_attn_linear_out_weight_to_fp16_palettized, x = input_771_cast_fp16)[name = string("linear_133_cast_fp16")]; + tensor input_775_cast_fp16 = add(x = input_765_cast_fp16, y = linear_133_cast_fp16)[name = string("input_775_cast_fp16")]; + tensor x_381_axes_0 = const()[name = string("x_381_axes_0"), val = tensor([-1])]; + tensor encoder_layers_14_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_14_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294569344)))]; + tensor encoder_layers_14_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_14_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294571456)))]; + tensor x_381_cast_fp16 = layer_norm(axes = x_381_axes_0, beta = encoder_layers_14_norm_conv_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_14_norm_conv_weight_to_fp16, x = input_775_cast_fp16)[name = string("x_381_cast_fp16")]; + tensor input_777_perm_0 = const()[name = string("input_777_perm_0"), val = tensor([0, 2, 1])]; + string input_779_pad_type_0 = const()[name = string("input_779_pad_type_0"), val = string("valid")]; + tensor input_779_strides_0 = const()[name = string("input_779_strides_0"), val = tensor([1])]; + tensor input_779_pad_0 = const()[name = string("input_779_pad_0"), val = tensor([0, 0])]; + tensor input_779_dilations_0 = const()[name = string("input_779_dilations_0"), val = tensor([1])]; + int32 input_779_groups_0 = const()[name = string("input_779_groups_0"), val = int32(1)]; + tensor encoder_layers_14_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294573568))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296670784))))[name = string("encoder_layers_14_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_777_cast_fp16 = transpose(perm = input_777_perm_0, x = x_381_cast_fp16)[name = string("transpose_231")]; + tensor input_779_cast_fp16 = conv(dilations = input_779_dilations_0, groups = input_779_groups_0, pad = input_779_pad_0, pad_type = input_779_pad_type_0, strides = input_779_strides_0, weight = encoder_layers_14_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_777_cast_fp16)[name = string("input_779_cast_fp16")]; + int32 x_383_split_num_splits_0 = const()[name = string("x_383_split_num_splits_0"), val = int32(2)]; + int32 x_383_split_axis_0 = const()[name = string("x_383_split_axis_0"), val = int32(1)]; + tensor x_383_split_cast_fp16_0, tensor x_383_split_cast_fp16_1 = split(axis = x_383_split_axis_0, num_splits = x_383_split_num_splits_0, x = input_779_cast_fp16)[name = string("x_383_split_cast_fp16")]; + tensor x_383_split_1_sigmoid_cast_fp16 = sigmoid(x = x_383_split_cast_fp16_1)[name = string("x_383_split_1_sigmoid_cast_fp16")]; + tensor x_383_cast_fp16 = mul(x = x_383_split_cast_fp16_0, y = x_383_split_1_sigmoid_cast_fp16)[name = string("x_383_cast_fp16")]; + tensor input_781_cast_fp16 = select(a = var_45_to_fp16, b = x_383_cast_fp16, cond = var_576)[name = string("input_781_cast_fp16")]; + bool new_x_59_interleave_0 = const()[name = string("new_x_59_interleave_0"), val = bool(false)]; + tensor new_x_59_cast_fp16 = concat(axis = var_60, interleave = new_x_59_interleave_0, values = (cache_59_cast_fp16, input_781_cast_fp16))[name = string("new_x_59_cast_fp16")]; + tensor var_3571_begin_0 = const()[name = string("op_3571_begin_0"), val = tensor([0, 0, 14])]; + tensor var_3571_end_0 = const()[name = string("op_3571_end_0"), val = tensor([1, 1024, 22])]; + tensor var_3571_end_mask_0 = const()[name = string("op_3571_end_mask_0"), val = tensor([true, true, true])]; + tensor var_3571_cast_fp16 = slice_by_index(begin = var_3571_begin_0, end = var_3571_end_0, end_mask = var_3571_end_mask_0, x = new_x_59_cast_fp16)[name = string("op_3571_cast_fp16")]; + string x_385_pad_type_0 = const()[name = string("x_385_pad_type_0"), val = string("valid")]; + int32 x_385_groups_0 = const()[name = string("x_385_groups_0"), val = int32(1024)]; + tensor x_385_strides_0 = const()[name = string("x_385_strides_0"), val = tensor([1])]; + tensor x_385_pad_0 = const()[name = string("x_385_pad_0"), val = tensor([0, 0])]; + tensor x_385_dilations_0 = const()[name = string("x_385_dilations_0"), val = tensor([1])]; + tensor encoder_layers_14_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296674944))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296684224))))[name = string("encoder_layers_14_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_385_cast_fp16 = conv(dilations = x_385_dilations_0, groups = x_385_groups_0, pad = x_385_pad_0, pad_type = x_385_pad_type_0, strides = x_385_strides_0, weight = encoder_layers_14_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_59_cast_fp16)[name = string("x_385_cast_fp16")]; + tensor input_783_perm_0 = const()[name = string("input_783_perm_0"), val = tensor([0, 2, 1])]; + tensor x_387_axes_0 = const()[name = string("x_387_axes_0"), val = tensor([-1])]; + tensor encoder_layers_14_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_14_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296686336)))]; + tensor encoder_layers_14_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_14_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296688448)))]; + tensor input_783_cast_fp16 = transpose(perm = input_783_perm_0, x = x_385_cast_fp16)[name = string("transpose_230")]; + tensor x_387_cast_fp16 = layer_norm(axes = x_387_axes_0, beta = encoder_layers_14_conv_batch_norm_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_14_conv_batch_norm_weight_to_fp16, x = input_783_cast_fp16)[name = string("x_387_cast_fp16")]; + tensor input_785_perm_0 = const()[name = string("input_785_perm_0"), val = tensor([0, 2, 1])]; + tensor input_785_cast_fp16 = transpose(perm = input_785_perm_0, x = x_387_cast_fp16)[name = string("transpose_229")]; + tensor input_787_cast_fp16 = silu(x = input_785_cast_fp16)[name = string("input_787_cast_fp16")]; + string x_389_pad_type_0 = const()[name = string("x_389_pad_type_0"), val = string("valid")]; + tensor x_389_strides_0 = const()[name = string("x_389_strides_0"), val = tensor([1])]; + tensor x_389_pad_0 = const()[name = string("x_389_pad_0"), val = tensor([0, 0])]; + tensor x_389_dilations_0 = const()[name = string("x_389_dilations_0"), val = tensor([1])]; + int32 x_389_groups_0 = const()[name = string("x_389_groups_0"), val = int32(1)]; + tensor encoder_layers_14_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296690560))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297739200))))[name = string("encoder_layers_14_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_389_cast_fp16 = conv(dilations = x_389_dilations_0, groups = x_389_groups_0, pad = x_389_pad_0, pad_type = x_389_pad_type_0, strides = x_389_strides_0, weight = encoder_layers_14_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_787_cast_fp16)[name = string("x_389_cast_fp16")]; + tensor input_789_perm_0 = const()[name = string("input_789_perm_0"), val = tensor([0, 2, 1])]; + tensor input_789_cast_fp16 = transpose(perm = input_789_perm_0, x = x_389_cast_fp16)[name = string("transpose_228")]; + tensor input_791_cast_fp16 = add(x = input_775_cast_fp16, y = input_789_cast_fp16)[name = string("input_791_cast_fp16")]; + tensor input_793_axes_0 = const()[name = string("input_793_axes_0"), val = tensor([-1])]; + tensor encoder_layers_14_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_14_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297741312)))]; + tensor encoder_layers_14_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_14_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297743424)))]; + tensor input_793_cast_fp16 = layer_norm(axes = input_793_axes_0, beta = encoder_layers_14_norm_feed_forward2_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_14_norm_feed_forward2_weight_to_fp16, x = input_791_cast_fp16)[name = string("input_793_cast_fp16")]; + tensor encoder_layers_14_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297745536))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(300891328))))[name = string("encoder_layers_14_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_14_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_14_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(300891520)))]; + tensor linear_134_cast_fp16 = linear(bias = encoder_layers_14_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_14_feed_forward2_linear1_weight_to_fp16_palettized, x = input_793_cast_fp16)[name = string("linear_134_cast_fp16")]; + tensor input_797_cast_fp16 = silu(x = linear_134_cast_fp16)[name = string("input_797_cast_fp16")]; + tensor encoder_layers_14_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(300899776))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(304045568))))[name = string("encoder_layers_14_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_14_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_14_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(304045760)))]; + tensor linear_135_cast_fp16 = linear(bias = encoder_layers_14_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_14_feed_forward2_linear2_weight_to_fp16_palettized, x = input_797_cast_fp16)[name = string("linear_135_cast_fp16")]; + fp16 var_3614_to_fp16 = const()[name = string("op_3614_to_fp16"), val = fp16(0x1p-1)]; + tensor var_3615_cast_fp16 = mul(x = linear_135_cast_fp16, y = var_3614_to_fp16)[name = string("op_3615_cast_fp16")]; + tensor input_803_cast_fp16 = add(x = input_791_cast_fp16, y = var_3615_cast_fp16)[name = string("input_803_cast_fp16")]; + tensor input_805_axes_0 = const()[name = string("input_805_axes_0"), val = tensor([-1])]; + tensor encoder_layers_14_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_14_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(304047872)))]; + tensor encoder_layers_14_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_14_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(304049984)))]; + tensor input_805_cast_fp16 = layer_norm(axes = input_805_axes_0, beta = encoder_layers_14_norm_out_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_14_norm_out_weight_to_fp16, x = input_803_cast_fp16)[name = string("input_805_cast_fp16")]; + tensor cache_61_begin_0 = const()[name = string("cache_61_begin_0"), val = tensor([15, 0, 0, 0])]; + tensor cache_61_end_0 = const()[name = string("cache_61_end_0"), val = tensor([16, 1, 42, 1024])]; + tensor cache_61_end_mask_0 = const()[name = string("cache_61_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_61_squeeze_mask_0 = const()[name = string("cache_61_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_61_cast_fp16 = slice_by_index(begin = cache_61_begin_0, end = cache_61_end_0, end_mask = cache_61_end_mask_0, squeeze_mask = cache_61_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_61_cast_fp16")]; + tensor cache_63_begin_0 = const()[name = string("cache_63_begin_0"), val = tensor([15, 0, 0, 0])]; + tensor cache_63_end_0 = const()[name = string("cache_63_end_0"), val = tensor([16, 1, 1024, 8])]; + tensor cache_63_end_mask_0 = const()[name = string("cache_63_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_63_squeeze_mask_0 = const()[name = string("cache_63_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_63_cast_fp16 = slice_by_index(begin = cache_63_begin_0, end = cache_63_end_0, end_mask = cache_63_end_mask_0, squeeze_mask = cache_63_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_63_cast_fp16")]; + tensor input_807_axes_0 = const()[name = string("input_807_axes_0"), val = tensor([-1])]; + tensor encoder_layers_15_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_15_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(304052096)))]; + tensor encoder_layers_15_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_15_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(304054208)))]; + tensor input_807_cast_fp16 = layer_norm(axes = input_807_axes_0, beta = encoder_layers_15_norm_feed_forward1_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_15_norm_feed_forward1_weight_to_fp16, x = input_805_cast_fp16)[name = string("input_807_cast_fp16")]; + tensor encoder_layers_15_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(304056320))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307202112))))[name = string("encoder_layers_15_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_15_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_15_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307202304)))]; + tensor linear_136_cast_fp16 = linear(bias = encoder_layers_15_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_15_feed_forward1_linear1_weight_to_fp16_palettized, x = input_807_cast_fp16)[name = string("linear_136_cast_fp16")]; + tensor input_811_cast_fp16 = silu(x = linear_136_cast_fp16)[name = string("input_811_cast_fp16")]; + tensor encoder_layers_15_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307210560))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310356352))))[name = string("encoder_layers_15_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_15_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_15_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310356544)))]; + tensor linear_137_cast_fp16 = linear(bias = encoder_layers_15_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_15_feed_forward1_linear2_weight_to_fp16_palettized, x = input_811_cast_fp16)[name = string("linear_137_cast_fp16")]; + fp16 var_3651_to_fp16 = const()[name = string("op_3651_to_fp16"), val = fp16(0x1p-1)]; + tensor var_3652_cast_fp16 = mul(x = linear_137_cast_fp16, y = var_3651_to_fp16)[name = string("op_3652_cast_fp16")]; + tensor input_817_cast_fp16 = add(x = input_805_cast_fp16, y = var_3652_cast_fp16)[name = string("input_817_cast_fp16")]; + tensor key_31_axes_0 = const()[name = string("key_31_axes_0"), val = tensor([-1])]; + tensor encoder_layers_15_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_15_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310358656)))]; + tensor encoder_layers_15_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_15_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310360768)))]; + tensor key_31_cast_fp16 = layer_norm(axes = key_31_axes_0, beta = encoder_layers_15_norm_self_att_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_15_norm_self_att_weight_to_fp16, x = input_817_cast_fp16)[name = string("key_31_cast_fp16")]; + bool input_819_interleave_0 = const()[name = string("input_819_interleave_0"), val = bool(false)]; + tensor input_819_cast_fp16 = concat(axis = var_69, interleave = input_819_interleave_0, values = (cache_61_cast_fp16, key_31_cast_fp16))[name = string("input_819_cast_fp16")]; + tensor var_3674_begin_0 = const()[name = string("op_3674_begin_0"), val = tensor([0, 14, 0])]; + tensor var_3674_end_0 = const()[name = string("op_3674_end_0"), val = tensor([1, 42, 1024])]; + tensor var_3674_end_mask_0 = const()[name = string("op_3674_end_mask_0"), val = tensor([true, true, true])]; + tensor var_3674_cast_fp16 = slice_by_index(begin = var_3674_begin_0, end = var_3674_end_0, end_mask = var_3674_end_mask_0, x = cache_61_cast_fp16)[name = string("op_3674_cast_fp16")]; + bool var_3680_interleave_0 = const()[name = string("op_3680_interleave_0"), val = bool(false)]; + tensor var_3680_cast_fp16 = concat(axis = var_69, interleave = var_3680_interleave_0, values = (var_3674_cast_fp16, key_31_cast_fp16))[name = string("op_3680_cast_fp16")]; + tensor encoder_layers_15_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310362880))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311149376))))[name = string("encoder_layers_15_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_15_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_15_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311149568)))]; + tensor linear_138_cast_fp16 = linear(bias = encoder_layers_15_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_15_self_attn_linear_q_weight_to_fp16_palettized, x = key_31_cast_fp16)[name = string("linear_138_cast_fp16")]; + tensor var_3685 = const()[name = string("op_3685"), val = tensor([1, -1, 8, 128])]; + tensor q_91_cast_fp16 = reshape(shape = var_3685, x = linear_138_cast_fp16)[name = string("q_91_cast_fp16")]; + tensor encoder_layers_15_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311151680))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311938176))))[name = string("encoder_layers_15_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_15_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_15_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311938368)))]; + tensor linear_139_cast_fp16 = linear(bias = encoder_layers_15_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_15_self_attn_linear_k_weight_to_fp16_palettized, x = input_819_cast_fp16)[name = string("linear_139_cast_fp16")]; + tensor var_3690 = const()[name = string("op_3690"), val = tensor([1, -1, 8, 128])]; + tensor k_61_cast_fp16 = reshape(shape = var_3690, x = linear_139_cast_fp16)[name = string("k_61_cast_fp16")]; + tensor encoder_layers_15_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311940480))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(312726976))))[name = string("encoder_layers_15_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_15_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_15_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(312727168)))]; + tensor linear_140_cast_fp16 = linear(bias = encoder_layers_15_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_15_self_attn_linear_v_weight_to_fp16_palettized, x = input_819_cast_fp16)[name = string("linear_140_cast_fp16")]; + tensor var_3695 = const()[name = string("op_3695"), val = tensor([1, -1, 8, 128])]; + tensor v_31_cast_fp16 = reshape(shape = var_3695, x = linear_140_cast_fp16)[name = string("v_31_cast_fp16")]; + tensor value_39_perm_0 = const()[name = string("value_39_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_15_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_15_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(312729280)))]; + tensor var_3708_cast_fp16 = add(x = q_91_cast_fp16, y = encoder_layers_15_self_attn_pos_bias_u_to_fp16)[name = string("op_3708_cast_fp16")]; + tensor encoder_layers_15_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_15_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(312731392)))]; + tensor var_3710_cast_fp16 = add(x = q_91_cast_fp16, y = encoder_layers_15_self_attn_pos_bias_v_to_fp16)[name = string("op_3710_cast_fp16")]; + tensor q_with_bias_v_31_perm_0 = const()[name = string("q_with_bias_v_31_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_397_transpose_x_0 = const()[name = string("x_397_transpose_x_0"), val = bool(false)]; + bool x_397_transpose_y_0 = const()[name = string("x_397_transpose_y_0"), val = bool(false)]; + tensor op_3712_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(312733504))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(312847232))))[name = string("op_3712_to_fp16_quantized")]; + tensor q_with_bias_v_31_cast_fp16 = transpose(perm = q_with_bias_v_31_perm_0, x = var_3710_cast_fp16)[name = string("transpose_227")]; + tensor x_397_cast_fp16 = matmul(transpose_x = x_397_transpose_x_0, transpose_y = x_397_transpose_y_0, x = q_with_bias_v_31_cast_fp16, y = op_3712_to_fp16_quantized)[name = string("x_397_cast_fp16")]; + tensor x_399_pad_0 = const()[name = string("x_399_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_399_mode_0 = const()[name = string("x_399_mode_0"), val = string("constant")]; + fp16 const_274_to_fp16 = const()[name = string("const_274_to_fp16"), val = fp16(0x0p+0)]; + tensor x_399_cast_fp16 = pad(constant_val = const_274_to_fp16, mode = x_399_mode_0, pad = x_399_pad_0, x = x_397_cast_fp16)[name = string("x_399_cast_fp16")]; + tensor var_3720 = const()[name = string("op_3720"), val = tensor([1, 8, -1, 14])]; + tensor x_401_cast_fp16 = reshape(shape = var_3720, x = x_399_cast_fp16)[name = string("x_401_cast_fp16")]; + tensor var_3724_begin_0 = const()[name = string("op_3724_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_3724_end_0 = const()[name = string("op_3724_end_0"), val = tensor([1, 8, 112, 14])]; + tensor var_3724_end_mask_0 = const()[name = string("op_3724_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_3724_cast_fp16 = slice_by_index(begin = var_3724_begin_0, end = var_3724_end_0, end_mask = var_3724_end_mask_0, x = x_401_cast_fp16)[name = string("op_3724_cast_fp16")]; + tensor var_3725 = const()[name = string("op_3725"), val = tensor([1, 8, 14, 111])]; + tensor matrix_bd_61_cast_fp16 = reshape(shape = var_3725, x = var_3724_cast_fp16)[name = string("matrix_bd_61_cast_fp16")]; + bool matrix_ac_31_transpose_x_0 = const()[name = string("matrix_ac_31_transpose_x_0"), val = bool(false)]; + bool matrix_ac_31_transpose_y_0 = const()[name = string("matrix_ac_31_transpose_y_0"), val = bool(false)]; + tensor transpose_126_perm_0 = const()[name = string("transpose_126_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_127_perm_0 = const()[name = string("transpose_127_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_127 = transpose(perm = transpose_127_perm_0, x = k_61_cast_fp16)[name = string("transpose_225")]; + tensor transpose_126 = transpose(perm = transpose_126_perm_0, x = var_3708_cast_fp16)[name = string("transpose_226")]; + tensor matrix_ac_31_cast_fp16 = matmul(transpose_x = matrix_ac_31_transpose_x_0, transpose_y = matrix_ac_31_transpose_y_0, x = transpose_126, y = transpose_127)[name = string("matrix_ac_31_cast_fp16")]; + tensor matrix_bd_63_begin_0 = const()[name = string("matrix_bd_63_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_63_end_0 = const()[name = string("matrix_bd_63_end_0"), val = tensor([1, 8, 14, 56])]; + tensor matrix_bd_63_end_mask_0 = const()[name = string("matrix_bd_63_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_63_cast_fp16 = slice_by_index(begin = matrix_bd_63_begin_0, end = matrix_bd_63_end_0, end_mask = matrix_bd_63_end_mask_0, x = matrix_bd_61_cast_fp16)[name = string("matrix_bd_63_cast_fp16")]; + tensor var_3734_cast_fp16 = add(x = matrix_ac_31_cast_fp16, y = matrix_bd_63_cast_fp16)[name = string("op_3734_cast_fp16")]; + fp16 _inversed_scores_61_y_0_to_fp16 = const()[name = string("_inversed_scores_61_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_61_cast_fp16 = mul(x = var_3734_cast_fp16, y = _inversed_scores_61_y_0_to_fp16)[name = string("_inversed_scores_61_cast_fp16")]; + tensor scores_63_cast_fp16 = select(a = var_46_to_fp16, b = _inversed_scores_61_cast_fp16, cond = mask_11)[name = string("scores_63_cast_fp16")]; + tensor var_3740_cast_fp16 = softmax(axis = var_60, x = scores_63_cast_fp16)[name = string("op_3740_cast_fp16")]; + tensor input_821_cast_fp16 = select(a = var_45_to_fp16, b = var_3740_cast_fp16, cond = mask_11)[name = string("input_821_cast_fp16")]; + bool x_403_transpose_x_0 = const()[name = string("x_403_transpose_x_0"), val = bool(false)]; + bool x_403_transpose_y_0 = const()[name = string("x_403_transpose_y_0"), val = bool(false)]; + tensor value_39_cast_fp16 = transpose(perm = value_39_perm_0, x = v_31_cast_fp16)[name = string("transpose_224")]; + tensor x_403_cast_fp16 = matmul(transpose_x = x_403_transpose_x_0, transpose_y = x_403_transpose_y_0, x = input_821_cast_fp16, y = value_39_cast_fp16)[name = string("x_403_cast_fp16")]; + tensor var_3744_perm_0 = const()[name = string("op_3744_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_3745 = const()[name = string("op_3745"), val = tensor([1, -1, 1024])]; + tensor var_3744_cast_fp16 = transpose(perm = var_3744_perm_0, x = x_403_cast_fp16)[name = string("transpose_223")]; + tensor input_823_cast_fp16 = reshape(shape = var_3745, x = var_3744_cast_fp16)[name = string("input_823_cast_fp16")]; + tensor encoder_layers_15_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(312847552))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(313634048))))[name = string("encoder_layers_15_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_15_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_15_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(313634240)))]; + tensor linear_142_cast_fp16 = linear(bias = encoder_layers_15_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_15_self_attn_linear_out_weight_to_fp16_palettized, x = input_823_cast_fp16)[name = string("linear_142_cast_fp16")]; + tensor input_827_cast_fp16 = add(x = input_817_cast_fp16, y = linear_142_cast_fp16)[name = string("input_827_cast_fp16")]; + tensor x_407_axes_0 = const()[name = string("x_407_axes_0"), val = tensor([-1])]; + tensor encoder_layers_15_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_15_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(313636352)))]; + tensor encoder_layers_15_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_15_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(313638464)))]; + tensor x_407_cast_fp16 = layer_norm(axes = x_407_axes_0, beta = encoder_layers_15_norm_conv_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_15_norm_conv_weight_to_fp16, x = input_827_cast_fp16)[name = string("x_407_cast_fp16")]; + tensor input_829_perm_0 = const()[name = string("input_829_perm_0"), val = tensor([0, 2, 1])]; + string input_831_pad_type_0 = const()[name = string("input_831_pad_type_0"), val = string("valid")]; + tensor input_831_strides_0 = const()[name = string("input_831_strides_0"), val = tensor([1])]; + tensor input_831_pad_0 = const()[name = string("input_831_pad_0"), val = tensor([0, 0])]; + tensor input_831_dilations_0 = const()[name = string("input_831_dilations_0"), val = tensor([1])]; + int32 input_831_groups_0 = const()[name = string("input_831_groups_0"), val = int32(1)]; + tensor encoder_layers_15_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(313640576))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315737792))))[name = string("encoder_layers_15_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_829_cast_fp16 = transpose(perm = input_829_perm_0, x = x_407_cast_fp16)[name = string("transpose_222")]; + tensor input_831_cast_fp16 = conv(dilations = input_831_dilations_0, groups = input_831_groups_0, pad = input_831_pad_0, pad_type = input_831_pad_type_0, strides = input_831_strides_0, weight = encoder_layers_15_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_829_cast_fp16)[name = string("input_831_cast_fp16")]; + int32 x_409_split_num_splits_0 = const()[name = string("x_409_split_num_splits_0"), val = int32(2)]; + int32 x_409_split_axis_0 = const()[name = string("x_409_split_axis_0"), val = int32(1)]; + tensor x_409_split_cast_fp16_0, tensor x_409_split_cast_fp16_1 = split(axis = x_409_split_axis_0, num_splits = x_409_split_num_splits_0, x = input_831_cast_fp16)[name = string("x_409_split_cast_fp16")]; + tensor x_409_split_1_sigmoid_cast_fp16 = sigmoid(x = x_409_split_cast_fp16_1)[name = string("x_409_split_1_sigmoid_cast_fp16")]; + tensor x_409_cast_fp16 = mul(x = x_409_split_cast_fp16_0, y = x_409_split_1_sigmoid_cast_fp16)[name = string("x_409_cast_fp16")]; + tensor input_833_cast_fp16 = select(a = var_45_to_fp16, b = x_409_cast_fp16, cond = var_576)[name = string("input_833_cast_fp16")]; + bool new_x_63_interleave_0 = const()[name = string("new_x_63_interleave_0"), val = bool(false)]; + tensor new_x_63_cast_fp16 = concat(axis = var_60, interleave = new_x_63_interleave_0, values = (cache_63_cast_fp16, input_833_cast_fp16))[name = string("new_x_63_cast_fp16")]; + tensor var_3784_begin_0 = const()[name = string("op_3784_begin_0"), val = tensor([0, 0, 14])]; + tensor var_3784_end_0 = const()[name = string("op_3784_end_0"), val = tensor([1, 1024, 22])]; + tensor var_3784_end_mask_0 = const()[name = string("op_3784_end_mask_0"), val = tensor([true, true, true])]; + tensor var_3784_cast_fp16 = slice_by_index(begin = var_3784_begin_0, end = var_3784_end_0, end_mask = var_3784_end_mask_0, x = new_x_63_cast_fp16)[name = string("op_3784_cast_fp16")]; + string x_411_pad_type_0 = const()[name = string("x_411_pad_type_0"), val = string("valid")]; + int32 x_411_groups_0 = const()[name = string("x_411_groups_0"), val = int32(1024)]; + tensor x_411_strides_0 = const()[name = string("x_411_strides_0"), val = tensor([1])]; + tensor x_411_pad_0 = const()[name = string("x_411_pad_0"), val = tensor([0, 0])]; + tensor x_411_dilations_0 = const()[name = string("x_411_dilations_0"), val = tensor([1])]; + tensor encoder_layers_15_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315741952))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315751232))))[name = string("encoder_layers_15_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_411_cast_fp16 = conv(dilations = x_411_dilations_0, groups = x_411_groups_0, pad = x_411_pad_0, pad_type = x_411_pad_type_0, strides = x_411_strides_0, weight = encoder_layers_15_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_63_cast_fp16)[name = string("x_411_cast_fp16")]; + tensor input_835_perm_0 = const()[name = string("input_835_perm_0"), val = tensor([0, 2, 1])]; + tensor x_413_axes_0 = const()[name = string("x_413_axes_0"), val = tensor([-1])]; + tensor encoder_layers_15_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_15_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315753344)))]; + tensor encoder_layers_15_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_15_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315755456)))]; + tensor input_835_cast_fp16 = transpose(perm = input_835_perm_0, x = x_411_cast_fp16)[name = string("transpose_221")]; + tensor x_413_cast_fp16 = layer_norm(axes = x_413_axes_0, beta = encoder_layers_15_conv_batch_norm_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_15_conv_batch_norm_weight_to_fp16, x = input_835_cast_fp16)[name = string("x_413_cast_fp16")]; + tensor input_837_perm_0 = const()[name = string("input_837_perm_0"), val = tensor([0, 2, 1])]; + tensor input_837_cast_fp16 = transpose(perm = input_837_perm_0, x = x_413_cast_fp16)[name = string("transpose_220")]; + tensor input_839_cast_fp16 = silu(x = input_837_cast_fp16)[name = string("input_839_cast_fp16")]; + string x_415_pad_type_0 = const()[name = string("x_415_pad_type_0"), val = string("valid")]; + tensor x_415_strides_0 = const()[name = string("x_415_strides_0"), val = tensor([1])]; + tensor x_415_pad_0 = const()[name = string("x_415_pad_0"), val = tensor([0, 0])]; + tensor x_415_dilations_0 = const()[name = string("x_415_dilations_0"), val = tensor([1])]; + int32 x_415_groups_0 = const()[name = string("x_415_groups_0"), val = int32(1)]; + tensor encoder_layers_15_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315757568))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316806208))))[name = string("encoder_layers_15_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_415_cast_fp16 = conv(dilations = x_415_dilations_0, groups = x_415_groups_0, pad = x_415_pad_0, pad_type = x_415_pad_type_0, strides = x_415_strides_0, weight = encoder_layers_15_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_839_cast_fp16)[name = string("x_415_cast_fp16")]; + tensor input_841_perm_0 = const()[name = string("input_841_perm_0"), val = tensor([0, 2, 1])]; + tensor input_841_cast_fp16 = transpose(perm = input_841_perm_0, x = x_415_cast_fp16)[name = string("transpose_219")]; + tensor input_843_cast_fp16 = add(x = input_827_cast_fp16, y = input_841_cast_fp16)[name = string("input_843_cast_fp16")]; + tensor input_845_axes_0 = const()[name = string("input_845_axes_0"), val = tensor([-1])]; + tensor encoder_layers_15_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_15_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316808320)))]; + tensor encoder_layers_15_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_15_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316810432)))]; + tensor input_845_cast_fp16 = layer_norm(axes = input_845_axes_0, beta = encoder_layers_15_norm_feed_forward2_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_15_norm_feed_forward2_weight_to_fp16, x = input_843_cast_fp16)[name = string("input_845_cast_fp16")]; + tensor encoder_layers_15_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316812544))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(319958336))))[name = string("encoder_layers_15_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_15_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_15_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(319958528)))]; + tensor linear_143_cast_fp16 = linear(bias = encoder_layers_15_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_15_feed_forward2_linear1_weight_to_fp16_palettized, x = input_845_cast_fp16)[name = string("linear_143_cast_fp16")]; + tensor input_849_cast_fp16 = silu(x = linear_143_cast_fp16)[name = string("input_849_cast_fp16")]; + tensor encoder_layers_15_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(319966784))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(323112576))))[name = string("encoder_layers_15_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_15_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_15_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(323112768)))]; + tensor linear_144_cast_fp16 = linear(bias = encoder_layers_15_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_15_feed_forward2_linear2_weight_to_fp16_palettized, x = input_849_cast_fp16)[name = string("linear_144_cast_fp16")]; + fp16 var_3827_to_fp16 = const()[name = string("op_3827_to_fp16"), val = fp16(0x1p-1)]; + tensor var_3828_cast_fp16 = mul(x = linear_144_cast_fp16, y = var_3827_to_fp16)[name = string("op_3828_cast_fp16")]; + tensor input_855_cast_fp16 = add(x = input_843_cast_fp16, y = var_3828_cast_fp16)[name = string("input_855_cast_fp16")]; + tensor input_857_axes_0 = const()[name = string("input_857_axes_0"), val = tensor([-1])]; + tensor encoder_layers_15_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_15_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(323114880)))]; + tensor encoder_layers_15_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_15_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(323116992)))]; + tensor input_857_cast_fp16 = layer_norm(axes = input_857_axes_0, beta = encoder_layers_15_norm_out_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_15_norm_out_weight_to_fp16, x = input_855_cast_fp16)[name = string("input_857_cast_fp16")]; + tensor cache_65_begin_0 = const()[name = string("cache_65_begin_0"), val = tensor([16, 0, 0, 0])]; + tensor cache_65_end_0 = const()[name = string("cache_65_end_0"), val = tensor([17, 1, 42, 1024])]; + tensor cache_65_end_mask_0 = const()[name = string("cache_65_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_65_squeeze_mask_0 = const()[name = string("cache_65_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_65_cast_fp16 = slice_by_index(begin = cache_65_begin_0, end = cache_65_end_0, end_mask = cache_65_end_mask_0, squeeze_mask = cache_65_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_65_cast_fp16")]; + tensor cache_67_begin_0 = const()[name = string("cache_67_begin_0"), val = tensor([16, 0, 0, 0])]; + tensor cache_67_end_0 = const()[name = string("cache_67_end_0"), val = tensor([17, 1, 1024, 8])]; + tensor cache_67_end_mask_0 = const()[name = string("cache_67_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_67_squeeze_mask_0 = const()[name = string("cache_67_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_67_cast_fp16 = slice_by_index(begin = cache_67_begin_0, end = cache_67_end_0, end_mask = cache_67_end_mask_0, squeeze_mask = cache_67_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_67_cast_fp16")]; + tensor input_859_axes_0 = const()[name = string("input_859_axes_0"), val = tensor([-1])]; + tensor encoder_layers_16_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_16_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(323119104)))]; + tensor encoder_layers_16_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_16_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(323121216)))]; + tensor input_859_cast_fp16 = layer_norm(axes = input_859_axes_0, beta = encoder_layers_16_norm_feed_forward1_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_16_norm_feed_forward1_weight_to_fp16, x = input_857_cast_fp16)[name = string("input_859_cast_fp16")]; + tensor encoder_layers_16_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(323123328))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(326269120))))[name = string("encoder_layers_16_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_16_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_16_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(326269312)))]; + tensor linear_145_cast_fp16 = linear(bias = encoder_layers_16_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_16_feed_forward1_linear1_weight_to_fp16_palettized, x = input_859_cast_fp16)[name = string("linear_145_cast_fp16")]; + tensor input_863_cast_fp16 = silu(x = linear_145_cast_fp16)[name = string("input_863_cast_fp16")]; + tensor encoder_layers_16_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(326277568))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(329423360))))[name = string("encoder_layers_16_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_16_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_16_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(329423552)))]; + tensor linear_146_cast_fp16 = linear(bias = encoder_layers_16_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_16_feed_forward1_linear2_weight_to_fp16_palettized, x = input_863_cast_fp16)[name = string("linear_146_cast_fp16")]; + fp16 var_3864_to_fp16 = const()[name = string("op_3864_to_fp16"), val = fp16(0x1p-1)]; + tensor var_3865_cast_fp16 = mul(x = linear_146_cast_fp16, y = var_3864_to_fp16)[name = string("op_3865_cast_fp16")]; + tensor input_869_cast_fp16 = add(x = input_857_cast_fp16, y = var_3865_cast_fp16)[name = string("input_869_cast_fp16")]; + tensor key_33_axes_0 = const()[name = string("key_33_axes_0"), val = tensor([-1])]; + tensor encoder_layers_16_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_16_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(329425664)))]; + tensor encoder_layers_16_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_16_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(329427776)))]; + tensor key_33_cast_fp16 = layer_norm(axes = key_33_axes_0, beta = encoder_layers_16_norm_self_att_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_16_norm_self_att_weight_to_fp16, x = input_869_cast_fp16)[name = string("key_33_cast_fp16")]; + bool input_871_interleave_0 = const()[name = string("input_871_interleave_0"), val = bool(false)]; + tensor input_871_cast_fp16 = concat(axis = var_69, interleave = input_871_interleave_0, values = (cache_65_cast_fp16, key_33_cast_fp16))[name = string("input_871_cast_fp16")]; + tensor var_3887_begin_0 = const()[name = string("op_3887_begin_0"), val = tensor([0, 14, 0])]; + tensor var_3887_end_0 = const()[name = string("op_3887_end_0"), val = tensor([1, 42, 1024])]; + tensor var_3887_end_mask_0 = const()[name = string("op_3887_end_mask_0"), val = tensor([true, true, true])]; + tensor var_3887_cast_fp16 = slice_by_index(begin = var_3887_begin_0, end = var_3887_end_0, end_mask = var_3887_end_mask_0, x = cache_65_cast_fp16)[name = string("op_3887_cast_fp16")]; + bool var_3893_interleave_0 = const()[name = string("op_3893_interleave_0"), val = bool(false)]; + tensor var_3893_cast_fp16 = concat(axis = var_69, interleave = var_3893_interleave_0, values = (var_3887_cast_fp16, key_33_cast_fp16))[name = string("op_3893_cast_fp16")]; + tensor encoder_layers_16_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(329429888))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(330216384))))[name = string("encoder_layers_16_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_16_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_16_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(330216576)))]; + tensor linear_147_cast_fp16 = linear(bias = encoder_layers_16_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_16_self_attn_linear_q_weight_to_fp16_palettized, x = key_33_cast_fp16)[name = string("linear_147_cast_fp16")]; + tensor var_3898 = const()[name = string("op_3898"), val = tensor([1, -1, 8, 128])]; + tensor q_97_cast_fp16 = reshape(shape = var_3898, x = linear_147_cast_fp16)[name = string("q_97_cast_fp16")]; + tensor encoder_layers_16_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(330218688))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331005184))))[name = string("encoder_layers_16_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_16_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_16_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331005376)))]; + tensor linear_148_cast_fp16 = linear(bias = encoder_layers_16_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_16_self_attn_linear_k_weight_to_fp16_palettized, x = input_871_cast_fp16)[name = string("linear_148_cast_fp16")]; + tensor var_3903 = const()[name = string("op_3903"), val = tensor([1, -1, 8, 128])]; + tensor k_65_cast_fp16 = reshape(shape = var_3903, x = linear_148_cast_fp16)[name = string("k_65_cast_fp16")]; + tensor encoder_layers_16_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331007488))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331793984))))[name = string("encoder_layers_16_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_16_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_16_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331794176)))]; + tensor linear_149_cast_fp16 = linear(bias = encoder_layers_16_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_16_self_attn_linear_v_weight_to_fp16_palettized, x = input_871_cast_fp16)[name = string("linear_149_cast_fp16")]; + tensor var_3908 = const()[name = string("op_3908"), val = tensor([1, -1, 8, 128])]; + tensor v_33_cast_fp16 = reshape(shape = var_3908, x = linear_149_cast_fp16)[name = string("v_33_cast_fp16")]; + tensor value_41_perm_0 = const()[name = string("value_41_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_16_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_16_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331796288)))]; + tensor var_3921_cast_fp16 = add(x = q_97_cast_fp16, y = encoder_layers_16_self_attn_pos_bias_u_to_fp16)[name = string("op_3921_cast_fp16")]; + tensor encoder_layers_16_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_16_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331798400)))]; + tensor var_3923_cast_fp16 = add(x = q_97_cast_fp16, y = encoder_layers_16_self_attn_pos_bias_v_to_fp16)[name = string("op_3923_cast_fp16")]; + tensor q_with_bias_v_33_perm_0 = const()[name = string("q_with_bias_v_33_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_423_transpose_x_0 = const()[name = string("x_423_transpose_x_0"), val = bool(false)]; + bool x_423_transpose_y_0 = const()[name = string("x_423_transpose_y_0"), val = bool(false)]; + tensor op_3925_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331800512))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331914240))))[name = string("op_3925_to_fp16_quantized")]; + tensor q_with_bias_v_33_cast_fp16 = transpose(perm = q_with_bias_v_33_perm_0, x = var_3923_cast_fp16)[name = string("transpose_218")]; + tensor x_423_cast_fp16 = matmul(transpose_x = x_423_transpose_x_0, transpose_y = x_423_transpose_y_0, x = q_with_bias_v_33_cast_fp16, y = op_3925_to_fp16_quantized)[name = string("x_423_cast_fp16")]; + tensor x_425_pad_0 = const()[name = string("x_425_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_425_mode_0 = const()[name = string("x_425_mode_0"), val = string("constant")]; + fp16 const_287_to_fp16 = const()[name = string("const_287_to_fp16"), val = fp16(0x0p+0)]; + tensor x_425_cast_fp16 = pad(constant_val = const_287_to_fp16, mode = x_425_mode_0, pad = x_425_pad_0, x = x_423_cast_fp16)[name = string("x_425_cast_fp16")]; + tensor var_3933 = const()[name = string("op_3933"), val = tensor([1, 8, -1, 14])]; + tensor x_427_cast_fp16 = reshape(shape = var_3933, x = x_425_cast_fp16)[name = string("x_427_cast_fp16")]; + tensor var_3937_begin_0 = const()[name = string("op_3937_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_3937_end_0 = const()[name = string("op_3937_end_0"), val = tensor([1, 8, 112, 14])]; + tensor var_3937_end_mask_0 = const()[name = string("op_3937_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_3937_cast_fp16 = slice_by_index(begin = var_3937_begin_0, end = var_3937_end_0, end_mask = var_3937_end_mask_0, x = x_427_cast_fp16)[name = string("op_3937_cast_fp16")]; + tensor var_3938 = const()[name = string("op_3938"), val = tensor([1, 8, 14, 111])]; + tensor matrix_bd_65_cast_fp16 = reshape(shape = var_3938, x = var_3937_cast_fp16)[name = string("matrix_bd_65_cast_fp16")]; + bool matrix_ac_33_transpose_x_0 = const()[name = string("matrix_ac_33_transpose_x_0"), val = bool(false)]; + bool matrix_ac_33_transpose_y_0 = const()[name = string("matrix_ac_33_transpose_y_0"), val = bool(false)]; + tensor transpose_128_perm_0 = const()[name = string("transpose_128_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_129_perm_0 = const()[name = string("transpose_129_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_129 = transpose(perm = transpose_129_perm_0, x = k_65_cast_fp16)[name = string("transpose_216")]; + tensor transpose_128 = transpose(perm = transpose_128_perm_0, x = var_3921_cast_fp16)[name = string("transpose_217")]; + tensor matrix_ac_33_cast_fp16 = matmul(transpose_x = matrix_ac_33_transpose_x_0, transpose_y = matrix_ac_33_transpose_y_0, x = transpose_128, y = transpose_129)[name = string("matrix_ac_33_cast_fp16")]; + tensor matrix_bd_67_begin_0 = const()[name = string("matrix_bd_67_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_67_end_0 = const()[name = string("matrix_bd_67_end_0"), val = tensor([1, 8, 14, 56])]; + tensor matrix_bd_67_end_mask_0 = const()[name = string("matrix_bd_67_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_67_cast_fp16 = slice_by_index(begin = matrix_bd_67_begin_0, end = matrix_bd_67_end_0, end_mask = matrix_bd_67_end_mask_0, x = matrix_bd_65_cast_fp16)[name = string("matrix_bd_67_cast_fp16")]; + tensor var_3947_cast_fp16 = add(x = matrix_ac_33_cast_fp16, y = matrix_bd_67_cast_fp16)[name = string("op_3947_cast_fp16")]; + fp16 _inversed_scores_65_y_0_to_fp16 = const()[name = string("_inversed_scores_65_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_65_cast_fp16 = mul(x = var_3947_cast_fp16, y = _inversed_scores_65_y_0_to_fp16)[name = string("_inversed_scores_65_cast_fp16")]; + tensor scores_67_cast_fp16 = select(a = var_46_to_fp16, b = _inversed_scores_65_cast_fp16, cond = mask_11)[name = string("scores_67_cast_fp16")]; + tensor var_3953_cast_fp16 = softmax(axis = var_60, x = scores_67_cast_fp16)[name = string("op_3953_cast_fp16")]; + tensor input_873_cast_fp16 = select(a = var_45_to_fp16, b = var_3953_cast_fp16, cond = mask_11)[name = string("input_873_cast_fp16")]; + bool x_429_transpose_x_0 = const()[name = string("x_429_transpose_x_0"), val = bool(false)]; + bool x_429_transpose_y_0 = const()[name = string("x_429_transpose_y_0"), val = bool(false)]; + tensor value_41_cast_fp16 = transpose(perm = value_41_perm_0, x = v_33_cast_fp16)[name = string("transpose_215")]; + tensor x_429_cast_fp16 = matmul(transpose_x = x_429_transpose_x_0, transpose_y = x_429_transpose_y_0, x = input_873_cast_fp16, y = value_41_cast_fp16)[name = string("x_429_cast_fp16")]; + tensor var_3957_perm_0 = const()[name = string("op_3957_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_3958 = const()[name = string("op_3958"), val = tensor([1, -1, 1024])]; + tensor var_3957_cast_fp16 = transpose(perm = var_3957_perm_0, x = x_429_cast_fp16)[name = string("transpose_214")]; + tensor input_875_cast_fp16 = reshape(shape = var_3958, x = var_3957_cast_fp16)[name = string("input_875_cast_fp16")]; + tensor encoder_layers_16_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331914560))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332701056))))[name = string("encoder_layers_16_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_16_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_16_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332701248)))]; + tensor linear_151_cast_fp16 = linear(bias = encoder_layers_16_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_16_self_attn_linear_out_weight_to_fp16_palettized, x = input_875_cast_fp16)[name = string("linear_151_cast_fp16")]; + tensor input_879_cast_fp16 = add(x = input_869_cast_fp16, y = linear_151_cast_fp16)[name = string("input_879_cast_fp16")]; + tensor x_433_axes_0 = const()[name = string("x_433_axes_0"), val = tensor([-1])]; + tensor encoder_layers_16_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_16_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332703360)))]; + tensor encoder_layers_16_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_16_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332705472)))]; + tensor x_433_cast_fp16 = layer_norm(axes = x_433_axes_0, beta = encoder_layers_16_norm_conv_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_16_norm_conv_weight_to_fp16, x = input_879_cast_fp16)[name = string("x_433_cast_fp16")]; + tensor input_881_perm_0 = const()[name = string("input_881_perm_0"), val = tensor([0, 2, 1])]; + string input_883_pad_type_0 = const()[name = string("input_883_pad_type_0"), val = string("valid")]; + tensor input_883_strides_0 = const()[name = string("input_883_strides_0"), val = tensor([1])]; + tensor input_883_pad_0 = const()[name = string("input_883_pad_0"), val = tensor([0, 0])]; + tensor input_883_dilations_0 = const()[name = string("input_883_dilations_0"), val = tensor([1])]; + int32 input_883_groups_0 = const()[name = string("input_883_groups_0"), val = int32(1)]; + tensor encoder_layers_16_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332707584))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(334804800))))[name = string("encoder_layers_16_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_881_cast_fp16 = transpose(perm = input_881_perm_0, x = x_433_cast_fp16)[name = string("transpose_213")]; + tensor input_883_cast_fp16 = conv(dilations = input_883_dilations_0, groups = input_883_groups_0, pad = input_883_pad_0, pad_type = input_883_pad_type_0, strides = input_883_strides_0, weight = encoder_layers_16_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_881_cast_fp16)[name = string("input_883_cast_fp16")]; + int32 x_435_split_num_splits_0 = const()[name = string("x_435_split_num_splits_0"), val = int32(2)]; + int32 x_435_split_axis_0 = const()[name = string("x_435_split_axis_0"), val = int32(1)]; + tensor x_435_split_cast_fp16_0, tensor x_435_split_cast_fp16_1 = split(axis = x_435_split_axis_0, num_splits = x_435_split_num_splits_0, x = input_883_cast_fp16)[name = string("x_435_split_cast_fp16")]; + tensor x_435_split_1_sigmoid_cast_fp16 = sigmoid(x = x_435_split_cast_fp16_1)[name = string("x_435_split_1_sigmoid_cast_fp16")]; + tensor x_435_cast_fp16 = mul(x = x_435_split_cast_fp16_0, y = x_435_split_1_sigmoid_cast_fp16)[name = string("x_435_cast_fp16")]; + tensor input_885_cast_fp16 = select(a = var_45_to_fp16, b = x_435_cast_fp16, cond = var_576)[name = string("input_885_cast_fp16")]; + bool new_x_67_interleave_0 = const()[name = string("new_x_67_interleave_0"), val = bool(false)]; + tensor new_x_67_cast_fp16 = concat(axis = var_60, interleave = new_x_67_interleave_0, values = (cache_67_cast_fp16, input_885_cast_fp16))[name = string("new_x_67_cast_fp16")]; + tensor var_3997_begin_0 = const()[name = string("op_3997_begin_0"), val = tensor([0, 0, 14])]; + tensor var_3997_end_0 = const()[name = string("op_3997_end_0"), val = tensor([1, 1024, 22])]; + tensor var_3997_end_mask_0 = const()[name = string("op_3997_end_mask_0"), val = tensor([true, true, true])]; + tensor var_3997_cast_fp16 = slice_by_index(begin = var_3997_begin_0, end = var_3997_end_0, end_mask = var_3997_end_mask_0, x = new_x_67_cast_fp16)[name = string("op_3997_cast_fp16")]; + string x_437_pad_type_0 = const()[name = string("x_437_pad_type_0"), val = string("valid")]; + int32 x_437_groups_0 = const()[name = string("x_437_groups_0"), val = int32(1024)]; + tensor x_437_strides_0 = const()[name = string("x_437_strides_0"), val = tensor([1])]; + tensor x_437_pad_0 = const()[name = string("x_437_pad_0"), val = tensor([0, 0])]; + tensor x_437_dilations_0 = const()[name = string("x_437_dilations_0"), val = tensor([1])]; + tensor encoder_layers_16_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(334808960))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(334818240))))[name = string("encoder_layers_16_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_437_cast_fp16 = conv(dilations = x_437_dilations_0, groups = x_437_groups_0, pad = x_437_pad_0, pad_type = x_437_pad_type_0, strides = x_437_strides_0, weight = encoder_layers_16_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_67_cast_fp16)[name = string("x_437_cast_fp16")]; + tensor input_887_perm_0 = const()[name = string("input_887_perm_0"), val = tensor([0, 2, 1])]; + tensor x_439_axes_0 = const()[name = string("x_439_axes_0"), val = tensor([-1])]; + tensor encoder_layers_16_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_16_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(334820352)))]; + tensor encoder_layers_16_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_16_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(334822464)))]; + tensor input_887_cast_fp16 = transpose(perm = input_887_perm_0, x = x_437_cast_fp16)[name = string("transpose_212")]; + tensor x_439_cast_fp16 = layer_norm(axes = x_439_axes_0, beta = encoder_layers_16_conv_batch_norm_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_16_conv_batch_norm_weight_to_fp16, x = input_887_cast_fp16)[name = string("x_439_cast_fp16")]; + tensor input_889_perm_0 = const()[name = string("input_889_perm_0"), val = tensor([0, 2, 1])]; + tensor input_889_cast_fp16 = transpose(perm = input_889_perm_0, x = x_439_cast_fp16)[name = string("transpose_211")]; + tensor input_891_cast_fp16 = silu(x = input_889_cast_fp16)[name = string("input_891_cast_fp16")]; + string x_441_pad_type_0 = const()[name = string("x_441_pad_type_0"), val = string("valid")]; + tensor x_441_strides_0 = const()[name = string("x_441_strides_0"), val = tensor([1])]; + tensor x_441_pad_0 = const()[name = string("x_441_pad_0"), val = tensor([0, 0])]; + tensor x_441_dilations_0 = const()[name = string("x_441_dilations_0"), val = tensor([1])]; + int32 x_441_groups_0 = const()[name = string("x_441_groups_0"), val = int32(1)]; + tensor encoder_layers_16_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(334824576))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(335873216))))[name = string("encoder_layers_16_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_441_cast_fp16 = conv(dilations = x_441_dilations_0, groups = x_441_groups_0, pad = x_441_pad_0, pad_type = x_441_pad_type_0, strides = x_441_strides_0, weight = encoder_layers_16_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_891_cast_fp16)[name = string("x_441_cast_fp16")]; + tensor input_893_perm_0 = const()[name = string("input_893_perm_0"), val = tensor([0, 2, 1])]; + tensor input_893_cast_fp16 = transpose(perm = input_893_perm_0, x = x_441_cast_fp16)[name = string("transpose_210")]; + tensor input_895_cast_fp16 = add(x = input_879_cast_fp16, y = input_893_cast_fp16)[name = string("input_895_cast_fp16")]; + tensor input_897_axes_0 = const()[name = string("input_897_axes_0"), val = tensor([-1])]; + tensor encoder_layers_16_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_16_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(335875328)))]; + tensor encoder_layers_16_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_16_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(335877440)))]; + tensor input_897_cast_fp16 = layer_norm(axes = input_897_axes_0, beta = encoder_layers_16_norm_feed_forward2_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_16_norm_feed_forward2_weight_to_fp16, x = input_895_cast_fp16)[name = string("input_897_cast_fp16")]; + tensor encoder_layers_16_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(335879552))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(339025344))))[name = string("encoder_layers_16_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_16_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_16_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(339025536)))]; + tensor linear_152_cast_fp16 = linear(bias = encoder_layers_16_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_16_feed_forward2_linear1_weight_to_fp16_palettized, x = input_897_cast_fp16)[name = string("linear_152_cast_fp16")]; + tensor input_901_cast_fp16 = silu(x = linear_152_cast_fp16)[name = string("input_901_cast_fp16")]; + tensor encoder_layers_16_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(339033792))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(342179584))))[name = string("encoder_layers_16_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_16_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_16_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(342179776)))]; + tensor linear_153_cast_fp16 = linear(bias = encoder_layers_16_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_16_feed_forward2_linear2_weight_to_fp16_palettized, x = input_901_cast_fp16)[name = string("linear_153_cast_fp16")]; + fp16 var_4040_to_fp16 = const()[name = string("op_4040_to_fp16"), val = fp16(0x1p-1)]; + tensor var_4041_cast_fp16 = mul(x = linear_153_cast_fp16, y = var_4040_to_fp16)[name = string("op_4041_cast_fp16")]; + tensor input_907_cast_fp16 = add(x = input_895_cast_fp16, y = var_4041_cast_fp16)[name = string("input_907_cast_fp16")]; + tensor input_909_axes_0 = const()[name = string("input_909_axes_0"), val = tensor([-1])]; + tensor encoder_layers_16_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_16_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(342181888)))]; + tensor encoder_layers_16_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_16_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(342184000)))]; + tensor input_909_cast_fp16 = layer_norm(axes = input_909_axes_0, beta = encoder_layers_16_norm_out_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_16_norm_out_weight_to_fp16, x = input_907_cast_fp16)[name = string("input_909_cast_fp16")]; + tensor cache_69_begin_0 = const()[name = string("cache_69_begin_0"), val = tensor([17, 0, 0, 0])]; + tensor cache_69_end_0 = const()[name = string("cache_69_end_0"), val = tensor([18, 1, 42, 1024])]; + tensor cache_69_end_mask_0 = const()[name = string("cache_69_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_69_squeeze_mask_0 = const()[name = string("cache_69_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_69_cast_fp16 = slice_by_index(begin = cache_69_begin_0, end = cache_69_end_0, end_mask = cache_69_end_mask_0, squeeze_mask = cache_69_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_69_cast_fp16")]; + tensor cache_71_begin_0 = const()[name = string("cache_71_begin_0"), val = tensor([17, 0, 0, 0])]; + tensor cache_71_end_0 = const()[name = string("cache_71_end_0"), val = tensor([18, 1, 1024, 8])]; + tensor cache_71_end_mask_0 = const()[name = string("cache_71_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_71_squeeze_mask_0 = const()[name = string("cache_71_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_71_cast_fp16 = slice_by_index(begin = cache_71_begin_0, end = cache_71_end_0, end_mask = cache_71_end_mask_0, squeeze_mask = cache_71_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_71_cast_fp16")]; + tensor input_911_axes_0 = const()[name = string("input_911_axes_0"), val = tensor([-1])]; + tensor encoder_layers_17_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_17_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(342186112)))]; + tensor encoder_layers_17_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_17_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(342188224)))]; + tensor input_911_cast_fp16 = layer_norm(axes = input_911_axes_0, beta = encoder_layers_17_norm_feed_forward1_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_17_norm_feed_forward1_weight_to_fp16, x = input_909_cast_fp16)[name = string("input_911_cast_fp16")]; + tensor encoder_layers_17_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(342190336))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(345336128))))[name = string("encoder_layers_17_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_17_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_17_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(345336320)))]; + tensor linear_154_cast_fp16 = linear(bias = encoder_layers_17_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_17_feed_forward1_linear1_weight_to_fp16_palettized, x = input_911_cast_fp16)[name = string("linear_154_cast_fp16")]; + tensor input_915_cast_fp16 = silu(x = linear_154_cast_fp16)[name = string("input_915_cast_fp16")]; + tensor encoder_layers_17_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(345344576))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(348490368))))[name = string("encoder_layers_17_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_17_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_17_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(348490560)))]; + tensor linear_155_cast_fp16 = linear(bias = encoder_layers_17_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_17_feed_forward1_linear2_weight_to_fp16_palettized, x = input_915_cast_fp16)[name = string("linear_155_cast_fp16")]; + fp16 var_4077_to_fp16 = const()[name = string("op_4077_to_fp16"), val = fp16(0x1p-1)]; + tensor var_4078_cast_fp16 = mul(x = linear_155_cast_fp16, y = var_4077_to_fp16)[name = string("op_4078_cast_fp16")]; + tensor input_921_cast_fp16 = add(x = input_909_cast_fp16, y = var_4078_cast_fp16)[name = string("input_921_cast_fp16")]; + tensor key_35_axes_0 = const()[name = string("key_35_axes_0"), val = tensor([-1])]; + tensor encoder_layers_17_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_17_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(348492672)))]; + tensor encoder_layers_17_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_17_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(348494784)))]; + tensor key_35_cast_fp16 = layer_norm(axes = key_35_axes_0, beta = encoder_layers_17_norm_self_att_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_17_norm_self_att_weight_to_fp16, x = input_921_cast_fp16)[name = string("key_35_cast_fp16")]; + bool input_923_interleave_0 = const()[name = string("input_923_interleave_0"), val = bool(false)]; + tensor input_923_cast_fp16 = concat(axis = var_69, interleave = input_923_interleave_0, values = (cache_69_cast_fp16, key_35_cast_fp16))[name = string("input_923_cast_fp16")]; + tensor var_4100_begin_0 = const()[name = string("op_4100_begin_0"), val = tensor([0, 14, 0])]; + tensor var_4100_end_0 = const()[name = string("op_4100_end_0"), val = tensor([1, 42, 1024])]; + tensor var_4100_end_mask_0 = const()[name = string("op_4100_end_mask_0"), val = tensor([true, true, true])]; + tensor var_4100_cast_fp16 = slice_by_index(begin = var_4100_begin_0, end = var_4100_end_0, end_mask = var_4100_end_mask_0, x = cache_69_cast_fp16)[name = string("op_4100_cast_fp16")]; + bool var_4106_interleave_0 = const()[name = string("op_4106_interleave_0"), val = bool(false)]; + tensor var_4106_cast_fp16 = concat(axis = var_69, interleave = var_4106_interleave_0, values = (var_4100_cast_fp16, key_35_cast_fp16))[name = string("op_4106_cast_fp16")]; + tensor encoder_layers_17_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(348496896))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(349283392))))[name = string("encoder_layers_17_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_17_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_17_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(349283584)))]; + tensor linear_156_cast_fp16 = linear(bias = encoder_layers_17_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_17_self_attn_linear_q_weight_to_fp16_palettized, x = key_35_cast_fp16)[name = string("linear_156_cast_fp16")]; + tensor var_4111 = const()[name = string("op_4111"), val = tensor([1, -1, 8, 128])]; + tensor q_103_cast_fp16 = reshape(shape = var_4111, x = linear_156_cast_fp16)[name = string("q_103_cast_fp16")]; + tensor encoder_layers_17_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(349285696))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350072192))))[name = string("encoder_layers_17_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_17_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_17_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350072384)))]; + tensor linear_157_cast_fp16 = linear(bias = encoder_layers_17_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_17_self_attn_linear_k_weight_to_fp16_palettized, x = input_923_cast_fp16)[name = string("linear_157_cast_fp16")]; + tensor var_4116 = const()[name = string("op_4116"), val = tensor([1, -1, 8, 128])]; + tensor k_69_cast_fp16 = reshape(shape = var_4116, x = linear_157_cast_fp16)[name = string("k_69_cast_fp16")]; + tensor encoder_layers_17_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350074496))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350860992))))[name = string("encoder_layers_17_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_17_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_17_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350861184)))]; + tensor linear_158_cast_fp16 = linear(bias = encoder_layers_17_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_17_self_attn_linear_v_weight_to_fp16_palettized, x = input_923_cast_fp16)[name = string("linear_158_cast_fp16")]; + tensor var_4121 = const()[name = string("op_4121"), val = tensor([1, -1, 8, 128])]; + tensor v_35_cast_fp16 = reshape(shape = var_4121, x = linear_158_cast_fp16)[name = string("v_35_cast_fp16")]; + tensor value_43_perm_0 = const()[name = string("value_43_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_17_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_17_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350863296)))]; + tensor var_4134_cast_fp16 = add(x = q_103_cast_fp16, y = encoder_layers_17_self_attn_pos_bias_u_to_fp16)[name = string("op_4134_cast_fp16")]; + tensor encoder_layers_17_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_17_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350865408)))]; + tensor var_4136_cast_fp16 = add(x = q_103_cast_fp16, y = encoder_layers_17_self_attn_pos_bias_v_to_fp16)[name = string("op_4136_cast_fp16")]; + tensor q_with_bias_v_35_perm_0 = const()[name = string("q_with_bias_v_35_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_449_transpose_x_0 = const()[name = string("x_449_transpose_x_0"), val = bool(false)]; + bool x_449_transpose_y_0 = const()[name = string("x_449_transpose_y_0"), val = bool(false)]; + tensor op_4138_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350867520))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350981248))))[name = string("op_4138_to_fp16_quantized")]; + tensor q_with_bias_v_35_cast_fp16 = transpose(perm = q_with_bias_v_35_perm_0, x = var_4136_cast_fp16)[name = string("transpose_209")]; + tensor x_449_cast_fp16 = matmul(transpose_x = x_449_transpose_x_0, transpose_y = x_449_transpose_y_0, x = q_with_bias_v_35_cast_fp16, y = op_4138_to_fp16_quantized)[name = string("x_449_cast_fp16")]; + tensor x_451_pad_0 = const()[name = string("x_451_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_451_mode_0 = const()[name = string("x_451_mode_0"), val = string("constant")]; + fp16 const_300_to_fp16 = const()[name = string("const_300_to_fp16"), val = fp16(0x0p+0)]; + tensor x_451_cast_fp16 = pad(constant_val = const_300_to_fp16, mode = x_451_mode_0, pad = x_451_pad_0, x = x_449_cast_fp16)[name = string("x_451_cast_fp16")]; + tensor var_4146 = const()[name = string("op_4146"), val = tensor([1, 8, -1, 14])]; + tensor x_453_cast_fp16 = reshape(shape = var_4146, x = x_451_cast_fp16)[name = string("x_453_cast_fp16")]; + tensor var_4150_begin_0 = const()[name = string("op_4150_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_4150_end_0 = const()[name = string("op_4150_end_0"), val = tensor([1, 8, 112, 14])]; + tensor var_4150_end_mask_0 = const()[name = string("op_4150_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_4150_cast_fp16 = slice_by_index(begin = var_4150_begin_0, end = var_4150_end_0, end_mask = var_4150_end_mask_0, x = x_453_cast_fp16)[name = string("op_4150_cast_fp16")]; + tensor var_4151 = const()[name = string("op_4151"), val = tensor([1, 8, 14, 111])]; + tensor matrix_bd_69_cast_fp16 = reshape(shape = var_4151, x = var_4150_cast_fp16)[name = string("matrix_bd_69_cast_fp16")]; + bool matrix_ac_35_transpose_x_0 = const()[name = string("matrix_ac_35_transpose_x_0"), val = bool(false)]; + bool matrix_ac_35_transpose_y_0 = const()[name = string("matrix_ac_35_transpose_y_0"), val = bool(false)]; + tensor transpose_130_perm_0 = const()[name = string("transpose_130_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_131_perm_0 = const()[name = string("transpose_131_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_131 = transpose(perm = transpose_131_perm_0, x = k_69_cast_fp16)[name = string("transpose_207")]; + tensor transpose_130 = transpose(perm = transpose_130_perm_0, x = var_4134_cast_fp16)[name = string("transpose_208")]; + tensor matrix_ac_35_cast_fp16 = matmul(transpose_x = matrix_ac_35_transpose_x_0, transpose_y = matrix_ac_35_transpose_y_0, x = transpose_130, y = transpose_131)[name = string("matrix_ac_35_cast_fp16")]; + tensor matrix_bd_71_begin_0 = const()[name = string("matrix_bd_71_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_71_end_0 = const()[name = string("matrix_bd_71_end_0"), val = tensor([1, 8, 14, 56])]; + tensor matrix_bd_71_end_mask_0 = const()[name = string("matrix_bd_71_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_71_cast_fp16 = slice_by_index(begin = matrix_bd_71_begin_0, end = matrix_bd_71_end_0, end_mask = matrix_bd_71_end_mask_0, x = matrix_bd_69_cast_fp16)[name = string("matrix_bd_71_cast_fp16")]; + tensor var_4160_cast_fp16 = add(x = matrix_ac_35_cast_fp16, y = matrix_bd_71_cast_fp16)[name = string("op_4160_cast_fp16")]; + fp16 _inversed_scores_69_y_0_to_fp16 = const()[name = string("_inversed_scores_69_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_69_cast_fp16 = mul(x = var_4160_cast_fp16, y = _inversed_scores_69_y_0_to_fp16)[name = string("_inversed_scores_69_cast_fp16")]; + tensor scores_71_cast_fp16 = select(a = var_46_to_fp16, b = _inversed_scores_69_cast_fp16, cond = mask_11)[name = string("scores_71_cast_fp16")]; + tensor var_4166_cast_fp16 = softmax(axis = var_60, x = scores_71_cast_fp16)[name = string("op_4166_cast_fp16")]; + tensor input_925_cast_fp16 = select(a = var_45_to_fp16, b = var_4166_cast_fp16, cond = mask_11)[name = string("input_925_cast_fp16")]; + bool x_455_transpose_x_0 = const()[name = string("x_455_transpose_x_0"), val = bool(false)]; + bool x_455_transpose_y_0 = const()[name = string("x_455_transpose_y_0"), val = bool(false)]; + tensor value_43_cast_fp16 = transpose(perm = value_43_perm_0, x = v_35_cast_fp16)[name = string("transpose_206")]; + tensor x_455_cast_fp16 = matmul(transpose_x = x_455_transpose_x_0, transpose_y = x_455_transpose_y_0, x = input_925_cast_fp16, y = value_43_cast_fp16)[name = string("x_455_cast_fp16")]; + tensor var_4170_perm_0 = const()[name = string("op_4170_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_4171 = const()[name = string("op_4171"), val = tensor([1, -1, 1024])]; + tensor var_4170_cast_fp16 = transpose(perm = var_4170_perm_0, x = x_455_cast_fp16)[name = string("transpose_205")]; + tensor input_927_cast_fp16 = reshape(shape = var_4171, x = var_4170_cast_fp16)[name = string("input_927_cast_fp16")]; + tensor encoder_layers_17_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350981568))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351768064))))[name = string("encoder_layers_17_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_17_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_17_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351768256)))]; + tensor linear_160_cast_fp16 = linear(bias = encoder_layers_17_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_17_self_attn_linear_out_weight_to_fp16_palettized, x = input_927_cast_fp16)[name = string("linear_160_cast_fp16")]; + tensor input_931_cast_fp16 = add(x = input_921_cast_fp16, y = linear_160_cast_fp16)[name = string("input_931_cast_fp16")]; + tensor x_459_axes_0 = const()[name = string("x_459_axes_0"), val = tensor([-1])]; + tensor encoder_layers_17_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_17_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351770368)))]; + tensor encoder_layers_17_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_17_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351772480)))]; + tensor x_459_cast_fp16 = layer_norm(axes = x_459_axes_0, beta = encoder_layers_17_norm_conv_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_17_norm_conv_weight_to_fp16, x = input_931_cast_fp16)[name = string("x_459_cast_fp16")]; + tensor input_933_perm_0 = const()[name = string("input_933_perm_0"), val = tensor([0, 2, 1])]; + string input_935_pad_type_0 = const()[name = string("input_935_pad_type_0"), val = string("valid")]; + tensor input_935_strides_0 = const()[name = string("input_935_strides_0"), val = tensor([1])]; + tensor input_935_pad_0 = const()[name = string("input_935_pad_0"), val = tensor([0, 0])]; + tensor input_935_dilations_0 = const()[name = string("input_935_dilations_0"), val = tensor([1])]; + int32 input_935_groups_0 = const()[name = string("input_935_groups_0"), val = int32(1)]; + tensor encoder_layers_17_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351774592))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(353871808))))[name = string("encoder_layers_17_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_933_cast_fp16 = transpose(perm = input_933_perm_0, x = x_459_cast_fp16)[name = string("transpose_204")]; + tensor input_935_cast_fp16 = conv(dilations = input_935_dilations_0, groups = input_935_groups_0, pad = input_935_pad_0, pad_type = input_935_pad_type_0, strides = input_935_strides_0, weight = encoder_layers_17_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_933_cast_fp16)[name = string("input_935_cast_fp16")]; + int32 x_461_split_num_splits_0 = const()[name = string("x_461_split_num_splits_0"), val = int32(2)]; + int32 x_461_split_axis_0 = const()[name = string("x_461_split_axis_0"), val = int32(1)]; + tensor x_461_split_cast_fp16_0, tensor x_461_split_cast_fp16_1 = split(axis = x_461_split_axis_0, num_splits = x_461_split_num_splits_0, x = input_935_cast_fp16)[name = string("x_461_split_cast_fp16")]; + tensor x_461_split_1_sigmoid_cast_fp16 = sigmoid(x = x_461_split_cast_fp16_1)[name = string("x_461_split_1_sigmoid_cast_fp16")]; + tensor x_461_cast_fp16 = mul(x = x_461_split_cast_fp16_0, y = x_461_split_1_sigmoid_cast_fp16)[name = string("x_461_cast_fp16")]; + tensor input_937_cast_fp16 = select(a = var_45_to_fp16, b = x_461_cast_fp16, cond = var_576)[name = string("input_937_cast_fp16")]; + bool new_x_71_interleave_0 = const()[name = string("new_x_71_interleave_0"), val = bool(false)]; + tensor new_x_71_cast_fp16 = concat(axis = var_60, interleave = new_x_71_interleave_0, values = (cache_71_cast_fp16, input_937_cast_fp16))[name = string("new_x_71_cast_fp16")]; + tensor var_4210_begin_0 = const()[name = string("op_4210_begin_0"), val = tensor([0, 0, 14])]; + tensor var_4210_end_0 = const()[name = string("op_4210_end_0"), val = tensor([1, 1024, 22])]; + tensor var_4210_end_mask_0 = const()[name = string("op_4210_end_mask_0"), val = tensor([true, true, true])]; + tensor var_4210_cast_fp16 = slice_by_index(begin = var_4210_begin_0, end = var_4210_end_0, end_mask = var_4210_end_mask_0, x = new_x_71_cast_fp16)[name = string("op_4210_cast_fp16")]; + string x_463_pad_type_0 = const()[name = string("x_463_pad_type_0"), val = string("valid")]; + int32 x_463_groups_0 = const()[name = string("x_463_groups_0"), val = int32(1024)]; + tensor x_463_strides_0 = const()[name = string("x_463_strides_0"), val = tensor([1])]; + tensor x_463_pad_0 = const()[name = string("x_463_pad_0"), val = tensor([0, 0])]; + tensor x_463_dilations_0 = const()[name = string("x_463_dilations_0"), val = tensor([1])]; + tensor encoder_layers_17_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(353875968))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(353885248))))[name = string("encoder_layers_17_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_463_cast_fp16 = conv(dilations = x_463_dilations_0, groups = x_463_groups_0, pad = x_463_pad_0, pad_type = x_463_pad_type_0, strides = x_463_strides_0, weight = encoder_layers_17_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_71_cast_fp16)[name = string("x_463_cast_fp16")]; + tensor input_939_perm_0 = const()[name = string("input_939_perm_0"), val = tensor([0, 2, 1])]; + tensor x_465_axes_0 = const()[name = string("x_465_axes_0"), val = tensor([-1])]; + tensor encoder_layers_17_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_17_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(353887360)))]; + tensor encoder_layers_17_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_17_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(353889472)))]; + tensor input_939_cast_fp16 = transpose(perm = input_939_perm_0, x = x_463_cast_fp16)[name = string("transpose_203")]; + tensor x_465_cast_fp16 = layer_norm(axes = x_465_axes_0, beta = encoder_layers_17_conv_batch_norm_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_17_conv_batch_norm_weight_to_fp16, x = input_939_cast_fp16)[name = string("x_465_cast_fp16")]; + tensor input_941_perm_0 = const()[name = string("input_941_perm_0"), val = tensor([0, 2, 1])]; + tensor input_941_cast_fp16 = transpose(perm = input_941_perm_0, x = x_465_cast_fp16)[name = string("transpose_202")]; + tensor input_943_cast_fp16 = silu(x = input_941_cast_fp16)[name = string("input_943_cast_fp16")]; + string x_467_pad_type_0 = const()[name = string("x_467_pad_type_0"), val = string("valid")]; + tensor x_467_strides_0 = const()[name = string("x_467_strides_0"), val = tensor([1])]; + tensor x_467_pad_0 = const()[name = string("x_467_pad_0"), val = tensor([0, 0])]; + tensor x_467_dilations_0 = const()[name = string("x_467_dilations_0"), val = tensor([1])]; + int32 x_467_groups_0 = const()[name = string("x_467_groups_0"), val = int32(1)]; + tensor encoder_layers_17_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(353891584))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(354940224))))[name = string("encoder_layers_17_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_467_cast_fp16 = conv(dilations = x_467_dilations_0, groups = x_467_groups_0, pad = x_467_pad_0, pad_type = x_467_pad_type_0, strides = x_467_strides_0, weight = encoder_layers_17_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_943_cast_fp16)[name = string("x_467_cast_fp16")]; + tensor input_945_perm_0 = const()[name = string("input_945_perm_0"), val = tensor([0, 2, 1])]; + tensor input_945_cast_fp16 = transpose(perm = input_945_perm_0, x = x_467_cast_fp16)[name = string("transpose_201")]; + tensor input_947_cast_fp16 = add(x = input_931_cast_fp16, y = input_945_cast_fp16)[name = string("input_947_cast_fp16")]; + tensor input_949_axes_0 = const()[name = string("input_949_axes_0"), val = tensor([-1])]; + tensor encoder_layers_17_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_17_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(354942336)))]; + tensor encoder_layers_17_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_17_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(354944448)))]; + tensor input_949_cast_fp16 = layer_norm(axes = input_949_axes_0, beta = encoder_layers_17_norm_feed_forward2_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_17_norm_feed_forward2_weight_to_fp16, x = input_947_cast_fp16)[name = string("input_949_cast_fp16")]; + tensor encoder_layers_17_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(354946560))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(358092352))))[name = string("encoder_layers_17_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_17_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_17_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(358092544)))]; + tensor linear_161_cast_fp16 = linear(bias = encoder_layers_17_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_17_feed_forward2_linear1_weight_to_fp16_palettized, x = input_949_cast_fp16)[name = string("linear_161_cast_fp16")]; + tensor input_953_cast_fp16 = silu(x = linear_161_cast_fp16)[name = string("input_953_cast_fp16")]; + tensor encoder_layers_17_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(358100800))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361246592))))[name = string("encoder_layers_17_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_17_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_17_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361246784)))]; + tensor linear_162_cast_fp16 = linear(bias = encoder_layers_17_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_17_feed_forward2_linear2_weight_to_fp16_palettized, x = input_953_cast_fp16)[name = string("linear_162_cast_fp16")]; + fp16 var_4253_to_fp16 = const()[name = string("op_4253_to_fp16"), val = fp16(0x1p-1)]; + tensor var_4254_cast_fp16 = mul(x = linear_162_cast_fp16, y = var_4253_to_fp16)[name = string("op_4254_cast_fp16")]; + tensor input_959_cast_fp16 = add(x = input_947_cast_fp16, y = var_4254_cast_fp16)[name = string("input_959_cast_fp16")]; + tensor input_961_axes_0 = const()[name = string("input_961_axes_0"), val = tensor([-1])]; + tensor encoder_layers_17_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_17_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361248896)))]; + tensor encoder_layers_17_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_17_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361251008)))]; + tensor input_961_cast_fp16 = layer_norm(axes = input_961_axes_0, beta = encoder_layers_17_norm_out_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_17_norm_out_weight_to_fp16, x = input_959_cast_fp16)[name = string("input_961_cast_fp16")]; + tensor cache_73_begin_0 = const()[name = string("cache_73_begin_0"), val = tensor([18, 0, 0, 0])]; + tensor cache_73_end_0 = const()[name = string("cache_73_end_0"), val = tensor([19, 1, 42, 1024])]; + tensor cache_73_end_mask_0 = const()[name = string("cache_73_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_73_squeeze_mask_0 = const()[name = string("cache_73_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_73_cast_fp16 = slice_by_index(begin = cache_73_begin_0, end = cache_73_end_0, end_mask = cache_73_end_mask_0, squeeze_mask = cache_73_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_73_cast_fp16")]; + tensor cache_75_begin_0 = const()[name = string("cache_75_begin_0"), val = tensor([18, 0, 0, 0])]; + tensor cache_75_end_0 = const()[name = string("cache_75_end_0"), val = tensor([19, 1, 1024, 8])]; + tensor cache_75_end_mask_0 = const()[name = string("cache_75_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_75_squeeze_mask_0 = const()[name = string("cache_75_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_75_cast_fp16 = slice_by_index(begin = cache_75_begin_0, end = cache_75_end_0, end_mask = cache_75_end_mask_0, squeeze_mask = cache_75_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_75_cast_fp16")]; + tensor input_963_axes_0 = const()[name = string("input_963_axes_0"), val = tensor([-1])]; + tensor encoder_layers_18_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_18_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361253120)))]; + tensor encoder_layers_18_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_18_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361255232)))]; + tensor input_963_cast_fp16 = layer_norm(axes = input_963_axes_0, beta = encoder_layers_18_norm_feed_forward1_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_18_norm_feed_forward1_weight_to_fp16, x = input_961_cast_fp16)[name = string("input_963_cast_fp16")]; + tensor encoder_layers_18_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361257344))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(364403136))))[name = string("encoder_layers_18_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_18_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_18_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(364403328)))]; + tensor linear_163_cast_fp16 = linear(bias = encoder_layers_18_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_18_feed_forward1_linear1_weight_to_fp16_palettized, x = input_963_cast_fp16)[name = string("linear_163_cast_fp16")]; + tensor input_967_cast_fp16 = silu(x = linear_163_cast_fp16)[name = string("input_967_cast_fp16")]; + tensor encoder_layers_18_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(364411584))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(367557376))))[name = string("encoder_layers_18_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_18_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_18_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(367557568)))]; + tensor linear_164_cast_fp16 = linear(bias = encoder_layers_18_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_18_feed_forward1_linear2_weight_to_fp16_palettized, x = input_967_cast_fp16)[name = string("linear_164_cast_fp16")]; + fp16 var_4290_to_fp16 = const()[name = string("op_4290_to_fp16"), val = fp16(0x1p-1)]; + tensor var_4291_cast_fp16 = mul(x = linear_164_cast_fp16, y = var_4290_to_fp16)[name = string("op_4291_cast_fp16")]; + tensor input_973_cast_fp16 = add(x = input_961_cast_fp16, y = var_4291_cast_fp16)[name = string("input_973_cast_fp16")]; + tensor key_37_axes_0 = const()[name = string("key_37_axes_0"), val = tensor([-1])]; + tensor encoder_layers_18_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_18_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(367559680)))]; + tensor encoder_layers_18_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_18_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(367561792)))]; + tensor key_37_cast_fp16 = layer_norm(axes = key_37_axes_0, beta = encoder_layers_18_norm_self_att_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_18_norm_self_att_weight_to_fp16, x = input_973_cast_fp16)[name = string("key_37_cast_fp16")]; + bool input_975_interleave_0 = const()[name = string("input_975_interleave_0"), val = bool(false)]; + tensor input_975_cast_fp16 = concat(axis = var_69, interleave = input_975_interleave_0, values = (cache_73_cast_fp16, key_37_cast_fp16))[name = string("input_975_cast_fp16")]; + tensor var_4313_begin_0 = const()[name = string("op_4313_begin_0"), val = tensor([0, 14, 0])]; + tensor var_4313_end_0 = const()[name = string("op_4313_end_0"), val = tensor([1, 42, 1024])]; + tensor var_4313_end_mask_0 = const()[name = string("op_4313_end_mask_0"), val = tensor([true, true, true])]; + tensor var_4313_cast_fp16 = slice_by_index(begin = var_4313_begin_0, end = var_4313_end_0, end_mask = var_4313_end_mask_0, x = cache_73_cast_fp16)[name = string("op_4313_cast_fp16")]; + bool var_4319_interleave_0 = const()[name = string("op_4319_interleave_0"), val = bool(false)]; + tensor var_4319_cast_fp16 = concat(axis = var_69, interleave = var_4319_interleave_0, values = (var_4313_cast_fp16, key_37_cast_fp16))[name = string("op_4319_cast_fp16")]; + tensor encoder_layers_18_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(367563904))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368350400))))[name = string("encoder_layers_18_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_18_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_18_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368350592)))]; + tensor linear_165_cast_fp16 = linear(bias = encoder_layers_18_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_18_self_attn_linear_q_weight_to_fp16_palettized, x = key_37_cast_fp16)[name = string("linear_165_cast_fp16")]; + tensor var_4324 = const()[name = string("op_4324"), val = tensor([1, -1, 8, 128])]; + tensor q_109_cast_fp16 = reshape(shape = var_4324, x = linear_165_cast_fp16)[name = string("q_109_cast_fp16")]; + tensor encoder_layers_18_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368352704))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369139200))))[name = string("encoder_layers_18_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_18_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_18_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369139392)))]; + tensor linear_166_cast_fp16 = linear(bias = encoder_layers_18_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_18_self_attn_linear_k_weight_to_fp16_palettized, x = input_975_cast_fp16)[name = string("linear_166_cast_fp16")]; + tensor var_4329 = const()[name = string("op_4329"), val = tensor([1, -1, 8, 128])]; + tensor k_73_cast_fp16 = reshape(shape = var_4329, x = linear_166_cast_fp16)[name = string("k_73_cast_fp16")]; + tensor encoder_layers_18_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369141504))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369928000))))[name = string("encoder_layers_18_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_18_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_18_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369928192)))]; + tensor linear_167_cast_fp16 = linear(bias = encoder_layers_18_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_18_self_attn_linear_v_weight_to_fp16_palettized, x = input_975_cast_fp16)[name = string("linear_167_cast_fp16")]; + tensor var_4334 = const()[name = string("op_4334"), val = tensor([1, -1, 8, 128])]; + tensor v_37_cast_fp16 = reshape(shape = var_4334, x = linear_167_cast_fp16)[name = string("v_37_cast_fp16")]; + tensor value_45_perm_0 = const()[name = string("value_45_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_18_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_18_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369930304)))]; + tensor var_4347_cast_fp16 = add(x = q_109_cast_fp16, y = encoder_layers_18_self_attn_pos_bias_u_to_fp16)[name = string("op_4347_cast_fp16")]; + tensor encoder_layers_18_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_18_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369932416)))]; + tensor var_4349_cast_fp16 = add(x = q_109_cast_fp16, y = encoder_layers_18_self_attn_pos_bias_v_to_fp16)[name = string("op_4349_cast_fp16")]; + tensor q_with_bias_v_37_perm_0 = const()[name = string("q_with_bias_v_37_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_475_transpose_x_0 = const()[name = string("x_475_transpose_x_0"), val = bool(false)]; + bool x_475_transpose_y_0 = const()[name = string("x_475_transpose_y_0"), val = bool(false)]; + tensor op_4351_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369934528))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(370048256))))[name = string("op_4351_to_fp16_quantized")]; + tensor q_with_bias_v_37_cast_fp16 = transpose(perm = q_with_bias_v_37_perm_0, x = var_4349_cast_fp16)[name = string("transpose_200")]; + tensor x_475_cast_fp16 = matmul(transpose_x = x_475_transpose_x_0, transpose_y = x_475_transpose_y_0, x = q_with_bias_v_37_cast_fp16, y = op_4351_to_fp16_quantized)[name = string("x_475_cast_fp16")]; + tensor x_477_pad_0 = const()[name = string("x_477_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_477_mode_0 = const()[name = string("x_477_mode_0"), val = string("constant")]; + fp16 const_313_to_fp16 = const()[name = string("const_313_to_fp16"), val = fp16(0x0p+0)]; + tensor x_477_cast_fp16 = pad(constant_val = const_313_to_fp16, mode = x_477_mode_0, pad = x_477_pad_0, x = x_475_cast_fp16)[name = string("x_477_cast_fp16")]; + tensor var_4359 = const()[name = string("op_4359"), val = tensor([1, 8, -1, 14])]; + tensor x_479_cast_fp16 = reshape(shape = var_4359, x = x_477_cast_fp16)[name = string("x_479_cast_fp16")]; + tensor var_4363_begin_0 = const()[name = string("op_4363_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_4363_end_0 = const()[name = string("op_4363_end_0"), val = tensor([1, 8, 112, 14])]; + tensor var_4363_end_mask_0 = const()[name = string("op_4363_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_4363_cast_fp16 = slice_by_index(begin = var_4363_begin_0, end = var_4363_end_0, end_mask = var_4363_end_mask_0, x = x_479_cast_fp16)[name = string("op_4363_cast_fp16")]; + tensor var_4364 = const()[name = string("op_4364"), val = tensor([1, 8, 14, 111])]; + tensor matrix_bd_73_cast_fp16 = reshape(shape = var_4364, x = var_4363_cast_fp16)[name = string("matrix_bd_73_cast_fp16")]; + bool matrix_ac_37_transpose_x_0 = const()[name = string("matrix_ac_37_transpose_x_0"), val = bool(false)]; + bool matrix_ac_37_transpose_y_0 = const()[name = string("matrix_ac_37_transpose_y_0"), val = bool(false)]; + tensor transpose_132_perm_0 = const()[name = string("transpose_132_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_133_perm_0 = const()[name = string("transpose_133_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_133 = transpose(perm = transpose_133_perm_0, x = k_73_cast_fp16)[name = string("transpose_198")]; + tensor transpose_132 = transpose(perm = transpose_132_perm_0, x = var_4347_cast_fp16)[name = string("transpose_199")]; + tensor matrix_ac_37_cast_fp16 = matmul(transpose_x = matrix_ac_37_transpose_x_0, transpose_y = matrix_ac_37_transpose_y_0, x = transpose_132, y = transpose_133)[name = string("matrix_ac_37_cast_fp16")]; + tensor matrix_bd_75_begin_0 = const()[name = string("matrix_bd_75_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_75_end_0 = const()[name = string("matrix_bd_75_end_0"), val = tensor([1, 8, 14, 56])]; + tensor matrix_bd_75_end_mask_0 = const()[name = string("matrix_bd_75_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_75_cast_fp16 = slice_by_index(begin = matrix_bd_75_begin_0, end = matrix_bd_75_end_0, end_mask = matrix_bd_75_end_mask_0, x = matrix_bd_73_cast_fp16)[name = string("matrix_bd_75_cast_fp16")]; + tensor var_4373_cast_fp16 = add(x = matrix_ac_37_cast_fp16, y = matrix_bd_75_cast_fp16)[name = string("op_4373_cast_fp16")]; + fp16 _inversed_scores_73_y_0_to_fp16 = const()[name = string("_inversed_scores_73_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_73_cast_fp16 = mul(x = var_4373_cast_fp16, y = _inversed_scores_73_y_0_to_fp16)[name = string("_inversed_scores_73_cast_fp16")]; + tensor scores_75_cast_fp16 = select(a = var_46_to_fp16, b = _inversed_scores_73_cast_fp16, cond = mask_11)[name = string("scores_75_cast_fp16")]; + tensor var_4379_cast_fp16 = softmax(axis = var_60, x = scores_75_cast_fp16)[name = string("op_4379_cast_fp16")]; + tensor input_977_cast_fp16 = select(a = var_45_to_fp16, b = var_4379_cast_fp16, cond = mask_11)[name = string("input_977_cast_fp16")]; + bool x_481_transpose_x_0 = const()[name = string("x_481_transpose_x_0"), val = bool(false)]; + bool x_481_transpose_y_0 = const()[name = string("x_481_transpose_y_0"), val = bool(false)]; + tensor value_45_cast_fp16 = transpose(perm = value_45_perm_0, x = v_37_cast_fp16)[name = string("transpose_197")]; + tensor x_481_cast_fp16 = matmul(transpose_x = x_481_transpose_x_0, transpose_y = x_481_transpose_y_0, x = input_977_cast_fp16, y = value_45_cast_fp16)[name = string("x_481_cast_fp16")]; + tensor var_4383_perm_0 = const()[name = string("op_4383_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_4384 = const()[name = string("op_4384"), val = tensor([1, -1, 1024])]; + tensor var_4383_cast_fp16 = transpose(perm = var_4383_perm_0, x = x_481_cast_fp16)[name = string("transpose_196")]; + tensor input_979_cast_fp16 = reshape(shape = var_4384, x = var_4383_cast_fp16)[name = string("input_979_cast_fp16")]; + tensor encoder_layers_18_self_attn_linear_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(370048576))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371097216))))[name = string("encoder_layers_18_self_attn_linear_out_weight_to_fp16_quantized")]; + tensor encoder_layers_18_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_18_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371099328)))]; + tensor linear_169_cast_fp16 = linear(bias = encoder_layers_18_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_18_self_attn_linear_out_weight_to_fp16_quantized, x = input_979_cast_fp16)[name = string("linear_169_cast_fp16")]; + tensor input_983_cast_fp16 = add(x = input_973_cast_fp16, y = linear_169_cast_fp16)[name = string("input_983_cast_fp16")]; + tensor x_485_axes_0 = const()[name = string("x_485_axes_0"), val = tensor([-1])]; + tensor encoder_layers_18_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_18_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371101440)))]; + tensor encoder_layers_18_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_18_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371103552)))]; + tensor x_485_cast_fp16 = layer_norm(axes = x_485_axes_0, beta = encoder_layers_18_norm_conv_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_18_norm_conv_weight_to_fp16, x = input_983_cast_fp16)[name = string("x_485_cast_fp16")]; + tensor input_985_perm_0 = const()[name = string("input_985_perm_0"), val = tensor([0, 2, 1])]; + string input_987_pad_type_0 = const()[name = string("input_987_pad_type_0"), val = string("valid")]; + tensor input_987_strides_0 = const()[name = string("input_987_strides_0"), val = tensor([1])]; + tensor input_987_pad_0 = const()[name = string("input_987_pad_0"), val = tensor([0, 0])]; + tensor input_987_dilations_0 = const()[name = string("input_987_dilations_0"), val = tensor([1])]; + int32 input_987_groups_0 = const()[name = string("input_987_groups_0"), val = int32(1)]; + tensor encoder_layers_18_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371105664))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(373202880))))[name = string("encoder_layers_18_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_985_cast_fp16 = transpose(perm = input_985_perm_0, x = x_485_cast_fp16)[name = string("transpose_195")]; + tensor input_987_cast_fp16 = conv(dilations = input_987_dilations_0, groups = input_987_groups_0, pad = input_987_pad_0, pad_type = input_987_pad_type_0, strides = input_987_strides_0, weight = encoder_layers_18_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_985_cast_fp16)[name = string("input_987_cast_fp16")]; + int32 x_487_split_num_splits_0 = const()[name = string("x_487_split_num_splits_0"), val = int32(2)]; + int32 x_487_split_axis_0 = const()[name = string("x_487_split_axis_0"), val = int32(1)]; + tensor x_487_split_cast_fp16_0, tensor x_487_split_cast_fp16_1 = split(axis = x_487_split_axis_0, num_splits = x_487_split_num_splits_0, x = input_987_cast_fp16)[name = string("x_487_split_cast_fp16")]; + tensor x_487_split_1_sigmoid_cast_fp16 = sigmoid(x = x_487_split_cast_fp16_1)[name = string("x_487_split_1_sigmoid_cast_fp16")]; + tensor x_487_cast_fp16 = mul(x = x_487_split_cast_fp16_0, y = x_487_split_1_sigmoid_cast_fp16)[name = string("x_487_cast_fp16")]; + tensor input_989_cast_fp16 = select(a = var_45_to_fp16, b = x_487_cast_fp16, cond = var_576)[name = string("input_989_cast_fp16")]; + bool new_x_75_interleave_0 = const()[name = string("new_x_75_interleave_0"), val = bool(false)]; + tensor new_x_75_cast_fp16 = concat(axis = var_60, interleave = new_x_75_interleave_0, values = (cache_75_cast_fp16, input_989_cast_fp16))[name = string("new_x_75_cast_fp16")]; + tensor var_4423_begin_0 = const()[name = string("op_4423_begin_0"), val = tensor([0, 0, 14])]; + tensor var_4423_end_0 = const()[name = string("op_4423_end_0"), val = tensor([1, 1024, 22])]; + tensor var_4423_end_mask_0 = const()[name = string("op_4423_end_mask_0"), val = tensor([true, true, true])]; + tensor var_4423_cast_fp16 = slice_by_index(begin = var_4423_begin_0, end = var_4423_end_0, end_mask = var_4423_end_mask_0, x = new_x_75_cast_fp16)[name = string("op_4423_cast_fp16")]; + string x_489_pad_type_0 = const()[name = string("x_489_pad_type_0"), val = string("valid")]; + int32 x_489_groups_0 = const()[name = string("x_489_groups_0"), val = int32(1024)]; + tensor x_489_strides_0 = const()[name = string("x_489_strides_0"), val = tensor([1])]; + tensor x_489_pad_0 = const()[name = string("x_489_pad_0"), val = tensor([0, 0])]; + tensor x_489_dilations_0 = const()[name = string("x_489_dilations_0"), val = tensor([1])]; + tensor encoder_layers_18_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(373207040))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(373216320))))[name = string("encoder_layers_18_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_489_cast_fp16 = conv(dilations = x_489_dilations_0, groups = x_489_groups_0, pad = x_489_pad_0, pad_type = x_489_pad_type_0, strides = x_489_strides_0, weight = encoder_layers_18_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_75_cast_fp16)[name = string("x_489_cast_fp16")]; + tensor input_991_perm_0 = const()[name = string("input_991_perm_0"), val = tensor([0, 2, 1])]; + tensor x_491_axes_0 = const()[name = string("x_491_axes_0"), val = tensor([-1])]; + tensor encoder_layers_18_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_18_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(373218432)))]; + tensor encoder_layers_18_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_18_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(373220544)))]; + tensor input_991_cast_fp16 = transpose(perm = input_991_perm_0, x = x_489_cast_fp16)[name = string("transpose_194")]; + tensor x_491_cast_fp16 = layer_norm(axes = x_491_axes_0, beta = encoder_layers_18_conv_batch_norm_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_18_conv_batch_norm_weight_to_fp16, x = input_991_cast_fp16)[name = string("x_491_cast_fp16")]; + tensor input_993_perm_0 = const()[name = string("input_993_perm_0"), val = tensor([0, 2, 1])]; + tensor input_993_cast_fp16 = transpose(perm = input_993_perm_0, x = x_491_cast_fp16)[name = string("transpose_193")]; + tensor input_995_cast_fp16 = silu(x = input_993_cast_fp16)[name = string("input_995_cast_fp16")]; + string x_493_pad_type_0 = const()[name = string("x_493_pad_type_0"), val = string("valid")]; + tensor x_493_strides_0 = const()[name = string("x_493_strides_0"), val = tensor([1])]; + tensor x_493_pad_0 = const()[name = string("x_493_pad_0"), val = tensor([0, 0])]; + tensor x_493_dilations_0 = const()[name = string("x_493_dilations_0"), val = tensor([1])]; + int32 x_493_groups_0 = const()[name = string("x_493_groups_0"), val = int32(1)]; + tensor encoder_layers_18_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(373222656))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(374271296))))[name = string("encoder_layers_18_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_493_cast_fp16 = conv(dilations = x_493_dilations_0, groups = x_493_groups_0, pad = x_493_pad_0, pad_type = x_493_pad_type_0, strides = x_493_strides_0, weight = encoder_layers_18_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_995_cast_fp16)[name = string("x_493_cast_fp16")]; + tensor input_997_perm_0 = const()[name = string("input_997_perm_0"), val = tensor([0, 2, 1])]; + tensor input_997_cast_fp16 = transpose(perm = input_997_perm_0, x = x_493_cast_fp16)[name = string("transpose_192")]; + tensor input_999_cast_fp16 = add(x = input_983_cast_fp16, y = input_997_cast_fp16)[name = string("input_999_cast_fp16")]; + tensor input_1001_axes_0 = const()[name = string("input_1001_axes_0"), val = tensor([-1])]; + tensor encoder_layers_18_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_18_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(374273408)))]; + tensor encoder_layers_18_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_18_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(374275520)))]; + tensor input_1001_cast_fp16 = layer_norm(axes = input_1001_axes_0, beta = encoder_layers_18_norm_feed_forward2_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_18_norm_feed_forward2_weight_to_fp16, x = input_999_cast_fp16)[name = string("input_1001_cast_fp16")]; + tensor encoder_layers_18_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(374277632))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(378472000))))[name = string("encoder_layers_18_feed_forward2_linear1_weight_to_fp16_quantized")]; + tensor encoder_layers_18_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_18_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(378480256)))]; + tensor linear_170_cast_fp16 = linear(bias = encoder_layers_18_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_18_feed_forward2_linear1_weight_to_fp16_quantized, x = input_1001_cast_fp16)[name = string("linear_170_cast_fp16")]; + tensor input_1005_cast_fp16 = silu(x = linear_170_cast_fp16)[name = string("input_1005_cast_fp16")]; + tensor encoder_layers_18_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(378488512))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(382682880))))[name = string("encoder_layers_18_feed_forward2_linear2_weight_to_fp16_quantized")]; + tensor encoder_layers_18_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_18_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(382684992)))]; + tensor linear_171_cast_fp16 = linear(bias = encoder_layers_18_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_18_feed_forward2_linear2_weight_to_fp16_quantized, x = input_1005_cast_fp16)[name = string("linear_171_cast_fp16")]; + fp16 var_4466_to_fp16 = const()[name = string("op_4466_to_fp16"), val = fp16(0x1p-1)]; + tensor var_4467_cast_fp16 = mul(x = linear_171_cast_fp16, y = var_4466_to_fp16)[name = string("op_4467_cast_fp16")]; + tensor input_1011_cast_fp16 = add(x = input_999_cast_fp16, y = var_4467_cast_fp16)[name = string("input_1011_cast_fp16")]; + tensor input_1013_axes_0 = const()[name = string("input_1013_axes_0"), val = tensor([-1])]; + tensor encoder_layers_18_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_18_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(382687104)))]; + tensor encoder_layers_18_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_18_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(382689216)))]; + tensor input_1013_cast_fp16 = layer_norm(axes = input_1013_axes_0, beta = encoder_layers_18_norm_out_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_18_norm_out_weight_to_fp16, x = input_1011_cast_fp16)[name = string("input_1013_cast_fp16")]; + tensor cache_77_begin_0 = const()[name = string("cache_77_begin_0"), val = tensor([19, 0, 0, 0])]; + tensor cache_77_end_0 = const()[name = string("cache_77_end_0"), val = tensor([20, 1, 42, 1024])]; + tensor cache_77_end_mask_0 = const()[name = string("cache_77_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_77_squeeze_mask_0 = const()[name = string("cache_77_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_77_cast_fp16 = slice_by_index(begin = cache_77_begin_0, end = cache_77_end_0, end_mask = cache_77_end_mask_0, squeeze_mask = cache_77_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_77_cast_fp16")]; + tensor cache_79_begin_0 = const()[name = string("cache_79_begin_0"), val = tensor([19, 0, 0, 0])]; + tensor cache_79_end_0 = const()[name = string("cache_79_end_0"), val = tensor([20, 1, 1024, 8])]; + tensor cache_79_end_mask_0 = const()[name = string("cache_79_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_79_squeeze_mask_0 = const()[name = string("cache_79_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_79_cast_fp16 = slice_by_index(begin = cache_79_begin_0, end = cache_79_end_0, end_mask = cache_79_end_mask_0, squeeze_mask = cache_79_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_79_cast_fp16")]; + tensor input_1015_axes_0 = const()[name = string("input_1015_axes_0"), val = tensor([-1])]; + tensor encoder_layers_19_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_19_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(382691328)))]; + tensor encoder_layers_19_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_19_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(382693440)))]; + tensor input_1015_cast_fp16 = layer_norm(axes = input_1015_axes_0, beta = encoder_layers_19_norm_feed_forward1_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_19_norm_feed_forward1_weight_to_fp16, x = input_1013_cast_fp16)[name = string("input_1015_cast_fp16")]; + tensor encoder_layers_19_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(382695552))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(386889920))))[name = string("encoder_layers_19_feed_forward1_linear1_weight_to_fp16_quantized")]; + tensor encoder_layers_19_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_19_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(386898176)))]; + tensor linear_172_cast_fp16 = linear(bias = encoder_layers_19_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_19_feed_forward1_linear1_weight_to_fp16_quantized, x = input_1015_cast_fp16)[name = string("linear_172_cast_fp16")]; + tensor input_1019_cast_fp16 = silu(x = linear_172_cast_fp16)[name = string("input_1019_cast_fp16")]; + tensor encoder_layers_19_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(386906432))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391100800))))[name = string("encoder_layers_19_feed_forward1_linear2_weight_to_fp16_quantized")]; + tensor encoder_layers_19_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_19_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391102912)))]; + tensor linear_173_cast_fp16 = linear(bias = encoder_layers_19_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_19_feed_forward1_linear2_weight_to_fp16_quantized, x = input_1019_cast_fp16)[name = string("linear_173_cast_fp16")]; + fp16 var_4503_to_fp16 = const()[name = string("op_4503_to_fp16"), val = fp16(0x1p-1)]; + tensor var_4504_cast_fp16 = mul(x = linear_173_cast_fp16, y = var_4503_to_fp16)[name = string("op_4504_cast_fp16")]; + tensor input_1025_cast_fp16 = add(x = input_1013_cast_fp16, y = var_4504_cast_fp16)[name = string("input_1025_cast_fp16")]; + tensor key_39_axes_0 = const()[name = string("key_39_axes_0"), val = tensor([-1])]; + tensor encoder_layers_19_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_19_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391105024)))]; + tensor encoder_layers_19_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_19_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391107136)))]; + tensor key_39_cast_fp16 = layer_norm(axes = key_39_axes_0, beta = encoder_layers_19_norm_self_att_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_19_norm_self_att_weight_to_fp16, x = input_1025_cast_fp16)[name = string("key_39_cast_fp16")]; + bool input_1027_interleave_0 = const()[name = string("input_1027_interleave_0"), val = bool(false)]; + tensor input_1027_cast_fp16 = concat(axis = var_69, interleave = input_1027_interleave_0, values = (cache_77_cast_fp16, key_39_cast_fp16))[name = string("input_1027_cast_fp16")]; + tensor var_4526_begin_0 = const()[name = string("op_4526_begin_0"), val = tensor([0, 14, 0])]; + tensor var_4526_end_0 = const()[name = string("op_4526_end_0"), val = tensor([1, 42, 1024])]; + tensor var_4526_end_mask_0 = const()[name = string("op_4526_end_mask_0"), val = tensor([true, true, true])]; + tensor var_4526_cast_fp16 = slice_by_index(begin = var_4526_begin_0, end = var_4526_end_0, end_mask = var_4526_end_mask_0, x = cache_77_cast_fp16)[name = string("op_4526_cast_fp16")]; + bool var_4532_interleave_0 = const()[name = string("op_4532_interleave_0"), val = bool(false)]; + tensor var_4532_cast_fp16 = concat(axis = var_69, interleave = var_4532_interleave_0, values = (var_4526_cast_fp16, key_39_cast_fp16))[name = string("op_4532_cast_fp16")]; + tensor encoder_layers_19_self_attn_linear_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391109248))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(392157888))))[name = string("encoder_layers_19_self_attn_linear_q_weight_to_fp16_quantized")]; + tensor encoder_layers_19_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_19_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(392160000)))]; + tensor linear_174_cast_fp16 = linear(bias = encoder_layers_19_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_19_self_attn_linear_q_weight_to_fp16_quantized, x = key_39_cast_fp16)[name = string("linear_174_cast_fp16")]; + tensor var_4537 = const()[name = string("op_4537"), val = tensor([1, -1, 8, 128])]; + tensor q_115_cast_fp16 = reshape(shape = var_4537, x = linear_174_cast_fp16)[name = string("q_115_cast_fp16")]; + tensor encoder_layers_19_self_attn_linear_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(392162112))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(393210752))))[name = string("encoder_layers_19_self_attn_linear_k_weight_to_fp16_quantized")]; + tensor encoder_layers_19_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_19_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(393212864)))]; + tensor linear_175_cast_fp16 = linear(bias = encoder_layers_19_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_19_self_attn_linear_k_weight_to_fp16_quantized, x = input_1027_cast_fp16)[name = string("linear_175_cast_fp16")]; + tensor var_4542 = const()[name = string("op_4542"), val = tensor([1, -1, 8, 128])]; + tensor k_77_cast_fp16 = reshape(shape = var_4542, x = linear_175_cast_fp16)[name = string("k_77_cast_fp16")]; + tensor encoder_layers_19_self_attn_linear_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(393214976))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(394263616))))[name = string("encoder_layers_19_self_attn_linear_v_weight_to_fp16_quantized")]; + tensor encoder_layers_19_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_19_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(394265728)))]; + tensor linear_176_cast_fp16 = linear(bias = encoder_layers_19_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_19_self_attn_linear_v_weight_to_fp16_quantized, x = input_1027_cast_fp16)[name = string("linear_176_cast_fp16")]; + tensor var_4547 = const()[name = string("op_4547"), val = tensor([1, -1, 8, 128])]; + tensor v_39_cast_fp16 = reshape(shape = var_4547, x = linear_176_cast_fp16)[name = string("v_39_cast_fp16")]; + tensor value_47_perm_0 = const()[name = string("value_47_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_19_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_19_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(394267840)))]; + tensor var_4560_cast_fp16 = add(x = q_115_cast_fp16, y = encoder_layers_19_self_attn_pos_bias_u_to_fp16)[name = string("op_4560_cast_fp16")]; + tensor encoder_layers_19_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_19_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(394269952)))]; + tensor var_4562_cast_fp16 = add(x = q_115_cast_fp16, y = encoder_layers_19_self_attn_pos_bias_v_to_fp16)[name = string("op_4562_cast_fp16")]; + tensor q_with_bias_v_39_perm_0 = const()[name = string("q_with_bias_v_39_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_501_transpose_x_0 = const()[name = string("x_501_transpose_x_0"), val = bool(false)]; + bool x_501_transpose_y_0 = const()[name = string("x_501_transpose_y_0"), val = bool(false)]; + tensor op_4564_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(394272064))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(394385792))))[name = string("op_4564_to_fp16_quantized")]; + tensor q_with_bias_v_39_cast_fp16 = transpose(perm = q_with_bias_v_39_perm_0, x = var_4562_cast_fp16)[name = string("transpose_191")]; + tensor x_501_cast_fp16 = matmul(transpose_x = x_501_transpose_x_0, transpose_y = x_501_transpose_y_0, x = q_with_bias_v_39_cast_fp16, y = op_4564_to_fp16_quantized)[name = string("x_501_cast_fp16")]; + tensor x_503_pad_0 = const()[name = string("x_503_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_503_mode_0 = const()[name = string("x_503_mode_0"), val = string("constant")]; + fp16 const_326_to_fp16 = const()[name = string("const_326_to_fp16"), val = fp16(0x0p+0)]; + tensor x_503_cast_fp16 = pad(constant_val = const_326_to_fp16, mode = x_503_mode_0, pad = x_503_pad_0, x = x_501_cast_fp16)[name = string("x_503_cast_fp16")]; + tensor var_4572 = const()[name = string("op_4572"), val = tensor([1, 8, -1, 14])]; + tensor x_505_cast_fp16 = reshape(shape = var_4572, x = x_503_cast_fp16)[name = string("x_505_cast_fp16")]; + tensor var_4576_begin_0 = const()[name = string("op_4576_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_4576_end_0 = const()[name = string("op_4576_end_0"), val = tensor([1, 8, 112, 14])]; + tensor var_4576_end_mask_0 = const()[name = string("op_4576_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_4576_cast_fp16 = slice_by_index(begin = var_4576_begin_0, end = var_4576_end_0, end_mask = var_4576_end_mask_0, x = x_505_cast_fp16)[name = string("op_4576_cast_fp16")]; + tensor var_4577 = const()[name = string("op_4577"), val = tensor([1, 8, 14, 111])]; + tensor matrix_bd_77_cast_fp16 = reshape(shape = var_4577, x = var_4576_cast_fp16)[name = string("matrix_bd_77_cast_fp16")]; + bool matrix_ac_39_transpose_x_0 = const()[name = string("matrix_ac_39_transpose_x_0"), val = bool(false)]; + bool matrix_ac_39_transpose_y_0 = const()[name = string("matrix_ac_39_transpose_y_0"), val = bool(false)]; + tensor transpose_134_perm_0 = const()[name = string("transpose_134_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_135_perm_0 = const()[name = string("transpose_135_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_135 = transpose(perm = transpose_135_perm_0, x = k_77_cast_fp16)[name = string("transpose_189")]; + tensor transpose_134 = transpose(perm = transpose_134_perm_0, x = var_4560_cast_fp16)[name = string("transpose_190")]; + tensor matrix_ac_39_cast_fp16 = matmul(transpose_x = matrix_ac_39_transpose_x_0, transpose_y = matrix_ac_39_transpose_y_0, x = transpose_134, y = transpose_135)[name = string("matrix_ac_39_cast_fp16")]; + tensor matrix_bd_79_begin_0 = const()[name = string("matrix_bd_79_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_79_end_0 = const()[name = string("matrix_bd_79_end_0"), val = tensor([1, 8, 14, 56])]; + tensor matrix_bd_79_end_mask_0 = const()[name = string("matrix_bd_79_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_79_cast_fp16 = slice_by_index(begin = matrix_bd_79_begin_0, end = matrix_bd_79_end_0, end_mask = matrix_bd_79_end_mask_0, x = matrix_bd_77_cast_fp16)[name = string("matrix_bd_79_cast_fp16")]; + tensor var_4586_cast_fp16 = add(x = matrix_ac_39_cast_fp16, y = matrix_bd_79_cast_fp16)[name = string("op_4586_cast_fp16")]; + fp16 _inversed_scores_77_y_0_to_fp16 = const()[name = string("_inversed_scores_77_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_77_cast_fp16 = mul(x = var_4586_cast_fp16, y = _inversed_scores_77_y_0_to_fp16)[name = string("_inversed_scores_77_cast_fp16")]; + tensor scores_79_cast_fp16 = select(a = var_46_to_fp16, b = _inversed_scores_77_cast_fp16, cond = mask_11)[name = string("scores_79_cast_fp16")]; + tensor var_4592_cast_fp16 = softmax(axis = var_60, x = scores_79_cast_fp16)[name = string("op_4592_cast_fp16")]; + tensor input_1029_cast_fp16 = select(a = var_45_to_fp16, b = var_4592_cast_fp16, cond = mask_11)[name = string("input_1029_cast_fp16")]; + bool x_507_transpose_x_0 = const()[name = string("x_507_transpose_x_0"), val = bool(false)]; + bool x_507_transpose_y_0 = const()[name = string("x_507_transpose_y_0"), val = bool(false)]; + tensor value_47_cast_fp16 = transpose(perm = value_47_perm_0, x = v_39_cast_fp16)[name = string("transpose_188")]; + tensor x_507_cast_fp16 = matmul(transpose_x = x_507_transpose_x_0, transpose_y = x_507_transpose_y_0, x = input_1029_cast_fp16, y = value_47_cast_fp16)[name = string("x_507_cast_fp16")]; + tensor var_4596_perm_0 = const()[name = string("op_4596_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_4597 = const()[name = string("op_4597"), val = tensor([1, -1, 1024])]; + tensor var_4596_cast_fp16 = transpose(perm = var_4596_perm_0, x = x_507_cast_fp16)[name = string("transpose_187")]; + tensor input_1031_cast_fp16 = reshape(shape = var_4597, x = var_4596_cast_fp16)[name = string("input_1031_cast_fp16")]; + tensor encoder_layers_19_self_attn_linear_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(394386112))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(395434752))))[name = string("encoder_layers_19_self_attn_linear_out_weight_to_fp16_quantized")]; + tensor encoder_layers_19_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_19_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(395436864)))]; + tensor linear_178_cast_fp16 = linear(bias = encoder_layers_19_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_19_self_attn_linear_out_weight_to_fp16_quantized, x = input_1031_cast_fp16)[name = string("linear_178_cast_fp16")]; + tensor input_1035_cast_fp16 = add(x = input_1025_cast_fp16, y = linear_178_cast_fp16)[name = string("input_1035_cast_fp16")]; + tensor x_511_axes_0 = const()[name = string("x_511_axes_0"), val = tensor([-1])]; + tensor encoder_layers_19_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_19_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(395438976)))]; + tensor encoder_layers_19_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_19_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(395441088)))]; + tensor x_511_cast_fp16 = layer_norm(axes = x_511_axes_0, beta = encoder_layers_19_norm_conv_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_19_norm_conv_weight_to_fp16, x = input_1035_cast_fp16)[name = string("x_511_cast_fp16")]; + tensor input_1037_perm_0 = const()[name = string("input_1037_perm_0"), val = tensor([0, 2, 1])]; + string input_1039_pad_type_0 = const()[name = string("input_1039_pad_type_0"), val = string("valid")]; + tensor input_1039_strides_0 = const()[name = string("input_1039_strides_0"), val = tensor([1])]; + tensor input_1039_pad_0 = const()[name = string("input_1039_pad_0"), val = tensor([0, 0])]; + tensor input_1039_dilations_0 = const()[name = string("input_1039_dilations_0"), val = tensor([1])]; + int32 input_1039_groups_0 = const()[name = string("input_1039_groups_0"), val = int32(1)]; + tensor encoder_layers_19_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(395443200))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397540416))))[name = string("encoder_layers_19_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_1037_cast_fp16 = transpose(perm = input_1037_perm_0, x = x_511_cast_fp16)[name = string("transpose_186")]; + tensor input_1039_cast_fp16 = conv(dilations = input_1039_dilations_0, groups = input_1039_groups_0, pad = input_1039_pad_0, pad_type = input_1039_pad_type_0, strides = input_1039_strides_0, weight = encoder_layers_19_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_1037_cast_fp16)[name = string("input_1039_cast_fp16")]; + int32 x_513_split_num_splits_0 = const()[name = string("x_513_split_num_splits_0"), val = int32(2)]; + int32 x_513_split_axis_0 = const()[name = string("x_513_split_axis_0"), val = int32(1)]; + tensor x_513_split_cast_fp16_0, tensor x_513_split_cast_fp16_1 = split(axis = x_513_split_axis_0, num_splits = x_513_split_num_splits_0, x = input_1039_cast_fp16)[name = string("x_513_split_cast_fp16")]; + tensor x_513_split_1_sigmoid_cast_fp16 = sigmoid(x = x_513_split_cast_fp16_1)[name = string("x_513_split_1_sigmoid_cast_fp16")]; + tensor x_513_cast_fp16 = mul(x = x_513_split_cast_fp16_0, y = x_513_split_1_sigmoid_cast_fp16)[name = string("x_513_cast_fp16")]; + tensor input_1041_cast_fp16 = select(a = var_45_to_fp16, b = x_513_cast_fp16, cond = var_576)[name = string("input_1041_cast_fp16")]; + bool new_x_79_interleave_0 = const()[name = string("new_x_79_interleave_0"), val = bool(false)]; + tensor new_x_79_cast_fp16 = concat(axis = var_60, interleave = new_x_79_interleave_0, values = (cache_79_cast_fp16, input_1041_cast_fp16))[name = string("new_x_79_cast_fp16")]; + tensor var_4636_begin_0 = const()[name = string("op_4636_begin_0"), val = tensor([0, 0, 14])]; + tensor var_4636_end_0 = const()[name = string("op_4636_end_0"), val = tensor([1, 1024, 22])]; + tensor var_4636_end_mask_0 = const()[name = string("op_4636_end_mask_0"), val = tensor([true, true, true])]; + tensor var_4636_cast_fp16 = slice_by_index(begin = var_4636_begin_0, end = var_4636_end_0, end_mask = var_4636_end_mask_0, x = new_x_79_cast_fp16)[name = string("op_4636_cast_fp16")]; + string x_515_pad_type_0 = const()[name = string("x_515_pad_type_0"), val = string("valid")]; + int32 x_515_groups_0 = const()[name = string("x_515_groups_0"), val = int32(1024)]; + tensor x_515_strides_0 = const()[name = string("x_515_strides_0"), val = tensor([1])]; + tensor x_515_pad_0 = const()[name = string("x_515_pad_0"), val = tensor([0, 0])]; + tensor x_515_dilations_0 = const()[name = string("x_515_dilations_0"), val = tensor([1])]; + tensor encoder_layers_19_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397544576))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397553856))))[name = string("encoder_layers_19_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_515_cast_fp16 = conv(dilations = x_515_dilations_0, groups = x_515_groups_0, pad = x_515_pad_0, pad_type = x_515_pad_type_0, strides = x_515_strides_0, weight = encoder_layers_19_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_79_cast_fp16)[name = string("x_515_cast_fp16")]; + tensor input_1043_perm_0 = const()[name = string("input_1043_perm_0"), val = tensor([0, 2, 1])]; + tensor x_517_axes_0 = const()[name = string("x_517_axes_0"), val = tensor([-1])]; + tensor encoder_layers_19_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_19_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397555968)))]; + tensor encoder_layers_19_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_19_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397558080)))]; + tensor input_1043_cast_fp16 = transpose(perm = input_1043_perm_0, x = x_515_cast_fp16)[name = string("transpose_185")]; + tensor x_517_cast_fp16 = layer_norm(axes = x_517_axes_0, beta = encoder_layers_19_conv_batch_norm_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_19_conv_batch_norm_weight_to_fp16, x = input_1043_cast_fp16)[name = string("x_517_cast_fp16")]; + tensor input_1045_perm_0 = const()[name = string("input_1045_perm_0"), val = tensor([0, 2, 1])]; + tensor input_1045_cast_fp16 = transpose(perm = input_1045_perm_0, x = x_517_cast_fp16)[name = string("transpose_184")]; + tensor input_1047_cast_fp16 = silu(x = input_1045_cast_fp16)[name = string("input_1047_cast_fp16")]; + string x_519_pad_type_0 = const()[name = string("x_519_pad_type_0"), val = string("valid")]; + tensor x_519_strides_0 = const()[name = string("x_519_strides_0"), val = tensor([1])]; + tensor x_519_pad_0 = const()[name = string("x_519_pad_0"), val = tensor([0, 0])]; + tensor x_519_dilations_0 = const()[name = string("x_519_dilations_0"), val = tensor([1])]; + int32 x_519_groups_0 = const()[name = string("x_519_groups_0"), val = int32(1)]; + tensor encoder_layers_19_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397560192))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(398608832))))[name = string("encoder_layers_19_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_519_cast_fp16 = conv(dilations = x_519_dilations_0, groups = x_519_groups_0, pad = x_519_pad_0, pad_type = x_519_pad_type_0, strides = x_519_strides_0, weight = encoder_layers_19_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_1047_cast_fp16)[name = string("x_519_cast_fp16")]; + tensor input_1049_perm_0 = const()[name = string("input_1049_perm_0"), val = tensor([0, 2, 1])]; + tensor input_1049_cast_fp16 = transpose(perm = input_1049_perm_0, x = x_519_cast_fp16)[name = string("transpose_183")]; + tensor input_1051_cast_fp16 = add(x = input_1035_cast_fp16, y = input_1049_cast_fp16)[name = string("input_1051_cast_fp16")]; + tensor input_1053_axes_0 = const()[name = string("input_1053_axes_0"), val = tensor([-1])]; + tensor encoder_layers_19_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_19_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(398610944)))]; + tensor encoder_layers_19_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_19_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(398613056)))]; + tensor input_1053_cast_fp16 = layer_norm(axes = input_1053_axes_0, beta = encoder_layers_19_norm_feed_forward2_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_19_norm_feed_forward2_weight_to_fp16, x = input_1051_cast_fp16)[name = string("input_1053_cast_fp16")]; + tensor encoder_layers_19_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(398615168))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(402809536))))[name = string("encoder_layers_19_feed_forward2_linear1_weight_to_fp16_quantized")]; + tensor encoder_layers_19_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_19_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(402817792)))]; + tensor linear_179_cast_fp16 = linear(bias = encoder_layers_19_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_19_feed_forward2_linear1_weight_to_fp16_quantized, x = input_1053_cast_fp16)[name = string("linear_179_cast_fp16")]; + tensor input_1057_cast_fp16 = silu(x = linear_179_cast_fp16)[name = string("input_1057_cast_fp16")]; + tensor encoder_layers_19_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(402826048))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407020416))))[name = string("encoder_layers_19_feed_forward2_linear2_weight_to_fp16_quantized")]; + tensor encoder_layers_19_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_19_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407022528)))]; + tensor linear_180_cast_fp16 = linear(bias = encoder_layers_19_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_19_feed_forward2_linear2_weight_to_fp16_quantized, x = input_1057_cast_fp16)[name = string("linear_180_cast_fp16")]; + fp16 var_4679_to_fp16 = const()[name = string("op_4679_to_fp16"), val = fp16(0x1p-1)]; + tensor var_4680_cast_fp16 = mul(x = linear_180_cast_fp16, y = var_4679_to_fp16)[name = string("op_4680_cast_fp16")]; + tensor input_1063_cast_fp16 = add(x = input_1051_cast_fp16, y = var_4680_cast_fp16)[name = string("input_1063_cast_fp16")]; + tensor input_1065_axes_0 = const()[name = string("input_1065_axes_0"), val = tensor([-1])]; + tensor encoder_layers_19_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_19_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407024640)))]; + tensor encoder_layers_19_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_19_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407026752)))]; + tensor input_1065_cast_fp16 = layer_norm(axes = input_1065_axes_0, beta = encoder_layers_19_norm_out_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_19_norm_out_weight_to_fp16, x = input_1063_cast_fp16)[name = string("input_1065_cast_fp16")]; + tensor cache_81_begin_0 = const()[name = string("cache_81_begin_0"), val = tensor([20, 0, 0, 0])]; + tensor cache_81_end_0 = const()[name = string("cache_81_end_0"), val = tensor([21, 1, 42, 1024])]; + tensor cache_81_end_mask_0 = const()[name = string("cache_81_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_81_squeeze_mask_0 = const()[name = string("cache_81_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_81_cast_fp16 = slice_by_index(begin = cache_81_begin_0, end = cache_81_end_0, end_mask = cache_81_end_mask_0, squeeze_mask = cache_81_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_81_cast_fp16")]; + tensor cache_83_begin_0 = const()[name = string("cache_83_begin_0"), val = tensor([20, 0, 0, 0])]; + tensor cache_83_end_0 = const()[name = string("cache_83_end_0"), val = tensor([21, 1, 1024, 8])]; + tensor cache_83_end_mask_0 = const()[name = string("cache_83_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_83_squeeze_mask_0 = const()[name = string("cache_83_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_83_cast_fp16 = slice_by_index(begin = cache_83_begin_0, end = cache_83_end_0, end_mask = cache_83_end_mask_0, squeeze_mask = cache_83_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_83_cast_fp16")]; + tensor input_1067_axes_0 = const()[name = string("input_1067_axes_0"), val = tensor([-1])]; + tensor encoder_layers_20_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_20_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407028864)))]; + tensor encoder_layers_20_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_20_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407030976)))]; + tensor input_1067_cast_fp16 = layer_norm(axes = input_1067_axes_0, beta = encoder_layers_20_norm_feed_forward1_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_20_norm_feed_forward1_weight_to_fp16, x = input_1065_cast_fp16)[name = string("input_1067_cast_fp16")]; + tensor encoder_layers_20_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407033088))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(411227456))))[name = string("encoder_layers_20_feed_forward1_linear1_weight_to_fp16_quantized")]; + tensor encoder_layers_20_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_20_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(411235712)))]; + tensor linear_181_cast_fp16 = linear(bias = encoder_layers_20_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_20_feed_forward1_linear1_weight_to_fp16_quantized, x = input_1067_cast_fp16)[name = string("linear_181_cast_fp16")]; + tensor input_1071_cast_fp16 = silu(x = linear_181_cast_fp16)[name = string("input_1071_cast_fp16")]; + tensor encoder_layers_20_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(411243968))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(415438336))))[name = string("encoder_layers_20_feed_forward1_linear2_weight_to_fp16_quantized")]; + tensor encoder_layers_20_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_20_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(415440448)))]; + tensor linear_182_cast_fp16 = linear(bias = encoder_layers_20_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_20_feed_forward1_linear2_weight_to_fp16_quantized, x = input_1071_cast_fp16)[name = string("linear_182_cast_fp16")]; + fp16 var_4716_to_fp16 = const()[name = string("op_4716_to_fp16"), val = fp16(0x1p-1)]; + tensor var_4717_cast_fp16 = mul(x = linear_182_cast_fp16, y = var_4716_to_fp16)[name = string("op_4717_cast_fp16")]; + tensor input_1077_cast_fp16 = add(x = input_1065_cast_fp16, y = var_4717_cast_fp16)[name = string("input_1077_cast_fp16")]; + tensor key_41_axes_0 = const()[name = string("key_41_axes_0"), val = tensor([-1])]; + tensor encoder_layers_20_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_20_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(415442560)))]; + tensor encoder_layers_20_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_20_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(415444672)))]; + tensor key_41_cast_fp16 = layer_norm(axes = key_41_axes_0, beta = encoder_layers_20_norm_self_att_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_20_norm_self_att_weight_to_fp16, x = input_1077_cast_fp16)[name = string("key_41_cast_fp16")]; + bool input_1079_interleave_0 = const()[name = string("input_1079_interleave_0"), val = bool(false)]; + tensor input_1079_cast_fp16 = concat(axis = var_69, interleave = input_1079_interleave_0, values = (cache_81_cast_fp16, key_41_cast_fp16))[name = string("input_1079_cast_fp16")]; + tensor var_4739_begin_0 = const()[name = string("op_4739_begin_0"), val = tensor([0, 14, 0])]; + tensor var_4739_end_0 = const()[name = string("op_4739_end_0"), val = tensor([1, 42, 1024])]; + tensor var_4739_end_mask_0 = const()[name = string("op_4739_end_mask_0"), val = tensor([true, true, true])]; + tensor var_4739_cast_fp16 = slice_by_index(begin = var_4739_begin_0, end = var_4739_end_0, end_mask = var_4739_end_mask_0, x = cache_81_cast_fp16)[name = string("op_4739_cast_fp16")]; + bool var_4745_interleave_0 = const()[name = string("op_4745_interleave_0"), val = bool(false)]; + tensor var_4745_cast_fp16 = concat(axis = var_69, interleave = var_4745_interleave_0, values = (var_4739_cast_fp16, key_41_cast_fp16))[name = string("op_4745_cast_fp16")]; + tensor encoder_layers_20_self_attn_linear_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(415446784))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(416495424))))[name = string("encoder_layers_20_self_attn_linear_q_weight_to_fp16_quantized")]; + tensor encoder_layers_20_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_20_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(416497536)))]; + tensor linear_183_cast_fp16 = linear(bias = encoder_layers_20_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_20_self_attn_linear_q_weight_to_fp16_quantized, x = key_41_cast_fp16)[name = string("linear_183_cast_fp16")]; + tensor var_4750 = const()[name = string("op_4750"), val = tensor([1, -1, 8, 128])]; + tensor q_121_cast_fp16 = reshape(shape = var_4750, x = linear_183_cast_fp16)[name = string("q_121_cast_fp16")]; + tensor encoder_layers_20_self_attn_linear_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(416499648))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(417548288))))[name = string("encoder_layers_20_self_attn_linear_k_weight_to_fp16_quantized")]; + tensor encoder_layers_20_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_20_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(417550400)))]; + tensor linear_184_cast_fp16 = linear(bias = encoder_layers_20_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_20_self_attn_linear_k_weight_to_fp16_quantized, x = input_1079_cast_fp16)[name = string("linear_184_cast_fp16")]; + tensor var_4755 = const()[name = string("op_4755"), val = tensor([1, -1, 8, 128])]; + tensor k_81_cast_fp16 = reshape(shape = var_4755, x = linear_184_cast_fp16)[name = string("k_81_cast_fp16")]; + tensor encoder_layers_20_self_attn_linear_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(417552512))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(418601152))))[name = string("encoder_layers_20_self_attn_linear_v_weight_to_fp16_quantized")]; + tensor encoder_layers_20_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_20_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(418603264)))]; + tensor linear_185_cast_fp16 = linear(bias = encoder_layers_20_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_20_self_attn_linear_v_weight_to_fp16_quantized, x = input_1079_cast_fp16)[name = string("linear_185_cast_fp16")]; + tensor var_4760 = const()[name = string("op_4760"), val = tensor([1, -1, 8, 128])]; + tensor v_41_cast_fp16 = reshape(shape = var_4760, x = linear_185_cast_fp16)[name = string("v_41_cast_fp16")]; + tensor value_49_perm_0 = const()[name = string("value_49_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_20_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_20_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(418605376)))]; + tensor var_4773_cast_fp16 = add(x = q_121_cast_fp16, y = encoder_layers_20_self_attn_pos_bias_u_to_fp16)[name = string("op_4773_cast_fp16")]; + tensor encoder_layers_20_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_20_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(418607488)))]; + tensor var_4775_cast_fp16 = add(x = q_121_cast_fp16, y = encoder_layers_20_self_attn_pos_bias_v_to_fp16)[name = string("op_4775_cast_fp16")]; + tensor q_with_bias_v_41_perm_0 = const()[name = string("q_with_bias_v_41_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_527_transpose_x_0 = const()[name = string("x_527_transpose_x_0"), val = bool(false)]; + bool x_527_transpose_y_0 = const()[name = string("x_527_transpose_y_0"), val = bool(false)]; + tensor op_4777_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(418609600))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(418723328))))[name = string("op_4777_to_fp16_quantized")]; + tensor q_with_bias_v_41_cast_fp16 = transpose(perm = q_with_bias_v_41_perm_0, x = var_4775_cast_fp16)[name = string("transpose_182")]; + tensor x_527_cast_fp16 = matmul(transpose_x = x_527_transpose_x_0, transpose_y = x_527_transpose_y_0, x = q_with_bias_v_41_cast_fp16, y = op_4777_to_fp16_quantized)[name = string("x_527_cast_fp16")]; + tensor x_529_pad_0 = const()[name = string("x_529_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_529_mode_0 = const()[name = string("x_529_mode_0"), val = string("constant")]; + fp16 const_339_to_fp16 = const()[name = string("const_339_to_fp16"), val = fp16(0x0p+0)]; + tensor x_529_cast_fp16 = pad(constant_val = const_339_to_fp16, mode = x_529_mode_0, pad = x_529_pad_0, x = x_527_cast_fp16)[name = string("x_529_cast_fp16")]; + tensor var_4785 = const()[name = string("op_4785"), val = tensor([1, 8, -1, 14])]; + tensor x_531_cast_fp16 = reshape(shape = var_4785, x = x_529_cast_fp16)[name = string("x_531_cast_fp16")]; + tensor var_4789_begin_0 = const()[name = string("op_4789_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_4789_end_0 = const()[name = string("op_4789_end_0"), val = tensor([1, 8, 112, 14])]; + tensor var_4789_end_mask_0 = const()[name = string("op_4789_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_4789_cast_fp16 = slice_by_index(begin = var_4789_begin_0, end = var_4789_end_0, end_mask = var_4789_end_mask_0, x = x_531_cast_fp16)[name = string("op_4789_cast_fp16")]; + tensor var_4790 = const()[name = string("op_4790"), val = tensor([1, 8, 14, 111])]; + tensor matrix_bd_81_cast_fp16 = reshape(shape = var_4790, x = var_4789_cast_fp16)[name = string("matrix_bd_81_cast_fp16")]; + bool matrix_ac_41_transpose_x_0 = const()[name = string("matrix_ac_41_transpose_x_0"), val = bool(false)]; + bool matrix_ac_41_transpose_y_0 = const()[name = string("matrix_ac_41_transpose_y_0"), val = bool(false)]; + tensor transpose_136_perm_0 = const()[name = string("transpose_136_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_137_perm_0 = const()[name = string("transpose_137_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_137 = transpose(perm = transpose_137_perm_0, x = k_81_cast_fp16)[name = string("transpose_180")]; + tensor transpose_136 = transpose(perm = transpose_136_perm_0, x = var_4773_cast_fp16)[name = string("transpose_181")]; + tensor matrix_ac_41_cast_fp16 = matmul(transpose_x = matrix_ac_41_transpose_x_0, transpose_y = matrix_ac_41_transpose_y_0, x = transpose_136, y = transpose_137)[name = string("matrix_ac_41_cast_fp16")]; + tensor matrix_bd_83_begin_0 = const()[name = string("matrix_bd_83_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_83_end_0 = const()[name = string("matrix_bd_83_end_0"), val = tensor([1, 8, 14, 56])]; + tensor matrix_bd_83_end_mask_0 = const()[name = string("matrix_bd_83_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_83_cast_fp16 = slice_by_index(begin = matrix_bd_83_begin_0, end = matrix_bd_83_end_0, end_mask = matrix_bd_83_end_mask_0, x = matrix_bd_81_cast_fp16)[name = string("matrix_bd_83_cast_fp16")]; + tensor var_4799_cast_fp16 = add(x = matrix_ac_41_cast_fp16, y = matrix_bd_83_cast_fp16)[name = string("op_4799_cast_fp16")]; + fp16 _inversed_scores_81_y_0_to_fp16 = const()[name = string("_inversed_scores_81_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_81_cast_fp16 = mul(x = var_4799_cast_fp16, y = _inversed_scores_81_y_0_to_fp16)[name = string("_inversed_scores_81_cast_fp16")]; + tensor scores_83_cast_fp16 = select(a = var_46_to_fp16, b = _inversed_scores_81_cast_fp16, cond = mask_11)[name = string("scores_83_cast_fp16")]; + tensor var_4805_cast_fp16 = softmax(axis = var_60, x = scores_83_cast_fp16)[name = string("op_4805_cast_fp16")]; + tensor input_1081_cast_fp16 = select(a = var_45_to_fp16, b = var_4805_cast_fp16, cond = mask_11)[name = string("input_1081_cast_fp16")]; + bool x_533_transpose_x_0 = const()[name = string("x_533_transpose_x_0"), val = bool(false)]; + bool x_533_transpose_y_0 = const()[name = string("x_533_transpose_y_0"), val = bool(false)]; + tensor value_49_cast_fp16 = transpose(perm = value_49_perm_0, x = v_41_cast_fp16)[name = string("transpose_179")]; + tensor x_533_cast_fp16 = matmul(transpose_x = x_533_transpose_x_0, transpose_y = x_533_transpose_y_0, x = input_1081_cast_fp16, y = value_49_cast_fp16)[name = string("x_533_cast_fp16")]; + tensor var_4809_perm_0 = const()[name = string("op_4809_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_4810 = const()[name = string("op_4810"), val = tensor([1, -1, 1024])]; + tensor var_4809_cast_fp16 = transpose(perm = var_4809_perm_0, x = x_533_cast_fp16)[name = string("transpose_178")]; + tensor input_1083_cast_fp16 = reshape(shape = var_4810, x = var_4809_cast_fp16)[name = string("input_1083_cast_fp16")]; + tensor encoder_layers_20_self_attn_linear_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(418723648))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(419772288))))[name = string("encoder_layers_20_self_attn_linear_out_weight_to_fp16_quantized")]; + tensor encoder_layers_20_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_20_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(419774400)))]; + tensor linear_187_cast_fp16 = linear(bias = encoder_layers_20_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_20_self_attn_linear_out_weight_to_fp16_quantized, x = input_1083_cast_fp16)[name = string("linear_187_cast_fp16")]; + tensor input_1087_cast_fp16 = add(x = input_1077_cast_fp16, y = linear_187_cast_fp16)[name = string("input_1087_cast_fp16")]; + tensor x_537_axes_0 = const()[name = string("x_537_axes_0"), val = tensor([-1])]; + tensor encoder_layers_20_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_20_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(419776512)))]; + tensor encoder_layers_20_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_20_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(419778624)))]; + tensor x_537_cast_fp16 = layer_norm(axes = x_537_axes_0, beta = encoder_layers_20_norm_conv_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_20_norm_conv_weight_to_fp16, x = input_1087_cast_fp16)[name = string("x_537_cast_fp16")]; + tensor input_1089_perm_0 = const()[name = string("input_1089_perm_0"), val = tensor([0, 2, 1])]; + string input_1091_pad_type_0 = const()[name = string("input_1091_pad_type_0"), val = string("valid")]; + tensor input_1091_strides_0 = const()[name = string("input_1091_strides_0"), val = tensor([1])]; + tensor input_1091_pad_0 = const()[name = string("input_1091_pad_0"), val = tensor([0, 0])]; + tensor input_1091_dilations_0 = const()[name = string("input_1091_dilations_0"), val = tensor([1])]; + int32 input_1091_groups_0 = const()[name = string("input_1091_groups_0"), val = int32(1)]; + tensor encoder_layers_20_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(419780736))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421877952))))[name = string("encoder_layers_20_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_1089_cast_fp16 = transpose(perm = input_1089_perm_0, x = x_537_cast_fp16)[name = string("transpose_177")]; + tensor input_1091_cast_fp16 = conv(dilations = input_1091_dilations_0, groups = input_1091_groups_0, pad = input_1091_pad_0, pad_type = input_1091_pad_type_0, strides = input_1091_strides_0, weight = encoder_layers_20_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_1089_cast_fp16)[name = string("input_1091_cast_fp16")]; + int32 x_539_split_num_splits_0 = const()[name = string("x_539_split_num_splits_0"), val = int32(2)]; + int32 x_539_split_axis_0 = const()[name = string("x_539_split_axis_0"), val = int32(1)]; + tensor x_539_split_cast_fp16_0, tensor x_539_split_cast_fp16_1 = split(axis = x_539_split_axis_0, num_splits = x_539_split_num_splits_0, x = input_1091_cast_fp16)[name = string("x_539_split_cast_fp16")]; + tensor x_539_split_1_sigmoid_cast_fp16 = sigmoid(x = x_539_split_cast_fp16_1)[name = string("x_539_split_1_sigmoid_cast_fp16")]; + tensor x_539_cast_fp16 = mul(x = x_539_split_cast_fp16_0, y = x_539_split_1_sigmoid_cast_fp16)[name = string("x_539_cast_fp16")]; + tensor input_1093_cast_fp16 = select(a = var_45_to_fp16, b = x_539_cast_fp16, cond = var_576)[name = string("input_1093_cast_fp16")]; + bool new_x_83_interleave_0 = const()[name = string("new_x_83_interleave_0"), val = bool(false)]; + tensor new_x_83_cast_fp16 = concat(axis = var_60, interleave = new_x_83_interleave_0, values = (cache_83_cast_fp16, input_1093_cast_fp16))[name = string("new_x_83_cast_fp16")]; + tensor var_4849_begin_0 = const()[name = string("op_4849_begin_0"), val = tensor([0, 0, 14])]; + tensor var_4849_end_0 = const()[name = string("op_4849_end_0"), val = tensor([1, 1024, 22])]; + tensor var_4849_end_mask_0 = const()[name = string("op_4849_end_mask_0"), val = tensor([true, true, true])]; + tensor var_4849_cast_fp16 = slice_by_index(begin = var_4849_begin_0, end = var_4849_end_0, end_mask = var_4849_end_mask_0, x = new_x_83_cast_fp16)[name = string("op_4849_cast_fp16")]; + string x_541_pad_type_0 = const()[name = string("x_541_pad_type_0"), val = string("valid")]; + int32 x_541_groups_0 = const()[name = string("x_541_groups_0"), val = int32(1024)]; + tensor x_541_strides_0 = const()[name = string("x_541_strides_0"), val = tensor([1])]; + tensor x_541_pad_0 = const()[name = string("x_541_pad_0"), val = tensor([0, 0])]; + tensor x_541_dilations_0 = const()[name = string("x_541_dilations_0"), val = tensor([1])]; + tensor encoder_layers_20_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421882112))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421891392))))[name = string("encoder_layers_20_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_541_cast_fp16 = conv(dilations = x_541_dilations_0, groups = x_541_groups_0, pad = x_541_pad_0, pad_type = x_541_pad_type_0, strides = x_541_strides_0, weight = encoder_layers_20_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_83_cast_fp16)[name = string("x_541_cast_fp16")]; + tensor input_1095_perm_0 = const()[name = string("input_1095_perm_0"), val = tensor([0, 2, 1])]; + tensor x_543_axes_0 = const()[name = string("x_543_axes_0"), val = tensor([-1])]; + tensor encoder_layers_20_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_20_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421893504)))]; + tensor encoder_layers_20_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_20_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421895616)))]; + tensor input_1095_cast_fp16 = transpose(perm = input_1095_perm_0, x = x_541_cast_fp16)[name = string("transpose_176")]; + tensor x_543_cast_fp16 = layer_norm(axes = x_543_axes_0, beta = encoder_layers_20_conv_batch_norm_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_20_conv_batch_norm_weight_to_fp16, x = input_1095_cast_fp16)[name = string("x_543_cast_fp16")]; + tensor input_1097_perm_0 = const()[name = string("input_1097_perm_0"), val = tensor([0, 2, 1])]; + tensor input_1097_cast_fp16 = transpose(perm = input_1097_perm_0, x = x_543_cast_fp16)[name = string("transpose_175")]; + tensor input_1099_cast_fp16 = silu(x = input_1097_cast_fp16)[name = string("input_1099_cast_fp16")]; + string x_545_pad_type_0 = const()[name = string("x_545_pad_type_0"), val = string("valid")]; + tensor x_545_strides_0 = const()[name = string("x_545_strides_0"), val = tensor([1])]; + tensor x_545_pad_0 = const()[name = string("x_545_pad_0"), val = tensor([0, 0])]; + tensor x_545_dilations_0 = const()[name = string("x_545_dilations_0"), val = tensor([1])]; + int32 x_545_groups_0 = const()[name = string("x_545_groups_0"), val = int32(1)]; + tensor encoder_layers_20_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421897728))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(422946368))))[name = string("encoder_layers_20_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_545_cast_fp16 = conv(dilations = x_545_dilations_0, groups = x_545_groups_0, pad = x_545_pad_0, pad_type = x_545_pad_type_0, strides = x_545_strides_0, weight = encoder_layers_20_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_1099_cast_fp16)[name = string("x_545_cast_fp16")]; + tensor input_1101_perm_0 = const()[name = string("input_1101_perm_0"), val = tensor([0, 2, 1])]; + tensor input_1101_cast_fp16 = transpose(perm = input_1101_perm_0, x = x_545_cast_fp16)[name = string("transpose_174")]; + tensor input_1103_cast_fp16 = add(x = input_1087_cast_fp16, y = input_1101_cast_fp16)[name = string("input_1103_cast_fp16")]; + tensor input_1105_axes_0 = const()[name = string("input_1105_axes_0"), val = tensor([-1])]; + tensor encoder_layers_20_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_20_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(422948480)))]; + tensor encoder_layers_20_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_20_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(422950592)))]; + tensor input_1105_cast_fp16 = layer_norm(axes = input_1105_axes_0, beta = encoder_layers_20_norm_feed_forward2_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_20_norm_feed_forward2_weight_to_fp16, x = input_1103_cast_fp16)[name = string("input_1105_cast_fp16")]; + tensor encoder_layers_20_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(422952704))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427147072))))[name = string("encoder_layers_20_feed_forward2_linear1_weight_to_fp16_quantized")]; + tensor encoder_layers_20_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_20_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427155328)))]; + tensor linear_188_cast_fp16 = linear(bias = encoder_layers_20_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_20_feed_forward2_linear1_weight_to_fp16_quantized, x = input_1105_cast_fp16)[name = string("linear_188_cast_fp16")]; + tensor input_1109_cast_fp16 = silu(x = linear_188_cast_fp16)[name = string("input_1109_cast_fp16")]; + tensor encoder_layers_20_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427163584))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(431357952))))[name = string("encoder_layers_20_feed_forward2_linear2_weight_to_fp16_quantized")]; + tensor encoder_layers_20_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_20_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(431360064)))]; + tensor linear_189_cast_fp16 = linear(bias = encoder_layers_20_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_20_feed_forward2_linear2_weight_to_fp16_quantized, x = input_1109_cast_fp16)[name = string("linear_189_cast_fp16")]; + fp16 var_4892_to_fp16 = const()[name = string("op_4892_to_fp16"), val = fp16(0x1p-1)]; + tensor var_4893_cast_fp16 = mul(x = linear_189_cast_fp16, y = var_4892_to_fp16)[name = string("op_4893_cast_fp16")]; + tensor input_1115_cast_fp16 = add(x = input_1103_cast_fp16, y = var_4893_cast_fp16)[name = string("input_1115_cast_fp16")]; + tensor input_1117_axes_0 = const()[name = string("input_1117_axes_0"), val = tensor([-1])]; + tensor encoder_layers_20_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_20_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(431362176)))]; + tensor encoder_layers_20_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_20_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(431364288)))]; + tensor input_1117_cast_fp16 = layer_norm(axes = input_1117_axes_0, beta = encoder_layers_20_norm_out_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_20_norm_out_weight_to_fp16, x = input_1115_cast_fp16)[name = string("input_1117_cast_fp16")]; + tensor cache_85_begin_0 = const()[name = string("cache_85_begin_0"), val = tensor([21, 0, 0, 0])]; + tensor cache_85_end_0 = const()[name = string("cache_85_end_0"), val = tensor([22, 1, 42, 1024])]; + tensor cache_85_end_mask_0 = const()[name = string("cache_85_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_85_squeeze_mask_0 = const()[name = string("cache_85_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_85_cast_fp16 = slice_by_index(begin = cache_85_begin_0, end = cache_85_end_0, end_mask = cache_85_end_mask_0, squeeze_mask = cache_85_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_85_cast_fp16")]; + tensor cache_87_begin_0 = const()[name = string("cache_87_begin_0"), val = tensor([21, 0, 0, 0])]; + tensor cache_87_end_0 = const()[name = string("cache_87_end_0"), val = tensor([22, 1, 1024, 8])]; + tensor cache_87_end_mask_0 = const()[name = string("cache_87_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_87_squeeze_mask_0 = const()[name = string("cache_87_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_87_cast_fp16 = slice_by_index(begin = cache_87_begin_0, end = cache_87_end_0, end_mask = cache_87_end_mask_0, squeeze_mask = cache_87_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_87_cast_fp16")]; + tensor input_1119_axes_0 = const()[name = string("input_1119_axes_0"), val = tensor([-1])]; + tensor encoder_layers_21_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_21_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(431366400)))]; + tensor encoder_layers_21_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_21_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(431368512)))]; + tensor input_1119_cast_fp16 = layer_norm(axes = input_1119_axes_0, beta = encoder_layers_21_norm_feed_forward1_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_21_norm_feed_forward1_weight_to_fp16, x = input_1117_cast_fp16)[name = string("input_1119_cast_fp16")]; + tensor encoder_layers_21_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(431370624))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(435564992))))[name = string("encoder_layers_21_feed_forward1_linear1_weight_to_fp16_quantized")]; + tensor encoder_layers_21_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_21_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(435573248)))]; + tensor linear_190_cast_fp16 = linear(bias = encoder_layers_21_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_21_feed_forward1_linear1_weight_to_fp16_quantized, x = input_1119_cast_fp16)[name = string("linear_190_cast_fp16")]; + tensor input_1123_cast_fp16 = silu(x = linear_190_cast_fp16)[name = string("input_1123_cast_fp16")]; + tensor encoder_layers_21_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(435581504))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(439775872))))[name = string("encoder_layers_21_feed_forward1_linear2_weight_to_fp16_quantized")]; + tensor encoder_layers_21_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_21_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(439777984)))]; + tensor linear_191_cast_fp16 = linear(bias = encoder_layers_21_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_21_feed_forward1_linear2_weight_to_fp16_quantized, x = input_1123_cast_fp16)[name = string("linear_191_cast_fp16")]; + fp16 var_4929_to_fp16 = const()[name = string("op_4929_to_fp16"), val = fp16(0x1p-1)]; + tensor var_4930_cast_fp16 = mul(x = linear_191_cast_fp16, y = var_4929_to_fp16)[name = string("op_4930_cast_fp16")]; + tensor input_1129_cast_fp16 = add(x = input_1117_cast_fp16, y = var_4930_cast_fp16)[name = string("input_1129_cast_fp16")]; + tensor key_43_axes_0 = const()[name = string("key_43_axes_0"), val = tensor([-1])]; + tensor encoder_layers_21_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_21_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(439780096)))]; + tensor encoder_layers_21_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_21_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(439782208)))]; + tensor key_43_cast_fp16 = layer_norm(axes = key_43_axes_0, beta = encoder_layers_21_norm_self_att_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_21_norm_self_att_weight_to_fp16, x = input_1129_cast_fp16)[name = string("key_43_cast_fp16")]; + bool input_1131_interleave_0 = const()[name = string("input_1131_interleave_0"), val = bool(false)]; + tensor input_1131_cast_fp16 = concat(axis = var_69, interleave = input_1131_interleave_0, values = (cache_85_cast_fp16, key_43_cast_fp16))[name = string("input_1131_cast_fp16")]; + tensor var_4952_begin_0 = const()[name = string("op_4952_begin_0"), val = tensor([0, 14, 0])]; + tensor var_4952_end_0 = const()[name = string("op_4952_end_0"), val = tensor([1, 42, 1024])]; + tensor var_4952_end_mask_0 = const()[name = string("op_4952_end_mask_0"), val = tensor([true, true, true])]; + tensor var_4952_cast_fp16 = slice_by_index(begin = var_4952_begin_0, end = var_4952_end_0, end_mask = var_4952_end_mask_0, x = cache_85_cast_fp16)[name = string("op_4952_cast_fp16")]; + bool var_4958_interleave_0 = const()[name = string("op_4958_interleave_0"), val = bool(false)]; + tensor var_4958_cast_fp16 = concat(axis = var_69, interleave = var_4958_interleave_0, values = (var_4952_cast_fp16, key_43_cast_fp16))[name = string("op_4958_cast_fp16")]; + tensor encoder_layers_21_self_attn_linear_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(439784320))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(440832960))))[name = string("encoder_layers_21_self_attn_linear_q_weight_to_fp16_quantized")]; + tensor encoder_layers_21_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_21_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(440835072)))]; + tensor linear_192_cast_fp16 = linear(bias = encoder_layers_21_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_21_self_attn_linear_q_weight_to_fp16_quantized, x = key_43_cast_fp16)[name = string("linear_192_cast_fp16")]; + tensor var_4963 = const()[name = string("op_4963"), val = tensor([1, -1, 8, 128])]; + tensor q_127_cast_fp16 = reshape(shape = var_4963, x = linear_192_cast_fp16)[name = string("q_127_cast_fp16")]; + tensor encoder_layers_21_self_attn_linear_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(440837184))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(441885824))))[name = string("encoder_layers_21_self_attn_linear_k_weight_to_fp16_quantized")]; + tensor encoder_layers_21_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_21_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(441887936)))]; + tensor linear_193_cast_fp16 = linear(bias = encoder_layers_21_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_21_self_attn_linear_k_weight_to_fp16_quantized, x = input_1131_cast_fp16)[name = string("linear_193_cast_fp16")]; + tensor var_4968 = const()[name = string("op_4968"), val = tensor([1, -1, 8, 128])]; + tensor k_85_cast_fp16 = reshape(shape = var_4968, x = linear_193_cast_fp16)[name = string("k_85_cast_fp16")]; + tensor encoder_layers_21_self_attn_linear_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(441890048))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(442938688))))[name = string("encoder_layers_21_self_attn_linear_v_weight_to_fp16_quantized")]; + tensor encoder_layers_21_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_21_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(442940800)))]; + tensor linear_194_cast_fp16 = linear(bias = encoder_layers_21_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_21_self_attn_linear_v_weight_to_fp16_quantized, x = input_1131_cast_fp16)[name = string("linear_194_cast_fp16")]; + tensor var_4973 = const()[name = string("op_4973"), val = tensor([1, -1, 8, 128])]; + tensor v_43_cast_fp16 = reshape(shape = var_4973, x = linear_194_cast_fp16)[name = string("v_43_cast_fp16")]; + tensor value_51_perm_0 = const()[name = string("value_51_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_21_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_21_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(442942912)))]; + tensor var_4986_cast_fp16 = add(x = q_127_cast_fp16, y = encoder_layers_21_self_attn_pos_bias_u_to_fp16)[name = string("op_4986_cast_fp16")]; + tensor encoder_layers_21_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_21_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(442945024)))]; + tensor var_4988_cast_fp16 = add(x = q_127_cast_fp16, y = encoder_layers_21_self_attn_pos_bias_v_to_fp16)[name = string("op_4988_cast_fp16")]; + tensor q_with_bias_v_43_perm_0 = const()[name = string("q_with_bias_v_43_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_553_transpose_x_0 = const()[name = string("x_553_transpose_x_0"), val = bool(false)]; + bool x_553_transpose_y_0 = const()[name = string("x_553_transpose_y_0"), val = bool(false)]; + tensor op_4990_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(442947136))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(443060864))))[name = string("op_4990_to_fp16_quantized")]; + tensor q_with_bias_v_43_cast_fp16 = transpose(perm = q_with_bias_v_43_perm_0, x = var_4988_cast_fp16)[name = string("transpose_173")]; + tensor x_553_cast_fp16 = matmul(transpose_x = x_553_transpose_x_0, transpose_y = x_553_transpose_y_0, x = q_with_bias_v_43_cast_fp16, y = op_4990_to_fp16_quantized)[name = string("x_553_cast_fp16")]; + tensor x_555_pad_0 = const()[name = string("x_555_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_555_mode_0 = const()[name = string("x_555_mode_0"), val = string("constant")]; + fp16 const_352_to_fp16 = const()[name = string("const_352_to_fp16"), val = fp16(0x0p+0)]; + tensor x_555_cast_fp16 = pad(constant_val = const_352_to_fp16, mode = x_555_mode_0, pad = x_555_pad_0, x = x_553_cast_fp16)[name = string("x_555_cast_fp16")]; + tensor var_4998 = const()[name = string("op_4998"), val = tensor([1, 8, -1, 14])]; + tensor x_557_cast_fp16 = reshape(shape = var_4998, x = x_555_cast_fp16)[name = string("x_557_cast_fp16")]; + tensor var_5002_begin_0 = const()[name = string("op_5002_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_5002_end_0 = const()[name = string("op_5002_end_0"), val = tensor([1, 8, 112, 14])]; + tensor var_5002_end_mask_0 = const()[name = string("op_5002_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_5002_cast_fp16 = slice_by_index(begin = var_5002_begin_0, end = var_5002_end_0, end_mask = var_5002_end_mask_0, x = x_557_cast_fp16)[name = string("op_5002_cast_fp16")]; + tensor var_5003 = const()[name = string("op_5003"), val = tensor([1, 8, 14, 111])]; + tensor matrix_bd_85_cast_fp16 = reshape(shape = var_5003, x = var_5002_cast_fp16)[name = string("matrix_bd_85_cast_fp16")]; + bool matrix_ac_43_transpose_x_0 = const()[name = string("matrix_ac_43_transpose_x_0"), val = bool(false)]; + bool matrix_ac_43_transpose_y_0 = const()[name = string("matrix_ac_43_transpose_y_0"), val = bool(false)]; + tensor transpose_138_perm_0 = const()[name = string("transpose_138_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_139_perm_0 = const()[name = string("transpose_139_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_139 = transpose(perm = transpose_139_perm_0, x = k_85_cast_fp16)[name = string("transpose_171")]; + tensor transpose_138 = transpose(perm = transpose_138_perm_0, x = var_4986_cast_fp16)[name = string("transpose_172")]; + tensor matrix_ac_43_cast_fp16 = matmul(transpose_x = matrix_ac_43_transpose_x_0, transpose_y = matrix_ac_43_transpose_y_0, x = transpose_138, y = transpose_139)[name = string("matrix_ac_43_cast_fp16")]; + tensor matrix_bd_87_begin_0 = const()[name = string("matrix_bd_87_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_87_end_0 = const()[name = string("matrix_bd_87_end_0"), val = tensor([1, 8, 14, 56])]; + tensor matrix_bd_87_end_mask_0 = const()[name = string("matrix_bd_87_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_87_cast_fp16 = slice_by_index(begin = matrix_bd_87_begin_0, end = matrix_bd_87_end_0, end_mask = matrix_bd_87_end_mask_0, x = matrix_bd_85_cast_fp16)[name = string("matrix_bd_87_cast_fp16")]; + tensor var_5012_cast_fp16 = add(x = matrix_ac_43_cast_fp16, y = matrix_bd_87_cast_fp16)[name = string("op_5012_cast_fp16")]; + fp16 _inversed_scores_85_y_0_to_fp16 = const()[name = string("_inversed_scores_85_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_85_cast_fp16 = mul(x = var_5012_cast_fp16, y = _inversed_scores_85_y_0_to_fp16)[name = string("_inversed_scores_85_cast_fp16")]; + tensor scores_87_cast_fp16 = select(a = var_46_to_fp16, b = _inversed_scores_85_cast_fp16, cond = mask_11)[name = string("scores_87_cast_fp16")]; + tensor var_5018_cast_fp16 = softmax(axis = var_60, x = scores_87_cast_fp16)[name = string("op_5018_cast_fp16")]; + tensor input_1133_cast_fp16 = select(a = var_45_to_fp16, b = var_5018_cast_fp16, cond = mask_11)[name = string("input_1133_cast_fp16")]; + bool x_559_transpose_x_0 = const()[name = string("x_559_transpose_x_0"), val = bool(false)]; + bool x_559_transpose_y_0 = const()[name = string("x_559_transpose_y_0"), val = bool(false)]; + tensor value_51_cast_fp16 = transpose(perm = value_51_perm_0, x = v_43_cast_fp16)[name = string("transpose_170")]; + tensor x_559_cast_fp16 = matmul(transpose_x = x_559_transpose_x_0, transpose_y = x_559_transpose_y_0, x = input_1133_cast_fp16, y = value_51_cast_fp16)[name = string("x_559_cast_fp16")]; + tensor var_5022_perm_0 = const()[name = string("op_5022_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_5023 = const()[name = string("op_5023"), val = tensor([1, -1, 1024])]; + tensor var_5022_cast_fp16 = transpose(perm = var_5022_perm_0, x = x_559_cast_fp16)[name = string("transpose_169")]; + tensor input_1135_cast_fp16 = reshape(shape = var_5023, x = var_5022_cast_fp16)[name = string("input_1135_cast_fp16")]; + tensor encoder_layers_21_self_attn_linear_out_weight_to_fp16 = const()[name = string("encoder_layers_21_self_attn_linear_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(443061184)))]; + tensor encoder_layers_21_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_21_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(445158400)))]; + tensor linear_196_cast_fp16 = linear(bias = encoder_layers_21_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_21_self_attn_linear_out_weight_to_fp16, x = input_1135_cast_fp16)[name = string("linear_196_cast_fp16")]; + tensor input_1139_cast_fp16 = add(x = input_1129_cast_fp16, y = linear_196_cast_fp16)[name = string("input_1139_cast_fp16")]; + tensor x_563_axes_0 = const()[name = string("x_563_axes_0"), val = tensor([-1])]; + tensor encoder_layers_21_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_21_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(445160512)))]; + tensor encoder_layers_21_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_21_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(445162624)))]; + tensor x_563_cast_fp16 = layer_norm(axes = x_563_axes_0, beta = encoder_layers_21_norm_conv_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_21_norm_conv_weight_to_fp16, x = input_1139_cast_fp16)[name = string("x_563_cast_fp16")]; + tensor input_1141_perm_0 = const()[name = string("input_1141_perm_0"), val = tensor([0, 2, 1])]; + string input_1143_pad_type_0 = const()[name = string("input_1143_pad_type_0"), val = string("valid")]; + tensor input_1143_strides_0 = const()[name = string("input_1143_strides_0"), val = tensor([1])]; + tensor input_1143_pad_0 = const()[name = string("input_1143_pad_0"), val = tensor([0, 0])]; + tensor input_1143_dilations_0 = const()[name = string("input_1143_dilations_0"), val = tensor([1])]; + int32 input_1143_groups_0 = const()[name = string("input_1143_groups_0"), val = int32(1)]; + tensor encoder_layers_21_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(445164736))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(447261952))))[name = string("encoder_layers_21_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_1141_cast_fp16 = transpose(perm = input_1141_perm_0, x = x_563_cast_fp16)[name = string("transpose_168")]; + tensor input_1143_cast_fp16 = conv(dilations = input_1143_dilations_0, groups = input_1143_groups_0, pad = input_1143_pad_0, pad_type = input_1143_pad_type_0, strides = input_1143_strides_0, weight = encoder_layers_21_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_1141_cast_fp16)[name = string("input_1143_cast_fp16")]; + int32 x_565_split_num_splits_0 = const()[name = string("x_565_split_num_splits_0"), val = int32(2)]; + int32 x_565_split_axis_0 = const()[name = string("x_565_split_axis_0"), val = int32(1)]; + tensor x_565_split_cast_fp16_0, tensor x_565_split_cast_fp16_1 = split(axis = x_565_split_axis_0, num_splits = x_565_split_num_splits_0, x = input_1143_cast_fp16)[name = string("x_565_split_cast_fp16")]; + tensor x_565_split_1_sigmoid_cast_fp16 = sigmoid(x = x_565_split_cast_fp16_1)[name = string("x_565_split_1_sigmoid_cast_fp16")]; + tensor x_565_cast_fp16 = mul(x = x_565_split_cast_fp16_0, y = x_565_split_1_sigmoid_cast_fp16)[name = string("x_565_cast_fp16")]; + tensor input_1145_cast_fp16 = select(a = var_45_to_fp16, b = x_565_cast_fp16, cond = var_576)[name = string("input_1145_cast_fp16")]; + bool new_x_87_interleave_0 = const()[name = string("new_x_87_interleave_0"), val = bool(false)]; + tensor new_x_87_cast_fp16 = concat(axis = var_60, interleave = new_x_87_interleave_0, values = (cache_87_cast_fp16, input_1145_cast_fp16))[name = string("new_x_87_cast_fp16")]; + tensor var_5062_begin_0 = const()[name = string("op_5062_begin_0"), val = tensor([0, 0, 14])]; + tensor var_5062_end_0 = const()[name = string("op_5062_end_0"), val = tensor([1, 1024, 22])]; + tensor var_5062_end_mask_0 = const()[name = string("op_5062_end_mask_0"), val = tensor([true, true, true])]; + tensor var_5062_cast_fp16 = slice_by_index(begin = var_5062_begin_0, end = var_5062_end_0, end_mask = var_5062_end_mask_0, x = new_x_87_cast_fp16)[name = string("op_5062_cast_fp16")]; + string x_567_pad_type_0 = const()[name = string("x_567_pad_type_0"), val = string("valid")]; + int32 x_567_groups_0 = const()[name = string("x_567_groups_0"), val = int32(1024)]; + tensor x_567_strides_0 = const()[name = string("x_567_strides_0"), val = tensor([1])]; + tensor x_567_pad_0 = const()[name = string("x_567_pad_0"), val = tensor([0, 0])]; + tensor x_567_dilations_0 = const()[name = string("x_567_dilations_0"), val = tensor([1])]; + tensor encoder_layers_21_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(447266112))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(447275392))))[name = string("encoder_layers_21_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_567_cast_fp16 = conv(dilations = x_567_dilations_0, groups = x_567_groups_0, pad = x_567_pad_0, pad_type = x_567_pad_type_0, strides = x_567_strides_0, weight = encoder_layers_21_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_87_cast_fp16)[name = string("x_567_cast_fp16")]; + tensor input_1147_perm_0 = const()[name = string("input_1147_perm_0"), val = tensor([0, 2, 1])]; + tensor x_569_axes_0 = const()[name = string("x_569_axes_0"), val = tensor([-1])]; + tensor encoder_layers_21_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_21_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(447277504)))]; + tensor encoder_layers_21_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_21_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(447279616)))]; + tensor input_1147_cast_fp16 = transpose(perm = input_1147_perm_0, x = x_567_cast_fp16)[name = string("transpose_167")]; + tensor x_569_cast_fp16 = layer_norm(axes = x_569_axes_0, beta = encoder_layers_21_conv_batch_norm_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_21_conv_batch_norm_weight_to_fp16, x = input_1147_cast_fp16)[name = string("x_569_cast_fp16")]; + tensor input_1149_perm_0 = const()[name = string("input_1149_perm_0"), val = tensor([0, 2, 1])]; + tensor input_1149_cast_fp16 = transpose(perm = input_1149_perm_0, x = x_569_cast_fp16)[name = string("transpose_166")]; + tensor input_1151_cast_fp16 = silu(x = input_1149_cast_fp16)[name = string("input_1151_cast_fp16")]; + string x_571_pad_type_0 = const()[name = string("x_571_pad_type_0"), val = string("valid")]; + tensor x_571_strides_0 = const()[name = string("x_571_strides_0"), val = tensor([1])]; + tensor x_571_pad_0 = const()[name = string("x_571_pad_0"), val = tensor([0, 0])]; + tensor x_571_dilations_0 = const()[name = string("x_571_dilations_0"), val = tensor([1])]; + int32 x_571_groups_0 = const()[name = string("x_571_groups_0"), val = int32(1)]; + tensor encoder_layers_21_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(447281728))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(448330368))))[name = string("encoder_layers_21_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_571_cast_fp16 = conv(dilations = x_571_dilations_0, groups = x_571_groups_0, pad = x_571_pad_0, pad_type = x_571_pad_type_0, strides = x_571_strides_0, weight = encoder_layers_21_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_1151_cast_fp16)[name = string("x_571_cast_fp16")]; + tensor input_1153_perm_0 = const()[name = string("input_1153_perm_0"), val = tensor([0, 2, 1])]; + tensor input_1153_cast_fp16 = transpose(perm = input_1153_perm_0, x = x_571_cast_fp16)[name = string("transpose_165")]; + tensor input_1155_cast_fp16 = add(x = input_1139_cast_fp16, y = input_1153_cast_fp16)[name = string("input_1155_cast_fp16")]; + tensor input_1157_axes_0 = const()[name = string("input_1157_axes_0"), val = tensor([-1])]; + tensor encoder_layers_21_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_21_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(448332480)))]; + tensor encoder_layers_21_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_21_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(448334592)))]; + tensor input_1157_cast_fp16 = layer_norm(axes = input_1157_axes_0, beta = encoder_layers_21_norm_feed_forward2_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_21_norm_feed_forward2_weight_to_fp16, x = input_1155_cast_fp16)[name = string("input_1157_cast_fp16")]; + tensor encoder_layers_21_feed_forward2_linear1_weight_to_fp16 = const()[name = string("encoder_layers_21_feed_forward2_linear1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(448336704)))]; + tensor encoder_layers_21_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_21_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(456725376)))]; + tensor linear_197_cast_fp16 = linear(bias = encoder_layers_21_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_21_feed_forward2_linear1_weight_to_fp16, x = input_1157_cast_fp16)[name = string("linear_197_cast_fp16")]; + tensor input_1161_cast_fp16 = silu(x = linear_197_cast_fp16)[name = string("input_1161_cast_fp16")]; + tensor encoder_layers_21_feed_forward2_linear2_weight_to_fp16 = const()[name = string("encoder_layers_21_feed_forward2_linear2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(456733632)))]; + tensor encoder_layers_21_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_21_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465122304)))]; + tensor linear_198_cast_fp16 = linear(bias = encoder_layers_21_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_21_feed_forward2_linear2_weight_to_fp16, x = input_1161_cast_fp16)[name = string("linear_198_cast_fp16")]; + fp16 var_5105_to_fp16 = const()[name = string("op_5105_to_fp16"), val = fp16(0x1p-1)]; + tensor var_5106_cast_fp16 = mul(x = linear_198_cast_fp16, y = var_5105_to_fp16)[name = string("op_5106_cast_fp16")]; + tensor input_1167_cast_fp16 = add(x = input_1155_cast_fp16, y = var_5106_cast_fp16)[name = string("input_1167_cast_fp16")]; + tensor input_1169_axes_0 = const()[name = string("input_1169_axes_0"), val = tensor([-1])]; + tensor encoder_layers_21_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_21_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465124416)))]; + tensor encoder_layers_21_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_21_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465126528)))]; + tensor input_1169_cast_fp16 = layer_norm(axes = input_1169_axes_0, beta = encoder_layers_21_norm_out_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_21_norm_out_weight_to_fp16, x = input_1167_cast_fp16)[name = string("input_1169_cast_fp16")]; + tensor cache_89_begin_0 = const()[name = string("cache_89_begin_0"), val = tensor([22, 0, 0, 0])]; + tensor cache_89_end_0 = const()[name = string("cache_89_end_0"), val = tensor([23, 1, 42, 1024])]; + tensor cache_89_end_mask_0 = const()[name = string("cache_89_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_89_squeeze_mask_0 = const()[name = string("cache_89_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_89_cast_fp16 = slice_by_index(begin = cache_89_begin_0, end = cache_89_end_0, end_mask = cache_89_end_mask_0, squeeze_mask = cache_89_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_89_cast_fp16")]; + tensor cache_91_begin_0 = const()[name = string("cache_91_begin_0"), val = tensor([22, 0, 0, 0])]; + tensor cache_91_end_0 = const()[name = string("cache_91_end_0"), val = tensor([23, 1, 1024, 8])]; + tensor cache_91_end_mask_0 = const()[name = string("cache_91_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_91_squeeze_mask_0 = const()[name = string("cache_91_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_91_cast_fp16 = slice_by_index(begin = cache_91_begin_0, end = cache_91_end_0, end_mask = cache_91_end_mask_0, squeeze_mask = cache_91_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_91_cast_fp16")]; + tensor input_1171_axes_0 = const()[name = string("input_1171_axes_0"), val = tensor([-1])]; + tensor encoder_layers_22_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_22_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465128640)))]; + tensor encoder_layers_22_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_22_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465130752)))]; + tensor input_1171_cast_fp16 = layer_norm(axes = input_1171_axes_0, beta = encoder_layers_22_norm_feed_forward1_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_22_norm_feed_forward1_weight_to_fp16, x = input_1169_cast_fp16)[name = string("input_1171_cast_fp16")]; + tensor encoder_layers_22_feed_forward1_linear1_weight_to_fp16 = const()[name = string("encoder_layers_22_feed_forward1_linear1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465132864)))]; + tensor encoder_layers_22_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_22_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(473521536)))]; + tensor linear_199_cast_fp16 = linear(bias = encoder_layers_22_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_22_feed_forward1_linear1_weight_to_fp16, x = input_1171_cast_fp16)[name = string("linear_199_cast_fp16")]; + tensor input_1175_cast_fp16 = silu(x = linear_199_cast_fp16)[name = string("input_1175_cast_fp16")]; + tensor encoder_layers_22_feed_forward1_linear2_weight_to_fp16 = const()[name = string("encoder_layers_22_feed_forward1_linear2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(473529792)))]; + tensor encoder_layers_22_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_22_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481918464)))]; + tensor linear_200_cast_fp16 = linear(bias = encoder_layers_22_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_22_feed_forward1_linear2_weight_to_fp16, x = input_1175_cast_fp16)[name = string("linear_200_cast_fp16")]; + fp16 var_5142_to_fp16 = const()[name = string("op_5142_to_fp16"), val = fp16(0x1p-1)]; + tensor var_5143_cast_fp16 = mul(x = linear_200_cast_fp16, y = var_5142_to_fp16)[name = string("op_5143_cast_fp16")]; + tensor input_1181_cast_fp16 = add(x = input_1169_cast_fp16, y = var_5143_cast_fp16)[name = string("input_1181_cast_fp16")]; + tensor key_45_axes_0 = const()[name = string("key_45_axes_0"), val = tensor([-1])]; + tensor encoder_layers_22_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_22_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481920576)))]; + tensor encoder_layers_22_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_22_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481922688)))]; + tensor key_45_cast_fp16 = layer_norm(axes = key_45_axes_0, beta = encoder_layers_22_norm_self_att_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_22_norm_self_att_weight_to_fp16, x = input_1181_cast_fp16)[name = string("key_45_cast_fp16")]; + bool input_1183_interleave_0 = const()[name = string("input_1183_interleave_0"), val = bool(false)]; + tensor input_1183_cast_fp16 = concat(axis = var_69, interleave = input_1183_interleave_0, values = (cache_89_cast_fp16, key_45_cast_fp16))[name = string("input_1183_cast_fp16")]; + tensor var_5165_begin_0 = const()[name = string("op_5165_begin_0"), val = tensor([0, 14, 0])]; + tensor var_5165_end_0 = const()[name = string("op_5165_end_0"), val = tensor([1, 42, 1024])]; + tensor var_5165_end_mask_0 = const()[name = string("op_5165_end_mask_0"), val = tensor([true, true, true])]; + tensor var_5165_cast_fp16 = slice_by_index(begin = var_5165_begin_0, end = var_5165_end_0, end_mask = var_5165_end_mask_0, x = cache_89_cast_fp16)[name = string("op_5165_cast_fp16")]; + bool var_5171_interleave_0 = const()[name = string("op_5171_interleave_0"), val = bool(false)]; + tensor var_5171_cast_fp16 = concat(axis = var_69, interleave = var_5171_interleave_0, values = (var_5165_cast_fp16, key_45_cast_fp16))[name = string("op_5171_cast_fp16")]; + tensor encoder_layers_22_self_attn_linear_q_weight_to_fp16 = const()[name = string("encoder_layers_22_self_attn_linear_q_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481924800)))]; + tensor encoder_layers_22_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_22_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(484022016)))]; + tensor linear_201_cast_fp16 = linear(bias = encoder_layers_22_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_22_self_attn_linear_q_weight_to_fp16, x = key_45_cast_fp16)[name = string("linear_201_cast_fp16")]; + tensor var_5176 = const()[name = string("op_5176"), val = tensor([1, -1, 8, 128])]; + tensor q_133_cast_fp16 = reshape(shape = var_5176, x = linear_201_cast_fp16)[name = string("q_133_cast_fp16")]; + tensor encoder_layers_22_self_attn_linear_k_weight_to_fp16 = const()[name = string("encoder_layers_22_self_attn_linear_k_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(484024128)))]; + tensor encoder_layers_22_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_22_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(486121344)))]; + tensor linear_202_cast_fp16 = linear(bias = encoder_layers_22_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_22_self_attn_linear_k_weight_to_fp16, x = input_1183_cast_fp16)[name = string("linear_202_cast_fp16")]; + tensor var_5181 = const()[name = string("op_5181"), val = tensor([1, -1, 8, 128])]; + tensor k_89_cast_fp16 = reshape(shape = var_5181, x = linear_202_cast_fp16)[name = string("k_89_cast_fp16")]; + tensor encoder_layers_22_self_attn_linear_v_weight_to_fp16 = const()[name = string("encoder_layers_22_self_attn_linear_v_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(486123456)))]; + tensor encoder_layers_22_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_22_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(488220672)))]; + tensor linear_203_cast_fp16 = linear(bias = encoder_layers_22_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_22_self_attn_linear_v_weight_to_fp16, x = input_1183_cast_fp16)[name = string("linear_203_cast_fp16")]; + tensor var_5186 = const()[name = string("op_5186"), val = tensor([1, -1, 8, 128])]; + tensor v_45_cast_fp16 = reshape(shape = var_5186, x = linear_203_cast_fp16)[name = string("v_45_cast_fp16")]; + tensor value_53_perm_0 = const()[name = string("value_53_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_22_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_22_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(488222784)))]; + tensor var_5199_cast_fp16 = add(x = q_133_cast_fp16, y = encoder_layers_22_self_attn_pos_bias_u_to_fp16)[name = string("op_5199_cast_fp16")]; + tensor encoder_layers_22_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_22_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(488224896)))]; + tensor var_5201_cast_fp16 = add(x = q_133_cast_fp16, y = encoder_layers_22_self_attn_pos_bias_v_to_fp16)[name = string("op_5201_cast_fp16")]; + tensor q_with_bias_v_45_perm_0 = const()[name = string("q_with_bias_v_45_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_579_transpose_x_0 = const()[name = string("x_579_transpose_x_0"), val = bool(false)]; + bool x_579_transpose_y_0 = const()[name = string("x_579_transpose_y_0"), val = bool(false)]; + tensor op_5203_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(488227008))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(488340736))))[name = string("op_5203_to_fp16_quantized")]; + tensor q_with_bias_v_45_cast_fp16 = transpose(perm = q_with_bias_v_45_perm_0, x = var_5201_cast_fp16)[name = string("transpose_164")]; + tensor x_579_cast_fp16 = matmul(transpose_x = x_579_transpose_x_0, transpose_y = x_579_transpose_y_0, x = q_with_bias_v_45_cast_fp16, y = op_5203_to_fp16_quantized)[name = string("x_579_cast_fp16")]; + tensor x_581_pad_0 = const()[name = string("x_581_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_581_mode_0 = const()[name = string("x_581_mode_0"), val = string("constant")]; + fp16 const_365_to_fp16 = const()[name = string("const_365_to_fp16"), val = fp16(0x0p+0)]; + tensor x_581_cast_fp16 = pad(constant_val = const_365_to_fp16, mode = x_581_mode_0, pad = x_581_pad_0, x = x_579_cast_fp16)[name = string("x_581_cast_fp16")]; + tensor var_5211 = const()[name = string("op_5211"), val = tensor([1, 8, -1, 14])]; + tensor x_583_cast_fp16 = reshape(shape = var_5211, x = x_581_cast_fp16)[name = string("x_583_cast_fp16")]; + tensor var_5215_begin_0 = const()[name = string("op_5215_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_5215_end_0 = const()[name = string("op_5215_end_0"), val = tensor([1, 8, 112, 14])]; + tensor var_5215_end_mask_0 = const()[name = string("op_5215_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_5215_cast_fp16 = slice_by_index(begin = var_5215_begin_0, end = var_5215_end_0, end_mask = var_5215_end_mask_0, x = x_583_cast_fp16)[name = string("op_5215_cast_fp16")]; + tensor var_5216 = const()[name = string("op_5216"), val = tensor([1, 8, 14, 111])]; + tensor matrix_bd_89_cast_fp16 = reshape(shape = var_5216, x = var_5215_cast_fp16)[name = string("matrix_bd_89_cast_fp16")]; + bool matrix_ac_45_transpose_x_0 = const()[name = string("matrix_ac_45_transpose_x_0"), val = bool(false)]; + bool matrix_ac_45_transpose_y_0 = const()[name = string("matrix_ac_45_transpose_y_0"), val = bool(false)]; + tensor transpose_140_perm_0 = const()[name = string("transpose_140_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_141_perm_0 = const()[name = string("transpose_141_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_141 = transpose(perm = transpose_141_perm_0, x = k_89_cast_fp16)[name = string("transpose_162")]; + tensor transpose_140 = transpose(perm = transpose_140_perm_0, x = var_5199_cast_fp16)[name = string("transpose_163")]; + tensor matrix_ac_45_cast_fp16 = matmul(transpose_x = matrix_ac_45_transpose_x_0, transpose_y = matrix_ac_45_transpose_y_0, x = transpose_140, y = transpose_141)[name = string("matrix_ac_45_cast_fp16")]; + tensor matrix_bd_91_begin_0 = const()[name = string("matrix_bd_91_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_91_end_0 = const()[name = string("matrix_bd_91_end_0"), val = tensor([1, 8, 14, 56])]; + tensor matrix_bd_91_end_mask_0 = const()[name = string("matrix_bd_91_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_91_cast_fp16 = slice_by_index(begin = matrix_bd_91_begin_0, end = matrix_bd_91_end_0, end_mask = matrix_bd_91_end_mask_0, x = matrix_bd_89_cast_fp16)[name = string("matrix_bd_91_cast_fp16")]; + tensor var_5225_cast_fp16 = add(x = matrix_ac_45_cast_fp16, y = matrix_bd_91_cast_fp16)[name = string("op_5225_cast_fp16")]; + fp16 _inversed_scores_89_y_0_to_fp16 = const()[name = string("_inversed_scores_89_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_89_cast_fp16 = mul(x = var_5225_cast_fp16, y = _inversed_scores_89_y_0_to_fp16)[name = string("_inversed_scores_89_cast_fp16")]; + tensor scores_91_cast_fp16 = select(a = var_46_to_fp16, b = _inversed_scores_89_cast_fp16, cond = mask_11)[name = string("scores_91_cast_fp16")]; + tensor var_5231_cast_fp16 = softmax(axis = var_60, x = scores_91_cast_fp16)[name = string("op_5231_cast_fp16")]; + tensor input_1185_cast_fp16 = select(a = var_45_to_fp16, b = var_5231_cast_fp16, cond = mask_11)[name = string("input_1185_cast_fp16")]; + bool x_585_transpose_x_0 = const()[name = string("x_585_transpose_x_0"), val = bool(false)]; + bool x_585_transpose_y_0 = const()[name = string("x_585_transpose_y_0"), val = bool(false)]; + tensor value_53_cast_fp16 = transpose(perm = value_53_perm_0, x = v_45_cast_fp16)[name = string("transpose_161")]; + tensor x_585_cast_fp16 = matmul(transpose_x = x_585_transpose_x_0, transpose_y = x_585_transpose_y_0, x = input_1185_cast_fp16, y = value_53_cast_fp16)[name = string("x_585_cast_fp16")]; + tensor var_5235_perm_0 = const()[name = string("op_5235_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_5236 = const()[name = string("op_5236"), val = tensor([1, -1, 1024])]; + tensor var_5235_cast_fp16 = transpose(perm = var_5235_perm_0, x = x_585_cast_fp16)[name = string("transpose_160")]; + tensor input_1187_cast_fp16 = reshape(shape = var_5236, x = var_5235_cast_fp16)[name = string("input_1187_cast_fp16")]; + tensor encoder_layers_22_self_attn_linear_out_weight_to_fp16 = const()[name = string("encoder_layers_22_self_attn_linear_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(488341056)))]; + tensor encoder_layers_22_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_22_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(490438272)))]; + tensor linear_205_cast_fp16 = linear(bias = encoder_layers_22_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_22_self_attn_linear_out_weight_to_fp16, x = input_1187_cast_fp16)[name = string("linear_205_cast_fp16")]; + tensor input_1191_cast_fp16 = add(x = input_1181_cast_fp16, y = linear_205_cast_fp16)[name = string("input_1191_cast_fp16")]; + tensor x_589_axes_0 = const()[name = string("x_589_axes_0"), val = tensor([-1])]; + tensor encoder_layers_22_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_22_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(490440384)))]; + tensor encoder_layers_22_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_22_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(490442496)))]; + tensor x_589_cast_fp16 = layer_norm(axes = x_589_axes_0, beta = encoder_layers_22_norm_conv_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_22_norm_conv_weight_to_fp16, x = input_1191_cast_fp16)[name = string("x_589_cast_fp16")]; + tensor input_1193_perm_0 = const()[name = string("input_1193_perm_0"), val = tensor([0, 2, 1])]; + string input_1195_pad_type_0 = const()[name = string("input_1195_pad_type_0"), val = string("valid")]; + tensor input_1195_strides_0 = const()[name = string("input_1195_strides_0"), val = tensor([1])]; + tensor input_1195_pad_0 = const()[name = string("input_1195_pad_0"), val = tensor([0, 0])]; + tensor input_1195_dilations_0 = const()[name = string("input_1195_dilations_0"), val = tensor([1])]; + int32 input_1195_groups_0 = const()[name = string("input_1195_groups_0"), val = int32(1)]; + tensor encoder_layers_22_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(490444608))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(492541824))))[name = string("encoder_layers_22_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_1193_cast_fp16 = transpose(perm = input_1193_perm_0, x = x_589_cast_fp16)[name = string("transpose_159")]; + tensor input_1195_cast_fp16 = conv(dilations = input_1195_dilations_0, groups = input_1195_groups_0, pad = input_1195_pad_0, pad_type = input_1195_pad_type_0, strides = input_1195_strides_0, weight = encoder_layers_22_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_1193_cast_fp16)[name = string("input_1195_cast_fp16")]; + int32 x_591_split_num_splits_0 = const()[name = string("x_591_split_num_splits_0"), val = int32(2)]; + int32 x_591_split_axis_0 = const()[name = string("x_591_split_axis_0"), val = int32(1)]; + tensor x_591_split_cast_fp16_0, tensor x_591_split_cast_fp16_1 = split(axis = x_591_split_axis_0, num_splits = x_591_split_num_splits_0, x = input_1195_cast_fp16)[name = string("x_591_split_cast_fp16")]; + tensor x_591_split_1_sigmoid_cast_fp16 = sigmoid(x = x_591_split_cast_fp16_1)[name = string("x_591_split_1_sigmoid_cast_fp16")]; + tensor x_591_cast_fp16 = mul(x = x_591_split_cast_fp16_0, y = x_591_split_1_sigmoid_cast_fp16)[name = string("x_591_cast_fp16")]; + tensor input_1197_cast_fp16 = select(a = var_45_to_fp16, b = x_591_cast_fp16, cond = var_576)[name = string("input_1197_cast_fp16")]; + bool new_x_91_interleave_0 = const()[name = string("new_x_91_interleave_0"), val = bool(false)]; + tensor new_x_91_cast_fp16 = concat(axis = var_60, interleave = new_x_91_interleave_0, values = (cache_91_cast_fp16, input_1197_cast_fp16))[name = string("new_x_91_cast_fp16")]; + tensor var_5275_begin_0 = const()[name = string("op_5275_begin_0"), val = tensor([0, 0, 14])]; + tensor var_5275_end_0 = const()[name = string("op_5275_end_0"), val = tensor([1, 1024, 22])]; + tensor var_5275_end_mask_0 = const()[name = string("op_5275_end_mask_0"), val = tensor([true, true, true])]; + tensor var_5275_cast_fp16 = slice_by_index(begin = var_5275_begin_0, end = var_5275_end_0, end_mask = var_5275_end_mask_0, x = new_x_91_cast_fp16)[name = string("op_5275_cast_fp16")]; + string x_593_pad_type_0 = const()[name = string("x_593_pad_type_0"), val = string("valid")]; + int32 x_593_groups_0 = const()[name = string("x_593_groups_0"), val = int32(1024)]; + tensor x_593_strides_0 = const()[name = string("x_593_strides_0"), val = tensor([1])]; + tensor x_593_pad_0 = const()[name = string("x_593_pad_0"), val = tensor([0, 0])]; + tensor x_593_dilations_0 = const()[name = string("x_593_dilations_0"), val = tensor([1])]; + tensor encoder_layers_22_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(492545984))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(492555264))))[name = string("encoder_layers_22_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_593_cast_fp16 = conv(dilations = x_593_dilations_0, groups = x_593_groups_0, pad = x_593_pad_0, pad_type = x_593_pad_type_0, strides = x_593_strides_0, weight = encoder_layers_22_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_91_cast_fp16)[name = string("x_593_cast_fp16")]; + tensor input_1199_perm_0 = const()[name = string("input_1199_perm_0"), val = tensor([0, 2, 1])]; + tensor x_595_axes_0 = const()[name = string("x_595_axes_0"), val = tensor([-1])]; + tensor encoder_layers_22_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_22_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(492557376)))]; + tensor encoder_layers_22_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_22_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(492559488)))]; + tensor input_1199_cast_fp16 = transpose(perm = input_1199_perm_0, x = x_593_cast_fp16)[name = string("transpose_158")]; + tensor x_595_cast_fp16 = layer_norm(axes = x_595_axes_0, beta = encoder_layers_22_conv_batch_norm_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_22_conv_batch_norm_weight_to_fp16, x = input_1199_cast_fp16)[name = string("x_595_cast_fp16")]; + tensor input_1201_perm_0 = const()[name = string("input_1201_perm_0"), val = tensor([0, 2, 1])]; + tensor input_1201_cast_fp16 = transpose(perm = input_1201_perm_0, x = x_595_cast_fp16)[name = string("transpose_157")]; + tensor input_1203_cast_fp16 = silu(x = input_1201_cast_fp16)[name = string("input_1203_cast_fp16")]; + string x_597_pad_type_0 = const()[name = string("x_597_pad_type_0"), val = string("valid")]; + tensor x_597_strides_0 = const()[name = string("x_597_strides_0"), val = tensor([1])]; + tensor x_597_pad_0 = const()[name = string("x_597_pad_0"), val = tensor([0, 0])]; + tensor x_597_dilations_0 = const()[name = string("x_597_dilations_0"), val = tensor([1])]; + int32 x_597_groups_0 = const()[name = string("x_597_groups_0"), val = int32(1)]; + tensor encoder_layers_22_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(492561600))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(493610240))))[name = string("encoder_layers_22_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_597_cast_fp16 = conv(dilations = x_597_dilations_0, groups = x_597_groups_0, pad = x_597_pad_0, pad_type = x_597_pad_type_0, strides = x_597_strides_0, weight = encoder_layers_22_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_1203_cast_fp16)[name = string("x_597_cast_fp16")]; + tensor input_1205_perm_0 = const()[name = string("input_1205_perm_0"), val = tensor([0, 2, 1])]; + tensor input_1205_cast_fp16 = transpose(perm = input_1205_perm_0, x = x_597_cast_fp16)[name = string("transpose_156")]; + tensor input_1207_cast_fp16 = add(x = input_1191_cast_fp16, y = input_1205_cast_fp16)[name = string("input_1207_cast_fp16")]; + tensor input_1209_axes_0 = const()[name = string("input_1209_axes_0"), val = tensor([-1])]; + tensor encoder_layers_22_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_22_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(493612352)))]; + tensor encoder_layers_22_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_22_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(493614464)))]; + tensor input_1209_cast_fp16 = layer_norm(axes = input_1209_axes_0, beta = encoder_layers_22_norm_feed_forward2_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_22_norm_feed_forward2_weight_to_fp16, x = input_1207_cast_fp16)[name = string("input_1209_cast_fp16")]; + tensor encoder_layers_22_feed_forward2_linear1_weight_to_fp16 = const()[name = string("encoder_layers_22_feed_forward2_linear1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(493616576)))]; + tensor encoder_layers_22_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_22_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(502005248)))]; + tensor linear_206_cast_fp16 = linear(bias = encoder_layers_22_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_22_feed_forward2_linear1_weight_to_fp16, x = input_1209_cast_fp16)[name = string("linear_206_cast_fp16")]; + tensor input_1213_cast_fp16 = silu(x = linear_206_cast_fp16)[name = string("input_1213_cast_fp16")]; + tensor encoder_layers_22_feed_forward2_linear2_weight_to_fp16 = const()[name = string("encoder_layers_22_feed_forward2_linear2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(502013504)))]; + tensor encoder_layers_22_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_22_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(510402176)))]; + tensor linear_207_cast_fp16 = linear(bias = encoder_layers_22_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_22_feed_forward2_linear2_weight_to_fp16, x = input_1213_cast_fp16)[name = string("linear_207_cast_fp16")]; + fp16 var_5318_to_fp16 = const()[name = string("op_5318_to_fp16"), val = fp16(0x1p-1)]; + tensor var_5319_cast_fp16 = mul(x = linear_207_cast_fp16, y = var_5318_to_fp16)[name = string("op_5319_cast_fp16")]; + tensor input_1219_cast_fp16 = add(x = input_1207_cast_fp16, y = var_5319_cast_fp16)[name = string("input_1219_cast_fp16")]; + tensor input_1221_axes_0 = const()[name = string("input_1221_axes_0"), val = tensor([-1])]; + tensor encoder_layers_22_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_22_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(510404288)))]; + tensor encoder_layers_22_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_22_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(510406400)))]; + tensor input_1221_cast_fp16 = layer_norm(axes = input_1221_axes_0, beta = encoder_layers_22_norm_out_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_22_norm_out_weight_to_fp16, x = input_1219_cast_fp16)[name = string("input_1221_cast_fp16")]; + tensor cache_93_begin_0 = const()[name = string("cache_93_begin_0"), val = tensor([23, 0, 0, 0])]; + tensor cache_93_end_0 = const()[name = string("cache_93_end_0"), val = tensor([24, 1, 42, 1024])]; + tensor cache_93_end_mask_0 = const()[name = string("cache_93_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_93_squeeze_mask_0 = const()[name = string("cache_93_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_93_cast_fp16 = slice_by_index(begin = cache_93_begin_0, end = cache_93_end_0, end_mask = cache_93_end_mask_0, squeeze_mask = cache_93_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_93_cast_fp16")]; + tensor cache_begin_0 = const()[name = string("cache_begin_0"), val = tensor([23, 0, 0, 0])]; + tensor cache_end_0 = const()[name = string("cache_end_0"), val = tensor([24, 1, 1024, 8])]; + tensor cache_end_mask_0 = const()[name = string("cache_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_squeeze_mask_0 = const()[name = string("cache_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_cast_fp16 = slice_by_index(begin = cache_begin_0, end = cache_end_0, end_mask = cache_end_mask_0, squeeze_mask = cache_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_cast_fp16")]; + tensor input_1223_axes_0 = const()[name = string("input_1223_axes_0"), val = tensor([-1])]; + tensor encoder_layers_23_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_23_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(510408512)))]; + tensor encoder_layers_23_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_23_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(510410624)))]; + tensor input_1223_cast_fp16 = layer_norm(axes = input_1223_axes_0, beta = encoder_layers_23_norm_feed_forward1_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_23_norm_feed_forward1_weight_to_fp16, x = input_1221_cast_fp16)[name = string("input_1223_cast_fp16")]; + tensor encoder_layers_23_feed_forward1_linear1_weight_to_fp16 = const()[name = string("encoder_layers_23_feed_forward1_linear1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(510412736)))]; + tensor encoder_layers_23_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_23_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(518801408)))]; + tensor linear_208_cast_fp16 = linear(bias = encoder_layers_23_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_23_feed_forward1_linear1_weight_to_fp16, x = input_1223_cast_fp16)[name = string("linear_208_cast_fp16")]; + tensor input_1227_cast_fp16 = silu(x = linear_208_cast_fp16)[name = string("input_1227_cast_fp16")]; + tensor encoder_layers_23_feed_forward1_linear2_weight_to_fp16 = const()[name = string("encoder_layers_23_feed_forward1_linear2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(518809664)))]; + tensor encoder_layers_23_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_23_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(527198336)))]; + tensor linear_209_cast_fp16 = linear(bias = encoder_layers_23_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_23_feed_forward1_linear2_weight_to_fp16, x = input_1227_cast_fp16)[name = string("linear_209_cast_fp16")]; + fp16 var_5355_to_fp16 = const()[name = string("op_5355_to_fp16"), val = fp16(0x1p-1)]; + tensor var_5356_cast_fp16 = mul(x = linear_209_cast_fp16, y = var_5355_to_fp16)[name = string("op_5356_cast_fp16")]; + tensor input_1233_cast_fp16 = add(x = input_1221_cast_fp16, y = var_5356_cast_fp16)[name = string("input_1233_cast_fp16")]; + tensor key_axes_0 = const()[name = string("key_axes_0"), val = tensor([-1])]; + tensor encoder_layers_23_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_23_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(527200448)))]; + tensor encoder_layers_23_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_23_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(527202560)))]; + tensor key_cast_fp16 = layer_norm(axes = key_axes_0, beta = encoder_layers_23_norm_self_att_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_23_norm_self_att_weight_to_fp16, x = input_1233_cast_fp16)[name = string("key_cast_fp16")]; + bool input_1235_interleave_0 = const()[name = string("input_1235_interleave_0"), val = bool(false)]; + tensor input_1235_cast_fp16 = concat(axis = var_69, interleave = input_1235_interleave_0, values = (cache_93_cast_fp16, key_cast_fp16))[name = string("input_1235_cast_fp16")]; + tensor var_5378_begin_0 = const()[name = string("op_5378_begin_0"), val = tensor([0, 14, 0])]; + tensor var_5378_end_0 = const()[name = string("op_5378_end_0"), val = tensor([1, 42, 1024])]; + tensor var_5378_end_mask_0 = const()[name = string("op_5378_end_mask_0"), val = tensor([true, true, true])]; + tensor var_5378_cast_fp16 = slice_by_index(begin = var_5378_begin_0, end = var_5378_end_0, end_mask = var_5378_end_mask_0, x = cache_93_cast_fp16)[name = string("op_5378_cast_fp16")]; + bool cache_last_channel_cur_interleave_0 = const()[name = string("cache_last_channel_cur_interleave_0"), val = bool(false)]; + tensor cache_last_channel_cur_cast_fp16 = concat(axis = var_69, interleave = cache_last_channel_cur_interleave_0, values = (var_5378_cast_fp16, key_cast_fp16))[name = string("cache_last_channel_cur_cast_fp16")]; + tensor encoder_layers_23_self_attn_linear_q_weight_to_fp16 = const()[name = string("encoder_layers_23_self_attn_linear_q_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(527204672)))]; + tensor encoder_layers_23_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_23_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(529301888)))]; + tensor linear_210_cast_fp16 = linear(bias = encoder_layers_23_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_23_self_attn_linear_q_weight_to_fp16, x = key_cast_fp16)[name = string("linear_210_cast_fp16")]; + tensor var_5389 = const()[name = string("op_5389"), val = tensor([1, -1, 8, 128])]; + tensor q_139_cast_fp16 = reshape(shape = var_5389, x = linear_210_cast_fp16)[name = string("q_139_cast_fp16")]; + tensor encoder_layers_23_self_attn_linear_k_weight_to_fp16 = const()[name = string("encoder_layers_23_self_attn_linear_k_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(529304000)))]; + tensor encoder_layers_23_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_23_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(531401216)))]; + tensor linear_211_cast_fp16 = linear(bias = encoder_layers_23_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_23_self_attn_linear_k_weight_to_fp16, x = input_1235_cast_fp16)[name = string("linear_211_cast_fp16")]; + tensor var_5394 = const()[name = string("op_5394"), val = tensor([1, -1, 8, 128])]; + tensor k_93_cast_fp16 = reshape(shape = var_5394, x = linear_211_cast_fp16)[name = string("k_93_cast_fp16")]; + tensor encoder_layers_23_self_attn_linear_v_weight_to_fp16 = const()[name = string("encoder_layers_23_self_attn_linear_v_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(531403328)))]; + tensor encoder_layers_23_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_23_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(533500544)))]; + tensor linear_212_cast_fp16 = linear(bias = encoder_layers_23_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_23_self_attn_linear_v_weight_to_fp16, x = input_1235_cast_fp16)[name = string("linear_212_cast_fp16")]; + tensor var_5399 = const()[name = string("op_5399"), val = tensor([1, -1, 8, 128])]; + tensor v_cast_fp16 = reshape(shape = var_5399, x = linear_212_cast_fp16)[name = string("v_cast_fp16")]; + tensor value_perm_0 = const()[name = string("value_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_23_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_23_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(533502656)))]; + tensor var_5412_cast_fp16 = add(x = q_139_cast_fp16, y = encoder_layers_23_self_attn_pos_bias_u_to_fp16)[name = string("op_5412_cast_fp16")]; + tensor encoder_layers_23_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_23_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(533504768)))]; + tensor var_5414_cast_fp16 = add(x = q_139_cast_fp16, y = encoder_layers_23_self_attn_pos_bias_v_to_fp16)[name = string("op_5414_cast_fp16")]; + tensor q_with_bias_v_perm_0 = const()[name = string("q_with_bias_v_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_605_transpose_x_0 = const()[name = string("x_605_transpose_x_0"), val = bool(false)]; + bool x_605_transpose_y_0 = const()[name = string("x_605_transpose_y_0"), val = bool(false)]; + tensor op_5416_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(533506880))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(533620608))))[name = string("op_5416_to_fp16_quantized")]; + tensor q_with_bias_v_cast_fp16 = transpose(perm = q_with_bias_v_perm_0, x = var_5414_cast_fp16)[name = string("transpose_155")]; + tensor x_605_cast_fp16 = matmul(transpose_x = x_605_transpose_x_0, transpose_y = x_605_transpose_y_0, x = q_with_bias_v_cast_fp16, y = op_5416_to_fp16_quantized)[name = string("x_605_cast_fp16")]; + tensor x_607_pad_0 = const()[name = string("x_607_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_607_mode_0 = const()[name = string("x_607_mode_0"), val = string("constant")]; + fp16 const_378_to_fp16 = const()[name = string("const_378_to_fp16"), val = fp16(0x0p+0)]; + tensor x_607_cast_fp16 = pad(constant_val = const_378_to_fp16, mode = x_607_mode_0, pad = x_607_pad_0, x = x_605_cast_fp16)[name = string("x_607_cast_fp16")]; + tensor var_5424 = const()[name = string("op_5424"), val = tensor([1, 8, -1, 14])]; + tensor x_609_cast_fp16 = reshape(shape = var_5424, x = x_607_cast_fp16)[name = string("x_609_cast_fp16")]; + tensor var_5428_begin_0 = const()[name = string("op_5428_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_5428_end_0 = const()[name = string("op_5428_end_0"), val = tensor([1, 8, 112, 14])]; + tensor var_5428_end_mask_0 = const()[name = string("op_5428_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_5428_cast_fp16 = slice_by_index(begin = var_5428_begin_0, end = var_5428_end_0, end_mask = var_5428_end_mask_0, x = x_609_cast_fp16)[name = string("op_5428_cast_fp16")]; + tensor var_5429 = const()[name = string("op_5429"), val = tensor([1, 8, 14, 111])]; + tensor matrix_bd_93_cast_fp16 = reshape(shape = var_5429, x = var_5428_cast_fp16)[name = string("matrix_bd_93_cast_fp16")]; + bool matrix_ac_transpose_x_0 = const()[name = string("matrix_ac_transpose_x_0"), val = bool(false)]; + bool matrix_ac_transpose_y_0 = const()[name = string("matrix_ac_transpose_y_0"), val = bool(false)]; + tensor transpose_142_perm_0 = const()[name = string("transpose_142_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_143_perm_0 = const()[name = string("transpose_143_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_143 = transpose(perm = transpose_143_perm_0, x = k_93_cast_fp16)[name = string("transpose_153")]; + tensor transpose_142 = transpose(perm = transpose_142_perm_0, x = var_5412_cast_fp16)[name = string("transpose_154")]; + tensor matrix_ac_cast_fp16 = matmul(transpose_x = matrix_ac_transpose_x_0, transpose_y = matrix_ac_transpose_y_0, x = transpose_142, y = transpose_143)[name = string("matrix_ac_cast_fp16")]; + tensor matrix_bd_begin_0 = const()[name = string("matrix_bd_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_end_0 = const()[name = string("matrix_bd_end_0"), val = tensor([1, 8, 14, 56])]; + tensor matrix_bd_end_mask_0 = const()[name = string("matrix_bd_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_cast_fp16 = slice_by_index(begin = matrix_bd_begin_0, end = matrix_bd_end_0, end_mask = matrix_bd_end_mask_0, x = matrix_bd_93_cast_fp16)[name = string("matrix_bd_cast_fp16")]; + tensor var_5438_cast_fp16 = add(x = matrix_ac_cast_fp16, y = matrix_bd_cast_fp16)[name = string("op_5438_cast_fp16")]; + fp16 _inversed_scores_93_y_0_to_fp16 = const()[name = string("_inversed_scores_93_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_93_cast_fp16 = mul(x = var_5438_cast_fp16, y = _inversed_scores_93_y_0_to_fp16)[name = string("_inversed_scores_93_cast_fp16")]; + tensor scores_cast_fp16 = select(a = var_46_to_fp16, b = _inversed_scores_93_cast_fp16, cond = mask_11)[name = string("scores_cast_fp16")]; + tensor var_5444_cast_fp16 = softmax(axis = var_60, x = scores_cast_fp16)[name = string("op_5444_cast_fp16")]; + tensor input_1237_cast_fp16 = select(a = var_45_to_fp16, b = var_5444_cast_fp16, cond = mask_11)[name = string("input_1237_cast_fp16")]; + bool x_611_transpose_x_0 = const()[name = string("x_611_transpose_x_0"), val = bool(false)]; + bool x_611_transpose_y_0 = const()[name = string("x_611_transpose_y_0"), val = bool(false)]; + tensor value_cast_fp16 = transpose(perm = value_perm_0, x = v_cast_fp16)[name = string("transpose_152")]; + tensor x_611_cast_fp16 = matmul(transpose_x = x_611_transpose_x_0, transpose_y = x_611_transpose_y_0, x = input_1237_cast_fp16, y = value_cast_fp16)[name = string("x_611_cast_fp16")]; + tensor var_5448_perm_0 = const()[name = string("op_5448_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_5449 = const()[name = string("op_5449"), val = tensor([1, -1, 1024])]; + tensor var_5448_cast_fp16 = transpose(perm = var_5448_perm_0, x = x_611_cast_fp16)[name = string("transpose_151")]; + tensor input_1239_cast_fp16 = reshape(shape = var_5449, x = var_5448_cast_fp16)[name = string("input_1239_cast_fp16")]; + tensor encoder_layers_23_self_attn_linear_out_weight_to_fp16 = const()[name = string("encoder_layers_23_self_attn_linear_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(533620928)))]; + tensor encoder_layers_23_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_23_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(535718144)))]; + tensor linear_214_cast_fp16 = linear(bias = encoder_layers_23_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_23_self_attn_linear_out_weight_to_fp16, x = input_1239_cast_fp16)[name = string("linear_214_cast_fp16")]; + tensor input_1243_cast_fp16 = add(x = input_1233_cast_fp16, y = linear_214_cast_fp16)[name = string("input_1243_cast_fp16")]; + tensor x_615_axes_0 = const()[name = string("x_615_axes_0"), val = tensor([-1])]; + tensor encoder_layers_23_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_23_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(535720256)))]; + tensor encoder_layers_23_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_23_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(535722368)))]; + tensor x_615_cast_fp16 = layer_norm(axes = x_615_axes_0, beta = encoder_layers_23_norm_conv_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_23_norm_conv_weight_to_fp16, x = input_1243_cast_fp16)[name = string("x_615_cast_fp16")]; + tensor input_1245_perm_0 = const()[name = string("input_1245_perm_0"), val = tensor([0, 2, 1])]; + string input_1247_pad_type_0 = const()[name = string("input_1247_pad_type_0"), val = string("valid")]; + tensor input_1247_strides_0 = const()[name = string("input_1247_strides_0"), val = tensor([1])]; + tensor input_1247_pad_0 = const()[name = string("input_1247_pad_0"), val = tensor([0, 0])]; + tensor input_1247_dilations_0 = const()[name = string("input_1247_dilations_0"), val = tensor([1])]; + int32 input_1247_groups_0 = const()[name = string("input_1247_groups_0"), val = int32(1)]; + tensor encoder_layers_23_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(535724480))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(537821696))))[name = string("encoder_layers_23_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_1245_cast_fp16 = transpose(perm = input_1245_perm_0, x = x_615_cast_fp16)[name = string("transpose_150")]; + tensor input_1247_cast_fp16 = conv(dilations = input_1247_dilations_0, groups = input_1247_groups_0, pad = input_1247_pad_0, pad_type = input_1247_pad_type_0, strides = input_1247_strides_0, weight = encoder_layers_23_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_1245_cast_fp16)[name = string("input_1247_cast_fp16")]; + int32 x_617_split_num_splits_0 = const()[name = string("x_617_split_num_splits_0"), val = int32(2)]; + int32 x_617_split_axis_0 = const()[name = string("x_617_split_axis_0"), val = int32(1)]; + tensor x_617_split_cast_fp16_0, tensor x_617_split_cast_fp16_1 = split(axis = x_617_split_axis_0, num_splits = x_617_split_num_splits_0, x = input_1247_cast_fp16)[name = string("x_617_split_cast_fp16")]; + tensor x_617_split_1_sigmoid_cast_fp16 = sigmoid(x = x_617_split_cast_fp16_1)[name = string("x_617_split_1_sigmoid_cast_fp16")]; + tensor x_617_cast_fp16 = mul(x = x_617_split_cast_fp16_0, y = x_617_split_1_sigmoid_cast_fp16)[name = string("x_617_cast_fp16")]; + tensor input_1249_cast_fp16 = select(a = var_45_to_fp16, b = x_617_cast_fp16, cond = var_576)[name = string("input_1249_cast_fp16")]; + bool new_x_interleave_0 = const()[name = string("new_x_interleave_0"), val = bool(false)]; + tensor new_x_cast_fp16 = concat(axis = var_60, interleave = new_x_interleave_0, values = (cache_cast_fp16, input_1249_cast_fp16))[name = string("new_x_cast_fp16")]; + tensor cache_last_time_cur_begin_0 = const()[name = string("cache_last_time_cur_begin_0"), val = tensor([0, 0, 14])]; + tensor cache_last_time_cur_end_0 = const()[name = string("cache_last_time_cur_end_0"), val = tensor([1, 1024, 22])]; + tensor cache_last_time_cur_end_mask_0 = const()[name = string("cache_last_time_cur_end_mask_0"), val = tensor([true, true, true])]; + tensor cache_last_time_cur_cast_fp16 = slice_by_index(begin = cache_last_time_cur_begin_0, end = cache_last_time_cur_end_0, end_mask = cache_last_time_cur_end_mask_0, x = new_x_cast_fp16)[name = string("cache_last_time_cur_cast_fp16")]; + string x_619_pad_type_0 = const()[name = string("x_619_pad_type_0"), val = string("valid")]; + int32 x_619_groups_0 = const()[name = string("x_619_groups_0"), val = int32(1024)]; + tensor x_619_strides_0 = const()[name = string("x_619_strides_0"), val = tensor([1])]; + tensor x_619_pad_0 = const()[name = string("x_619_pad_0"), val = tensor([0, 0])]; + tensor x_619_dilations_0 = const()[name = string("x_619_dilations_0"), val = tensor([1])]; + tensor encoder_layers_23_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(537825856))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(537835136))))[name = string("encoder_layers_23_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_619_cast_fp16 = conv(dilations = x_619_dilations_0, groups = x_619_groups_0, pad = x_619_pad_0, pad_type = x_619_pad_type_0, strides = x_619_strides_0, weight = encoder_layers_23_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_cast_fp16)[name = string("x_619_cast_fp16")]; + tensor input_1251_perm_0 = const()[name = string("input_1251_perm_0"), val = tensor([0, 2, 1])]; + tensor x_621_axes_0 = const()[name = string("x_621_axes_0"), val = tensor([-1])]; + tensor encoder_layers_23_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_23_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(537837248)))]; + tensor encoder_layers_23_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_23_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(537839360)))]; + tensor input_1251_cast_fp16 = transpose(perm = input_1251_perm_0, x = x_619_cast_fp16)[name = string("transpose_149")]; + tensor x_621_cast_fp16 = layer_norm(axes = x_621_axes_0, beta = encoder_layers_23_conv_batch_norm_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_23_conv_batch_norm_weight_to_fp16, x = input_1251_cast_fp16)[name = string("x_621_cast_fp16")]; + tensor input_1253_perm_0 = const()[name = string("input_1253_perm_0"), val = tensor([0, 2, 1])]; + tensor input_1253_cast_fp16 = transpose(perm = input_1253_perm_0, x = x_621_cast_fp16)[name = string("transpose_148")]; + tensor input_1255_cast_fp16 = silu(x = input_1253_cast_fp16)[name = string("input_1255_cast_fp16")]; + string x_623_pad_type_0 = const()[name = string("x_623_pad_type_0"), val = string("valid")]; + tensor x_623_strides_0 = const()[name = string("x_623_strides_0"), val = tensor([1])]; + tensor x_623_pad_0 = const()[name = string("x_623_pad_0"), val = tensor([0, 0])]; + tensor x_623_dilations_0 = const()[name = string("x_623_dilations_0"), val = tensor([1])]; + int32 x_623_groups_0 = const()[name = string("x_623_groups_0"), val = int32(1)]; + tensor encoder_layers_23_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(537841472))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(538890112))))[name = string("encoder_layers_23_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_623_cast_fp16 = conv(dilations = x_623_dilations_0, groups = x_623_groups_0, pad = x_623_pad_0, pad_type = x_623_pad_type_0, strides = x_623_strides_0, weight = encoder_layers_23_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_1255_cast_fp16)[name = string("x_623_cast_fp16")]; + tensor input_1257_perm_0 = const()[name = string("input_1257_perm_0"), val = tensor([0, 2, 1])]; + tensor input_1257_cast_fp16 = transpose(perm = input_1257_perm_0, x = x_623_cast_fp16)[name = string("transpose_147")]; + tensor input_1259_cast_fp16 = add(x = input_1243_cast_fp16, y = input_1257_cast_fp16)[name = string("input_1259_cast_fp16")]; + tensor input_1261_axes_0 = const()[name = string("input_1261_axes_0"), val = tensor([-1])]; + tensor encoder_layers_23_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_23_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(538892224)))]; + tensor encoder_layers_23_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_23_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(538894336)))]; + tensor input_1261_cast_fp16 = layer_norm(axes = input_1261_axes_0, beta = encoder_layers_23_norm_feed_forward2_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_23_norm_feed_forward2_weight_to_fp16, x = input_1259_cast_fp16)[name = string("input_1261_cast_fp16")]; + tensor encoder_layers_23_feed_forward2_linear1_weight_to_fp16 = const()[name = string("encoder_layers_23_feed_forward2_linear1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(538896448)))]; + tensor encoder_layers_23_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_23_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(547285120)))]; + tensor linear_215_cast_fp16 = linear(bias = encoder_layers_23_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_23_feed_forward2_linear1_weight_to_fp16, x = input_1261_cast_fp16)[name = string("linear_215_cast_fp16")]; + tensor input_1265_cast_fp16 = silu(x = linear_215_cast_fp16)[name = string("input_1265_cast_fp16")]; + tensor encoder_layers_23_feed_forward2_linear2_weight_to_fp16 = const()[name = string("encoder_layers_23_feed_forward2_linear2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(547293376)))]; + tensor encoder_layers_23_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_23_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(555682048)))]; + tensor linear_216_cast_fp16 = linear(bias = encoder_layers_23_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_23_feed_forward2_linear2_weight_to_fp16, x = input_1265_cast_fp16)[name = string("linear_216_cast_fp16")]; + fp16 var_5531_to_fp16 = const()[name = string("op_5531_to_fp16"), val = fp16(0x1p-1)]; + tensor var_5532_cast_fp16 = mul(x = linear_216_cast_fp16, y = var_5531_to_fp16)[name = string("op_5532_cast_fp16")]; + tensor input_1271_cast_fp16 = add(x = input_1259_cast_fp16, y = var_5532_cast_fp16)[name = string("input_1271_cast_fp16")]; + tensor audio_signal_axes_0 = const()[name = string("audio_signal_axes_0"), val = tensor([-1])]; + tensor encoder_layers_23_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_23_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(555684160)))]; + tensor encoder_layers_23_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_23_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(555686272)))]; + tensor audio_signal_cast_fp16 = layer_norm(axes = audio_signal_axes_0, beta = encoder_layers_23_norm_out_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_23_norm_out_weight_to_fp16, x = input_1271_cast_fp16)[name = string("audio_signal_cast_fp16")]; + int32 obj_5_axis_0 = const()[name = string("obj_5_axis_0"), val = int32(0)]; + tensor obj_5_cast_fp16 = stack(axis = obj_5_axis_0, values = (var_485_cast_fp16, var_698_cast_fp16, var_911_cast_fp16, var_1124_cast_fp16, var_1337_cast_fp16, var_1550_cast_fp16, var_1763_cast_fp16, var_1976_cast_fp16, var_2189_cast_fp16, var_2402_cast_fp16, var_2615_cast_fp16, var_2828_cast_fp16, var_3041_cast_fp16, var_3254_cast_fp16, var_3467_cast_fp16, var_3680_cast_fp16, var_3893_cast_fp16, var_4106_cast_fp16, var_4319_cast_fp16, var_4532_cast_fp16, var_4745_cast_fp16, var_4958_cast_fp16, var_5171_cast_fp16, cache_last_channel_cur_cast_fp16))[name = string("obj_5_cast_fp16")]; + int32 obj_7_axis_0 = const()[name = string("obj_7_axis_0"), val = int32(0)]; + tensor obj_7_cast_fp16 = stack(axis = obj_7_axis_0, values = (var_589_cast_fp16, var_802_cast_fp16, var_1015_cast_fp16, var_1228_cast_fp16, var_1441_cast_fp16, var_1654_cast_fp16, var_1867_cast_fp16, var_2080_cast_fp16, var_2293_cast_fp16, var_2506_cast_fp16, var_2719_cast_fp16, var_2932_cast_fp16, var_3145_cast_fp16, var_3358_cast_fp16, var_3571_cast_fp16, var_3784_cast_fp16, var_3997_cast_fp16, var_4210_cast_fp16, var_4423_cast_fp16, var_4636_cast_fp16, var_4849_cast_fp16, var_5062_cast_fp16, var_5275_cast_fp16, cache_last_time_cur_cast_fp16))[name = string("obj_7_cast_fp16")]; + tensor var_5548 = add(x = cache_len, y = max_audio_length_1)[name = string("op_5548")]; + string var_5548_promoted_to_fp16_dtype_0 = const()[name = string("op_5548_promoted_to_fp16_dtype_0"), val = string("fp16")]; + fp16 const_384_to_fp16 = const()[name = string("const_384_to_fp16"), val = fp16(-inf)]; + fp16 var_50_promoted_to_fp16 = const()[name = string("op_50_promoted_to_fp16"), val = fp16(0x1.5p+5)]; + tensor var_5548_to_fp16 = cast(dtype = var_5548_promoted_to_fp16_dtype_0, x = var_5548)[name = string("cast_10")]; + tensor clip_1_cast_fp16 = clip(alpha = const_384_to_fp16, beta = var_50_promoted_to_fp16, x = var_5548_to_fp16)[name = string("clip_1_cast_fp16")]; + int32 one_hot_1_batch_dims_0 = const()[name = string("one_hot_1_batch_dims_0"), val = int32(0)]; + bool one_hot_1_validate_indices_0 = const()[name = string("one_hot_1_validate_indices_0"), val = bool(false)]; + tensor to_onehot_identity_table_to_fp16 = const()[name = string("to_onehot_identity_table_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(555688384)))]; + string prompt_id_to_int16_dtype_0 = const()[name = string("prompt_id_to_int16_dtype_0"), val = string("int16")]; + string cast_230_dtype_0 = const()[name = string("cast_230_dtype_0"), val = string("int32")]; + int32 greater_equal_0_y_0 = const()[name = string("greater_equal_0_y_0"), val = int32(0)]; + tensor prompt_id_to_int16 = cast(dtype = prompt_id_to_int16_dtype_0, x = prompt_id)[name = string("cast_9")]; + tensor cast_230 = cast(dtype = cast_230_dtype_0, x = prompt_id_to_int16)[name = string("cast_8")]; + tensor greater_equal_0 = greater_equal(x = cast_230, y = greater_equal_0_y_0)[name = string("greater_equal_0")]; + int32 slice_by_index_2 = const()[name = string("slice_by_index_2"), val = int32(128)]; + tensor add_0 = add(x = cast_230, y = slice_by_index_2)[name = string("add_0")]; + tensor select_0 = select(a = cast_230, b = add_0, cond = greater_equal_0)[name = string("select_0")]; + string select_0_to_int16_dtype_0 = const()[name = string("select_0_to_int16_dtype_0"), val = string("int16")]; + string cast_0_dtype_0 = const()[name = string("cast_0_dtype_0"), val = string("int32")]; + int32 greater_equal_0_y_0_1 = const()[name = string("greater_equal_0_y_0_1"), val = int32(0)]; + tensor select_0_to_int16 = cast(dtype = select_0_to_int16_dtype_0, x = select_0)[name = string("cast_7")]; + tensor cast_0 = cast(dtype = cast_0_dtype_0, x = select_0_to_int16)[name = string("cast_6")]; + tensor greater_equal_0_1 = greater_equal(x = cast_0, y = greater_equal_0_y_0_1)[name = string("greater_equal_0_1")]; + int32 slice_by_index_0 = const()[name = string("slice_by_index_0"), val = int32(128)]; + tensor add_0_1 = add(x = cast_0, y = slice_by_index_0)[name = string("add_0_1")]; + tensor select_0_1 = select(a = cast_0, b = add_0_1, cond = greater_equal_0_1)[name = string("select_0_1")]; + int32 greater_equal_0_y_0_2 = const()[name = string("greater_equal_0_y_0_2"), val = int32(0)]; + tensor greater_equal_0_2 = greater_equal(x = select_0_1, y = greater_equal_0_y_0_2)[name = string("greater_equal_0_2")]; + int32 slice_by_index_0_1 = const()[name = string("slice_by_index_0_1"), val = int32(128)]; + tensor add_0_2 = add(x = select_0_1, y = slice_by_index_0_1)[name = string("add_0_2")]; + tensor select_0_2 = select(a = select_0_1, b = add_0_2, cond = greater_equal_0_2)[name = string("select_0_2")]; + int32 one_hot_1_cast_fp16_cast_uint16_cast_uint16_axis_0 = const()[name = string("one_hot_1_cast_fp16_cast_uint16_cast_uint16_axis_0"), val = int32(0)]; + tensor one_hot_1_cast_fp16_cast_uint16_cast_uint16 = gather(axis = one_hot_1_cast_fp16_cast_uint16_cast_uint16_axis_0, batch_dims = one_hot_1_batch_dims_0, indices = select_0_2, validate_indices = one_hot_1_validate_indices_0, x = to_onehot_identity_table_to_fp16)[name = string("one_hot_1_cast_fp16_cast_uint16_cast_uint16")]; + tensor var_5594_axes_0 = const()[name = string("op_5594_axes_0"), val = tensor([1])]; + tensor var_5594_cast_fp16 = expand_dims(axes = var_5594_axes_0, x = one_hot_1_cast_fp16_cast_uint16_cast_uint16)[name = string("op_5594_cast_fp16")]; + tensor one_hot_reps_0 = const()[name = string("one_hot_reps_0"), val = tensor([1, 14, 1])]; + tensor one_hot_cast_fp16 = tile(reps = one_hot_reps_0, x = var_5594_cast_fp16)[name = string("one_hot_cast_fp16")]; + int32 var_5603 = const()[name = string("op_5603"), val = int32(-1)]; + bool input_1273_interleave_0 = const()[name = string("input_1273_interleave_0"), val = bool(false)]; + tensor input_1273_cast_fp16 = concat(axis = var_5603, interleave = input_1273_interleave_0, values = (audio_signal_cast_fp16, one_hot_cast_fp16))[name = string("input_1273_cast_fp16")]; + tensor prompt_kernel_0_weight_to_fp16 = const()[name = string("prompt_kernel_0_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(555721216)))]; + tensor prompt_kernel_0_bias_to_fp16 = const()[name = string("prompt_kernel_0_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(560439872)))]; + tensor linear_217_cast_fp16 = linear(bias = prompt_kernel_0_bias_to_fp16, weight = prompt_kernel_0_weight_to_fp16, x = input_1273_cast_fp16)[name = string("linear_217_cast_fp16")]; + tensor input_1277_cast_fp16 = relu(x = linear_217_cast_fp16)[name = string("input_1277_cast_fp16")]; + tensor prompt_kernel_2_weight_to_fp16 = const()[name = string("prompt_kernel_2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(560444032)))]; + tensor prompt_kernel_2_bias_to_fp16 = const()[name = string("prompt_kernel_2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(564638400)))]; + tensor linear_218_cast_fp16 = linear(bias = prompt_kernel_2_bias_to_fp16, weight = prompt_kernel_2_weight_to_fp16, x = input_1277_cast_fp16)[name = string("linear_218_cast_fp16")]; + string conditioned_cast_fp16_to_fp32_dtype_0 = const()[name = string("conditioned_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + tensor transpose_72_perm_0_1 = const()[name = string("transpose_72_perm_0_1"), val = tensor([0, 2, 1])]; + string var_5621_dtype_0 = const()[name = string("op_5621_dtype_0"), val = string("int32")]; + tensor var_5624_perm_0 = const()[name = string("op_5624_perm_0"), val = tensor([1, 0, 2, 3])]; + string var_5624_cast_fp16_to_fp32_dtype_0 = const()[name = string("op_5624_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + tensor var_5627_perm_0 = const()[name = string("op_5627_perm_0"), val = tensor([1, 0, 2, 3])]; + string var_5627_cast_fp16_to_fp32_dtype_0 = const()[name = string("op_5627_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + string var_5632_dtype_0 = const()[name = string("op_5632_dtype_0"), val = string("int32")]; + tensor joint_enc_weight_to_fp16 = const()[name = string("joint_enc_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(564640512)))]; + tensor joint_enc_bias_to_fp16 = const()[name = string("joint_enc_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(565951296)))]; + tensor linear_219_cast_fp16 = linear(bias = joint_enc_bias_to_fp16, weight = joint_enc_weight_to_fp16, x = linear_218_cast_fp16)[name = string("linear_219_cast_fp16")]; + string linear_219_cast_fp16_to_fp32_dtype_0 = const()[name = string("linear_219_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + tensor encoder_proj = cast(dtype = linear_219_cast_fp16_to_fp32_dtype_0, x = linear_219_cast_fp16)[name = string("cast_0")]; + tensor cache_len_out = cast(dtype = var_5632_dtype_0, x = clip_1_cast_fp16)[name = string("cast_1")]; + tensor var_5627_cast_fp16 = transpose(perm = var_5627_perm_0, x = obj_7_cast_fp16)[name = string("transpose_144")]; + tensor cache_time_out = cast(dtype = var_5627_cast_fp16_to_fp32_dtype_0, x = var_5627_cast_fp16)[name = string("cast_2")]; + tensor var_5624_cast_fp16 = transpose(perm = var_5624_perm_0, x = obj_5_cast_fp16)[name = string("transpose_145")]; + tensor cache_channel_out = cast(dtype = var_5624_cast_fp16_to_fp32_dtype_0, x = var_5624_cast_fp16)[name = string("cast_3")]; + tensor encoded_length = cast(dtype = var_5621_dtype_0, x = clip_0_cast_fp16)[name = string("cast_4")]; + tensor transpose_72_1 = transpose(perm = transpose_72_perm_0_1, x = linear_218_cast_fp16)[name = string("transpose_146")]; + tensor encoded = cast(dtype = conditioned_cast_fp16_to_fp32_dtype_0, x = transpose_72_1)[name = string("cast_5")]; + } -> (encoded, encoded_length, cache_channel_out, cache_time_out, cache_len_out, encoder_proj); +} \ No newline at end of file diff --git a/ja/1120ms/encoder.mlmodelc/weights/weight.bin b/ja/1120ms/encoder.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..db719d65717ba13a2abcc1fd6915682c3eeef841 --- /dev/null +++ b/ja/1120ms/encoder.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:011a5342cc1e2901633a8f0468be561f50dfc289ebe51851cddb4831f9a6a23f +size 565952640 diff --git a/ja/1120ms/encoder.mlpackage/Data/com.apple.CoreML/model.mlmodel b/ja/1120ms/encoder.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..d7e63d6d9ce257a9bcf94ee5f3694e8fe97c90ab --- /dev/null +++ b/ja/1120ms/encoder.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:14655ea0dbccc666356247f05afc87bbea77ff469344bfb084526e00c0ce6d15 +size 804512 diff --git a/ja/1120ms/encoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/ja/1120ms/encoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..db719d65717ba13a2abcc1fd6915682c3eeef841 --- /dev/null +++ b/ja/1120ms/encoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:011a5342cc1e2901633a8f0468be561f50dfc289ebe51851cddb4831f9a6a23f +size 565952640 diff --git a/ja/1120ms/encoder.mlpackage/Manifest.json b/ja/1120ms/encoder.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..2ca731dbb685fe1350d224dc0e899e22cb3c0055 --- /dev/null +++ b/ja/1120ms/encoder.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "5A1F3899-93C0-4456-9DE6-B2D78CDDE258": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + }, + "E2FF78FE-1F88-4C4D-820E-E59B536101F4": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + } + }, + "rootModelIdentifier": "E2FF78FE-1F88-4C4D-820E-E59B536101F4" +} diff --git a/ja/1120ms/joint.mlmodelc/analytics/coremldata.bin b/ja/1120ms/joint.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..a7ba5a90671da5c40e03362f44f23df528bc6d93 --- /dev/null +++ b/ja/1120ms/joint.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e342ce20383866520d2c6c860c2bf14d887b9e7fef53606661b41a23ad09472e +size 243 diff --git a/ja/1120ms/joint.mlmodelc/coremldata.bin b/ja/1120ms/joint.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..219c9bac9ed82b5626e34705f215643647b64d90 --- /dev/null +++ b/ja/1120ms/joint.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1df3d770386f24c28a1a187cd341e4f2527d0b1c7f2959e5f606383e2ba9ddc6 +size 401 diff --git a/ja/1120ms/joint.mlmodelc/model.mil b/ja/1120ms/joint.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..827f6cd71b5910ea07d4f6ba43462967d8b86410 --- /dev/null +++ b/ja/1120ms/joint.mlmodelc/model.mil @@ -0,0 +1,31 @@ +program(1.3) +[buildInfo = dict({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.10.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor decoder, tensor encoder) { + tensor input_1_perm_0 = const()[name = string("input_1_perm_0"), val = tensor([0, 2, 1])]; + string encoder_to_fp16_dtype_0 = const()[name = string("encoder_to_fp16_dtype_0"), val = string("fp16")]; + tensor input_3_perm_0 = const()[name = string("input_3_perm_0"), val = tensor([0, 2, 1])]; + string decoder_to_fp16_dtype_0 = const()[name = string("decoder_to_fp16_dtype_0"), val = string("fp16")]; + tensor module_enc_weight_to_fp16 = const()[name = string("module_enc_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))]; + tensor module_enc_bias_to_fp16 = const()[name = string("module_enc_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1310848)))]; + tensor encoder_to_fp16 = cast(dtype = encoder_to_fp16_dtype_0, x = encoder)[name = string("cast_2")]; + tensor input_1_cast_fp16 = transpose(perm = input_1_perm_0, x = encoder_to_fp16)[name = string("transpose_1")]; + tensor linear_0_cast_fp16 = linear(bias = module_enc_bias_to_fp16, weight = module_enc_weight_to_fp16, x = input_1_cast_fp16)[name = string("linear_0_cast_fp16")]; + tensor module_pred_weight_to_fp16 = const()[name = string("module_pred_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1312192)))]; + tensor module_pred_bias_to_fp16 = const()[name = string("module_pred_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2131456)))]; + tensor decoder_to_fp16 = cast(dtype = decoder_to_fp16_dtype_0, x = decoder)[name = string("cast_1")]; + tensor input_3_cast_fp16 = transpose(perm = input_3_perm_0, x = decoder_to_fp16)[name = string("transpose_0")]; + tensor linear_1_cast_fp16 = linear(bias = module_pred_bias_to_fp16, weight = module_pred_weight_to_fp16, x = input_3_cast_fp16)[name = string("linear_1_cast_fp16")]; + tensor var_23_axes_0 = const()[name = string("op_23_axes_0"), val = tensor([2])]; + tensor var_23_cast_fp16 = expand_dims(axes = var_23_axes_0, x = linear_0_cast_fp16)[name = string("op_23_cast_fp16")]; + tensor var_25_axes_0 = const()[name = string("op_25_axes_0"), val = tensor([1])]; + tensor var_25_cast_fp16 = expand_dims(axes = var_25_axes_0, x = linear_1_cast_fp16)[name = string("op_25_cast_fp16")]; + tensor input_5_cast_fp16 = add(x = var_23_cast_fp16, y = var_25_cast_fp16)[name = string("input_5_cast_fp16")]; + tensor input_7_cast_fp16 = relu(x = input_5_cast_fp16)[name = string("input_7_cast_fp16")]; + tensor module_joint_net_2_weight_to_fp16 = const()[name = string("module_joint_net_2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2132800)))]; + tensor module_joint_net_2_bias_to_fp16 = const()[name = string("module_joint_net_2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3929984)))]; + tensor linear_2_cast_fp16 = linear(bias = module_joint_net_2_bias_to_fp16, weight = module_joint_net_2_weight_to_fp16, x = input_7_cast_fp16)[name = string("linear_2_cast_fp16")]; + string linear_2_cast_fp16_to_fp32_dtype_0 = const()[name = string("linear_2_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + tensor logits = cast(dtype = linear_2_cast_fp16_to_fp32_dtype_0, x = linear_2_cast_fp16)[name = string("cast_0")]; + } -> (logits); +} \ No newline at end of file diff --git a/ja/1120ms/joint.mlmodelc/weights/weight.bin b/ja/1120ms/joint.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..31da4412e1214e0b52e77c023e0490150a12e242 --- /dev/null +++ b/ja/1120ms/joint.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:44c15c8be3c89e5d531258b840e560e467084d55ed400453a9d21325757fb111 +size 3932856 diff --git a/ja/1120ms/joint.mlpackage/Data/com.apple.CoreML/model.mlmodel b/ja/1120ms/joint.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..1bc98711fb995d49c835ce43242e33bc518a943d --- /dev/null +++ b/ja/1120ms/joint.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1df49e8dd61d86e19d95b935020748db2eda7ae1273f39988001a30535c5be45 +size 4545 diff --git a/ja/1120ms/joint.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/ja/1120ms/joint.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..31da4412e1214e0b52e77c023e0490150a12e242 --- /dev/null +++ b/ja/1120ms/joint.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:44c15c8be3c89e5d531258b840e560e467084d55ed400453a9d21325757fb111 +size 3932856 diff --git a/ja/1120ms/joint.mlpackage/Manifest.json b/ja/1120ms/joint.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..ef5aca0b538b9de8e541f27a29fff913203df8c7 --- /dev/null +++ b/ja/1120ms/joint.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "8348189F-7CB9-4961-A35F-4049C53D63B6": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + }, + "AA6A8B4F-747E-4EC1-87E1-2B387F1149D8": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + } + }, + "rootModelIdentifier": "8348189F-7CB9-4961-A35F-4049C53D63B6" +} diff --git a/ja/1120ms/metadata.json b/ja/1120ms/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..58f34774474c70ef96d67c7e085aa6ca642518b7 --- /dev/null +++ b/ja/1120ms/metadata.json @@ -0,0 +1,199 @@ +{ + "model": "nvidia/nemotron-asr-streaming-multilingual-0.6b", + "model_class": "nemo.collections.asr.models.rnnt_bpe_models_prompt.EncDecRNNTBPEModelWithPrompt", + "sample_rate": 16000, + "mel_features": 128, + "chunk_mel_frames": 112, + "pre_encode_cache": 9, + "total_mel_frames": 121, + "att_context_size": [ + 42, + 13 + ], + "vocab_size": 1403, + "blank_idx": 1403, + "vocab_pruned": true, + "vocab_pruned_original_size": 13087, + "cache_channel_shape": [ + 1, + 24, + 42, + 1024 + ], + "cache_time_shape": [ + 1, + 24, + 1024, + 8 + ], + "decoder_hidden": 640, + "decoder_layers": 2, + "encoder_dim": 1024, + "num_prompts": 128, + "prompt_dictionary": { + "af-ZA": 54, + "am-ET": 49, + "ar": 7, + "ar-AR": 7, + "auto": 101, + "ay-BO": 81, + "az-AZ": 66, + "bg": 30, + "bg-BG": 30, + "bn-IN": 36, + "cs": 22, + "cs-CZ": 22, + "da": 25, + "da-DK": 25, + "de": 9, + "de-DE": 9, + "el": 21, + "el-GR": 21, + "en": 0, + "en-GB": 1, + "en-US": 0, + "enGB": 1, + "es": 3, + "es-ES": 2, + "es-US": 3, + "esES": 2, + "et": 60, + "et-EE": 60, + "fa-IR": 38, + "fi": 26, + "fi-FI": 26, + "fr": 8, + "fr-CA": 100, + "fr-FR": 8, + "gn-PY": 82, + "gu-IN": 42, + "ha-NG": 50, + "haw-US": 97, + "he-IL": 64, + "hi": 6, + "hi-HI": 6, + "hi-IN": 6, + "hr": 29, + "hr-HR": 29, + "hu": 23, + "hu-HU": 23, + "hy-AM": 68, + "id-ID": 34, + "ig-NG": 53, + "it": 15, + "it-IT": 15, + "ja-JA": 10, + "ja-JP": 10, + "ka-GE": 67, + "km-KH": 47, + "kn-IN": 43, + "ko": 14, + "ko-KO": 14, + "ko-KR": 14, + "ku-TR": 65, + "ky-KG": 71, + "ln-CD": 58, + "lt": 31, + "lt-LT": 31, + "lv": 61, + "lv-LV": 61, + "mi-NZ": 96, + "ml-IN": 44, + "mr-IN": 41, + "ms-MY": 35, + "mt-MT": 102, + "nah-MX": 83, + "nb": 103, + "nb-NO": 103, + "ne-NP": 46, + "nl": 16, + "nl-NL": 16, + "nn": 104, + "nn-NO": 104, + "no": 27, + "no-NO": 27, + "ny-MW": 57, + "or-KE": 59, + "pl": 17, + "pl-PL": 17, + "pt": 13, + "pt-BR": 12, + "pt-PT": 13, + "qu-PE": 80, + "ro": 20, + "ro-RO": 20, + "ru": 11, + "ru-RU": 11, + "rw-RW": 55, + "si-LK": 45, + "sk": 28, + "sk-SK": 28, + "sl": 62, + "sl-SI": 62, + "sm-WS": 98, + "so-SO": 56, + "sv": 24, + "sv-SE": 24, + "sw-KE": 48, + "ta-IN": 39, + "te-IN": 40, + "tg-TJ": 70, + "th-TH": 32, + "to-TO": 99, + "tr": 18, + "tr-TR": 18, + "uk": 19, + "uk-UA": 19, + "ur-PK": 37, + "uz-UZ": 69, + "vi-VN": 33, + "yo-NG": 52, + "zh-CN": 4, + "zh-TW": 5, + "zh-ZH": 4, + "zu-ZA": 51 + }, + "default_prompt_id": 101, + "lang_tag_token_ids": [ + 1, + 52, + 62, + 66, + 69, + 70, + 75, + 76, + 77, + 79, + 81, + 83, + 86, + 88, + 89, + 90, + 92, + 94, + 95, + 96, + 97, + 99, + 100, + 103, + 107, + 109, + 111, + 112, + 114, + 115, + 117, + 1389, + 1390, + 1391, + 1392, + 1393, + 1394, + 1395, + 1402 + ], + "chunk_ms": 1120 +} \ No newline at end of file diff --git a/ja/1120ms/preprocessor.mlmodelc/analytics/coremldata.bin b/ja/1120ms/preprocessor.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..e87594f68006bef8db4fcfe2a2379a3c1197ba56 --- /dev/null +++ b/ja/1120ms/preprocessor.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:422a77fc3b64b27260a8ae2031d287abaa1630d1ebc70343e9dcd280dd4c7e5c +size 243 diff --git a/ja/1120ms/preprocessor.mlmodelc/coremldata.bin b/ja/1120ms/preprocessor.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..9f88d2976124fbdeae6f6c4c492443eb5f32c97e --- /dev/null +++ b/ja/1120ms/preprocessor.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:41ec4b4c1059ff1f2ac7c71d90b1da0caf9244499d06f6c9de175a1ef992bec1 +size 371 diff --git a/ja/1120ms/preprocessor.mlmodelc/model.mil b/ja/1120ms/preprocessor.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..0b8261362f9cbf465b530a0d2d0ee9a2b2f462cd --- /dev/null +++ b/ja/1120ms/preprocessor.mlmodelc/model.mil @@ -0,0 +1,122 @@ +program(1.3) +[buildInfo = dict({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.5.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.3.0"}})] +{ + func main(tensor audio, tensor audio_length) [FlexibleShapeInformation = tuple>>, tuple, ?>>>>((("DefaultShapes", {{"audio", [1, 1]}}), ("RangeDims", {{"audio", [[1, 1], [1, 480000]]}})))] { + int32 var_9 = const()[name = string("op_9"), val = int32(1)]; + int32 var_10 = const()[name = string("op_10"), val = int32(160)]; + int32 var_12 = const()[name = string("op_12"), val = int32(0)]; + int32 var_33 = const()[name = string("op_33"), val = int32(512)]; + tensor var_34 = add(x = audio_length, y = var_33)[name = string("op_34")]; + int32 var_35 = const()[name = string("op_35"), val = int32(512)]; + tensor var_36 = sub(x = var_34, y = var_35)[name = string("op_36")]; + tensor floor_div_0 = floor_div(x = var_36, y = var_10)[name = string("floor_div_0")]; + tensor var_39 = equal(x = audio_length, y = var_12)[name = string("op_39")]; + tensor var_40 = const()[name = string("op_40"), val = tensor([0])]; + tensor mel_length = select(a = var_40, b = floor_div_0, cond = var_39)[name = string("seq_len")]; + string audio_to_fp16_dtype_0 = const()[name = string("audio_to_fp16_dtype_0"), val = string("fp16")]; + tensor audio_to_fp16 = cast(dtype = audio_to_fp16_dtype_0, x = audio)[name = string("cast_14")]; + tensor var_42_shape_cast_fp16 = shape(x = audio_to_fp16)[name = string("op_42_shape_cast_fp16")]; + int32 gather_0_axis_0 = const()[name = string("gather_0_axis_0"), val = int32(0)]; + int32 gather_0_batch_dims_0 = const()[name = string("gather_0_batch_dims_0"), val = int32(0)]; + bool gather_0_validate_indices_0 = const()[name = string("gather_0_validate_indices_0"), val = bool(false)]; + string var_42_shape_cast_fp16_to_int16_dtype_0 = const()[name = string("op_42_shape_cast_fp16_to_int16_dtype_0"), val = string("int16")]; + uint16 select_0_to_uint16 = const()[name = string("select_0_to_uint16"), val = uint16(1)]; + tensor var_42_shape_cast_fp16_to_int16 = cast(dtype = var_42_shape_cast_fp16_to_int16_dtype_0, x = var_42_shape_cast_fp16)[name = string("cast_13")]; + int16 gather_0_cast_uint16 = gather(axis = gather_0_axis_0, batch_dims = gather_0_batch_dims_0, indices = select_0_to_uint16, validate_indices = gather_0_validate_indices_0, x = var_42_shape_cast_fp16_to_int16)[name = string("gather_0_cast_uint16")]; + string gather_0_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_0_cast_uint16_to_int32_dtype_0"), val = string("int32")]; + int32 const_0 = const()[name = string("const_0"), val = int32(0)]; + int32 const_1 = const()[name = string("const_1"), val = int32(1)]; + int32 gather_0_cast_uint16_to_int32 = cast(dtype = gather_0_cast_uint16_to_int32_dtype_0, x = gather_0_cast_uint16)[name = string("cast_12")]; + tensor var_43 = range_1d(end = gather_0_cast_uint16_to_int32, start = const_0, step = const_1)[name = string("op_43")]; + tensor var_44_axes_0 = const()[name = string("op_44_axes_0"), val = tensor([0])]; + tensor var_44 = expand_dims(axes = var_44_axes_0, x = var_43)[name = string("op_44")]; + tensor var_45_axes_0 = const()[name = string("op_45_axes_0"), val = tensor([1])]; + tensor var_45 = expand_dims(axes = var_45_axes_0, x = audio_length)[name = string("op_45")]; + tensor timemask = less(x = var_44, y = var_45)[name = string("timemask")]; + tensor var_48_begin_0 = const()[name = string("op_48_begin_0"), val = tensor([0, 0])]; + tensor var_48_end_0 = const()[name = string("op_48_end_0"), val = tensor([1, 1])]; + tensor var_48_end_mask_0 = const()[name = string("op_48_end_mask_0"), val = tensor([true, false])]; + tensor var_48_squeeze_mask_0 = const()[name = string("op_48_squeeze_mask_0"), val = tensor([false, true])]; + tensor var_48_cast_fp16 = slice_by_index(begin = var_48_begin_0, end = var_48_end_0, end_mask = var_48_end_mask_0, squeeze_mask = var_48_squeeze_mask_0, x = audio_to_fp16)[name = string("op_48_cast_fp16")]; + tensor var_49_axes_0 = const()[name = string("op_49_axes_0"), val = tensor([1])]; + tensor var_49_cast_fp16 = expand_dims(axes = var_49_axes_0, x = var_48_cast_fp16)[name = string("op_49_cast_fp16")]; + tensor var_51_begin_0 = const()[name = string("op_51_begin_0"), val = tensor([0, 1])]; + tensor var_51_end_0 = const()[name = string("op_51_end_0"), val = tensor([1, 0])]; + tensor var_51_end_mask_0 = const()[name = string("op_51_end_mask_0"), val = tensor([true, true])]; + tensor var_51_cast_fp16 = slice_by_index(begin = var_51_begin_0, end = var_51_end_0, end_mask = var_51_end_mask_0, x = audio_to_fp16)[name = string("op_51_cast_fp16")]; + tensor var_53_begin_0 = const()[name = string("op_53_begin_0"), val = tensor([0, 0])]; + tensor var_53_end_0 = const()[name = string("op_53_end_0"), val = tensor([1, -1])]; + tensor var_53_end_mask_0 = const()[name = string("op_53_end_mask_0"), val = tensor([true, false])]; + tensor var_53_cast_fp16 = slice_by_index(begin = var_53_begin_0, end = var_53_end_0, end_mask = var_53_end_mask_0, x = audio_to_fp16)[name = string("op_53_cast_fp16")]; + fp16 var_54_to_fp16 = const()[name = string("op_54_to_fp16"), val = fp16(0x1.f0cp-1)]; + tensor var_55_cast_fp16 = mul(x = var_53_cast_fp16, y = var_54_to_fp16)[name = string("op_55_cast_fp16")]; + tensor var_56_cast_fp16 = sub(x = var_51_cast_fp16, y = var_55_cast_fp16)[name = string("op_56_cast_fp16")]; + bool x_3_interleave_0 = const()[name = string("x_3_interleave_0"), val = bool(false)]; + tensor x_3_cast_fp16 = concat(axis = var_9, interleave = x_3_interleave_0, values = (var_49_cast_fp16, var_56_cast_fp16))[name = string("x_3_cast_fp16")]; + tensor var_59 = logical_not(x = timemask)[name = string("op_59")]; + fp16 var_16_to_fp16 = const()[name = string("op_16_to_fp16"), val = fp16(0x0p+0)]; + tensor input_1_cast_fp16 = select(a = var_16_to_fp16, b = x_3_cast_fp16, cond = var_59)[name = string("input_1_cast_fp16")]; + tensor concat_1x = const()[name = string("concat_1x"), val = tensor([1, 1, -1])]; + tensor input_3_cast_fp16 = reshape(shape = concat_1x, x = input_1_cast_fp16)[name = string("input_3_cast_fp16")]; + tensor input_5_pad_0 = const()[name = string("input_5_pad_0"), val = tensor([0, 0, 0, 0, 256, 256])]; + string input_5_mode_0 = const()[name = string("input_5_mode_0"), val = string("constant")]; + fp16 const_3_to_fp16 = const()[name = string("const_3_to_fp16"), val = fp16(0x0p+0)]; + tensor input_5_cast_fp16 = pad(constant_val = const_3_to_fp16, mode = input_5_mode_0, pad = input_5_pad_0, x = input_3_cast_fp16)[name = string("input_5_cast_fp16")]; + tensor concat_2x = const()[name = string("concat_2x"), val = tensor([1, -1])]; + tensor input_cast_fp16 = reshape(shape = concat_2x, x = input_5_cast_fp16)[name = string("input_cast_fp16")]; + tensor expand_dims_3 = const()[name = string("expand_dims_3"), val = tensor([160])]; + tensor expand_dims_4_axes_0 = const()[name = string("expand_dims_4_axes_0"), val = tensor([1])]; + tensor expand_dims_4_cast_fp16 = expand_dims(axes = expand_dims_4_axes_0, x = input_cast_fp16)[name = string("expand_dims_4_cast_fp16")]; + string conv_0_pad_type_0 = const()[name = string("conv_0_pad_type_0"), val = string("valid")]; + tensor conv_0_pad_0 = const()[name = string("conv_0_pad_0"), val = tensor([0, 0])]; + tensor conv_0_dilations_0 = const()[name = string("conv_0_dilations_0"), val = tensor([1])]; + int32 conv_0_groups_0 = const()[name = string("conv_0_groups_0"), val = int32(1)]; + tensor expand_dims_1_to_fp16 = const()[name = string("expand_dims_1_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))]; + tensor conv_0_cast_fp16 = conv(dilations = conv_0_dilations_0, groups = conv_0_groups_0, pad = conv_0_pad_0, pad_type = conv_0_pad_type_0, strides = expand_dims_3, weight = expand_dims_1_to_fp16, x = expand_dims_4_cast_fp16)[name = string("conv_0_cast_fp16")]; + string conv_1_pad_type_0 = const()[name = string("conv_1_pad_type_0"), val = string("valid")]; + tensor conv_1_pad_0 = const()[name = string("conv_1_pad_0"), val = tensor([0, 0])]; + tensor conv_1_dilations_0 = const()[name = string("conv_1_dilations_0"), val = tensor([1])]; + int32 conv_1_groups_0 = const()[name = string("conv_1_groups_0"), val = int32(1)]; + tensor expand_dims_2_to_fp16 = const()[name = string("expand_dims_2_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(263296)))]; + tensor conv_1_cast_fp16 = conv(dilations = conv_1_dilations_0, groups = conv_1_groups_0, pad = conv_1_pad_0, pad_type = conv_1_pad_type_0, strides = expand_dims_3, weight = expand_dims_2_to_fp16, x = expand_dims_4_cast_fp16)[name = string("conv_1_cast_fp16")]; + int32 stack_0_axis_0 = const()[name = string("stack_0_axis_0"), val = int32(-1)]; + tensor stack_0_cast_fp16 = stack(axis = stack_0_axis_0, values = (conv_0_cast_fp16, conv_1_cast_fp16))[name = string("stack_0_cast_fp16")]; + fp16 var_19_promoted_to_fp16 = const()[name = string("op_19_promoted_to_fp16"), val = fp16(0x1p+1)]; + tensor var_74_cast_fp16 = pow(x = stack_0_cast_fp16, y = var_19_promoted_to_fp16)[name = string("op_74_cast_fp16")]; + tensor var_76_axes_0 = const()[name = string("op_76_axes_0"), val = tensor([-1])]; + bool var_76_keep_dims_0 = const()[name = string("op_76_keep_dims_0"), val = bool(false)]; + tensor var_76_cast_fp16 = reduce_sum(axes = var_76_axes_0, keep_dims = var_76_keep_dims_0, x = var_74_cast_fp16)[name = string("op_76_cast_fp16")]; + tensor x_11_cast_fp16 = identity(x = var_76_cast_fp16)[name = string("x_11_cast_fp16")]; + bool x_13_transpose_x_0 = const()[name = string("x_13_transpose_x_0"), val = bool(false)]; + bool x_13_transpose_y_0 = const()[name = string("x_13_transpose_y_0"), val = bool(false)]; + tensor const_4_to_fp16 = const()[name = string("const_4_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(526528)))]; + tensor x_13_cast_fp16 = matmul(transpose_x = x_13_transpose_x_0, transpose_y = x_13_transpose_y_0, x = const_4_to_fp16, y = x_11_cast_fp16)[name = string("x_13_cast_fp16")]; + fp16 var_83_to_fp16 = const()[name = string("op_83_to_fp16"), val = fp16(0x1p-24)]; + tensor var_84_cast_fp16 = add(x = x_13_cast_fp16, y = var_83_to_fp16)[name = string("op_84_cast_fp16")]; + fp32 x_epsilon_0 = const()[name = string("x_epsilon_0"), val = fp32(0x1p-149)]; + tensor x_cast_fp16 = log(epsilon = x_epsilon_0, x = var_84_cast_fp16)[name = string("x_cast_fp16")]; + tensor var_86_shape_cast_fp16 = shape(x = x_cast_fp16)[name = string("op_86_shape_cast_fp16")]; + int32 gather_5_axis_0 = const()[name = string("gather_5_axis_0"), val = int32(0)]; + int32 gather_5_batch_dims_0 = const()[name = string("gather_5_batch_dims_0"), val = int32(0)]; + bool gather_5_validate_indices_0 = const()[name = string("gather_5_validate_indices_0"), val = bool(false)]; + string var_86_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_86_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")]; + uint16 select_5_to_uint16 = const()[name = string("select_5_to_uint16"), val = uint16(2)]; + tensor var_86_shape_cast_fp16_to_uint16 = cast(dtype = var_86_shape_cast_fp16_to_uint16_dtype_0, x = var_86_shape_cast_fp16)[name = string("cast_11")]; + uint16 gather_5_cast_uint16 = gather(axis = gather_5_axis_0, batch_dims = gather_5_batch_dims_0, indices = select_5_to_uint16, validate_indices = gather_5_validate_indices_0, x = var_86_shape_cast_fp16_to_uint16)[name = string("gather_5_cast_uint16")]; + string gather_5_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_5_cast_uint16_to_int32_dtype_0"), val = string("int32")]; + int32 const_5 = const()[name = string("const_5"), val = int32(0)]; + int32 const_6 = const()[name = string("const_6"), val = int32(1)]; + int32 gather_5_cast_uint16_to_int32 = cast(dtype = gather_5_cast_uint16_to_int32_dtype_0, x = gather_5_cast_uint16)[name = string("cast_10")]; + tensor mask_1 = range_1d(end = gather_5_cast_uint16_to_int32, start = const_5, step = const_6)[name = string("mask_1")]; + tensor expand_dims_0_axes_0 = const()[name = string("expand_dims_0_axes_0"), val = tensor([0])]; + tensor expand_dims_0 = expand_dims(axes = expand_dims_0_axes_0, x = mask_1)[name = string("expand_dims_0")]; + tensor var_91_axes_0 = const()[name = string("op_91_axes_0"), val = tensor([1])]; + tensor var_91 = expand_dims(axes = var_91_axes_0, x = mel_length)[name = string("op_91")]; + tensor mask = greater_equal(x = expand_dims_0, y = var_91)[name = string("mask")]; + tensor var_93_axes_0 = const()[name = string("op_93_axes_0"), val = tensor([1])]; + tensor var_93 = expand_dims(axes = var_93_axes_0, x = mask)[name = string("op_93")]; + tensor processed_signal_cast_fp16 = select(a = var_16_to_fp16, b = x_cast_fp16, cond = var_93)[name = string("processed_signal_cast_fp16")]; + string processed_signal_cast_fp16_to_fp32_dtype_0 = const()[name = string("processed_signal_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + tensor mel = cast(dtype = processed_signal_cast_fp16_to_fp32_dtype_0, x = processed_signal_cast_fp16)[name = string("cast_9")]; + } -> (mel, mel_length); +} \ No newline at end of file diff --git a/ja/1120ms/preprocessor.mlmodelc/weights/weight.bin b/ja/1120ms/preprocessor.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..86dd375f6649d262d58c9cf8b89006ceab216411 --- /dev/null +++ b/ja/1120ms/preprocessor.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:297514e2b211d14b0e53cb97193d679bb89ead98d28e578f3f1d049ddbcc36b3 +size 592384 diff --git a/ja/1120ms/preprocessor.mlpackage/Data/com.apple.CoreML/model.mlmodel b/ja/1120ms/preprocessor.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..961c08f6554bdf851377e1e59750c9ee4a63de19 --- /dev/null +++ b/ja/1120ms/preprocessor.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:912e39c12c0ddeb4d3716693621d0140bb7de556056ab8dc9e5cd9336ebf1baa +size 15878 diff --git a/ja/1120ms/preprocessor.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/ja/1120ms/preprocessor.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..86dd375f6649d262d58c9cf8b89006ceab216411 --- /dev/null +++ b/ja/1120ms/preprocessor.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:297514e2b211d14b0e53cb97193d679bb89ead98d28e578f3f1d049ddbcc36b3 +size 592384 diff --git a/ja/1120ms/preprocessor.mlpackage/Manifest.json b/ja/1120ms/preprocessor.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..3bfc4c45d1b99a3d127a61aff3d1d0ab4d95963c --- /dev/null +++ b/ja/1120ms/preprocessor.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "470B2EDA-A2F1-4777-8F1D-55E0D6A275A8": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + }, + "D83A1966-0388-45B9-9CA7-1DACB258C26C": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + } + }, + "rootModelIdentifier": "470B2EDA-A2F1-4777-8F1D-55E0D6A275A8" +} diff --git a/ja/1120ms/tokenizer.json b/ja/1120ms/tokenizer.json new file mode 100644 index 0000000000000000000000000000000000000000..b05c35042625729d82c27359c8b8cf9ae4e2e4c9 --- /dev/null +++ b/ja/1120ms/tokenizer.json @@ -0,0 +1,1406 @@ +{ + "0": "", + "1": "", + "2": "▁", + "3": ".", + "4": ",", + "5": "e", + "6": "t", + "7": "a", + "8": "s", + "9": "o", + "10": "i", + "11": "r", + "12": "h", + "13": "m", + "14": "p", + "15": "n", + "16": "g", + "17": "en", + "18": "on", + "19": "y", + "20": "0", + "21": "b", + "22": "2", + "23": "1", + "24": "k", + "25": "3", + "26": "5", + "27": "I", + "28": "4", + "29": "z", + "30": "A", + "31": "6", + "32": "E", + "33": "7", + "34": "T", + "35": "8", + "36": "9", + "37": "S", + "38": "C", + "39": "M", + "40": "P", + "41": "D", + "42": "B", + "43": "U", + "44": "N", + "45": "O", + "46": "G", + "47": "F", + "48": "L", + "49": "R", + "50": "H", + "51": "K", + "52": "", + "53": "le", + "54": "li", + "55": "to", + "56": "na", + "57": "ci", + "58": "▁A", + "59": "vi", + "60": "Z", + "61": "X", + "62": "", + "63": "et", + "64": "ar", + "65": "el", + "66": "", + "67": "ie", + "68": "▁G", + "69": "", + "70": "", + "71": "ta", + "72": "ks", + "73": "gi", + "74": "di", + "75": "", + "76": "", + "77": "", + "78": "cs", + "79": "", + "80": "sti", + "81": "", + "82": "pe", + "83": "", + "84": "as", + "85": "▁Tai", + "86": "", + "87": "▁20", + "88": "", + "89": "", + "90": "", + "91": "-", + "92": "", + "93": "▁M", + "94": "", + "95": "", + "96": "", + "97": "", + "98": "pp", + "99": "", + "100": "", + "101": "▁N", + "102": "mm", + "103": "", + "104": "▁and", + "105": "th", + "106": "ex", + "107": "", + "108": "ng", + "109": "", + "110": "▁La", + "111": "", + "112": "", + "113": "avec", + "114": "", + "115": "", + "116": "!", + "117": "", + "118": "▁あ", + "119": "▁す", + "120": "▁し", + "121": "▁そ", + "122": "▁ま", + "123": "▁こ", + "124": "以", + "125": "仕", + "126": "▁文", + "127": "▁強", + "128": "▁月", + "129": "▁連", + "130": "▁ペ", + "131": "▁海", + "132": "▁科", + "133": "▁ゴ", + "134": "▁原", + "135": "▁ギ", + "136": "▁ヤ", + "137": "▁特", + "138": "▁参", + "139": "▁男", + "140": "普", + "141": "受", + "142": "▁都", + "143": "▁水", + "144": "▁終", + "145": "▁南", + "146": "▁例", + "147": "▁利", + "148": "▁現", + "149": "▁真", + "150": "▁古", + "151": "▁彼", + "152": "▁交", + "153": "専", + "154": "▁悪", + "155": "▁旅", + "156": "▁技", + "157": "▁再", + "158": "企", + "159": "登", + "160": "総", + "161": "改", + "162": "探", + "163": "異", + "164": "完", + "165": "頑", + "166": "深", + "167": "推", + "168": "昨", + "169": "危", + "170": "詳", + "171": "余", + "172": "描", + "173": "被", + "174": "固", + "175": "副", + "176": "混", + "177": "救", + "178": "互", + "179": "閉", + "180": "亜", + "181": "植", + "182": "責", + "183": "仮", + "184": "脱", + "185": "恐", + "186": "契", + "187": "隠", + "188": "麻", + "189": "哲", + "190": "宣", + "191": "浮", + "192": "諸", + "193": "巨", + "194": "午", + "195": "占", + "196": "距", + "197": "汚", + "198": "延", + "199": "訪", + "200": "沿", + "201": "唯", + "202": "批", + "203": "傾", + "204": "邦", + "205": "購", + "206": "維", + "207": "燃", + "208": "既", + "209": "緊", + "210": "獲", + "211": "炭", + "212": "搭", + "213": "努", + "214": "顧", + "215": "狭", + "216": "雰", + "217": "塗", + "218": "排", + "219": "奪", + "220": "埋", + "221": "履", + "222": "侵", + "223": "該", + "224": "頻", + "225": "勧", + "226": "魅", + "227": "抑", + "228": "驚", + "229": "衝", + "230": "耐", + "231": "披", + "232": "透", + "233": "促", + "234": "撤", + "235": "潜", + "236": "吐", + "237": "睡", + "238": "孤", + "239": "昆", + "240": "遭", + "241": "溶", + "242": "貿", + "243": "疾", + "244": "拒", + "245": "脅", + "246": "挟", + "247": "漏", + "248": "覆", + "249": "畜", + "250": "紛", + "251": "絞", + "252": "擬", + "253": "疎", + "254": "跳", + "255": "廷", + "256": "郊", + "257": "克", + "258": "措", + "259": "虐", + "260": "陥", + "261": "徐", + "262": "漠", + "263": "愚", + "264": "噴", + "265": "卿", + "266": "胡", + "267": "溢", + "268": "妨", + "269": "沸", + "270": "阻", + "271": "妃", + "272": "舶", + "273": "濡", + "274": "剥", + "275": "洪", + "276": "鋭", + "277": "娯", + "278": "腎", + "279": "訂", + "280": "肖", + "281": "枢", + "282": "髄", + "283": "傑", + "284": "砕", + "285": "幣", + "286": "勾", + "287": "臆", + "288": "歪", + "289": "喝", + "290": "彗", + "291": "礁", + "292": "蚊", + "293": "焉", + "294": "塵", + "295": "脊", + "296": "鱗", + "297": "瞑", + "298": "峻", + "299": "徊", + "300": "徘", + "301": "猥", + "302": "竣", + "303": "穣", + "304": "隻", + "305": "睨", + "306": "瀕", + "307": "褻", + "308": "踵", + "309": "奢", + "310": "董", + "311": "毅", + "312": "醒", + "313": "墜", + "314": "窟", + "315": "綿", + "316": "遮", + "317": "騰", + "318": "顎", + "319": "遍", + "320": "蓋", + "321": "澄", + "322": "棲", + "323": "虎", + "324": "粒", + "325": "胎", + "326": "爪", + "327": "殻", + "328": "鉛", + "329": "枕", + "330": "堤", + "331": "膝", + "332": "峠", + "333": "搬", + "334": "晋", + "335": "餌", + "336": "狼", + "337": "衰", + "338": "閲", + "339": "筒", + "340": "軸", + "341": "緯", + "342": "暦", + "343": "刃", + "344": "潔", + "345": "裂", + "346": "叫", + "347": "殖", + "348": "媒", + "349": "癌", + "350": "鎖", + "351": "磁", + "352": "洞", + "353": "遇", + "354": "枝", + "355": "紫", + "356": "縦", + "357": "胞", + "358": "釈", + "359": "威", + "360": "晶", + "361": "砲", + "362": "焦", + "363": "尿", + "364": "魂", + "365": "潮", + "366": "旬", + "367": "慎", + "368": "噂", + "369": "隔", + "370": "穴", + "371": "慮", + "372": "即", + "373": "滑", + "374": "雷", + "375": "摘", + "376": "鏡", + "377": "棒", + "378": "悟", + "379": "葬", + "380": "序", + "381": "貫", + "382": "氷", + "383": "針", + "384": "煮", + "385": "棄", + "386": "銃", + "387": "汁", + "388": "封", + "389": "湿", + "390": "靴", + "391": "豚", + "392": "締", + "393": "豪", + "394": "票", + "395": "皮", + "396": "縮", + "397": "徹", + "398": "較", + "399": "忍", + "400": "核", + "401": "儀", + "402": "到", + "403": "削", + "404": "駆", + "405": "繁", + "406": "陰", + "407": "浄", + "408": "脈", + "409": "滞", + "410": "至", + "411": "枚", + "412": "偉", + "413": "致", + "414": "貨", + "415": "漢", + "416": "己", + "417": "握", + "418": "欧", + "419": "薄", + "420": "献", + "421": "預", + "422": "龍", + "423": "快", + "424": "句", + "425": "縁", + "426": "微", + "427": "妙", + "428": "晩", + "429": "粉", + "430": "卓", + "431": "圏", + "432": "兼", + "433": "脳", + "434": "竜", + "435": "鳴", + "436": "騒", + "437": "請", + "438": "卵", + "439": "唱", + "440": "嵐", + "441": "臓", + "442": "箱", + "443": "祖", + "444": "浴", + "445": "壁", + "446": "析", + "447": "厚", + "448": "筆", + "449": "承", + "450": "均", + "451": "律", + "452": "否", + "453": "脚", + "454": "湖", + "455": "乳", + "456": "揮", + "457": "滅", + "458": "乾", + "459": "羽", + "460": "候", + "461": "拡", + "462": "貸", + "463": "砂", + "464": "敬", + "465": "庁", + "466": "煙", + "467": "底", + "468": "露", + "469": "骨", + "470": "倍", + "471": "殿", + "472": "易", + "473": "層", + "474": "幕", + "475": "毛", + "476": "爆", + "477": "暇", + "478": "械", + "479": "隣", + "480": "輸", + "481": "柄", + "482": "範", + "483": "掲", + "484": "嘘", + "485": "剤", + "486": "墓", + "487": "衣", + "488": "射", + "489": "菓", + "490": "募", + "491": "乱", + "492": "迎", + "493": "抱", + "494": "酸", + "495": "雄", + "496": "虫", + "497": "複", + "498": "為", + "499": "泳", + "500": "宝", + "501": "激", + "502": "暑", + "503": "疑", + "504": "誘", + "505": "暴", + "506": "聖", + "507": "捨", + "508": "破", + "509": "革", + "510": "希", + "511": "折", + "512": "惑", + "513": "測", + "514": "紀", + "515": "舎", + "516": "署", + "517": "患", + "518": "岸", + "519": "秀", + "520": "免", + "521": "禁", + "522": "躍", + "523": "聴", + "524": "抗", + "525": "税", + "526": "奏", + "527": "弾", + "528": "礼", + "529": "童", + "530": "裏", + "531": "吹", + "532": "眠", + "533": "歯", + "534": "拠", + "535": "慣", + "536": "触", + "537": "飼", + "538": "群", + "539": "宗", + "540": "傷", + "541": "額", + "542": "塩", + "543": "静", + "544": "留", + "545": "罪", + "546": "純", + "547": "壊", + "548": "闘", + "549": "弱", + "550": "刻", + "551": "航", + "552": "栄", + "553": "姿", + "554": "亡", + "555": "織", + "556": "敗", + "557": "章", + "558": "吸", + "559": "令", + "560": "捜", + "561": "模", + "562": "絵", + "563": "申", + "564": "盤", + "565": "積", + "566": "標", + "567": "階", + "568": "省", + "569": "項", + "570": "猫", + "571": "従", + "572": "非", + "573": "季", + "574": "捕", + "575": "党", + "576": "圧", + "577": "香", + "578": "操", + "579": "暗", + "580": "症", + "581": "散", + "582": "突", + "583": "適", + "584": "雑", + "585": "跡", + "586": "厳", + "587": "鳥", + "588": "逃", + "589": "講", + "590": "晴", + "591": "徴", + "592": "困", + "593": "短", + "594": "婦", + "595": "略", + "596": "齢", + "597": "震", + "598": "敵", + "599": "博", + "600": "血", + "601": "満", + "602": "舗", + "603": "宙", + "604": "寿", + "605": "遺", + "606": "極", + "607": "里", + "608": "因", + "609": "典", + "610": "染", + "611": "徒", + "612": "巻", + "613": "頂", + "614": "超", + "615": "河", + "616": "盛", + "617": "犬", + "618": "豊", + "619": "端", + "620": "紹", + "621": "首", + "622": "陽", + "623": "歳", + "624": "印", + "625": "紙", + "626": "払", + "627": "求", + "628": "障", + "629": "簡", + "630": "途", + "631": "創", + "632": "船", + "633": "菜", + "634": "ゥ", + "635": "勤", + "636": "痛", + "637": "並", + "638": "景", + "639": "雪", + "640": "節", + "641": "浜", + "642": "清", + "643": "抜", + "644": "勢", + "645": "暮", + "646": "銀", + "647": "盟", + "648": "魚", + "649": "率", + "650": "洋", + "651": "渡", + "652": "順", + "653": "況", + "654": "談", + "655": "舞", + "656": "案", + "657": "岩", + "658": "負", + "659": "旧", + "660": "財", + "661": "故", + "662": "冬", + "663": "横", + "664": "奥", + "665": "比", + "666": "囲", + "667": "停", + "668": "築", + "669": "波", + "670": "林", + "671": "暖", + "672": "索", + "673": "赤", + "674": "給", + "675": "末", + "676": "催", + "677": "遅", + "678": "述", + "679": "黒", + "680": "細", + "681": "与", + "682": "減", + "683": "級", + "684": "費", + "685": "越", + "686": "差", + "687": "領", + "688": "衛", + "689": "隊", + "690": "薬", + "691": "氏", + "692": "望", + "693": "似", + "694": "就", + "695": "条", + "696": "処", + "697": "谷", + "698": "策", + "699": "効", + "700": "熱", + "701": "復", + "702": "ヌ", + "703": "振", + "704": "規", + "705": "港", + "706": "注", + "707": "森", + "708": "防", + "709": "継", + "710": "退", + "711": "火", + "712": "陸", + "713": "去", + "714": "視", + "715": "整", + "716": "準", + "717": "庭", + "718": "ゾ", + "719": "独", + "720": "撃", + "721": "児", + "722": "橋", + "723": "換", + "724": "念", + "725": "識", + "726": "打", + "727": "津", + "728": "雨", + "729": "幸", + "730": "含", + "731": "響", + "732": "労", + "733": "官", + "734": "追", + "735": "遠", + "736": "未", + "737": "販", + "738": "街", + "739": "曜", + "740": "程", + "741": "提", + "742": "玉", + "743": "判", + "744": "移", + "745": "攻", + "746": "低", + "747": "装", + "748": "断", + "749": "及", + "750": "証", + "751": "象", + "752": "守", + "753": "戻", + "754": "詞", + "755": "投", + "756": "載", + "757": "具", + "758": "除", + "759": "環", + "760": "展", + "761": "争", + "762": "失", + "763": "春", + "764": "挙", + "765": "返", + "766": "馬", + "767": "欲", + "768": "材", + "769": "図", + "770": "養", + "771": "焼", + "772": "導", + "773": "夢", + "774": "米", + "775": "冷", + "776": "息", + "777": "兵", + "778": "済", + "779": "劇", + "780": "央", + "781": "険", + "782": "服", + "783": "態", + "784": "走", + "785": "評", + "786": "権", + "787": "論", + "788": "ゅ", + "789": "境", + "790": "察", + "791": "授", + "792": "頼", + "793": "派", + "794": "撮", + "795": "素", + "796": "修", + "797": "第", + "798": "質", + "799": "告", + "800": "興", + "801": "秒", + "802": "宇", + "803": "肉", + "804": "像", + "805": "称", + "806": "値", + "807": "頭", + "808": "週", + "809": "督", + "810": "消", + "811": "芸", + "812": "顔", + "813": "読", + "814": "仲", + "815": "遊", + "816": "試", + "817": "酒", + "818": "離", + "819": "増", + "820": "殺", + "821": "鉄", + "822": "害", + "823": "割", + "824": "石", + "825": "夏", + "826": "助", + "827": "英", + "828": "想", + "829": "管", + "830": "急", + "831": "頃", + "832": "づ", + "833": "造", + "834": "史", + "835": "量", + "836": "製", + "837": "府", + "838": "足", + "839": "王", + "840": "委", + "841": "両", + "842": "辺", + "843": "残", + "844": "逆", + "845": "備", + "846": "軍", + "847": "警", + "848": "査", + "849": "列", + "850": "編", + "851": "段", + "852": "反", + "853": "ゼ", + "854": "携", + "855": "歩", + "856": "座", + "857": "飛", + "858": "丈", + "859": "価", + "860": "監", + "861": "ヘ", + "862": "周", + "863": "毎", + "864": "統", + "865": "収", + "866": "落", + "867": "星", + "868": "降", + "869": "側", + "870": "療", + "871": "師", + "872": "写", + "873": "類", + "874": "命", + "875": "介", + "876": "護", + "877": "死", + "878": "果", + "879": "任", + "880": "更", + "881": "常", + "882": "検", + "883": "過", + "884": "資", + "885": "働", + "886": "認", + "887": "般", + "888": "示", + "889": "客", + "890": "習", + "891": "究", + "892": "半", + "893": "録", + "894": "字", + "895": "昔", + "896": "影", + "897": "覚", + "898": "型", + "899": "声", + "900": "件", + "901": "義", + "902": "施", + "903": "容", + "904": "路", + "905": "呼", + "906": "役", + "907": "単", + "908": "状", + "909": "建", + "910": "由", + "911": "属", + "912": "土", + "913": "葉", + "914": "起", + "915": "覧", + "916": "配", + "917": "張", + "918": "接", + "919": "込", + "920": "待", + "921": "室", + "922": "病", + "923": "帯", + "924": "婚", + "925": "光", + "926": "個", + "927": "職", + "928": "営", + "929": "ぼ", + "930": "研", + "931": "計", + "932": "直", + "933": "難", + "934": "絶", + "935": "ヨ", + "936": "照", + "937": "西", + "938": "約", + "939": "存", + "940": "験", + "941": "治", + "942": "解", + "943": "転", + "944": "商", + "945": "進", + "946": "係", + "947": "説", + "948": "観", + "949": "球", + "950": "乗", + "951": "支", + "952": "得", + "953": "議", + "954": "門", + "955": "止", + "956": "重", + "957": "温", + "958": "着", + "959": "飲", + "960": "母", + "961": "士", + "962": "ざ", + "963": "集", + "964": "万", + "965": "太", + "966": "続", + "967": "線", + "968": "種", + "969": "格", + "970": "位", + "971": "ユ", + "972": "歌", + "973": "夜", + "974": "共", + "975": "正", + "976": "必", + "977": "ヒ", + "978": "色", + "979": "問", + "980": "再", + "981": "域", + "982": "ゆ", + "983": "勝", + "984": "台", + "985": "技", + "986": "旅", + "987": "引", + "988": "系", + "989": "院", + "990": "悪", + "991": "基", + "992": "神", + "993": "産", + "994": "決", + "995": "民", + "996": "交", + "997": "政", + "998": "賞", + "999": "空", + "1000": "医", + "1001": "彼", + "1002": "夫", + "1003": "可", + "1004": "誰", + "1005": "古", + "1006": "帰", + "1007": "術", + "1008": "相", + "1009": "団", + "1010": "伝", + "1011": "住", + "1012": "題", + "1013": "平", + "1014": "予", + "1015": "音", + "1016": "朝", + "1017": "指", + "1018": "真", + "1019": "ヴ", + "1020": "務", + "1021": "点", + "1022": "各", + "1023": "館", + "1024": "応", + "1025": "現", + "1026": "利", + "1027": "天", + "1028": "等", + "1029": "木", + "1030": "白", + "1031": "形", + "1032": "供", + "1033": "経", + "1034": "族", + "1035": "早", + "1036": "例", + "1037": "不", + "1038": "切", + "1039": "南", + "1040": "加", + "1041": "際", + "1042": "終", + "1043": "様", + "1044": "放", + "1045": "和", + "1046": "州", + "1047": "水", + "1048": "協", + "1049": "在", + "1050": "組", + "1051": "向", + "1052": "広", + "1053": "身", + "1054": "界", + "1055": "工", + "1056": "選", + "1057": "始", + "1058": "元", + "1059": "々", + "1060": "親", + "1061": "美", + "1062": "信", + "1063": "都", + "1064": "置", + "1065": "局", + "1066": "運", + "1067": "送", + "1068": "風", + "1069": "口", + "1070": "演", + "1071": "調", + "1072": "ぎ", + "1073": "優", + "1074": "次", + "1075": "ォ", + "1076": "他", + "1077": "園", + "1078": "保", + "1079": "男", + "1080": "参", + "1081": "少", + "1082": "百", + "1083": "特", + "1084": "考", + "1085": "無", + "1086": "七", + "1087": "ヤ", + "1088": "ギ", + "1089": "良", + "1090": "ザ", + "1091": "制", + "1092": "売", + "1093": "能", + "1094": "原", + "1095": "ゲ", + "1096": "有", + "1097": "安", + "1098": "ゴ", + "1099": "育", + "1100": "科", + "1101": "要", + "1102": "料", + "1103": "書", + "1104": "語", + "1105": "設", + "1106": "海", + "1107": "期", + "1108": "流", + "1109": "確", + "1110": "ペ", + "1111": "区", + "1112": "む", + "1113": "連", + "1114": "買", + "1115": "ひ", + "1116": "ふ", + "1117": "付", + "1118": "町", + "1119": "活", + "1120": "情", + "1121": "月", + "1122": "表", + "1123": "曲", + "1124": "強", + "1125": "世", + "1126": "明", + "1127": "成", + "1128": "ノ", + "1129": "ァ", + "1130": "文", + "1131": "違", + "1132": "東", + "1133": "友", + "1134": "意", + "1135": "力", + "1136": "式", + "1137": "法", + "1138": "報", + "1139": "員", + "1140": "心", + "1141": "屋", + "1142": "品", + "1143": "北", + "1144": "先", + "1145": "島", + "1146": "味", + "1147": "川", + "1148": "開", + "1149": "千", + "1150": "関", + "1151": "電", + "1152": "然", + "1153": "度", + "1154": "達", + "1155": "面", + "1156": "九", + "1157": "数", + "1158": "取", + "1159": "楽", + "1160": "金", + "1161": "性", + "1162": "野", + "1163": "別", + "1164": "戦", + "1165": "公", + "1166": "機", + "1167": "道", + "1168": "目", + "1169": "記", + "1170": "び", + "1171": "発", + "1172": "対", + "1173": "立", + "1174": "初", + "1175": "化", + "1176": "ソ", + "1177": "ワ", + "1178": "田", + "1179": "持", + "1180": "ガ", + "1181": "車", + "1182": "番", + "1183": "ピ", + "1184": "聞", + "1185": "回", + "1186": "ぶ", + "1187": "ベ", + "1188": "げ", + "1189": "実", + "1190": "ボ", + "1191": "店", + "1192": "小", + "1193": "定", + "1194": "モ", + "1195": "長", + "1196": "新", + "1197": "ハ", + "1198": "ケ", + "1199": "外", + "1200": "ポ", + "1201": "近", + "1202": "所", + "1203": "へ", + "1204": "同", + "1205": "ネ", + "1206": "内", + "1207": "女", + "1208": "ホ", + "1209": "体", + "1210": "好", + "1211": "ツ", + "1212": "セ", + "1213": "知", + "1214": "山", + "1215": "来", + "1216": "ェ", + "1217": "使", + "1218": "ョ", + "1219": "ズ", + "1220": "主", + "1221": "動", + "1222": "理", + "1223": "物", + "1224": "映", + "1225": "者", + "1226": "ぐ", + "1227": "的", + "1228": "代", + "1229": "変", + "1230": "教", + "1231": "社", + "1232": "用", + "1233": "話", + "1234": "名", + "1235": "構", + "1236": "高", + "1237": "最", + "1238": "ず", + "1239": "ミ", + "1240": "校", + "1241": "ダ", + "1242": "食", + "1243": "後", + "1244": "手", + "1245": "三", + "1246": "通", + "1247": "感", + "1248": "合", + "1249": "多", + "1250": "業", + "1251": "入", + "1252": "エ", + "1253": "場", + "1254": "べ", + "1255": "上", + "1256": "家", + "1257": "私", + "1258": "年", + "1259": "間", + "1260": "画", + "1261": "前", + "1262": "下", + "1263": "ャ", + "1264": "地", + "1265": "二", + "1266": "ウ", + "1267": "ナ", + "1268": "ビ", + "1269": "自", + "1270": "全", + "1271": "パ", + "1272": "結", + "1273": "ブ", + "1274": "ュ", + "1275": "市", + "1276": "サ", + "1277": "気", + "1278": "方", + "1279": "デ", + "1280": "十", + "1281": "キ", + "1282": "当", + "1283": "国", + "1284": "作", + "1285": "ィ", + "1286": "部", + "1287": "オ", + "1288": "ニ", + "1289": "チ", + "1290": "ム", + "1291": "グ", + "1292": "メ", + "1293": "ご", + "1294": "子", + "1295": "ば", + "1296": "生", + "1297": "ほ", + "1298": "せ", + "1299": "何", + "1300": "出", + "1301": "言", + "1302": "今", + "1303": "バ", + "1304": "事", + "1305": "中", + "1306": "プ", + "1307": "時", + "1308": "コ", + "1309": "見", + "1310": "テ", + "1311": "会", + "1312": "マ", + "1313": "カ", + "1314": "思", + "1315": "ロ", + "1316": "ジ", + "1317": "フ", + "1318": "シ", + "1319": "め", + "1320": "レ", + "1321": "ド", + "1322": "分", + "1323": "ょ", + "1324": "ろ", + "1325": "学", + "1326": "行", + "1327": "タ", + "1328": "大", + "1329": "つ", + "1330": "本", + "1331": "日", + "1332": "わ", + "1333": "一", + "1334": "ク", + "1335": "み", + "1336": "リ", + "1337": "ア", + "1338": "ッ", + "1339": "人", + "1340": "ラ", + "1341": "お", + "1342": "じ", + "1343": "イ", + "1344": "ル", + "1345": "ト", + "1346": "ゃ", + "1347": "き", + "1348": "さ", + "1349": "ち", + "1350": "や", + "1351": "ス", + "1352": "ど", + "1353": "け", + "1354": "く", + "1355": "え", + "1356": "を", + "1357": "り", + "1358": "よ", + "1359": "こ", + "1360": "ン", + "1361": "だ", + "1362": "れ", + "1363": "ら", + "1364": "ね", + "1365": "が", + "1366": "ま", + "1367": "ー", + "1368": "も", + "1369": "そ", + "1370": "し", + "1371": "に", + "1372": "は", + "1373": "る", + "1374": "す", + "1375": "と", + "1376": "た", + "1377": "あ", + "1378": "て", + "1379": "っ", + "1380": "で", + "1381": "か", + "1382": "な", + "1383": "ん", + "1384": "う", + "1385": "の", + "1386": "、", + "1387": "。", + "1388": "い", + "1389": "", + "1390": "", + "1391": "", + "1392": "", + "1393": "", + "1394": "", + "1395": "", + "1396": "▁香", + "1397": "▁群", + "1398": "▁米", + "1399": "咆", + "1400": "哮", + "1401": "翅", + "1402": "", + "1403": "" +} \ No newline at end of file diff --git a/ja/2240ms/decoder.mlmodelc/analytics/coremldata.bin b/ja/2240ms/decoder.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..a23f14dd8e4d2bccc2844d3d81c6c9ca86ea3cba --- /dev/null +++ b/ja/2240ms/decoder.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2fcae710f3db79230f47be6daadc8af085539067285a96f89b2a4c0fd0cb3808 +size 243 diff --git a/ja/2240ms/decoder.mlmodelc/coremldata.bin b/ja/2240ms/decoder.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..ba0a40877c3fec55be46433278caccefbc29e03a --- /dev/null +++ b/ja/2240ms/decoder.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:650d41536fbbf813bacd98330acb53ec0ee6dc87c0350bc0bdd5531c3a9f7ea4 +size 493 diff --git a/ja/2240ms/decoder.mlmodelc/model.mil b/ja/2240ms/decoder.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..9f4b6b4cebc16f759164ca05a77b06fb57dedbce --- /dev/null +++ b/ja/2240ms/decoder.mlmodelc/model.mil @@ -0,0 +1,73 @@ +program(1.3) +[buildInfo = dict({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.10.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor c_in, tensor h_in, tensor token, tensor token_length) { + int32 y_batch_dims_0 = const()[name = string("y_batch_dims_0"), val = int32(0)]; + bool y_validate_indices_0 = const()[name = string("y_validate_indices_0"), val = bool(false)]; + tensor module_prediction_embed_weight_to_fp16 = const()[name = string("module_prediction_embed_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))]; + string token_to_int16_dtype_0 = const()[name = string("token_to_int16_dtype_0"), val = string("int16")]; + string cast_1_dtype_0 = const()[name = string("cast_1_dtype_0"), val = string("int32")]; + int32 greater_equal_0_y_0 = const()[name = string("greater_equal_0_y_0"), val = int32(0)]; + tensor token_to_int16 = cast(dtype = token_to_int16_dtype_0, x = token)[name = string("cast_9")]; + tensor cast_1 = cast(dtype = cast_1_dtype_0, x = token_to_int16)[name = string("cast_8")]; + tensor greater_equal_0 = greater_equal(x = cast_1, y = greater_equal_0_y_0)[name = string("greater_equal_0")]; + int32 slice_by_index_0 = const()[name = string("slice_by_index_0"), val = int32(1404)]; + tensor add_2 = add(x = cast_1, y = slice_by_index_0)[name = string("add_2")]; + tensor select_0 = select(a = cast_1, b = add_2, cond = greater_equal_0)[name = string("select_0")]; + int32 y_cast_fp16_cast_uint16_axis_0 = const()[name = string("y_cast_fp16_cast_uint16_axis_0"), val = int32(0)]; + string select_0_to_int16_dtype_0 = const()[name = string("select_0_to_int16_dtype_0"), val = string("int16")]; + tensor select_0_to_int16 = cast(dtype = select_0_to_int16_dtype_0, x = select_0)[name = string("cast_7")]; + tensor y_cast_fp16_cast_uint16_cast_uint16 = gather(axis = y_cast_fp16_cast_uint16_axis_0, batch_dims = y_batch_dims_0, indices = select_0_to_int16, validate_indices = y_validate_indices_0, x = module_prediction_embed_weight_to_fp16)[name = string("y_cast_fp16_cast_uint16_cast_uint16")]; + tensor input_3_perm_0 = const()[name = string("input_3_perm_0"), val = tensor([1, 0, 2])]; + int32 split_0_num_splits_0 = const()[name = string("split_0_num_splits_0"), val = int32(2)]; + int32 split_0_axis_0 = const()[name = string("split_0_axis_0"), val = int32(0)]; + string h_in_to_fp16_dtype_0 = const()[name = string("h_in_to_fp16_dtype_0"), val = string("fp16")]; + tensor h_in_to_fp16 = cast(dtype = h_in_to_fp16_dtype_0, x = h_in)[name = string("cast_6")]; + tensor split_0_cast_fp16_0, tensor split_0_cast_fp16_1 = split(axis = split_0_axis_0, num_splits = split_0_num_splits_0, x = h_in_to_fp16)[name = string("split_0_cast_fp16")]; + int32 split_1_num_splits_0 = const()[name = string("split_1_num_splits_0"), val = int32(2)]; + int32 split_1_axis_0 = const()[name = string("split_1_axis_0"), val = int32(0)]; + string c_in_to_fp16_dtype_0 = const()[name = string("c_in_to_fp16_dtype_0"), val = string("fp16")]; + tensor c_in_to_fp16 = cast(dtype = c_in_to_fp16_dtype_0, x = c_in)[name = string("cast_5")]; + tensor split_1_cast_fp16_0, tensor split_1_cast_fp16_1 = split(axis = split_1_axis_0, num_splits = split_1_num_splits_0, x = c_in_to_fp16)[name = string("split_1_cast_fp16")]; + tensor input_lstm_layer_0_lstm_h0_squeeze_axes_0 = const()[name = string("input_lstm_layer_0_lstm_h0_squeeze_axes_0"), val = tensor([0])]; + tensor input_lstm_layer_0_lstm_h0_squeeze_cast_fp16 = squeeze(axes = input_lstm_layer_0_lstm_h0_squeeze_axes_0, x = split_0_cast_fp16_0)[name = string("input_lstm_layer_0_lstm_h0_squeeze_cast_fp16")]; + tensor input_lstm_layer_0_lstm_c0_squeeze_axes_0 = const()[name = string("input_lstm_layer_0_lstm_c0_squeeze_axes_0"), val = tensor([0])]; + tensor input_lstm_layer_0_lstm_c0_squeeze_cast_fp16 = squeeze(axes = input_lstm_layer_0_lstm_c0_squeeze_axes_0, x = split_1_cast_fp16_0)[name = string("input_lstm_layer_0_lstm_c0_squeeze_cast_fp16")]; + string input_lstm_layer_0_direction_0 = const()[name = string("input_lstm_layer_0_direction_0"), val = string("forward")]; + bool input_lstm_layer_0_output_sequence_0 = const()[name = string("input_lstm_layer_0_output_sequence_0"), val = bool(true)]; + string input_lstm_layer_0_recurrent_activation_0 = const()[name = string("input_lstm_layer_0_recurrent_activation_0"), val = string("sigmoid")]; + string input_lstm_layer_0_cell_activation_0 = const()[name = string("input_lstm_layer_0_cell_activation_0"), val = string("tanh")]; + string input_lstm_layer_0_activation_0 = const()[name = string("input_lstm_layer_0_activation_0"), val = string("tanh")]; + tensor concat_1_to_fp16 = const()[name = string("concat_1_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1797248)))]; + tensor concat_2_to_fp16 = const()[name = string("concat_2_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5074112)))]; + tensor concat_0_to_fp16 = const()[name = string("concat_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8350976)))]; + tensor input_3_cast_fp16 = transpose(perm = input_3_perm_0, x = y_cast_fp16_cast_uint16_cast_uint16)[name = string("transpose_2")]; + tensor input_lstm_layer_0_cast_fp16_0, tensor input_lstm_layer_0_cast_fp16_1, tensor input_lstm_layer_0_cast_fp16_2 = lstm(activation = input_lstm_layer_0_activation_0, bias = concat_0_to_fp16, cell_activation = input_lstm_layer_0_cell_activation_0, direction = input_lstm_layer_0_direction_0, initial_c = input_lstm_layer_0_lstm_c0_squeeze_cast_fp16, initial_h = input_lstm_layer_0_lstm_h0_squeeze_cast_fp16, output_sequence = input_lstm_layer_0_output_sequence_0, recurrent_activation = input_lstm_layer_0_recurrent_activation_0, weight_hh = concat_2_to_fp16, weight_ih = concat_1_to_fp16, x = input_3_cast_fp16)[name = string("input_lstm_layer_0_cast_fp16")]; + tensor input_lstm_h0_squeeze_axes_0 = const()[name = string("input_lstm_h0_squeeze_axes_0"), val = tensor([0])]; + tensor input_lstm_h0_squeeze_cast_fp16 = squeeze(axes = input_lstm_h0_squeeze_axes_0, x = split_0_cast_fp16_1)[name = string("input_lstm_h0_squeeze_cast_fp16")]; + tensor input_lstm_c0_squeeze_axes_0 = const()[name = string("input_lstm_c0_squeeze_axes_0"), val = tensor([0])]; + tensor input_lstm_c0_squeeze_cast_fp16 = squeeze(axes = input_lstm_c0_squeeze_axes_0, x = split_1_cast_fp16_1)[name = string("input_lstm_c0_squeeze_cast_fp16")]; + string input_direction_0 = const()[name = string("input_direction_0"), val = string("forward")]; + bool input_output_sequence_0 = const()[name = string("input_output_sequence_0"), val = bool(true)]; + string input_recurrent_activation_0 = const()[name = string("input_recurrent_activation_0"), val = string("sigmoid")]; + string input_cell_activation_0 = const()[name = string("input_cell_activation_0"), val = string("tanh")]; + string input_activation_0 = const()[name = string("input_activation_0"), val = string("tanh")]; + tensor concat_4_to_fp16 = const()[name = string("concat_4_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8356160)))]; + tensor concat_5_to_fp16 = const()[name = string("concat_5_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11633024)))]; + tensor concat_3_to_fp16 = const()[name = string("concat_3_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14909888)))]; + tensor input_cast_fp16_0, tensor input_cast_fp16_1, tensor input_cast_fp16_2 = lstm(activation = input_activation_0, bias = concat_3_to_fp16, cell_activation = input_cell_activation_0, direction = input_direction_0, initial_c = input_lstm_c0_squeeze_cast_fp16, initial_h = input_lstm_h0_squeeze_cast_fp16, output_sequence = input_output_sequence_0, recurrent_activation = input_recurrent_activation_0, weight_hh = concat_5_to_fp16, weight_ih = concat_4_to_fp16, x = input_lstm_layer_0_cast_fp16_0)[name = string("input_cast_fp16")]; + int32 obj_3_axis_0 = const()[name = string("obj_3_axis_0"), val = int32(0)]; + tensor obj_3_cast_fp16 = stack(axis = obj_3_axis_0, values = (input_lstm_layer_0_cast_fp16_1, input_cast_fp16_1))[name = string("obj_3_cast_fp16")]; + string obj_3_cast_fp16_to_fp32_dtype_0 = const()[name = string("obj_3_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + int32 obj_axis_0 = const()[name = string("obj_axis_0"), val = int32(0)]; + tensor obj_cast_fp16 = stack(axis = obj_axis_0, values = (input_lstm_layer_0_cast_fp16_2, input_cast_fp16_2))[name = string("obj_cast_fp16")]; + string obj_cast_fp16_to_fp32_dtype_0 = const()[name = string("obj_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + tensor transpose_0_perm_0 = const()[name = string("transpose_0_perm_0"), val = tensor([1, 2, 0])]; + string transpose_0_cast_fp16_to_fp32_dtype_0 = const()[name = string("transpose_0_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + tensor transpose_0_cast_fp16 = transpose(perm = transpose_0_perm_0, x = input_cast_fp16_0)[name = string("transpose_1")]; + tensor decoder_out = cast(dtype = transpose_0_cast_fp16_to_fp32_dtype_0, x = transpose_0_cast_fp16)[name = string("cast_2")]; + tensor c_out = cast(dtype = obj_cast_fp16_to_fp32_dtype_0, x = obj_cast_fp16)[name = string("cast_3")]; + tensor h_out = cast(dtype = obj_3_cast_fp16_to_fp32_dtype_0, x = obj_3_cast_fp16)[name = string("cast_4")]; + tensor token_length_tmp = identity(x = token_length)[name = string("token_length_tmp")]; + } -> (decoder_out, h_out, c_out); +} \ No newline at end of file diff --git a/ja/2240ms/decoder.mlmodelc/weights/weight.bin b/ja/2240ms/decoder.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..9bcbce3e617e1135c0460941dfc723b71230d030 --- /dev/null +++ b/ja/2240ms/decoder.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d5eeb17b5de822e941f66ce3bd6739aae939dfa054f61b69ff6ce89d0ac9e778 +size 14915072 diff --git a/ja/2240ms/decoder.mlpackage/Data/com.apple.CoreML/model.mlmodel b/ja/2240ms/decoder.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..931ee2253d124627cf1b2689c6e01d5cf3746838 --- /dev/null +++ b/ja/2240ms/decoder.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:25c461bd45595f33022b4ce50bf3d493d5b70ae73c50bd0a98598336bd38864a +size 11598 diff --git a/ja/2240ms/decoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/ja/2240ms/decoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..9bcbce3e617e1135c0460941dfc723b71230d030 --- /dev/null +++ b/ja/2240ms/decoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d5eeb17b5de822e941f66ce3bd6739aae939dfa054f61b69ff6ce89d0ac9e778 +size 14915072 diff --git a/ja/2240ms/decoder.mlpackage/Manifest.json b/ja/2240ms/decoder.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..b4431e1d0753f3273eeff30f89de1232349486a7 --- /dev/null +++ b/ja/2240ms/decoder.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "8C20B369-4E12-4E4E-B3E8-A79B91D9CAFC": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + }, + "9356FC01-CF91-4D74-A142-118AF15703DD": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + } + }, + "rootModelIdentifier": "9356FC01-CF91-4D74-A142-118AF15703DD" +} diff --git a/ja/2240ms/decoder_joint.mlmodelc/analytics/coremldata.bin b/ja/2240ms/decoder_joint.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..30ee1bc4e73ed57bede1d9e6315c983146d06e8c --- /dev/null +++ b/ja/2240ms/decoder_joint.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:adba69d0e8e1547064d062072f64e8a9f1da383a6d09e2986a28268dd78cb23c +size 243 diff --git a/ja/2240ms/decoder_joint.mlmodelc/coremldata.bin b/ja/2240ms/decoder_joint.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..d6e5e288b0420308e0064f494bc40e62272f2dd8 --- /dev/null +++ b/ja/2240ms/decoder_joint.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8e2759de1b9d6040c9ae7d30cea3d8579fd7588d2ad4aeb711c42e014a2af203 +size 514 diff --git a/ja/2240ms/decoder_joint.mlmodelc/model.mil b/ja/2240ms/decoder_joint.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..2cb099751b91d7bd911aeac28092cc495bcaf315 --- /dev/null +++ b/ja/2240ms/decoder_joint.mlmodelc/model.mil @@ -0,0 +1,92 @@ +program(1.3) +[buildInfo = dict({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.10.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor c_in, tensor encoder, tensor h_in, tensor token, tensor token_length) { + int32 y_batch_dims_0 = const()[name = string("y_batch_dims_0"), val = int32(0)]; + bool y_validate_indices_0 = const()[name = string("y_validate_indices_0"), val = bool(false)]; + tensor decoder_module_prediction_embed_weight_to_fp16 = const()[name = string("decoder_module_prediction_embed_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))]; + string token_to_int16_dtype_0 = const()[name = string("token_to_int16_dtype_0"), val = string("int16")]; + string cast_1_dtype_0 = const()[name = string("cast_1_dtype_0"), val = string("int32")]; + int32 greater_equal_0_y_0 = const()[name = string("greater_equal_0_y_0"), val = int32(0)]; + tensor token_to_int16 = cast(dtype = token_to_int16_dtype_0, x = token)[name = string("cast_10")]; + tensor cast_1 = cast(dtype = cast_1_dtype_0, x = token_to_int16)[name = string("cast_9")]; + tensor greater_equal_0 = greater_equal(x = cast_1, y = greater_equal_0_y_0)[name = string("greater_equal_0")]; + int32 slice_by_index_0 = const()[name = string("slice_by_index_0"), val = int32(1404)]; + tensor add_2 = add(x = cast_1, y = slice_by_index_0)[name = string("add_2")]; + tensor select_0 = select(a = cast_1, b = add_2, cond = greater_equal_0)[name = string("select_0")]; + int32 y_cast_fp16_cast_uint16_axis_0 = const()[name = string("y_cast_fp16_cast_uint16_axis_0"), val = int32(0)]; + string select_0_to_int16_dtype_0 = const()[name = string("select_0_to_int16_dtype_0"), val = string("int16")]; + tensor select_0_to_int16 = cast(dtype = select_0_to_int16_dtype_0, x = select_0)[name = string("cast_8")]; + tensor y_cast_fp16_cast_uint16_cast_uint16 = gather(axis = y_cast_fp16_cast_uint16_axis_0, batch_dims = y_batch_dims_0, indices = select_0_to_int16, validate_indices = y_validate_indices_0, x = decoder_module_prediction_embed_weight_to_fp16)[name = string("y_cast_fp16_cast_uint16_cast_uint16")]; + tensor input_3_perm_0 = const()[name = string("input_3_perm_0"), val = tensor([1, 0, 2])]; + int32 split_0_num_splits_0 = const()[name = string("split_0_num_splits_0"), val = int32(2)]; + int32 split_0_axis_0 = const()[name = string("split_0_axis_0"), val = int32(0)]; + string h_in_to_fp16_dtype_0 = const()[name = string("h_in_to_fp16_dtype_0"), val = string("fp16")]; + tensor h_in_to_fp16 = cast(dtype = h_in_to_fp16_dtype_0, x = h_in)[name = string("cast_7")]; + tensor split_0_cast_fp16_0, tensor split_0_cast_fp16_1 = split(axis = split_0_axis_0, num_splits = split_0_num_splits_0, x = h_in_to_fp16)[name = string("split_0_cast_fp16")]; + int32 split_1_num_splits_0 = const()[name = string("split_1_num_splits_0"), val = int32(2)]; + int32 split_1_axis_0 = const()[name = string("split_1_axis_0"), val = int32(0)]; + string c_in_to_fp16_dtype_0 = const()[name = string("c_in_to_fp16_dtype_0"), val = string("fp16")]; + tensor c_in_to_fp16 = cast(dtype = c_in_to_fp16_dtype_0, x = c_in)[name = string("cast_6")]; + tensor split_1_cast_fp16_0, tensor split_1_cast_fp16_1 = split(axis = split_1_axis_0, num_splits = split_1_num_splits_0, x = c_in_to_fp16)[name = string("split_1_cast_fp16")]; + tensor input_5_lstm_layer_0_lstm_h0_squeeze_axes_0 = const()[name = string("input_5_lstm_layer_0_lstm_h0_squeeze_axes_0"), val = tensor([0])]; + tensor input_5_lstm_layer_0_lstm_h0_squeeze_cast_fp16 = squeeze(axes = input_5_lstm_layer_0_lstm_h0_squeeze_axes_0, x = split_0_cast_fp16_0)[name = string("input_5_lstm_layer_0_lstm_h0_squeeze_cast_fp16")]; + tensor input_5_lstm_layer_0_lstm_c0_squeeze_axes_0 = const()[name = string("input_5_lstm_layer_0_lstm_c0_squeeze_axes_0"), val = tensor([0])]; + tensor input_5_lstm_layer_0_lstm_c0_squeeze_cast_fp16 = squeeze(axes = input_5_lstm_layer_0_lstm_c0_squeeze_axes_0, x = split_1_cast_fp16_0)[name = string("input_5_lstm_layer_0_lstm_c0_squeeze_cast_fp16")]; + string input_5_lstm_layer_0_direction_0 = const()[name = string("input_5_lstm_layer_0_direction_0"), val = string("forward")]; + bool input_5_lstm_layer_0_output_sequence_0 = const()[name = string("input_5_lstm_layer_0_output_sequence_0"), val = bool(true)]; + string input_5_lstm_layer_0_recurrent_activation_0 = const()[name = string("input_5_lstm_layer_0_recurrent_activation_0"), val = string("sigmoid")]; + string input_5_lstm_layer_0_cell_activation_0 = const()[name = string("input_5_lstm_layer_0_cell_activation_0"), val = string("tanh")]; + string input_5_lstm_layer_0_activation_0 = const()[name = string("input_5_lstm_layer_0_activation_0"), val = string("tanh")]; + tensor concat_1_to_fp16 = const()[name = string("concat_1_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1797248)))]; + tensor concat_2_to_fp16 = const()[name = string("concat_2_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5074112)))]; + tensor concat_0_to_fp16 = const()[name = string("concat_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8350976)))]; + tensor input_3_cast_fp16 = transpose(perm = input_3_perm_0, x = y_cast_fp16_cast_uint16_cast_uint16)[name = string("transpose_4")]; + tensor input_5_lstm_layer_0_cast_fp16_0, tensor input_5_lstm_layer_0_cast_fp16_1, tensor input_5_lstm_layer_0_cast_fp16_2 = lstm(activation = input_5_lstm_layer_0_activation_0, bias = concat_0_to_fp16, cell_activation = input_5_lstm_layer_0_cell_activation_0, direction = input_5_lstm_layer_0_direction_0, initial_c = input_5_lstm_layer_0_lstm_c0_squeeze_cast_fp16, initial_h = input_5_lstm_layer_0_lstm_h0_squeeze_cast_fp16, output_sequence = input_5_lstm_layer_0_output_sequence_0, recurrent_activation = input_5_lstm_layer_0_recurrent_activation_0, weight_hh = concat_2_to_fp16, weight_ih = concat_1_to_fp16, x = input_3_cast_fp16)[name = string("input_5_lstm_layer_0_cast_fp16")]; + tensor input_5_lstm_h0_squeeze_axes_0 = const()[name = string("input_5_lstm_h0_squeeze_axes_0"), val = tensor([0])]; + tensor input_5_lstm_h0_squeeze_cast_fp16 = squeeze(axes = input_5_lstm_h0_squeeze_axes_0, x = split_0_cast_fp16_1)[name = string("input_5_lstm_h0_squeeze_cast_fp16")]; + tensor input_5_lstm_c0_squeeze_axes_0 = const()[name = string("input_5_lstm_c0_squeeze_axes_0"), val = tensor([0])]; + tensor input_5_lstm_c0_squeeze_cast_fp16 = squeeze(axes = input_5_lstm_c0_squeeze_axes_0, x = split_1_cast_fp16_1)[name = string("input_5_lstm_c0_squeeze_cast_fp16")]; + string input_5_direction_0 = const()[name = string("input_5_direction_0"), val = string("forward")]; + bool input_5_output_sequence_0 = const()[name = string("input_5_output_sequence_0"), val = bool(true)]; + string input_5_recurrent_activation_0 = const()[name = string("input_5_recurrent_activation_0"), val = string("sigmoid")]; + string input_5_cell_activation_0 = const()[name = string("input_5_cell_activation_0"), val = string("tanh")]; + string input_5_activation_0 = const()[name = string("input_5_activation_0"), val = string("tanh")]; + tensor concat_4_to_fp16 = const()[name = string("concat_4_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8356160)))]; + tensor concat_5_to_fp16 = const()[name = string("concat_5_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11633024)))]; + tensor concat_3_to_fp16 = const()[name = string("concat_3_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14909888)))]; + tensor input_5_cast_fp16_0, tensor input_5_cast_fp16_1, tensor input_5_cast_fp16_2 = lstm(activation = input_5_activation_0, bias = concat_3_to_fp16, cell_activation = input_5_cell_activation_0, direction = input_5_direction_0, initial_c = input_5_lstm_c0_squeeze_cast_fp16, initial_h = input_5_lstm_h0_squeeze_cast_fp16, output_sequence = input_5_output_sequence_0, recurrent_activation = input_5_recurrent_activation_0, weight_hh = concat_5_to_fp16, weight_ih = concat_4_to_fp16, x = input_5_lstm_layer_0_cast_fp16_0)[name = string("input_5_cast_fp16")]; + int32 obj_3_axis_0 = const()[name = string("obj_3_axis_0"), val = int32(0)]; + tensor obj_3_cast_fp16 = stack(axis = obj_3_axis_0, values = (input_5_lstm_layer_0_cast_fp16_1, input_5_cast_fp16_1))[name = string("obj_3_cast_fp16")]; + string obj_3_cast_fp16_to_fp32_dtype_0 = const()[name = string("obj_3_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + int32 obj_axis_0 = const()[name = string("obj_axis_0"), val = int32(0)]; + tensor obj_cast_fp16 = stack(axis = obj_axis_0, values = (input_5_lstm_layer_0_cast_fp16_2, input_5_cast_fp16_2))[name = string("obj_cast_fp16")]; + string obj_cast_fp16_to_fp32_dtype_0 = const()[name = string("obj_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + tensor transpose_1_perm_0 = const()[name = string("transpose_1_perm_0"), val = tensor([1, 0, 2])]; + tensor input_7_perm_0 = const()[name = string("input_7_perm_0"), val = tensor([0, 2, 1])]; + string encoder_to_fp16_dtype_0 = const()[name = string("encoder_to_fp16_dtype_0"), val = string("fp16")]; + tensor joint_module_enc_weight_to_fp16 = const()[name = string("joint_module_enc_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14915072)))]; + tensor joint_module_enc_bias_to_fp16 = const()[name = string("joint_module_enc_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16225856)))]; + tensor encoder_to_fp16 = cast(dtype = encoder_to_fp16_dtype_0, x = encoder)[name = string("cast_3")]; + tensor input_7_cast_fp16 = transpose(perm = input_7_perm_0, x = encoder_to_fp16)[name = string("transpose_2")]; + tensor linear_0_cast_fp16 = linear(bias = joint_module_enc_bias_to_fp16, weight = joint_module_enc_weight_to_fp16, x = input_7_cast_fp16)[name = string("linear_0_cast_fp16")]; + tensor joint_module_pred_weight_to_fp16 = const()[name = string("joint_module_pred_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16227200)))]; + tensor joint_module_pred_bias_to_fp16 = const()[name = string("joint_module_pred_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17046464)))]; + tensor transpose_1_cast_fp16 = transpose(perm = transpose_1_perm_0, x = input_5_cast_fp16_0)[name = string("transpose_3")]; + tensor linear_1_cast_fp16 = linear(bias = joint_module_pred_bias_to_fp16, weight = joint_module_pred_weight_to_fp16, x = transpose_1_cast_fp16)[name = string("linear_1_cast_fp16")]; + tensor var_79_axes_0 = const()[name = string("op_79_axes_0"), val = tensor([2])]; + tensor var_79_cast_fp16 = expand_dims(axes = var_79_axes_0, x = linear_0_cast_fp16)[name = string("op_79_cast_fp16")]; + tensor var_80_axes_0 = const()[name = string("op_80_axes_0"), val = tensor([1])]; + tensor var_80_cast_fp16 = expand_dims(axes = var_80_axes_0, x = linear_1_cast_fp16)[name = string("op_80_cast_fp16")]; + tensor input_11_cast_fp16 = add(x = var_79_cast_fp16, y = var_80_cast_fp16)[name = string("input_11_cast_fp16")]; + tensor input_13_cast_fp16 = relu(x = input_11_cast_fp16)[name = string("input_13_cast_fp16")]; + tensor joint_module_joint_net_2_weight_to_fp16 = const()[name = string("joint_module_joint_net_2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17047808)))]; + tensor joint_module_joint_net_2_bias_to_fp16 = const()[name = string("joint_module_joint_net_2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18844992)))]; + tensor linear_2_cast_fp16 = linear(bias = joint_module_joint_net_2_bias_to_fp16, weight = joint_module_joint_net_2_weight_to_fp16, x = input_13_cast_fp16)[name = string("linear_2_cast_fp16")]; + string linear_2_cast_fp16_to_fp32_dtype_0 = const()[name = string("linear_2_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + tensor logits = cast(dtype = linear_2_cast_fp16_to_fp32_dtype_0, x = linear_2_cast_fp16)[name = string("cast_2")]; + tensor c_out = cast(dtype = obj_cast_fp16_to_fp32_dtype_0, x = obj_cast_fp16)[name = string("cast_4")]; + tensor h_out = cast(dtype = obj_3_cast_fp16_to_fp32_dtype_0, x = obj_3_cast_fp16)[name = string("cast_5")]; + tensor token_length_tmp = identity(x = token_length)[name = string("token_length_tmp")]; + } -> (logits, h_out, c_out); +} \ No newline at end of file diff --git a/ja/2240ms/decoder_joint.mlmodelc/weights/weight.bin b/ja/2240ms/decoder_joint.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..c3bb2e494d21dfd602e504fdfe76da274071d914 --- /dev/null +++ b/ja/2240ms/decoder_joint.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bfc1c768dd0e0e61c0ab8806894ecc03902d2a5028e9c30f5d0a5e38d5139fd9 +size 18847864 diff --git a/ja/2240ms/decoder_joint.mlpackage/Data/com.apple.CoreML/model.mlmodel b/ja/2240ms/decoder_joint.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..4b6b70ea10ec445416345cff7ae90b536c38ce69 --- /dev/null +++ b/ja/2240ms/decoder_joint.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dad926552d028aa609be27dde32ede7e7e0687ce48d057ee7bae6d472f15229e +size 14986 diff --git a/ja/2240ms/decoder_joint.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/ja/2240ms/decoder_joint.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..c3bb2e494d21dfd602e504fdfe76da274071d914 --- /dev/null +++ b/ja/2240ms/decoder_joint.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bfc1c768dd0e0e61c0ab8806894ecc03902d2a5028e9c30f5d0a5e38d5139fd9 +size 18847864 diff --git a/ja/2240ms/decoder_joint.mlpackage/Manifest.json b/ja/2240ms/decoder_joint.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..8a2ce629ffcccdfce8e228506afd499877ee4507 --- /dev/null +++ b/ja/2240ms/decoder_joint.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "4DAC4947-89EA-4167-8E25-F5740051D450": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + }, + "D23EE53F-5E4C-405E-B75A-E4553E8229D1": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + } + }, + "rootModelIdentifier": "D23EE53F-5E4C-405E-B75A-E4553E8229D1" +} diff --git a/ja/2240ms/encoder.mlmodelc/analytics/coremldata.bin b/ja/2240ms/encoder.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..cf222b05b3795ceca6d95b114574c50f82d5a9fc --- /dev/null +++ b/ja/2240ms/encoder.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cb008a0254143d32e93073d5db48272287719969c062c9ac55514b11dc699a3f +size 243 diff --git a/ja/2240ms/encoder.mlmodelc/coremldata.bin b/ja/2240ms/encoder.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..2c99cfb335452b709279229f59f111685be4df81 --- /dev/null +++ b/ja/2240ms/encoder.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a4dd2996ab890fc119ac75286ff5412a4c046dd0d692d23909139d30b0171926 +size 633 diff --git a/ja/2240ms/encoder.mlmodelc/model.mil b/ja/2240ms/encoder.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..0b61772b118bff8a3d2b83d31f02ad558e04cde4 --- /dev/null +++ b/ja/2240ms/encoder.mlmodelc/model.mil @@ -0,0 +1,4434 @@ +program(1.3) +[buildInfo = dict({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}})] +{ + func main(tensor cache_channel, tensor cache_len, tensor cache_time, tensor mel, tensor mel_length, tensor prompt_id) { + tensor value_3_perm_0 = const()[name = string("value_3_perm_0"), val = tensor([1, 0, 2, 3])]; + string cache_channel_to_fp16_dtype_0 = const()[name = string("cache_channel_to_fp16_dtype_0"), val = string("fp16")]; + tensor value_5_perm_0 = const()[name = string("value_5_perm_0"), val = tensor([1, 0, 2, 3])]; + string cache_time_to_fp16_dtype_0 = const()[name = string("cache_time_to_fp16_dtype_0"), val = string("fp16")]; + int32 var_59 = const()[name = string("op_59"), val = int32(-1)]; + int32 var_68 = const()[name = string("op_68"), val = int32(1)]; + tensor x_1_perm_0 = const()[name = string("x_1_perm_0"), val = tensor([0, 2, 1])]; + string mel_to_fp16_dtype_0 = const()[name = string("mel_to_fp16_dtype_0"), val = string("fp16")]; + tensor tensor_1_axes_0 = const()[name = string("tensor_1_axes_0"), val = tensor([1])]; + tensor mel_to_fp16 = cast(dtype = mel_to_fp16_dtype_0, x = mel)[name = string("cast_21")]; + tensor x_1_cast_fp16 = transpose(perm = x_1_perm_0, x = mel_to_fp16)[name = string("transpose_367")]; + tensor tensor_1_cast_fp16 = expand_dims(axes = tensor_1_axes_0, x = x_1_cast_fp16)[name = string("tensor_1_cast_fp16")]; + tensor expand_dims_0 = const()[name = string("expand_dims_0"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))]; + tensor var_137_axes_0 = const()[name = string("op_137_axes_0"), val = tensor([1])]; + tensor var_137 = expand_dims(axes = var_137_axes_0, x = mel_length)[name = string("op_137")]; + tensor time_mask_1 = less(x = expand_dims_0, y = var_137)[name = string("time_mask_1")]; + tensor var_139_axes_0 = const()[name = string("op_139_axes_0"), val = tensor([-1])]; + tensor var_139 = expand_dims(axes = var_139_axes_0, x = time_mask_1)[name = string("op_139")]; + tensor var_141_reps_0 = const()[name = string("op_141_reps_0"), val = tensor([1, 1, 128])]; + tensor var_141 = tile(reps = var_141_reps_0, x = var_139)[name = string("op_141")]; + tensor var_147_axes_0 = const()[name = string("op_147_axes_0"), val = tensor([1])]; + string mask_1_to_fp16_dtype_0 = const()[name = string("mask_1_to_fp16_dtype_0"), val = string("fp16")]; + tensor var_141_to_fp16 = cast(dtype = mask_1_to_fp16_dtype_0, x = var_141)[name = string("cast_20")]; + tensor var_147_cast_fp16 = expand_dims(axes = var_147_axes_0, x = var_141_to_fp16)[name = string("op_147_cast_fp16")]; + tensor input_1_cast_fp16 = mul(x = tensor_1_cast_fp16, y = var_147_cast_fp16)[name = string("input_1_cast_fp16")]; + tensor input_3_pad_0 = const()[name = string("input_3_pad_0"), val = tensor([0, 0, 0, 0, 2, 1, 2, 1])]; + string input_3_mode_0 = const()[name = string("input_3_mode_0"), val = string("constant")]; + fp16 const_9_to_fp16 = const()[name = string("const_9_to_fp16"), val = fp16(0x0p+0)]; + tensor input_3_cast_fp16 = pad(constant_val = const_9_to_fp16, mode = input_3_mode_0, pad = input_3_pad_0, x = input_1_cast_fp16)[name = string("input_3_cast_fp16")]; + string tensor_3_pad_type_0 = const()[name = string("tensor_3_pad_type_0"), val = string("valid")]; + tensor tensor_3_strides_0 = const()[name = string("tensor_3_strides_0"), val = tensor([2, 2])]; + tensor tensor_3_pad_0 = const()[name = string("tensor_3_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor tensor_3_dilations_0 = const()[name = string("tensor_3_dilations_0"), val = tensor([1, 1])]; + int32 tensor_3_groups_0 = const()[name = string("tensor_3_groups_0"), val = int32(1)]; + tensor encoder_pre_encode_conv_0_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1088))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3456))))[name = string("encoder_pre_encode_conv_0_weight_to_fp16_quantized")]; + tensor encoder_pre_encode_conv_0_bias_to_fp16 = const()[name = string("encoder_pre_encode_conv_0_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4032)))]; + tensor tensor_3_cast_fp16 = conv(bias = encoder_pre_encode_conv_0_bias_to_fp16, dilations = tensor_3_dilations_0, groups = tensor_3_groups_0, pad = tensor_3_pad_0, pad_type = tensor_3_pad_type_0, strides = tensor_3_strides_0, weight = encoder_pre_encode_conv_0_weight_to_fp16_quantized, x = input_3_cast_fp16)[name = string("tensor_3_cast_fp16")]; + string current_lengths_1_to_fp16_dtype_0 = const()[name = string("current_lengths_1_to_fp16_dtype_0"), val = string("fp16")]; + fp16 var_160_promoted_to_fp16 = const()[name = string("op_160_promoted_to_fp16"), val = fp16(0x1p+1)]; + tensor mel_length_to_fp16 = cast(dtype = current_lengths_1_to_fp16_dtype_0, x = mel_length)[name = string("cast_19")]; + tensor var_161_cast_fp16 = add(x = mel_length_to_fp16, y = var_160_promoted_to_fp16)[name = string("op_161_cast_fp16")]; + fp16 var_162_promoted_to_fp16 = const()[name = string("op_162_promoted_to_fp16"), val = fp16(0x1p+0)]; + tensor var_163_cast_fp16 = add(x = var_161_cast_fp16, y = var_162_promoted_to_fp16)[name = string("op_163_cast_fp16")]; + fp16 var_164_promoted_to_fp16 = const()[name = string("op_164_promoted_to_fp16"), val = fp16(0x1.8p+1)]; + tensor var_165_cast_fp16 = sub(x = var_163_cast_fp16, y = var_164_promoted_to_fp16)[name = string("op_165_cast_fp16")]; + fp16 var_56_promoted_to_fp16 = const()[name = string("op_56_promoted_to_fp16"), val = fp16(0x1p+1)]; + tensor floor_div_0_cast_fp16 = floor_div(x = var_165_cast_fp16, y = var_56_promoted_to_fp16)[name = string("floor_div_0_cast_fp16")]; + fp16 var_167_promoted_to_fp16 = const()[name = string("op_167_promoted_to_fp16"), val = fp16(0x1p+0)]; + tensor current_lengths_3_cast_fp16 = add(x = floor_div_0_cast_fp16, y = var_167_promoted_to_fp16)[name = string("current_lengths_3_cast_fp16")]; + string lengths_19_dtype_0 = const()[name = string("lengths_19_dtype_0"), val = string("int32")]; + tensor expand_dims_1 = const()[name = string("expand_dims_1"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4608)))]; + tensor var_176_axes_0 = const()[name = string("op_176_axes_0"), val = tensor([1])]; + tensor current_lengths_3_cast_fp16_to_int32 = cast(dtype = lengths_19_dtype_0, x = current_lengths_3_cast_fp16)[name = string("cast_18")]; + tensor var_176 = expand_dims(axes = var_176_axes_0, x = current_lengths_3_cast_fp16_to_int32)[name = string("op_176")]; + tensor time_mask_3 = less(x = expand_dims_1, y = var_176)[name = string("time_mask_3")]; + tensor var_178_axes_0 = const()[name = string("op_178_axes_0"), val = tensor([-1])]; + tensor var_178 = expand_dims(axes = var_178_axes_0, x = time_mask_3)[name = string("op_178")]; + tensor var_180_reps_0 = const()[name = string("op_180_reps_0"), val = tensor([1, 1, 65])]; + tensor var_180 = tile(reps = var_180_reps_0, x = var_178)[name = string("op_180")]; + tensor var_186_axes_0 = const()[name = string("op_186_axes_0"), val = tensor([1])]; + string mask_3_to_fp16_dtype_0 = const()[name = string("mask_3_to_fp16_dtype_0"), val = string("fp16")]; + tensor var_180_to_fp16 = cast(dtype = mask_3_to_fp16_dtype_0, x = var_180)[name = string("cast_17")]; + tensor var_186_cast_fp16 = expand_dims(axes = var_186_axes_0, x = var_180_to_fp16)[name = string("op_186_cast_fp16")]; + tensor expanded_mask_3_reps_0 = const()[name = string("expanded_mask_3_reps_0"), val = tensor([1, 256, 1, 1])]; + tensor expanded_mask_3_cast_fp16 = tile(reps = expanded_mask_3_reps_0, x = var_186_cast_fp16)[name = string("expanded_mask_3_cast_fp16")]; + tensor input_5_cast_fp16 = mul(x = tensor_3_cast_fp16, y = expanded_mask_3_cast_fp16)[name = string("input_5_cast_fp16")]; + tensor tensor_5_cast_fp16 = relu(x = input_5_cast_fp16)[name = string("tensor_5_cast_fp16")]; + tensor input_7_cast_fp16 = mul(x = tensor_5_cast_fp16, y = expanded_mask_3_cast_fp16)[name = string("input_7_cast_fp16")]; + tensor input_9_pad_0 = const()[name = string("input_9_pad_0"), val = tensor([0, 0, 0, 0, 2, 1, 2, 1])]; + string input_9_mode_0 = const()[name = string("input_9_mode_0"), val = string("constant")]; + fp16 const_23_to_fp16 = const()[name = string("const_23_to_fp16"), val = fp16(0x0p+0)]; + tensor input_9_cast_fp16 = pad(constant_val = const_23_to_fp16, mode = input_9_mode_0, pad = input_9_pad_0, x = input_7_cast_fp16)[name = string("input_9_cast_fp16")]; + string tensor_7_pad_type_0 = const()[name = string("tensor_7_pad_type_0"), val = string("valid")]; + tensor tensor_7_strides_0 = const()[name = string("tensor_7_strides_0"), val = tensor([2, 2])]; + int32 tensor_7_groups_0 = const()[name = string("tensor_7_groups_0"), val = int32(256)]; + tensor tensor_7_pad_0 = const()[name = string("tensor_7_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor tensor_7_dilations_0 = const()[name = string("tensor_7_dilations_0"), val = tensor([1, 1])]; + tensor encoder_pre_encode_conv_2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5184))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7552))))[name = string("encoder_pre_encode_conv_2_weight_to_fp16_quantized")]; + tensor encoder_pre_encode_conv_2_bias_to_fp16 = const()[name = string("encoder_pre_encode_conv_2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8128)))]; + tensor tensor_7_cast_fp16 = conv(bias = encoder_pre_encode_conv_2_bias_to_fp16, dilations = tensor_7_dilations_0, groups = tensor_7_groups_0, pad = tensor_7_pad_0, pad_type = tensor_7_pad_type_0, strides = tensor_7_strides_0, weight = encoder_pre_encode_conv_2_weight_to_fp16_quantized, x = input_9_cast_fp16)[name = string("tensor_7_cast_fp16")]; + fp16 var_208_promoted_to_fp16 = const()[name = string("op_208_promoted_to_fp16"), val = fp16(0x1p+1)]; + tensor var_209_cast_fp16 = add(x = current_lengths_3_cast_fp16, y = var_208_promoted_to_fp16)[name = string("op_209_cast_fp16")]; + fp16 var_210_promoted_to_fp16 = const()[name = string("op_210_promoted_to_fp16"), val = fp16(0x1p+0)]; + tensor var_211_cast_fp16 = add(x = var_209_cast_fp16, y = var_210_promoted_to_fp16)[name = string("op_211_cast_fp16")]; + fp16 var_212_promoted_to_fp16 = const()[name = string("op_212_promoted_to_fp16"), val = fp16(0x1.8p+1)]; + tensor var_213_cast_fp16 = sub(x = var_211_cast_fp16, y = var_212_promoted_to_fp16)[name = string("op_213_cast_fp16")]; + fp16 var_56_promoted_1_to_fp16 = const()[name = string("op_56_promoted_1_to_fp16"), val = fp16(0x1p+1)]; + tensor floor_div_1_cast_fp16 = floor_div(x = var_213_cast_fp16, y = var_56_promoted_1_to_fp16)[name = string("floor_div_1_cast_fp16")]; + fp16 var_215_promoted_to_fp16 = const()[name = string("op_215_promoted_to_fp16"), val = fp16(0x1p+0)]; + tensor current_lengths_5_cast_fp16 = add(x = floor_div_1_cast_fp16, y = var_215_promoted_to_fp16)[name = string("current_lengths_5_cast_fp16")]; + string lengths_21_dtype_0 = const()[name = string("lengths_21_dtype_0"), val = string("int32")]; + tensor expand_dims_2 = const()[name = string("expand_dims_2"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8704)))]; + tensor var_224_axes_0 = const()[name = string("op_224_axes_0"), val = tensor([1])]; + tensor current_lengths_5_cast_fp16_to_int32 = cast(dtype = lengths_21_dtype_0, x = current_lengths_5_cast_fp16)[name = string("cast_16")]; + tensor var_224 = expand_dims(axes = var_224_axes_0, x = current_lengths_5_cast_fp16_to_int32)[name = string("op_224")]; + tensor time_mask_5 = less(x = expand_dims_2, y = var_224)[name = string("time_mask_5")]; + tensor var_226_axes_0 = const()[name = string("op_226_axes_0"), val = tensor([-1])]; + tensor var_226 = expand_dims(axes = var_226_axes_0, x = time_mask_5)[name = string("op_226")]; + tensor var_228_reps_0 = const()[name = string("op_228_reps_0"), val = tensor([1, 1, 33])]; + tensor var_228 = tile(reps = var_228_reps_0, x = var_226)[name = string("op_228")]; + tensor var_234_axes_0 = const()[name = string("op_234_axes_0"), val = tensor([1])]; + string mask_5_to_fp16_dtype_0 = const()[name = string("mask_5_to_fp16_dtype_0"), val = string("fp16")]; + tensor var_228_to_fp16 = cast(dtype = mask_5_to_fp16_dtype_0, x = var_228)[name = string("cast_15")]; + tensor var_234_cast_fp16 = expand_dims(axes = var_234_axes_0, x = var_228_to_fp16)[name = string("op_234_cast_fp16")]; + tensor expanded_mask_7_reps_0 = const()[name = string("expanded_mask_7_reps_0"), val = tensor([1, 256, 1, 1])]; + tensor expanded_mask_7_cast_fp16 = tile(reps = expanded_mask_7_reps_0, x = var_234_cast_fp16)[name = string("expanded_mask_7_cast_fp16")]; + tensor input_11_cast_fp16 = mul(x = tensor_7_cast_fp16, y = expanded_mask_7_cast_fp16)[name = string("input_11_cast_fp16")]; + string tensor_9_pad_type_0 = const()[name = string("tensor_9_pad_type_0"), val = string("valid")]; + tensor tensor_9_strides_0 = const()[name = string("tensor_9_strides_0"), val = tensor([1, 1])]; + tensor tensor_9_pad_0 = const()[name = string("tensor_9_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor tensor_9_dilations_0 = const()[name = string("tensor_9_dilations_0"), val = tensor([1, 1])]; + int32 tensor_9_groups_0 = const()[name = string("tensor_9_groups_0"), val = int32(1)]; + tensor encoder_pre_encode_conv_3_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(9024))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(74624))))[name = string("encoder_pre_encode_conv_3_weight_to_fp16_quantized")]; + tensor encoder_pre_encode_conv_3_bias_to_fp16 = const()[name = string("encoder_pre_encode_conv_3_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75200)))]; + tensor tensor_9_cast_fp16 = conv(bias = encoder_pre_encode_conv_3_bias_to_fp16, dilations = tensor_9_dilations_0, groups = tensor_9_groups_0, pad = tensor_9_pad_0, pad_type = tensor_9_pad_type_0, strides = tensor_9_strides_0, weight = encoder_pre_encode_conv_3_weight_to_fp16_quantized, x = input_11_cast_fp16)[name = string("tensor_9_cast_fp16")]; + tensor input_13_cast_fp16 = mul(x = tensor_9_cast_fp16, y = expanded_mask_7_cast_fp16)[name = string("input_13_cast_fp16")]; + tensor tensor_11_cast_fp16 = relu(x = input_13_cast_fp16)[name = string("tensor_11_cast_fp16")]; + tensor input_15_cast_fp16 = mul(x = tensor_11_cast_fp16, y = expanded_mask_7_cast_fp16)[name = string("input_15_cast_fp16")]; + tensor input_17_pad_0 = const()[name = string("input_17_pad_0"), val = tensor([0, 0, 0, 0, 2, 1, 2, 1])]; + string input_17_mode_0 = const()[name = string("input_17_mode_0"), val = string("constant")]; + fp16 const_41_to_fp16 = const()[name = string("const_41_to_fp16"), val = fp16(0x0p+0)]; + tensor input_17_cast_fp16 = pad(constant_val = const_41_to_fp16, mode = input_17_mode_0, pad = input_17_pad_0, x = input_15_cast_fp16)[name = string("input_17_cast_fp16")]; + string tensor_13_pad_type_0 = const()[name = string("tensor_13_pad_type_0"), val = string("valid")]; + tensor tensor_13_strides_0 = const()[name = string("tensor_13_strides_0"), val = tensor([2, 2])]; + int32 tensor_13_groups_0 = const()[name = string("tensor_13_groups_0"), val = int32(256)]; + tensor tensor_13_pad_0 = const()[name = string("tensor_13_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor tensor_13_dilations_0 = const()[name = string("tensor_13_dilations_0"), val = tensor([1, 1])]; + tensor encoder_pre_encode_conv_5_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75776))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78144))))[name = string("encoder_pre_encode_conv_5_weight_to_fp16_quantized")]; + tensor encoder_pre_encode_conv_5_bias_to_fp16 = const()[name = string("encoder_pre_encode_conv_5_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78720)))]; + tensor tensor_13_cast_fp16 = conv(bias = encoder_pre_encode_conv_5_bias_to_fp16, dilations = tensor_13_dilations_0, groups = tensor_13_groups_0, pad = tensor_13_pad_0, pad_type = tensor_13_pad_type_0, strides = tensor_13_strides_0, weight = encoder_pre_encode_conv_5_weight_to_fp16_quantized, x = input_17_cast_fp16)[name = string("tensor_13_cast_fp16")]; + fp16 var_271_promoted_to_fp16 = const()[name = string("op_271_promoted_to_fp16"), val = fp16(0x1p+1)]; + tensor var_272_cast_fp16 = add(x = current_lengths_5_cast_fp16, y = var_271_promoted_to_fp16)[name = string("op_272_cast_fp16")]; + fp16 var_273_promoted_to_fp16 = const()[name = string("op_273_promoted_to_fp16"), val = fp16(0x1p+0)]; + tensor var_274_cast_fp16 = add(x = var_272_cast_fp16, y = var_273_promoted_to_fp16)[name = string("op_274_cast_fp16")]; + fp16 var_275_promoted_to_fp16 = const()[name = string("op_275_promoted_to_fp16"), val = fp16(0x1.8p+1)]; + tensor var_276_cast_fp16 = sub(x = var_274_cast_fp16, y = var_275_promoted_to_fp16)[name = string("op_276_cast_fp16")]; + fp16 var_56_promoted_2_to_fp16 = const()[name = string("op_56_promoted_2_to_fp16"), val = fp16(0x1p+1)]; + tensor floor_div_2_cast_fp16 = floor_div(x = var_276_cast_fp16, y = var_56_promoted_2_to_fp16)[name = string("floor_div_2_cast_fp16")]; + fp16 var_278_promoted_to_fp16 = const()[name = string("op_278_promoted_to_fp16"), val = fp16(0x1p+0)]; + tensor current_lengths_cast_fp16 = add(x = floor_div_2_cast_fp16, y = var_278_promoted_to_fp16)[name = string("current_lengths_cast_fp16")]; + string lengths_dtype_0 = const()[name = string("lengths_dtype_0"), val = string("int32")]; + tensor expand_dims_3 = const()[name = string("expand_dims_3"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(79296)))]; + tensor var_287_axes_0 = const()[name = string("op_287_axes_0"), val = tensor([1])]; + tensor current_lengths_cast_fp16_to_int32 = cast(dtype = lengths_dtype_0, x = current_lengths_cast_fp16)[name = string("cast_14")]; + tensor var_287 = expand_dims(axes = var_287_axes_0, x = current_lengths_cast_fp16_to_int32)[name = string("op_287")]; + tensor time_mask = less(x = expand_dims_3, y = var_287)[name = string("time_mask")]; + tensor var_289_axes_0 = const()[name = string("op_289_axes_0"), val = tensor([-1])]; + tensor var_289 = expand_dims(axes = var_289_axes_0, x = time_mask)[name = string("op_289")]; + tensor var_291_reps_0 = const()[name = string("op_291_reps_0"), val = tensor([1, 1, 17])]; + tensor var_291 = tile(reps = var_291_reps_0, x = var_289)[name = string("op_291")]; + tensor var_297_axes_0 = const()[name = string("op_297_axes_0"), val = tensor([1])]; + string mask_7_to_fp16_dtype_0 = const()[name = string("mask_7_to_fp16_dtype_0"), val = string("fp16")]; + tensor var_291_to_fp16 = cast(dtype = mask_7_to_fp16_dtype_0, x = var_291)[name = string("cast_13")]; + tensor var_297_cast_fp16 = expand_dims(axes = var_297_axes_0, x = var_291_to_fp16)[name = string("op_297_cast_fp16")]; + tensor expanded_mask_13_reps_0 = const()[name = string("expanded_mask_13_reps_0"), val = tensor([1, 256, 1, 1])]; + tensor expanded_mask_13_cast_fp16 = tile(reps = expanded_mask_13_reps_0, x = var_297_cast_fp16)[name = string("expanded_mask_13_cast_fp16")]; + tensor input_19_cast_fp16 = mul(x = tensor_13_cast_fp16, y = expanded_mask_13_cast_fp16)[name = string("input_19_cast_fp16")]; + string tensor_15_pad_type_0 = const()[name = string("tensor_15_pad_type_0"), val = string("valid")]; + tensor tensor_15_strides_0 = const()[name = string("tensor_15_strides_0"), val = tensor([1, 1])]; + tensor tensor_15_pad_0 = const()[name = string("tensor_15_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor tensor_15_dilations_0 = const()[name = string("tensor_15_dilations_0"), val = tensor([1, 1])]; + int32 tensor_15_groups_0 = const()[name = string("tensor_15_groups_0"), val = int32(1)]; + tensor encoder_pre_encode_conv_6_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(79488))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(145088))))[name = string("encoder_pre_encode_conv_6_weight_to_fp16_quantized")]; + tensor encoder_pre_encode_conv_6_bias_to_fp16 = const()[name = string("encoder_pre_encode_conv_6_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(145664)))]; + tensor tensor_15_cast_fp16 = conv(bias = encoder_pre_encode_conv_6_bias_to_fp16, dilations = tensor_15_dilations_0, groups = tensor_15_groups_0, pad = tensor_15_pad_0, pad_type = tensor_15_pad_type_0, strides = tensor_15_strides_0, weight = encoder_pre_encode_conv_6_weight_to_fp16_quantized, x = input_19_cast_fp16)[name = string("tensor_15_cast_fp16")]; + tensor input_21_cast_fp16 = mul(x = tensor_15_cast_fp16, y = expanded_mask_13_cast_fp16)[name = string("input_21_cast_fp16")]; + tensor tensor_cast_fp16 = relu(x = input_21_cast_fp16)[name = string("tensor_cast_fp16")]; + tensor x_3_cast_fp16 = mul(x = tensor_cast_fp16, y = expanded_mask_13_cast_fp16)[name = string("x_3_cast_fp16")]; + tensor var_331_perm_0 = const()[name = string("op_331_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_332 = const()[name = string("op_332"), val = tensor([1, 30, -1])]; + tensor var_331_cast_fp16 = transpose(perm = var_331_perm_0, x = x_3_cast_fp16)[name = string("transpose_366")]; + tensor input_23_cast_fp16 = reshape(shape = var_332, x = var_331_cast_fp16)[name = string("input_23_cast_fp16")]; + tensor encoder_pre_encode_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(146240))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4602752))))[name = string("encoder_pre_encode_out_weight_to_fp16_quantized")]; + tensor encoder_pre_encode_out_bias_to_fp16 = const()[name = string("encoder_pre_encode_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4604864)))]; + tensor linear_0_cast_fp16 = linear(bias = encoder_pre_encode_out_bias_to_fp16, weight = encoder_pre_encode_out_weight_to_fp16_quantized, x = input_23_cast_fp16)[name = string("linear_0_cast_fp16")]; + tensor var_342_begin_0 = const()[name = string("op_342_begin_0"), val = tensor([0, 2, 0])]; + tensor var_342_end_0 = const()[name = string("op_342_end_0"), val = tensor([1, 30, 1024])]; + tensor var_342_end_mask_0 = const()[name = string("op_342_end_mask_0"), val = tensor([true, true, true])]; + tensor var_342_cast_fp16 = slice_by_index(begin = var_342_begin_0, end = var_342_end_0, end_mask = var_342_end_mask_0, x = linear_0_cast_fp16)[name = string("op_342_cast_fp16")]; + int32 var_344 = const()[name = string("op_344"), val = int32(2)]; + tensor var_345 = sub(x = current_lengths_cast_fp16_to_int32, y = var_344)[name = string("op_345")]; + string var_345_promoted_to_fp16_dtype_0 = const()[name = string("op_345_promoted_to_fp16_dtype_0"), val = string("fp16")]; + fp16 var_62_promoted_to_fp16 = const()[name = string("op_62_promoted_to_fp16"), val = fp16(0x0p+0)]; + fp16 const_61_to_fp16 = const()[name = string("const_61_to_fp16"), val = fp16(inf)]; + tensor var_345_to_fp16 = cast(dtype = var_345_promoted_to_fp16_dtype_0, x = var_345)[name = string("cast_12")]; + tensor clip_0_cast_fp16 = clip(alpha = var_62_promoted_to_fp16, beta = const_61_to_fp16, x = var_345_to_fp16)[name = string("clip_0_cast_fp16")]; + tensor max_audio_length_1 = const()[name = string("max_audio_length_1"), val = tensor([28])]; + fp16 var_361_promoted_to_fp16 = const()[name = string("op_361_promoted_to_fp16"), val = fp16(0x1.5p+5)]; + tensor padding_length_cast_fp16 = add(x = clip_0_cast_fp16, y = var_361_promoted_to_fp16)[name = string("padding_length_cast_fp16")]; + int32 const_63 = const()[name = string("const_63"), val = int32(-1)]; + tensor var_363 = mul(x = cache_len, y = const_63)[name = string("op_363")]; + int32 var_364 = const()[name = string("op_364"), val = int32(42)]; + tensor offset = add(x = var_363, y = var_364)[name = string("offset")]; + tensor var_404_axes_0 = const()[name = string("op_404_axes_0"), val = tensor([-1])]; + tensor var_404_cast_fp16 = expand_dims(axes = var_404_axes_0, x = padding_length_cast_fp16)[name = string("op_404_cast_fp16")]; + tensor var_403_promoted_to_fp16 = const()[name = string("op_403_promoted_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4606976)))]; + tensor pad_mask_1_cast_fp16 = less(x = var_403_promoted_to_fp16, y = var_404_cast_fp16)[name = string("pad_mask_1_cast_fp16")]; + tensor expand_dims_5 = const()[name = string("expand_dims_5"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4607232)))]; + tensor var_410_axes_0 = const()[name = string("op_410_axes_0"), val = tensor([-1])]; + tensor var_410 = expand_dims(axes = var_410_axes_0, x = offset)[name = string("op_410")]; + tensor pad_mask_off = greater_equal(x = expand_dims_5, y = var_410)[name = string("pad_mask_off")]; + tensor pad_mask_3 = logical_and(x = pad_mask_off, y = pad_mask_1_cast_fp16)[name = string("pad_mask_3")]; + tensor var_413_axes_0 = const()[name = string("op_413_axes_0"), val = tensor([1])]; + tensor var_413 = expand_dims(axes = var_413_axes_0, x = pad_mask_3)[name = string("op_413")]; + tensor var_414 = const()[name = string("op_414"), val = tensor([1, 70, 1])]; + tensor pad_mask_for_att_mask_1 = tile(reps = var_414, x = var_413)[name = string("pad_mask_for_att_mask_1")]; + tensor var_416_perm_0 = const()[name = string("op_416_perm_0"), val = tensor([0, 2, 1])]; + tensor var_416 = transpose(perm = var_416_perm_0, x = pad_mask_for_att_mask_1)[name = string("transpose_365")]; + tensor pad_mask_for_att_mask = logical_and(x = pad_mask_for_att_mask_1, y = var_416)[name = string("pad_mask_for_att_mask")]; + tensor const_71 = const()[name = string("const_71"), val = tensor([[[true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true]]])]; + tensor att_mask_9 = logical_and(x = pad_mask_for_att_mask, y = const_71)[name = string("att_mask_9")]; + tensor att_mask = logical_not(x = att_mask_9)[name = string("att_mask")]; + tensor pad_mask_5 = logical_not(x = pad_mask_3)[name = string("pad_mask_5")]; + tensor pad_mask_begin_0 = const()[name = string("pad_mask_begin_0"), val = tensor([0, 42])]; + tensor pad_mask_end_0 = const()[name = string("pad_mask_end_0"), val = tensor([1, 70])]; + tensor pad_mask_end_mask_0 = const()[name = string("pad_mask_end_mask_0"), val = tensor([true, true])]; + tensor pad_mask = slice_by_index(begin = pad_mask_begin_0, end = pad_mask_end_0, end_mask = pad_mask_end_mask_0, x = pad_mask_5)[name = string("pad_mask")]; + tensor mask_9_begin_0 = const()[name = string("mask_9_begin_0"), val = tensor([0, 42, 0])]; + tensor mask_9_end_0 = const()[name = string("mask_9_end_0"), val = tensor([1, 70, 70])]; + tensor mask_9_end_mask_0 = const()[name = string("mask_9_end_mask_0"), val = tensor([true, true, true])]; + tensor mask_9 = slice_by_index(begin = mask_9_begin_0, end = mask_9_end_0, end_mask = mask_9_end_mask_0, x = att_mask)[name = string("mask_9")]; + tensor cache_1_begin_0 = const()[name = string("cache_1_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor cache_1_end_0 = const()[name = string("cache_1_end_0"), val = tensor([1, 1, 42, 1024])]; + tensor cache_1_end_mask_0 = const()[name = string("cache_1_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_1_squeeze_mask_0 = const()[name = string("cache_1_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_channel_to_fp16 = cast(dtype = cache_channel_to_fp16_dtype_0, x = cache_channel)[name = string("cast_11")]; + tensor value_3_cast_fp16 = transpose(perm = value_3_perm_0, x = cache_channel_to_fp16)[name = string("transpose_364")]; + tensor cache_1_cast_fp16 = slice_by_index(begin = cache_1_begin_0, end = cache_1_end_0, end_mask = cache_1_end_mask_0, squeeze_mask = cache_1_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_1_cast_fp16")]; + tensor cache_3_begin_0 = const()[name = string("cache_3_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor cache_3_end_0 = const()[name = string("cache_3_end_0"), val = tensor([1, 1, 1024, 8])]; + tensor cache_3_end_mask_0 = const()[name = string("cache_3_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_3_squeeze_mask_0 = const()[name = string("cache_3_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_time_to_fp16 = cast(dtype = cache_time_to_fp16_dtype_0, x = cache_time)[name = string("cast_10")]; + tensor value_5_cast_fp16 = transpose(perm = value_5_perm_0, x = cache_time_to_fp16)[name = string("transpose_363")]; + tensor cache_3_cast_fp16 = slice_by_index(begin = cache_3_begin_0, end = cache_3_end_0, end_mask = cache_3_end_mask_0, squeeze_mask = cache_3_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_3_cast_fp16")]; + tensor input_27_axes_0 = const()[name = string("input_27_axes_0"), val = tensor([-1])]; + tensor encoder_layers_0_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_0_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4607616)))]; + tensor encoder_layers_0_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_0_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4609728)))]; + fp16 var_42_to_fp16 = const()[name = string("op_42_to_fp16"), val = fp16(0x1.5p-17)]; + tensor input_27_cast_fp16 = layer_norm(axes = input_27_axes_0, beta = encoder_layers_0_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_0_norm_feed_forward1_weight_to_fp16, x = var_342_cast_fp16)[name = string("input_27_cast_fp16")]; + tensor encoder_layers_0_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4611840))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8806208))))[name = string("encoder_layers_0_feed_forward1_linear1_weight_to_fp16_quantized")]; + tensor encoder_layers_0_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_0_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8814464)))]; + tensor linear_1_cast_fp16 = linear(bias = encoder_layers_0_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_0_feed_forward1_linear1_weight_to_fp16_quantized, x = input_27_cast_fp16)[name = string("linear_1_cast_fp16")]; + tensor input_31_cast_fp16 = silu(x = linear_1_cast_fp16)[name = string("input_31_cast_fp16")]; + tensor encoder_layers_0_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8822720))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13017088))))[name = string("encoder_layers_0_feed_forward1_linear2_weight_to_fp16_quantized")]; + tensor encoder_layers_0_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_0_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13019200)))]; + tensor linear_2_cast_fp16 = linear(bias = encoder_layers_0_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_0_feed_forward1_linear2_weight_to_fp16_quantized, x = input_31_cast_fp16)[name = string("linear_2_cast_fp16")]; + fp16 var_455_to_fp16 = const()[name = string("op_455_to_fp16"), val = fp16(0x1p-1)]; + tensor var_456_cast_fp16 = mul(x = linear_2_cast_fp16, y = var_455_to_fp16)[name = string("op_456_cast_fp16")]; + tensor input_37_cast_fp16 = add(x = var_342_cast_fp16, y = var_456_cast_fp16)[name = string("input_37_cast_fp16")]; + tensor key_1_axes_0 = const()[name = string("key_1_axes_0"), val = tensor([-1])]; + tensor encoder_layers_0_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_0_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13021312)))]; + tensor encoder_layers_0_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_0_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13023424)))]; + tensor key_1_cast_fp16 = layer_norm(axes = key_1_axes_0, beta = encoder_layers_0_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_0_norm_self_att_weight_to_fp16, x = input_37_cast_fp16)[name = string("key_1_cast_fp16")]; + bool input_39_interleave_0 = const()[name = string("input_39_interleave_0"), val = bool(false)]; + tensor input_39_cast_fp16 = concat(axis = var_68, interleave = input_39_interleave_0, values = (cache_1_cast_fp16, key_1_cast_fp16))[name = string("input_39_cast_fp16")]; + tensor var_478_begin_0 = const()[name = string("op_478_begin_0"), val = tensor([0, 28, 0])]; + tensor var_478_end_0 = const()[name = string("op_478_end_0"), val = tensor([1, 42, 1024])]; + tensor var_478_end_mask_0 = const()[name = string("op_478_end_mask_0"), val = tensor([true, true, true])]; + tensor var_478_cast_fp16 = slice_by_index(begin = var_478_begin_0, end = var_478_end_0, end_mask = var_478_end_mask_0, x = cache_1_cast_fp16)[name = string("op_478_cast_fp16")]; + bool var_484_interleave_0 = const()[name = string("op_484_interleave_0"), val = bool(false)]; + tensor var_484_cast_fp16 = concat(axis = var_68, interleave = var_484_interleave_0, values = (var_478_cast_fp16, key_1_cast_fp16))[name = string("op_484_cast_fp16")]; + tensor encoder_layers_0_self_attn_linear_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13025536))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14074176))))[name = string("encoder_layers_0_self_attn_linear_q_weight_to_fp16_quantized")]; + tensor encoder_layers_0_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_0_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14076288)))]; + tensor linear_3_cast_fp16 = linear(bias = encoder_layers_0_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_0_self_attn_linear_q_weight_to_fp16_quantized, x = key_1_cast_fp16)[name = string("linear_3_cast_fp16")]; + tensor var_489 = const()[name = string("op_489"), val = tensor([1, -1, 8, 128])]; + tensor q_1_cast_fp16 = reshape(shape = var_489, x = linear_3_cast_fp16)[name = string("q_1_cast_fp16")]; + tensor encoder_layers_0_self_attn_linear_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14078400))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15127040))))[name = string("encoder_layers_0_self_attn_linear_k_weight_to_fp16_quantized")]; + tensor encoder_layers_0_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_0_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15129152)))]; + tensor linear_4_cast_fp16 = linear(bias = encoder_layers_0_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_0_self_attn_linear_k_weight_to_fp16_quantized, x = input_39_cast_fp16)[name = string("linear_4_cast_fp16")]; + tensor var_494 = const()[name = string("op_494"), val = tensor([1, -1, 8, 128])]; + tensor k_1_cast_fp16 = reshape(shape = var_494, x = linear_4_cast_fp16)[name = string("k_1_cast_fp16")]; + tensor encoder_layers_0_self_attn_linear_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15131264))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16179904))))[name = string("encoder_layers_0_self_attn_linear_v_weight_to_fp16_quantized")]; + tensor encoder_layers_0_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_0_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16182016)))]; + tensor linear_5_cast_fp16 = linear(bias = encoder_layers_0_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_0_self_attn_linear_v_weight_to_fp16_quantized, x = input_39_cast_fp16)[name = string("linear_5_cast_fp16")]; + tensor var_499 = const()[name = string("op_499"), val = tensor([1, -1, 8, 128])]; + tensor v_1_cast_fp16 = reshape(shape = var_499, x = linear_5_cast_fp16)[name = string("v_1_cast_fp16")]; + tensor value_9_perm_0 = const()[name = string("value_9_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_0_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_0_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16184128)))]; + tensor var_512_cast_fp16 = add(x = q_1_cast_fp16, y = encoder_layers_0_self_attn_pos_bias_u_to_fp16)[name = string("op_512_cast_fp16")]; + tensor encoder_layers_0_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_0_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16186240)))]; + tensor var_514_cast_fp16 = add(x = q_1_cast_fp16, y = encoder_layers_0_self_attn_pos_bias_v_to_fp16)[name = string("op_514_cast_fp16")]; + tensor q_with_bias_v_1_perm_0 = const()[name = string("q_with_bias_v_1_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_7_transpose_x_0 = const()[name = string("x_7_transpose_x_0"), val = bool(false)]; + bool x_7_transpose_y_0 = const()[name = string("x_7_transpose_y_0"), val = bool(false)]; + tensor op_516_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16188352))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16330752))))[name = string("op_516_to_fp16_quantized")]; + tensor q_with_bias_v_1_cast_fp16 = transpose(perm = q_with_bias_v_1_perm_0, x = var_514_cast_fp16)[name = string("transpose_362")]; + tensor x_7_cast_fp16 = matmul(transpose_x = x_7_transpose_x_0, transpose_y = x_7_transpose_y_0, x = q_with_bias_v_1_cast_fp16, y = op_516_to_fp16_quantized)[name = string("x_7_cast_fp16")]; + tensor x_9_pad_0 = const()[name = string("x_9_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_9_mode_0 = const()[name = string("x_9_mode_0"), val = string("constant")]; + fp16 const_79_to_fp16 = const()[name = string("const_79_to_fp16"), val = fp16(0x0p+0)]; + tensor x_9_cast_fp16 = pad(constant_val = const_79_to_fp16, mode = x_9_mode_0, pad = x_9_pad_0, x = x_7_cast_fp16)[name = string("x_9_cast_fp16")]; + tensor var_524 = const()[name = string("op_524"), val = tensor([1, 8, -1, 28])]; + tensor x_11_cast_fp16 = reshape(shape = var_524, x = x_9_cast_fp16)[name = string("x_11_cast_fp16")]; + tensor var_528_begin_0 = const()[name = string("op_528_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_528_end_0 = const()[name = string("op_528_end_0"), val = tensor([1, 8, 140, 28])]; + tensor var_528_end_mask_0 = const()[name = string("op_528_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_528_cast_fp16 = slice_by_index(begin = var_528_begin_0, end = var_528_end_0, end_mask = var_528_end_mask_0, x = x_11_cast_fp16)[name = string("op_528_cast_fp16")]; + tensor var_529 = const()[name = string("op_529"), val = tensor([1, 8, 28, 139])]; + tensor matrix_bd_1_cast_fp16 = reshape(shape = var_529, x = var_528_cast_fp16)[name = string("matrix_bd_1_cast_fp16")]; + bool matrix_ac_1_transpose_x_0 = const()[name = string("matrix_ac_1_transpose_x_0"), val = bool(false)]; + bool matrix_ac_1_transpose_y_0 = const()[name = string("matrix_ac_1_transpose_y_0"), val = bool(false)]; + tensor transpose_96_perm_0 = const()[name = string("transpose_96_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_97_perm_0 = const()[name = string("transpose_97_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_97 = transpose(perm = transpose_97_perm_0, x = k_1_cast_fp16)[name = string("transpose_360")]; + tensor transpose_96 = transpose(perm = transpose_96_perm_0, x = var_512_cast_fp16)[name = string("transpose_361")]; + tensor matrix_ac_1_cast_fp16 = matmul(transpose_x = matrix_ac_1_transpose_x_0, transpose_y = matrix_ac_1_transpose_y_0, x = transpose_96, y = transpose_97)[name = string("matrix_ac_1_cast_fp16")]; + tensor matrix_bd_3_begin_0 = const()[name = string("matrix_bd_3_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_3_end_0 = const()[name = string("matrix_bd_3_end_0"), val = tensor([1, 8, 28, 70])]; + tensor matrix_bd_3_end_mask_0 = const()[name = string("matrix_bd_3_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_3_cast_fp16 = slice_by_index(begin = matrix_bd_3_begin_0, end = matrix_bd_3_end_0, end_mask = matrix_bd_3_end_mask_0, x = matrix_bd_1_cast_fp16)[name = string("matrix_bd_3_cast_fp16")]; + tensor var_538_cast_fp16 = add(x = matrix_ac_1_cast_fp16, y = matrix_bd_3_cast_fp16)[name = string("op_538_cast_fp16")]; + fp16 _inversed_scores_1_y_0_to_fp16 = const()[name = string("_inversed_scores_1_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_1_cast_fp16 = mul(x = var_538_cast_fp16, y = _inversed_scores_1_y_0_to_fp16)[name = string("_inversed_scores_1_cast_fp16")]; + tensor mask_11_axes_0 = const()[name = string("mask_11_axes_0"), val = tensor([1])]; + tensor mask_11 = expand_dims(axes = mask_11_axes_0, x = mask_9)[name = string("mask_11")]; + fp16 var_45_to_fp16 = const()[name = string("op_45_to_fp16"), val = fp16(-0x1.388p+13)]; + tensor scores_3_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_1_cast_fp16, cond = mask_11)[name = string("scores_3_cast_fp16")]; + tensor var_544_cast_fp16 = softmax(axis = var_59, x = scores_3_cast_fp16)[name = string("op_544_cast_fp16")]; + fp16 var_44_to_fp16 = const()[name = string("op_44_to_fp16"), val = fp16(0x0p+0)]; + tensor input_41_cast_fp16 = select(a = var_44_to_fp16, b = var_544_cast_fp16, cond = mask_11)[name = string("input_41_cast_fp16")]; + bool x_13_transpose_x_0 = const()[name = string("x_13_transpose_x_0"), val = bool(false)]; + bool x_13_transpose_y_0 = const()[name = string("x_13_transpose_y_0"), val = bool(false)]; + tensor value_9_cast_fp16 = transpose(perm = value_9_perm_0, x = v_1_cast_fp16)[name = string("transpose_359")]; + tensor x_13_cast_fp16 = matmul(transpose_x = x_13_transpose_x_0, transpose_y = x_13_transpose_y_0, x = input_41_cast_fp16, y = value_9_cast_fp16)[name = string("x_13_cast_fp16")]; + tensor var_548_perm_0 = const()[name = string("op_548_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_549 = const()[name = string("op_549"), val = tensor([1, -1, 1024])]; + tensor var_548_cast_fp16 = transpose(perm = var_548_perm_0, x = x_13_cast_fp16)[name = string("transpose_358")]; + tensor input_43_cast_fp16 = reshape(shape = var_549, x = var_548_cast_fp16)[name = string("input_43_cast_fp16")]; + tensor encoder_layers_0_self_attn_linear_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16331136))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17379776))))[name = string("encoder_layers_0_self_attn_linear_out_weight_to_fp16_quantized")]; + tensor encoder_layers_0_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_0_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17381888)))]; + tensor linear_7_cast_fp16 = linear(bias = encoder_layers_0_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_0_self_attn_linear_out_weight_to_fp16_quantized, x = input_43_cast_fp16)[name = string("linear_7_cast_fp16")]; + tensor input_47_cast_fp16 = add(x = input_37_cast_fp16, y = linear_7_cast_fp16)[name = string("input_47_cast_fp16")]; + tensor x_17_axes_0 = const()[name = string("x_17_axes_0"), val = tensor([-1])]; + tensor encoder_layers_0_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_0_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17384000)))]; + tensor encoder_layers_0_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_0_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17386112)))]; + tensor x_17_cast_fp16 = layer_norm(axes = x_17_axes_0, beta = encoder_layers_0_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_0_norm_conv_weight_to_fp16, x = input_47_cast_fp16)[name = string("x_17_cast_fp16")]; + tensor input_49_perm_0 = const()[name = string("input_49_perm_0"), val = tensor([0, 2, 1])]; + string input_51_pad_type_0 = const()[name = string("input_51_pad_type_0"), val = string("valid")]; + tensor input_51_strides_0 = const()[name = string("input_51_strides_0"), val = tensor([1])]; + tensor input_51_pad_0 = const()[name = string("input_51_pad_0"), val = tensor([0, 0])]; + tensor input_51_dilations_0 = const()[name = string("input_51_dilations_0"), val = tensor([1])]; + int32 input_51_groups_0 = const()[name = string("input_51_groups_0"), val = int32(1)]; + tensor encoder_layers_0_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17388224))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19485440))))[name = string("encoder_layers_0_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_49_cast_fp16 = transpose(perm = input_49_perm_0, x = x_17_cast_fp16)[name = string("transpose_357")]; + tensor input_51_cast_fp16 = conv(dilations = input_51_dilations_0, groups = input_51_groups_0, pad = input_51_pad_0, pad_type = input_51_pad_type_0, strides = input_51_strides_0, weight = encoder_layers_0_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_49_cast_fp16)[name = string("input_51_cast_fp16")]; + int32 x_19_split_num_splits_0 = const()[name = string("x_19_split_num_splits_0"), val = int32(2)]; + int32 x_19_split_axis_0 = const()[name = string("x_19_split_axis_0"), val = int32(1)]; + tensor x_19_split_cast_fp16_0, tensor x_19_split_cast_fp16_1 = split(axis = x_19_split_axis_0, num_splits = x_19_split_num_splits_0, x = input_51_cast_fp16)[name = string("x_19_split_cast_fp16")]; + tensor x_19_split_1_sigmoid_cast_fp16 = sigmoid(x = x_19_split_cast_fp16_1)[name = string("x_19_split_1_sigmoid_cast_fp16")]; + tensor x_19_cast_fp16 = mul(x = x_19_split_cast_fp16_0, y = x_19_split_1_sigmoid_cast_fp16)[name = string("x_19_cast_fp16")]; + tensor var_575_axes_0 = const()[name = string("op_575_axes_0"), val = tensor([1])]; + tensor var_575 = expand_dims(axes = var_575_axes_0, x = pad_mask)[name = string("op_575")]; + tensor input_53_cast_fp16 = select(a = var_44_to_fp16, b = x_19_cast_fp16, cond = var_575)[name = string("input_53_cast_fp16")]; + bool new_x_3_interleave_0 = const()[name = string("new_x_3_interleave_0"), val = bool(false)]; + tensor new_x_3_cast_fp16 = concat(axis = var_59, interleave = new_x_3_interleave_0, values = (cache_3_cast_fp16, input_53_cast_fp16))[name = string("new_x_3_cast_fp16")]; + tensor var_588_begin_0 = const()[name = string("op_588_begin_0"), val = tensor([0, 0, 28])]; + tensor var_588_end_0 = const()[name = string("op_588_end_0"), val = tensor([1, 1024, 36])]; + tensor var_588_end_mask_0 = const()[name = string("op_588_end_mask_0"), val = tensor([true, true, true])]; + tensor var_588_cast_fp16 = slice_by_index(begin = var_588_begin_0, end = var_588_end_0, end_mask = var_588_end_mask_0, x = new_x_3_cast_fp16)[name = string("op_588_cast_fp16")]; + string x_21_pad_type_0 = const()[name = string("x_21_pad_type_0"), val = string("valid")]; + int32 x_21_groups_0 = const()[name = string("x_21_groups_0"), val = int32(1024)]; + tensor x_21_strides_0 = const()[name = string("x_21_strides_0"), val = tensor([1])]; + tensor x_21_pad_0 = const()[name = string("x_21_pad_0"), val = tensor([0, 0])]; + tensor x_21_dilations_0 = const()[name = string("x_21_dilations_0"), val = tensor([1])]; + tensor encoder_layers_0_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19489600))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19498880))))[name = string("encoder_layers_0_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_21_cast_fp16 = conv(dilations = x_21_dilations_0, groups = x_21_groups_0, pad = x_21_pad_0, pad_type = x_21_pad_type_0, strides = x_21_strides_0, weight = encoder_layers_0_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_3_cast_fp16)[name = string("x_21_cast_fp16")]; + tensor input_55_perm_0 = const()[name = string("input_55_perm_0"), val = tensor([0, 2, 1])]; + tensor x_23_axes_0 = const()[name = string("x_23_axes_0"), val = tensor([-1])]; + tensor encoder_layers_0_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_0_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19500992)))]; + tensor encoder_layers_0_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_0_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19503104)))]; + tensor input_55_cast_fp16 = transpose(perm = input_55_perm_0, x = x_21_cast_fp16)[name = string("transpose_356")]; + tensor x_23_cast_fp16 = layer_norm(axes = x_23_axes_0, beta = encoder_layers_0_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_0_conv_batch_norm_weight_to_fp16, x = input_55_cast_fp16)[name = string("x_23_cast_fp16")]; + tensor input_57_perm_0 = const()[name = string("input_57_perm_0"), val = tensor([0, 2, 1])]; + tensor input_57_cast_fp16 = transpose(perm = input_57_perm_0, x = x_23_cast_fp16)[name = string("transpose_355")]; + tensor input_59_cast_fp16 = silu(x = input_57_cast_fp16)[name = string("input_59_cast_fp16")]; + string x_25_pad_type_0 = const()[name = string("x_25_pad_type_0"), val = string("valid")]; + tensor x_25_strides_0 = const()[name = string("x_25_strides_0"), val = tensor([1])]; + tensor x_25_pad_0 = const()[name = string("x_25_pad_0"), val = tensor([0, 0])]; + tensor x_25_dilations_0 = const()[name = string("x_25_dilations_0"), val = tensor([1])]; + int32 x_25_groups_0 = const()[name = string("x_25_groups_0"), val = int32(1)]; + tensor encoder_layers_0_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19505216))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20553856))))[name = string("encoder_layers_0_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_25_cast_fp16 = conv(dilations = x_25_dilations_0, groups = x_25_groups_0, pad = x_25_pad_0, pad_type = x_25_pad_type_0, strides = x_25_strides_0, weight = encoder_layers_0_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_59_cast_fp16)[name = string("x_25_cast_fp16")]; + tensor input_61_perm_0 = const()[name = string("input_61_perm_0"), val = tensor([0, 2, 1])]; + tensor input_61_cast_fp16 = transpose(perm = input_61_perm_0, x = x_25_cast_fp16)[name = string("transpose_354")]; + tensor input_63_cast_fp16 = add(x = input_47_cast_fp16, y = input_61_cast_fp16)[name = string("input_63_cast_fp16")]; + tensor input_65_axes_0 = const()[name = string("input_65_axes_0"), val = tensor([-1])]; + tensor encoder_layers_0_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_0_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20555968)))]; + tensor encoder_layers_0_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_0_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20558080)))]; + tensor input_65_cast_fp16 = layer_norm(axes = input_65_axes_0, beta = encoder_layers_0_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_0_norm_feed_forward2_weight_to_fp16, x = input_63_cast_fp16)[name = string("input_65_cast_fp16")]; + tensor encoder_layers_0_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20560192))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24754560))))[name = string("encoder_layers_0_feed_forward2_linear1_weight_to_fp16_quantized")]; + tensor encoder_layers_0_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_0_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24762816)))]; + tensor linear_8_cast_fp16 = linear(bias = encoder_layers_0_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_0_feed_forward2_linear1_weight_to_fp16_quantized, x = input_65_cast_fp16)[name = string("linear_8_cast_fp16")]; + tensor input_69_cast_fp16 = silu(x = linear_8_cast_fp16)[name = string("input_69_cast_fp16")]; + tensor encoder_layers_0_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24771072))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28965440))))[name = string("encoder_layers_0_feed_forward2_linear2_weight_to_fp16_quantized")]; + tensor encoder_layers_0_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_0_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28967552)))]; + tensor linear_9_cast_fp16 = linear(bias = encoder_layers_0_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_0_feed_forward2_linear2_weight_to_fp16_quantized, x = input_69_cast_fp16)[name = string("linear_9_cast_fp16")]; + fp16 var_631_to_fp16 = const()[name = string("op_631_to_fp16"), val = fp16(0x1p-1)]; + tensor var_632_cast_fp16 = mul(x = linear_9_cast_fp16, y = var_631_to_fp16)[name = string("op_632_cast_fp16")]; + tensor input_75_cast_fp16 = add(x = input_63_cast_fp16, y = var_632_cast_fp16)[name = string("input_75_cast_fp16")]; + tensor input_77_axes_0 = const()[name = string("input_77_axes_0"), val = tensor([-1])]; + tensor encoder_layers_0_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_0_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28969664)))]; + tensor encoder_layers_0_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_0_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28971776)))]; + tensor input_77_cast_fp16 = layer_norm(axes = input_77_axes_0, beta = encoder_layers_0_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_0_norm_out_weight_to_fp16, x = input_75_cast_fp16)[name = string("input_77_cast_fp16")]; + tensor cache_5_begin_0 = const()[name = string("cache_5_begin_0"), val = tensor([1, 0, 0, 0])]; + tensor cache_5_end_0 = const()[name = string("cache_5_end_0"), val = tensor([2, 1, 42, 1024])]; + tensor cache_5_end_mask_0 = const()[name = string("cache_5_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_5_squeeze_mask_0 = const()[name = string("cache_5_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_5_cast_fp16 = slice_by_index(begin = cache_5_begin_0, end = cache_5_end_0, end_mask = cache_5_end_mask_0, squeeze_mask = cache_5_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_5_cast_fp16")]; + tensor cache_7_begin_0 = const()[name = string("cache_7_begin_0"), val = tensor([1, 0, 0, 0])]; + tensor cache_7_end_0 = const()[name = string("cache_7_end_0"), val = tensor([2, 1, 1024, 8])]; + tensor cache_7_end_mask_0 = const()[name = string("cache_7_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_7_squeeze_mask_0 = const()[name = string("cache_7_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_7_cast_fp16 = slice_by_index(begin = cache_7_begin_0, end = cache_7_end_0, end_mask = cache_7_end_mask_0, squeeze_mask = cache_7_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_7_cast_fp16")]; + tensor input_79_axes_0 = const()[name = string("input_79_axes_0"), val = tensor([-1])]; + tensor encoder_layers_1_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_1_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28973888)))]; + tensor encoder_layers_1_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_1_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28976000)))]; + tensor input_79_cast_fp16 = layer_norm(axes = input_79_axes_0, beta = encoder_layers_1_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_1_norm_feed_forward1_weight_to_fp16, x = input_77_cast_fp16)[name = string("input_79_cast_fp16")]; + tensor encoder_layers_1_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28978112))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33172480))))[name = string("encoder_layers_1_feed_forward1_linear1_weight_to_fp16_quantized")]; + tensor encoder_layers_1_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_1_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33180736)))]; + tensor linear_10_cast_fp16 = linear(bias = encoder_layers_1_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_1_feed_forward1_linear1_weight_to_fp16_quantized, x = input_79_cast_fp16)[name = string("linear_10_cast_fp16")]; + tensor input_83_cast_fp16 = silu(x = linear_10_cast_fp16)[name = string("input_83_cast_fp16")]; + tensor encoder_layers_1_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33188992))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37383360))))[name = string("encoder_layers_1_feed_forward1_linear2_weight_to_fp16_quantized")]; + tensor encoder_layers_1_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_1_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37385472)))]; + tensor linear_11_cast_fp16 = linear(bias = encoder_layers_1_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_1_feed_forward1_linear2_weight_to_fp16_quantized, x = input_83_cast_fp16)[name = string("linear_11_cast_fp16")]; + fp16 var_668_to_fp16 = const()[name = string("op_668_to_fp16"), val = fp16(0x1p-1)]; + tensor var_669_cast_fp16 = mul(x = linear_11_cast_fp16, y = var_668_to_fp16)[name = string("op_669_cast_fp16")]; + tensor input_89_cast_fp16 = add(x = input_77_cast_fp16, y = var_669_cast_fp16)[name = string("input_89_cast_fp16")]; + tensor key_3_axes_0 = const()[name = string("key_3_axes_0"), val = tensor([-1])]; + tensor encoder_layers_1_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_1_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37387584)))]; + tensor encoder_layers_1_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_1_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37389696)))]; + tensor key_3_cast_fp16 = layer_norm(axes = key_3_axes_0, beta = encoder_layers_1_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_1_norm_self_att_weight_to_fp16, x = input_89_cast_fp16)[name = string("key_3_cast_fp16")]; + bool input_91_interleave_0 = const()[name = string("input_91_interleave_0"), val = bool(false)]; + tensor input_91_cast_fp16 = concat(axis = var_68, interleave = input_91_interleave_0, values = (cache_5_cast_fp16, key_3_cast_fp16))[name = string("input_91_cast_fp16")]; + tensor var_691_begin_0 = const()[name = string("op_691_begin_0"), val = tensor([0, 28, 0])]; + tensor var_691_end_0 = const()[name = string("op_691_end_0"), val = tensor([1, 42, 1024])]; + tensor var_691_end_mask_0 = const()[name = string("op_691_end_mask_0"), val = tensor([true, true, true])]; + tensor var_691_cast_fp16 = slice_by_index(begin = var_691_begin_0, end = var_691_end_0, end_mask = var_691_end_mask_0, x = cache_5_cast_fp16)[name = string("op_691_cast_fp16")]; + bool var_697_interleave_0 = const()[name = string("op_697_interleave_0"), val = bool(false)]; + tensor var_697_cast_fp16 = concat(axis = var_68, interleave = var_697_interleave_0, values = (var_691_cast_fp16, key_3_cast_fp16))[name = string("op_697_cast_fp16")]; + tensor encoder_layers_1_self_attn_linear_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37391808))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38440448))))[name = string("encoder_layers_1_self_attn_linear_q_weight_to_fp16_quantized")]; + tensor encoder_layers_1_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_1_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38442560)))]; + tensor linear_12_cast_fp16 = linear(bias = encoder_layers_1_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_1_self_attn_linear_q_weight_to_fp16_quantized, x = key_3_cast_fp16)[name = string("linear_12_cast_fp16")]; + tensor var_702 = const()[name = string("op_702"), val = tensor([1, -1, 8, 128])]; + tensor q_7_cast_fp16 = reshape(shape = var_702, x = linear_12_cast_fp16)[name = string("q_7_cast_fp16")]; + tensor encoder_layers_1_self_attn_linear_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38444672))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39493312))))[name = string("encoder_layers_1_self_attn_linear_k_weight_to_fp16_quantized")]; + tensor encoder_layers_1_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_1_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39495424)))]; + tensor linear_13_cast_fp16 = linear(bias = encoder_layers_1_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_1_self_attn_linear_k_weight_to_fp16_quantized, x = input_91_cast_fp16)[name = string("linear_13_cast_fp16")]; + tensor var_707 = const()[name = string("op_707"), val = tensor([1, -1, 8, 128])]; + tensor k_5_cast_fp16 = reshape(shape = var_707, x = linear_13_cast_fp16)[name = string("k_5_cast_fp16")]; + tensor encoder_layers_1_self_attn_linear_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39497536))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40546176))))[name = string("encoder_layers_1_self_attn_linear_v_weight_to_fp16_quantized")]; + tensor encoder_layers_1_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_1_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40548288)))]; + tensor linear_14_cast_fp16 = linear(bias = encoder_layers_1_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_1_self_attn_linear_v_weight_to_fp16_quantized, x = input_91_cast_fp16)[name = string("linear_14_cast_fp16")]; + tensor var_712 = const()[name = string("op_712"), val = tensor([1, -1, 8, 128])]; + tensor v_3_cast_fp16 = reshape(shape = var_712, x = linear_14_cast_fp16)[name = string("v_3_cast_fp16")]; + tensor value_11_perm_0 = const()[name = string("value_11_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_1_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_1_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40550400)))]; + tensor var_725_cast_fp16 = add(x = q_7_cast_fp16, y = encoder_layers_1_self_attn_pos_bias_u_to_fp16)[name = string("op_725_cast_fp16")]; + tensor encoder_layers_1_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_1_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40552512)))]; + tensor var_727_cast_fp16 = add(x = q_7_cast_fp16, y = encoder_layers_1_self_attn_pos_bias_v_to_fp16)[name = string("op_727_cast_fp16")]; + tensor q_with_bias_v_3_perm_0 = const()[name = string("q_with_bias_v_3_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_33_transpose_x_0 = const()[name = string("x_33_transpose_x_0"), val = bool(false)]; + bool x_33_transpose_y_0 = const()[name = string("x_33_transpose_y_0"), val = bool(false)]; + tensor op_729_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40554624))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40697024))))[name = string("op_729_to_fp16_quantized")]; + tensor q_with_bias_v_3_cast_fp16 = transpose(perm = q_with_bias_v_3_perm_0, x = var_727_cast_fp16)[name = string("transpose_353")]; + tensor x_33_cast_fp16 = matmul(transpose_x = x_33_transpose_x_0, transpose_y = x_33_transpose_y_0, x = q_with_bias_v_3_cast_fp16, y = op_729_to_fp16_quantized)[name = string("x_33_cast_fp16")]; + tensor x_35_pad_0 = const()[name = string("x_35_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_35_mode_0 = const()[name = string("x_35_mode_0"), val = string("constant")]; + fp16 const_92_to_fp16 = const()[name = string("const_92_to_fp16"), val = fp16(0x0p+0)]; + tensor x_35_cast_fp16 = pad(constant_val = const_92_to_fp16, mode = x_35_mode_0, pad = x_35_pad_0, x = x_33_cast_fp16)[name = string("x_35_cast_fp16")]; + tensor var_737 = const()[name = string("op_737"), val = tensor([1, 8, -1, 28])]; + tensor x_37_cast_fp16 = reshape(shape = var_737, x = x_35_cast_fp16)[name = string("x_37_cast_fp16")]; + tensor var_741_begin_0 = const()[name = string("op_741_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_741_end_0 = const()[name = string("op_741_end_0"), val = tensor([1, 8, 140, 28])]; + tensor var_741_end_mask_0 = const()[name = string("op_741_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_741_cast_fp16 = slice_by_index(begin = var_741_begin_0, end = var_741_end_0, end_mask = var_741_end_mask_0, x = x_37_cast_fp16)[name = string("op_741_cast_fp16")]; + tensor var_742 = const()[name = string("op_742"), val = tensor([1, 8, 28, 139])]; + tensor matrix_bd_5_cast_fp16 = reshape(shape = var_742, x = var_741_cast_fp16)[name = string("matrix_bd_5_cast_fp16")]; + bool matrix_ac_3_transpose_x_0 = const()[name = string("matrix_ac_3_transpose_x_0"), val = bool(false)]; + bool matrix_ac_3_transpose_y_0 = const()[name = string("matrix_ac_3_transpose_y_0"), val = bool(false)]; + tensor transpose_98_perm_0 = const()[name = string("transpose_98_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_99_perm_0 = const()[name = string("transpose_99_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_99 = transpose(perm = transpose_99_perm_0, x = k_5_cast_fp16)[name = string("transpose_351")]; + tensor transpose_98 = transpose(perm = transpose_98_perm_0, x = var_725_cast_fp16)[name = string("transpose_352")]; + tensor matrix_ac_3_cast_fp16 = matmul(transpose_x = matrix_ac_3_transpose_x_0, transpose_y = matrix_ac_3_transpose_y_0, x = transpose_98, y = transpose_99)[name = string("matrix_ac_3_cast_fp16")]; + tensor matrix_bd_7_begin_0 = const()[name = string("matrix_bd_7_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_7_end_0 = const()[name = string("matrix_bd_7_end_0"), val = tensor([1, 8, 28, 70])]; + tensor matrix_bd_7_end_mask_0 = const()[name = string("matrix_bd_7_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_7_cast_fp16 = slice_by_index(begin = matrix_bd_7_begin_0, end = matrix_bd_7_end_0, end_mask = matrix_bd_7_end_mask_0, x = matrix_bd_5_cast_fp16)[name = string("matrix_bd_7_cast_fp16")]; + tensor var_751_cast_fp16 = add(x = matrix_ac_3_cast_fp16, y = matrix_bd_7_cast_fp16)[name = string("op_751_cast_fp16")]; + fp16 _inversed_scores_5_y_0_to_fp16 = const()[name = string("_inversed_scores_5_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_5_cast_fp16 = mul(x = var_751_cast_fp16, y = _inversed_scores_5_y_0_to_fp16)[name = string("_inversed_scores_5_cast_fp16")]; + tensor scores_7_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_5_cast_fp16, cond = mask_11)[name = string("scores_7_cast_fp16")]; + tensor var_757_cast_fp16 = softmax(axis = var_59, x = scores_7_cast_fp16)[name = string("op_757_cast_fp16")]; + tensor input_93_cast_fp16 = select(a = var_44_to_fp16, b = var_757_cast_fp16, cond = mask_11)[name = string("input_93_cast_fp16")]; + bool x_39_transpose_x_0 = const()[name = string("x_39_transpose_x_0"), val = bool(false)]; + bool x_39_transpose_y_0 = const()[name = string("x_39_transpose_y_0"), val = bool(false)]; + tensor value_11_cast_fp16 = transpose(perm = value_11_perm_0, x = v_3_cast_fp16)[name = string("transpose_350")]; + tensor x_39_cast_fp16 = matmul(transpose_x = x_39_transpose_x_0, transpose_y = x_39_transpose_y_0, x = input_93_cast_fp16, y = value_11_cast_fp16)[name = string("x_39_cast_fp16")]; + tensor var_761_perm_0 = const()[name = string("op_761_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_762 = const()[name = string("op_762"), val = tensor([1, -1, 1024])]; + tensor var_761_cast_fp16 = transpose(perm = var_761_perm_0, x = x_39_cast_fp16)[name = string("transpose_349")]; + tensor input_95_cast_fp16 = reshape(shape = var_762, x = var_761_cast_fp16)[name = string("input_95_cast_fp16")]; + tensor encoder_layers_1_self_attn_linear_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40697408))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41746048))))[name = string("encoder_layers_1_self_attn_linear_out_weight_to_fp16_quantized")]; + tensor encoder_layers_1_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_1_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41748160)))]; + tensor linear_16_cast_fp16 = linear(bias = encoder_layers_1_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_1_self_attn_linear_out_weight_to_fp16_quantized, x = input_95_cast_fp16)[name = string("linear_16_cast_fp16")]; + tensor input_99_cast_fp16 = add(x = input_89_cast_fp16, y = linear_16_cast_fp16)[name = string("input_99_cast_fp16")]; + tensor x_43_axes_0 = const()[name = string("x_43_axes_0"), val = tensor([-1])]; + tensor encoder_layers_1_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_1_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41750272)))]; + tensor encoder_layers_1_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_1_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41752384)))]; + tensor x_43_cast_fp16 = layer_norm(axes = x_43_axes_0, beta = encoder_layers_1_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_1_norm_conv_weight_to_fp16, x = input_99_cast_fp16)[name = string("x_43_cast_fp16")]; + tensor input_101_perm_0 = const()[name = string("input_101_perm_0"), val = tensor([0, 2, 1])]; + string input_103_pad_type_0 = const()[name = string("input_103_pad_type_0"), val = string("valid")]; + tensor input_103_strides_0 = const()[name = string("input_103_strides_0"), val = tensor([1])]; + tensor input_103_pad_0 = const()[name = string("input_103_pad_0"), val = tensor([0, 0])]; + tensor input_103_dilations_0 = const()[name = string("input_103_dilations_0"), val = tensor([1])]; + int32 input_103_groups_0 = const()[name = string("input_103_groups_0"), val = int32(1)]; + tensor encoder_layers_1_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41754496))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43851712))))[name = string("encoder_layers_1_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_101_cast_fp16 = transpose(perm = input_101_perm_0, x = x_43_cast_fp16)[name = string("transpose_348")]; + tensor input_103_cast_fp16 = conv(dilations = input_103_dilations_0, groups = input_103_groups_0, pad = input_103_pad_0, pad_type = input_103_pad_type_0, strides = input_103_strides_0, weight = encoder_layers_1_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_101_cast_fp16)[name = string("input_103_cast_fp16")]; + int32 x_45_split_num_splits_0 = const()[name = string("x_45_split_num_splits_0"), val = int32(2)]; + int32 x_45_split_axis_0 = const()[name = string("x_45_split_axis_0"), val = int32(1)]; + tensor x_45_split_cast_fp16_0, tensor x_45_split_cast_fp16_1 = split(axis = x_45_split_axis_0, num_splits = x_45_split_num_splits_0, x = input_103_cast_fp16)[name = string("x_45_split_cast_fp16")]; + tensor x_45_split_1_sigmoid_cast_fp16 = sigmoid(x = x_45_split_cast_fp16_1)[name = string("x_45_split_1_sigmoid_cast_fp16")]; + tensor x_45_cast_fp16 = mul(x = x_45_split_cast_fp16_0, y = x_45_split_1_sigmoid_cast_fp16)[name = string("x_45_cast_fp16")]; + tensor input_105_cast_fp16 = select(a = var_44_to_fp16, b = x_45_cast_fp16, cond = var_575)[name = string("input_105_cast_fp16")]; + bool new_x_7_interleave_0 = const()[name = string("new_x_7_interleave_0"), val = bool(false)]; + tensor new_x_7_cast_fp16 = concat(axis = var_59, interleave = new_x_7_interleave_0, values = (cache_7_cast_fp16, input_105_cast_fp16))[name = string("new_x_7_cast_fp16")]; + tensor var_801_begin_0 = const()[name = string("op_801_begin_0"), val = tensor([0, 0, 28])]; + tensor var_801_end_0 = const()[name = string("op_801_end_0"), val = tensor([1, 1024, 36])]; + tensor var_801_end_mask_0 = const()[name = string("op_801_end_mask_0"), val = tensor([true, true, true])]; + tensor var_801_cast_fp16 = slice_by_index(begin = var_801_begin_0, end = var_801_end_0, end_mask = var_801_end_mask_0, x = new_x_7_cast_fp16)[name = string("op_801_cast_fp16")]; + string x_47_pad_type_0 = const()[name = string("x_47_pad_type_0"), val = string("valid")]; + int32 x_47_groups_0 = const()[name = string("x_47_groups_0"), val = int32(1024)]; + tensor x_47_strides_0 = const()[name = string("x_47_strides_0"), val = tensor([1])]; + tensor x_47_pad_0 = const()[name = string("x_47_pad_0"), val = tensor([0, 0])]; + tensor x_47_dilations_0 = const()[name = string("x_47_dilations_0"), val = tensor([1])]; + tensor encoder_layers_1_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43855872))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43865152))))[name = string("encoder_layers_1_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_47_cast_fp16 = conv(dilations = x_47_dilations_0, groups = x_47_groups_0, pad = x_47_pad_0, pad_type = x_47_pad_type_0, strides = x_47_strides_0, weight = encoder_layers_1_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_7_cast_fp16)[name = string("x_47_cast_fp16")]; + tensor input_107_perm_0 = const()[name = string("input_107_perm_0"), val = tensor([0, 2, 1])]; + tensor x_49_axes_0 = const()[name = string("x_49_axes_0"), val = tensor([-1])]; + tensor encoder_layers_1_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_1_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43867264)))]; + tensor encoder_layers_1_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_1_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43869376)))]; + tensor input_107_cast_fp16 = transpose(perm = input_107_perm_0, x = x_47_cast_fp16)[name = string("transpose_347")]; + tensor x_49_cast_fp16 = layer_norm(axes = x_49_axes_0, beta = encoder_layers_1_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_1_conv_batch_norm_weight_to_fp16, x = input_107_cast_fp16)[name = string("x_49_cast_fp16")]; + tensor input_109_perm_0 = const()[name = string("input_109_perm_0"), val = tensor([0, 2, 1])]; + tensor input_109_cast_fp16 = transpose(perm = input_109_perm_0, x = x_49_cast_fp16)[name = string("transpose_346")]; + tensor input_111_cast_fp16 = silu(x = input_109_cast_fp16)[name = string("input_111_cast_fp16")]; + string x_51_pad_type_0 = const()[name = string("x_51_pad_type_0"), val = string("valid")]; + tensor x_51_strides_0 = const()[name = string("x_51_strides_0"), val = tensor([1])]; + tensor x_51_pad_0 = const()[name = string("x_51_pad_0"), val = tensor([0, 0])]; + tensor x_51_dilations_0 = const()[name = string("x_51_dilations_0"), val = tensor([1])]; + int32 x_51_groups_0 = const()[name = string("x_51_groups_0"), val = int32(1)]; + tensor encoder_layers_1_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43871488))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44920128))))[name = string("encoder_layers_1_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_51_cast_fp16 = conv(dilations = x_51_dilations_0, groups = x_51_groups_0, pad = x_51_pad_0, pad_type = x_51_pad_type_0, strides = x_51_strides_0, weight = encoder_layers_1_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_111_cast_fp16)[name = string("x_51_cast_fp16")]; + tensor input_113_perm_0 = const()[name = string("input_113_perm_0"), val = tensor([0, 2, 1])]; + tensor input_113_cast_fp16 = transpose(perm = input_113_perm_0, x = x_51_cast_fp16)[name = string("transpose_345")]; + tensor input_115_cast_fp16 = add(x = input_99_cast_fp16, y = input_113_cast_fp16)[name = string("input_115_cast_fp16")]; + tensor input_117_axes_0 = const()[name = string("input_117_axes_0"), val = tensor([-1])]; + tensor encoder_layers_1_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_1_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44922240)))]; + tensor encoder_layers_1_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_1_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44924352)))]; + tensor input_117_cast_fp16 = layer_norm(axes = input_117_axes_0, beta = encoder_layers_1_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_1_norm_feed_forward2_weight_to_fp16, x = input_115_cast_fp16)[name = string("input_117_cast_fp16")]; + tensor encoder_layers_1_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44926464))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49120832))))[name = string("encoder_layers_1_feed_forward2_linear1_weight_to_fp16_quantized")]; + tensor encoder_layers_1_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_1_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49129088)))]; + tensor linear_17_cast_fp16 = linear(bias = encoder_layers_1_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_1_feed_forward2_linear1_weight_to_fp16_quantized, x = input_117_cast_fp16)[name = string("linear_17_cast_fp16")]; + tensor input_121_cast_fp16 = silu(x = linear_17_cast_fp16)[name = string("input_121_cast_fp16")]; + tensor encoder_layers_1_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49137344))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53331712))))[name = string("encoder_layers_1_feed_forward2_linear2_weight_to_fp16_quantized")]; + tensor encoder_layers_1_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_1_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53333824)))]; + tensor linear_18_cast_fp16 = linear(bias = encoder_layers_1_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_1_feed_forward2_linear2_weight_to_fp16_quantized, x = input_121_cast_fp16)[name = string("linear_18_cast_fp16")]; + fp16 var_844_to_fp16 = const()[name = string("op_844_to_fp16"), val = fp16(0x1p-1)]; + tensor var_845_cast_fp16 = mul(x = linear_18_cast_fp16, y = var_844_to_fp16)[name = string("op_845_cast_fp16")]; + tensor input_127_cast_fp16 = add(x = input_115_cast_fp16, y = var_845_cast_fp16)[name = string("input_127_cast_fp16")]; + tensor input_129_axes_0 = const()[name = string("input_129_axes_0"), val = tensor([-1])]; + tensor encoder_layers_1_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_1_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53335936)))]; + tensor encoder_layers_1_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_1_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53338048)))]; + tensor input_129_cast_fp16 = layer_norm(axes = input_129_axes_0, beta = encoder_layers_1_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_1_norm_out_weight_to_fp16, x = input_127_cast_fp16)[name = string("input_129_cast_fp16")]; + tensor cache_9_begin_0 = const()[name = string("cache_9_begin_0"), val = tensor([2, 0, 0, 0])]; + tensor cache_9_end_0 = const()[name = string("cache_9_end_0"), val = tensor([3, 1, 42, 1024])]; + tensor cache_9_end_mask_0 = const()[name = string("cache_9_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_9_squeeze_mask_0 = const()[name = string("cache_9_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_9_cast_fp16 = slice_by_index(begin = cache_9_begin_0, end = cache_9_end_0, end_mask = cache_9_end_mask_0, squeeze_mask = cache_9_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_9_cast_fp16")]; + tensor cache_11_begin_0 = const()[name = string("cache_11_begin_0"), val = tensor([2, 0, 0, 0])]; + tensor cache_11_end_0 = const()[name = string("cache_11_end_0"), val = tensor([3, 1, 1024, 8])]; + tensor cache_11_end_mask_0 = const()[name = string("cache_11_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_11_squeeze_mask_0 = const()[name = string("cache_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_11_cast_fp16 = slice_by_index(begin = cache_11_begin_0, end = cache_11_end_0, end_mask = cache_11_end_mask_0, squeeze_mask = cache_11_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_11_cast_fp16")]; + tensor input_131_axes_0 = const()[name = string("input_131_axes_0"), val = tensor([-1])]; + tensor encoder_layers_2_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_2_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53340160)))]; + tensor encoder_layers_2_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_2_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53342272)))]; + tensor input_131_cast_fp16 = layer_norm(axes = input_131_axes_0, beta = encoder_layers_2_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_2_norm_feed_forward1_weight_to_fp16, x = input_129_cast_fp16)[name = string("input_131_cast_fp16")]; + tensor encoder_layers_2_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53344384))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57538752))))[name = string("encoder_layers_2_feed_forward1_linear1_weight_to_fp16_quantized")]; + tensor encoder_layers_2_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_2_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57547008)))]; + tensor linear_19_cast_fp16 = linear(bias = encoder_layers_2_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_2_feed_forward1_linear1_weight_to_fp16_quantized, x = input_131_cast_fp16)[name = string("linear_19_cast_fp16")]; + tensor input_135_cast_fp16 = silu(x = linear_19_cast_fp16)[name = string("input_135_cast_fp16")]; + tensor encoder_layers_2_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57555264))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61749632))))[name = string("encoder_layers_2_feed_forward1_linear2_weight_to_fp16_quantized")]; + tensor encoder_layers_2_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_2_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61751744)))]; + tensor linear_20_cast_fp16 = linear(bias = encoder_layers_2_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_2_feed_forward1_linear2_weight_to_fp16_quantized, x = input_135_cast_fp16)[name = string("linear_20_cast_fp16")]; + fp16 var_881_to_fp16 = const()[name = string("op_881_to_fp16"), val = fp16(0x1p-1)]; + tensor var_882_cast_fp16 = mul(x = linear_20_cast_fp16, y = var_881_to_fp16)[name = string("op_882_cast_fp16")]; + tensor input_141_cast_fp16 = add(x = input_129_cast_fp16, y = var_882_cast_fp16)[name = string("input_141_cast_fp16")]; + tensor key_5_axes_0 = const()[name = string("key_5_axes_0"), val = tensor([-1])]; + tensor encoder_layers_2_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_2_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61753856)))]; + tensor encoder_layers_2_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_2_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61755968)))]; + tensor key_5_cast_fp16 = layer_norm(axes = key_5_axes_0, beta = encoder_layers_2_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_2_norm_self_att_weight_to_fp16, x = input_141_cast_fp16)[name = string("key_5_cast_fp16")]; + bool input_143_interleave_0 = const()[name = string("input_143_interleave_0"), val = bool(false)]; + tensor input_143_cast_fp16 = concat(axis = var_68, interleave = input_143_interleave_0, values = (cache_9_cast_fp16, key_5_cast_fp16))[name = string("input_143_cast_fp16")]; + tensor var_904_begin_0 = const()[name = string("op_904_begin_0"), val = tensor([0, 28, 0])]; + tensor var_904_end_0 = const()[name = string("op_904_end_0"), val = tensor([1, 42, 1024])]; + tensor var_904_end_mask_0 = const()[name = string("op_904_end_mask_0"), val = tensor([true, true, true])]; + tensor var_904_cast_fp16 = slice_by_index(begin = var_904_begin_0, end = var_904_end_0, end_mask = var_904_end_mask_0, x = cache_9_cast_fp16)[name = string("op_904_cast_fp16")]; + bool var_910_interleave_0 = const()[name = string("op_910_interleave_0"), val = bool(false)]; + tensor var_910_cast_fp16 = concat(axis = var_68, interleave = var_910_interleave_0, values = (var_904_cast_fp16, key_5_cast_fp16))[name = string("op_910_cast_fp16")]; + tensor encoder_layers_2_self_attn_linear_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61758080))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62806720))))[name = string("encoder_layers_2_self_attn_linear_q_weight_to_fp16_quantized")]; + tensor encoder_layers_2_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_2_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62808832)))]; + tensor linear_21_cast_fp16 = linear(bias = encoder_layers_2_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_2_self_attn_linear_q_weight_to_fp16_quantized, x = key_5_cast_fp16)[name = string("linear_21_cast_fp16")]; + tensor var_915 = const()[name = string("op_915"), val = tensor([1, -1, 8, 128])]; + tensor q_13_cast_fp16 = reshape(shape = var_915, x = linear_21_cast_fp16)[name = string("q_13_cast_fp16")]; + tensor encoder_layers_2_self_attn_linear_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62810944))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(63859584))))[name = string("encoder_layers_2_self_attn_linear_k_weight_to_fp16_quantized")]; + tensor encoder_layers_2_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_2_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(63861696)))]; + tensor linear_22_cast_fp16 = linear(bias = encoder_layers_2_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_2_self_attn_linear_k_weight_to_fp16_quantized, x = input_143_cast_fp16)[name = string("linear_22_cast_fp16")]; + tensor var_920 = const()[name = string("op_920"), val = tensor([1, -1, 8, 128])]; + tensor k_9_cast_fp16 = reshape(shape = var_920, x = linear_22_cast_fp16)[name = string("k_9_cast_fp16")]; + tensor encoder_layers_2_self_attn_linear_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(63863808))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64912448))))[name = string("encoder_layers_2_self_attn_linear_v_weight_to_fp16_quantized")]; + tensor encoder_layers_2_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_2_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64914560)))]; + tensor linear_23_cast_fp16 = linear(bias = encoder_layers_2_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_2_self_attn_linear_v_weight_to_fp16_quantized, x = input_143_cast_fp16)[name = string("linear_23_cast_fp16")]; + tensor var_925 = const()[name = string("op_925"), val = tensor([1, -1, 8, 128])]; + tensor v_5_cast_fp16 = reshape(shape = var_925, x = linear_23_cast_fp16)[name = string("v_5_cast_fp16")]; + tensor value_13_perm_0 = const()[name = string("value_13_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_2_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_2_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64916672)))]; + tensor var_938_cast_fp16 = add(x = q_13_cast_fp16, y = encoder_layers_2_self_attn_pos_bias_u_to_fp16)[name = string("op_938_cast_fp16")]; + tensor encoder_layers_2_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_2_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64918784)))]; + tensor var_940_cast_fp16 = add(x = q_13_cast_fp16, y = encoder_layers_2_self_attn_pos_bias_v_to_fp16)[name = string("op_940_cast_fp16")]; + tensor q_with_bias_v_5_perm_0 = const()[name = string("q_with_bias_v_5_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_59_transpose_x_0 = const()[name = string("x_59_transpose_x_0"), val = bool(false)]; + bool x_59_transpose_y_0 = const()[name = string("x_59_transpose_y_0"), val = bool(false)]; + tensor op_942_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64920896))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65063296))))[name = string("op_942_to_fp16_quantized")]; + tensor q_with_bias_v_5_cast_fp16 = transpose(perm = q_with_bias_v_5_perm_0, x = var_940_cast_fp16)[name = string("transpose_344")]; + tensor x_59_cast_fp16 = matmul(transpose_x = x_59_transpose_x_0, transpose_y = x_59_transpose_y_0, x = q_with_bias_v_5_cast_fp16, y = op_942_to_fp16_quantized)[name = string("x_59_cast_fp16")]; + tensor x_61_pad_0 = const()[name = string("x_61_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_61_mode_0 = const()[name = string("x_61_mode_0"), val = string("constant")]; + fp16 const_105_to_fp16 = const()[name = string("const_105_to_fp16"), val = fp16(0x0p+0)]; + tensor x_61_cast_fp16 = pad(constant_val = const_105_to_fp16, mode = x_61_mode_0, pad = x_61_pad_0, x = x_59_cast_fp16)[name = string("x_61_cast_fp16")]; + tensor var_950 = const()[name = string("op_950"), val = tensor([1, 8, -1, 28])]; + tensor x_63_cast_fp16 = reshape(shape = var_950, x = x_61_cast_fp16)[name = string("x_63_cast_fp16")]; + tensor var_954_begin_0 = const()[name = string("op_954_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_954_end_0 = const()[name = string("op_954_end_0"), val = tensor([1, 8, 140, 28])]; + tensor var_954_end_mask_0 = const()[name = string("op_954_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_954_cast_fp16 = slice_by_index(begin = var_954_begin_0, end = var_954_end_0, end_mask = var_954_end_mask_0, x = x_63_cast_fp16)[name = string("op_954_cast_fp16")]; + tensor var_955 = const()[name = string("op_955"), val = tensor([1, 8, 28, 139])]; + tensor matrix_bd_9_cast_fp16 = reshape(shape = var_955, x = var_954_cast_fp16)[name = string("matrix_bd_9_cast_fp16")]; + bool matrix_ac_5_transpose_x_0 = const()[name = string("matrix_ac_5_transpose_x_0"), val = bool(false)]; + bool matrix_ac_5_transpose_y_0 = const()[name = string("matrix_ac_5_transpose_y_0"), val = bool(false)]; + tensor transpose_100_perm_0 = const()[name = string("transpose_100_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_101_perm_0 = const()[name = string("transpose_101_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_101 = transpose(perm = transpose_101_perm_0, x = k_9_cast_fp16)[name = string("transpose_342")]; + tensor transpose_100 = transpose(perm = transpose_100_perm_0, x = var_938_cast_fp16)[name = string("transpose_343")]; + tensor matrix_ac_5_cast_fp16 = matmul(transpose_x = matrix_ac_5_transpose_x_0, transpose_y = matrix_ac_5_transpose_y_0, x = transpose_100, y = transpose_101)[name = string("matrix_ac_5_cast_fp16")]; + tensor matrix_bd_11_begin_0 = const()[name = string("matrix_bd_11_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_11_end_0 = const()[name = string("matrix_bd_11_end_0"), val = tensor([1, 8, 28, 70])]; + tensor matrix_bd_11_end_mask_0 = const()[name = string("matrix_bd_11_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_11_cast_fp16 = slice_by_index(begin = matrix_bd_11_begin_0, end = matrix_bd_11_end_0, end_mask = matrix_bd_11_end_mask_0, x = matrix_bd_9_cast_fp16)[name = string("matrix_bd_11_cast_fp16")]; + tensor var_964_cast_fp16 = add(x = matrix_ac_5_cast_fp16, y = matrix_bd_11_cast_fp16)[name = string("op_964_cast_fp16")]; + fp16 _inversed_scores_9_y_0_to_fp16 = const()[name = string("_inversed_scores_9_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_9_cast_fp16 = mul(x = var_964_cast_fp16, y = _inversed_scores_9_y_0_to_fp16)[name = string("_inversed_scores_9_cast_fp16")]; + tensor scores_11_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_9_cast_fp16, cond = mask_11)[name = string("scores_11_cast_fp16")]; + tensor var_970_cast_fp16 = softmax(axis = var_59, x = scores_11_cast_fp16)[name = string("op_970_cast_fp16")]; + tensor input_145_cast_fp16 = select(a = var_44_to_fp16, b = var_970_cast_fp16, cond = mask_11)[name = string("input_145_cast_fp16")]; + bool x_65_transpose_x_0 = const()[name = string("x_65_transpose_x_0"), val = bool(false)]; + bool x_65_transpose_y_0 = const()[name = string("x_65_transpose_y_0"), val = bool(false)]; + tensor value_13_cast_fp16 = transpose(perm = value_13_perm_0, x = v_5_cast_fp16)[name = string("transpose_341")]; + tensor x_65_cast_fp16 = matmul(transpose_x = x_65_transpose_x_0, transpose_y = x_65_transpose_y_0, x = input_145_cast_fp16, y = value_13_cast_fp16)[name = string("x_65_cast_fp16")]; + tensor var_974_perm_0 = const()[name = string("op_974_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_975 = const()[name = string("op_975"), val = tensor([1, -1, 1024])]; + tensor var_974_cast_fp16 = transpose(perm = var_974_perm_0, x = x_65_cast_fp16)[name = string("transpose_340")]; + tensor input_147_cast_fp16 = reshape(shape = var_975, x = var_974_cast_fp16)[name = string("input_147_cast_fp16")]; + tensor encoder_layers_2_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65063680))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65850176))))[name = string("encoder_layers_2_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_2_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_2_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65850368)))]; + tensor linear_25_cast_fp16 = linear(bias = encoder_layers_2_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_2_self_attn_linear_out_weight_to_fp16_palettized, x = input_147_cast_fp16)[name = string("linear_25_cast_fp16")]; + tensor input_151_cast_fp16 = add(x = input_141_cast_fp16, y = linear_25_cast_fp16)[name = string("input_151_cast_fp16")]; + tensor x_69_axes_0 = const()[name = string("x_69_axes_0"), val = tensor([-1])]; + tensor encoder_layers_2_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_2_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65852480)))]; + tensor encoder_layers_2_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_2_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65854592)))]; + tensor x_69_cast_fp16 = layer_norm(axes = x_69_axes_0, beta = encoder_layers_2_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_2_norm_conv_weight_to_fp16, x = input_151_cast_fp16)[name = string("x_69_cast_fp16")]; + tensor input_153_perm_0 = const()[name = string("input_153_perm_0"), val = tensor([0, 2, 1])]; + string input_155_pad_type_0 = const()[name = string("input_155_pad_type_0"), val = string("valid")]; + tensor input_155_strides_0 = const()[name = string("input_155_strides_0"), val = tensor([1])]; + tensor input_155_pad_0 = const()[name = string("input_155_pad_0"), val = tensor([0, 0])]; + tensor input_155_dilations_0 = const()[name = string("input_155_dilations_0"), val = tensor([1])]; + int32 input_155_groups_0 = const()[name = string("input_155_groups_0"), val = int32(1)]; + tensor encoder_layers_2_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65856704))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67953920))))[name = string("encoder_layers_2_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_153_cast_fp16 = transpose(perm = input_153_perm_0, x = x_69_cast_fp16)[name = string("transpose_339")]; + tensor input_155_cast_fp16 = conv(dilations = input_155_dilations_0, groups = input_155_groups_0, pad = input_155_pad_0, pad_type = input_155_pad_type_0, strides = input_155_strides_0, weight = encoder_layers_2_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_153_cast_fp16)[name = string("input_155_cast_fp16")]; + int32 x_71_split_num_splits_0 = const()[name = string("x_71_split_num_splits_0"), val = int32(2)]; + int32 x_71_split_axis_0 = const()[name = string("x_71_split_axis_0"), val = int32(1)]; + tensor x_71_split_cast_fp16_0, tensor x_71_split_cast_fp16_1 = split(axis = x_71_split_axis_0, num_splits = x_71_split_num_splits_0, x = input_155_cast_fp16)[name = string("x_71_split_cast_fp16")]; + tensor x_71_split_1_sigmoid_cast_fp16 = sigmoid(x = x_71_split_cast_fp16_1)[name = string("x_71_split_1_sigmoid_cast_fp16")]; + tensor x_71_cast_fp16 = mul(x = x_71_split_cast_fp16_0, y = x_71_split_1_sigmoid_cast_fp16)[name = string("x_71_cast_fp16")]; + tensor input_157_cast_fp16 = select(a = var_44_to_fp16, b = x_71_cast_fp16, cond = var_575)[name = string("input_157_cast_fp16")]; + bool new_x_11_interleave_0 = const()[name = string("new_x_11_interleave_0"), val = bool(false)]; + tensor new_x_11_cast_fp16 = concat(axis = var_59, interleave = new_x_11_interleave_0, values = (cache_11_cast_fp16, input_157_cast_fp16))[name = string("new_x_11_cast_fp16")]; + tensor var_1014_begin_0 = const()[name = string("op_1014_begin_0"), val = tensor([0, 0, 28])]; + tensor var_1014_end_0 = const()[name = string("op_1014_end_0"), val = tensor([1, 1024, 36])]; + tensor var_1014_end_mask_0 = const()[name = string("op_1014_end_mask_0"), val = tensor([true, true, true])]; + tensor var_1014_cast_fp16 = slice_by_index(begin = var_1014_begin_0, end = var_1014_end_0, end_mask = var_1014_end_mask_0, x = new_x_11_cast_fp16)[name = string("op_1014_cast_fp16")]; + string x_73_pad_type_0 = const()[name = string("x_73_pad_type_0"), val = string("valid")]; + int32 x_73_groups_0 = const()[name = string("x_73_groups_0"), val = int32(1024)]; + tensor x_73_strides_0 = const()[name = string("x_73_strides_0"), val = tensor([1])]; + tensor x_73_pad_0 = const()[name = string("x_73_pad_0"), val = tensor([0, 0])]; + tensor x_73_dilations_0 = const()[name = string("x_73_dilations_0"), val = tensor([1])]; + tensor encoder_layers_2_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67958080))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67967360))))[name = string("encoder_layers_2_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_73_cast_fp16 = conv(dilations = x_73_dilations_0, groups = x_73_groups_0, pad = x_73_pad_0, pad_type = x_73_pad_type_0, strides = x_73_strides_0, weight = encoder_layers_2_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_11_cast_fp16)[name = string("x_73_cast_fp16")]; + tensor input_159_perm_0 = const()[name = string("input_159_perm_0"), val = tensor([0, 2, 1])]; + tensor x_75_axes_0 = const()[name = string("x_75_axes_0"), val = tensor([-1])]; + tensor encoder_layers_2_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_2_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67969472)))]; + tensor encoder_layers_2_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_2_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67971584)))]; + tensor input_159_cast_fp16 = transpose(perm = input_159_perm_0, x = x_73_cast_fp16)[name = string("transpose_338")]; + tensor x_75_cast_fp16 = layer_norm(axes = x_75_axes_0, beta = encoder_layers_2_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_2_conv_batch_norm_weight_to_fp16, x = input_159_cast_fp16)[name = string("x_75_cast_fp16")]; + tensor input_161_perm_0 = const()[name = string("input_161_perm_0"), val = tensor([0, 2, 1])]; + tensor input_161_cast_fp16 = transpose(perm = input_161_perm_0, x = x_75_cast_fp16)[name = string("transpose_337")]; + tensor input_163_cast_fp16 = silu(x = input_161_cast_fp16)[name = string("input_163_cast_fp16")]; + string x_77_pad_type_0 = const()[name = string("x_77_pad_type_0"), val = string("valid")]; + tensor x_77_strides_0 = const()[name = string("x_77_strides_0"), val = tensor([1])]; + tensor x_77_pad_0 = const()[name = string("x_77_pad_0"), val = tensor([0, 0])]; + tensor x_77_dilations_0 = const()[name = string("x_77_dilations_0"), val = tensor([1])]; + int32 x_77_groups_0 = const()[name = string("x_77_groups_0"), val = int32(1)]; + tensor encoder_layers_2_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67973696))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69022336))))[name = string("encoder_layers_2_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_77_cast_fp16 = conv(dilations = x_77_dilations_0, groups = x_77_groups_0, pad = x_77_pad_0, pad_type = x_77_pad_type_0, strides = x_77_strides_0, weight = encoder_layers_2_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_163_cast_fp16)[name = string("x_77_cast_fp16")]; + tensor input_165_perm_0 = const()[name = string("input_165_perm_0"), val = tensor([0, 2, 1])]; + tensor input_165_cast_fp16 = transpose(perm = input_165_perm_0, x = x_77_cast_fp16)[name = string("transpose_336")]; + tensor input_167_cast_fp16 = add(x = input_151_cast_fp16, y = input_165_cast_fp16)[name = string("input_167_cast_fp16")]; + tensor input_169_axes_0 = const()[name = string("input_169_axes_0"), val = tensor([-1])]; + tensor encoder_layers_2_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_2_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69024448)))]; + tensor encoder_layers_2_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_2_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69026560)))]; + tensor input_169_cast_fp16 = layer_norm(axes = input_169_axes_0, beta = encoder_layers_2_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_2_norm_feed_forward2_weight_to_fp16, x = input_167_cast_fp16)[name = string("input_169_cast_fp16")]; + tensor encoder_layers_2_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69028672))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72174464))))[name = string("encoder_layers_2_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_2_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_2_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72174656)))]; + tensor linear_26_cast_fp16 = linear(bias = encoder_layers_2_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_2_feed_forward2_linear1_weight_to_fp16_palettized, x = input_169_cast_fp16)[name = string("linear_26_cast_fp16")]; + tensor input_173_cast_fp16 = silu(x = linear_26_cast_fp16)[name = string("input_173_cast_fp16")]; + tensor encoder_layers_2_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72182912))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75328704))))[name = string("encoder_layers_2_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_2_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_2_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75328896)))]; + tensor linear_27_cast_fp16 = linear(bias = encoder_layers_2_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_2_feed_forward2_linear2_weight_to_fp16_palettized, x = input_173_cast_fp16)[name = string("linear_27_cast_fp16")]; + fp16 var_1057_to_fp16 = const()[name = string("op_1057_to_fp16"), val = fp16(0x1p-1)]; + tensor var_1058_cast_fp16 = mul(x = linear_27_cast_fp16, y = var_1057_to_fp16)[name = string("op_1058_cast_fp16")]; + tensor input_179_cast_fp16 = add(x = input_167_cast_fp16, y = var_1058_cast_fp16)[name = string("input_179_cast_fp16")]; + tensor input_181_axes_0 = const()[name = string("input_181_axes_0"), val = tensor([-1])]; + tensor encoder_layers_2_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_2_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75331008)))]; + tensor encoder_layers_2_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_2_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75333120)))]; + tensor input_181_cast_fp16 = layer_norm(axes = input_181_axes_0, beta = encoder_layers_2_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_2_norm_out_weight_to_fp16, x = input_179_cast_fp16)[name = string("input_181_cast_fp16")]; + tensor cache_13_begin_0 = const()[name = string("cache_13_begin_0"), val = tensor([3, 0, 0, 0])]; + tensor cache_13_end_0 = const()[name = string("cache_13_end_0"), val = tensor([4, 1, 42, 1024])]; + tensor cache_13_end_mask_0 = const()[name = string("cache_13_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_13_squeeze_mask_0 = const()[name = string("cache_13_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_13_cast_fp16 = slice_by_index(begin = cache_13_begin_0, end = cache_13_end_0, end_mask = cache_13_end_mask_0, squeeze_mask = cache_13_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_13_cast_fp16")]; + tensor cache_15_begin_0 = const()[name = string("cache_15_begin_0"), val = tensor([3, 0, 0, 0])]; + tensor cache_15_end_0 = const()[name = string("cache_15_end_0"), val = tensor([4, 1, 1024, 8])]; + tensor cache_15_end_mask_0 = const()[name = string("cache_15_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_15_squeeze_mask_0 = const()[name = string("cache_15_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_15_cast_fp16 = slice_by_index(begin = cache_15_begin_0, end = cache_15_end_0, end_mask = cache_15_end_mask_0, squeeze_mask = cache_15_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_15_cast_fp16")]; + tensor input_183_axes_0 = const()[name = string("input_183_axes_0"), val = tensor([-1])]; + tensor encoder_layers_3_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_3_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75335232)))]; + tensor encoder_layers_3_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_3_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75337344)))]; + tensor input_183_cast_fp16 = layer_norm(axes = input_183_axes_0, beta = encoder_layers_3_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_3_norm_feed_forward1_weight_to_fp16, x = input_181_cast_fp16)[name = string("input_183_cast_fp16")]; + tensor encoder_layers_3_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75339456))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78485248))))[name = string("encoder_layers_3_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_3_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_3_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78485440)))]; + tensor linear_28_cast_fp16 = linear(bias = encoder_layers_3_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_3_feed_forward1_linear1_weight_to_fp16_palettized, x = input_183_cast_fp16)[name = string("linear_28_cast_fp16")]; + tensor input_187_cast_fp16 = silu(x = linear_28_cast_fp16)[name = string("input_187_cast_fp16")]; + tensor encoder_layers_3_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78493696))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81639488))))[name = string("encoder_layers_3_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_3_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_3_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81639680)))]; + tensor linear_29_cast_fp16 = linear(bias = encoder_layers_3_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_3_feed_forward1_linear2_weight_to_fp16_palettized, x = input_187_cast_fp16)[name = string("linear_29_cast_fp16")]; + fp16 var_1094_to_fp16 = const()[name = string("op_1094_to_fp16"), val = fp16(0x1p-1)]; + tensor var_1095_cast_fp16 = mul(x = linear_29_cast_fp16, y = var_1094_to_fp16)[name = string("op_1095_cast_fp16")]; + tensor input_193_cast_fp16 = add(x = input_181_cast_fp16, y = var_1095_cast_fp16)[name = string("input_193_cast_fp16")]; + tensor key_7_axes_0 = const()[name = string("key_7_axes_0"), val = tensor([-1])]; + tensor encoder_layers_3_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_3_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81641792)))]; + tensor encoder_layers_3_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_3_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81643904)))]; + tensor key_7_cast_fp16 = layer_norm(axes = key_7_axes_0, beta = encoder_layers_3_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_3_norm_self_att_weight_to_fp16, x = input_193_cast_fp16)[name = string("key_7_cast_fp16")]; + bool input_195_interleave_0 = const()[name = string("input_195_interleave_0"), val = bool(false)]; + tensor input_195_cast_fp16 = concat(axis = var_68, interleave = input_195_interleave_0, values = (cache_13_cast_fp16, key_7_cast_fp16))[name = string("input_195_cast_fp16")]; + tensor var_1117_begin_0 = const()[name = string("op_1117_begin_0"), val = tensor([0, 28, 0])]; + tensor var_1117_end_0 = const()[name = string("op_1117_end_0"), val = tensor([1, 42, 1024])]; + tensor var_1117_end_mask_0 = const()[name = string("op_1117_end_mask_0"), val = tensor([true, true, true])]; + tensor var_1117_cast_fp16 = slice_by_index(begin = var_1117_begin_0, end = var_1117_end_0, end_mask = var_1117_end_mask_0, x = cache_13_cast_fp16)[name = string("op_1117_cast_fp16")]; + bool var_1123_interleave_0 = const()[name = string("op_1123_interleave_0"), val = bool(false)]; + tensor var_1123_cast_fp16 = concat(axis = var_68, interleave = var_1123_interleave_0, values = (var_1117_cast_fp16, key_7_cast_fp16))[name = string("op_1123_cast_fp16")]; + tensor encoder_layers_3_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81646016))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82432512))))[name = string("encoder_layers_3_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_3_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_3_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82432704)))]; + tensor linear_30_cast_fp16 = linear(bias = encoder_layers_3_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_3_self_attn_linear_q_weight_to_fp16_palettized, x = key_7_cast_fp16)[name = string("linear_30_cast_fp16")]; + tensor var_1128 = const()[name = string("op_1128"), val = tensor([1, -1, 8, 128])]; + tensor q_19_cast_fp16 = reshape(shape = var_1128, x = linear_30_cast_fp16)[name = string("q_19_cast_fp16")]; + tensor encoder_layers_3_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82434816))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83221312))))[name = string("encoder_layers_3_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_3_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_3_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83221504)))]; + tensor linear_31_cast_fp16 = linear(bias = encoder_layers_3_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_3_self_attn_linear_k_weight_to_fp16_palettized, x = input_195_cast_fp16)[name = string("linear_31_cast_fp16")]; + tensor var_1133 = const()[name = string("op_1133"), val = tensor([1, -1, 8, 128])]; + tensor k_13_cast_fp16 = reshape(shape = var_1133, x = linear_31_cast_fp16)[name = string("k_13_cast_fp16")]; + tensor encoder_layers_3_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83223616))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84010112))))[name = string("encoder_layers_3_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_3_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_3_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84010304)))]; + tensor linear_32_cast_fp16 = linear(bias = encoder_layers_3_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_3_self_attn_linear_v_weight_to_fp16_palettized, x = input_195_cast_fp16)[name = string("linear_32_cast_fp16")]; + tensor var_1138 = const()[name = string("op_1138"), val = tensor([1, -1, 8, 128])]; + tensor v_7_cast_fp16 = reshape(shape = var_1138, x = linear_32_cast_fp16)[name = string("v_7_cast_fp16")]; + tensor value_15_perm_0 = const()[name = string("value_15_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_3_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_3_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84012416)))]; + tensor var_1151_cast_fp16 = add(x = q_19_cast_fp16, y = encoder_layers_3_self_attn_pos_bias_u_to_fp16)[name = string("op_1151_cast_fp16")]; + tensor encoder_layers_3_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_3_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84014528)))]; + tensor var_1153_cast_fp16 = add(x = q_19_cast_fp16, y = encoder_layers_3_self_attn_pos_bias_v_to_fp16)[name = string("op_1153_cast_fp16")]; + tensor q_with_bias_v_7_perm_0 = const()[name = string("q_with_bias_v_7_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_85_transpose_x_0 = const()[name = string("x_85_transpose_x_0"), val = bool(false)]; + bool x_85_transpose_y_0 = const()[name = string("x_85_transpose_y_0"), val = bool(false)]; + tensor op_1155_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84016640))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84159040))))[name = string("op_1155_to_fp16_quantized")]; + tensor q_with_bias_v_7_cast_fp16 = transpose(perm = q_with_bias_v_7_perm_0, x = var_1153_cast_fp16)[name = string("transpose_335")]; + tensor x_85_cast_fp16 = matmul(transpose_x = x_85_transpose_x_0, transpose_y = x_85_transpose_y_0, x = q_with_bias_v_7_cast_fp16, y = op_1155_to_fp16_quantized)[name = string("x_85_cast_fp16")]; + tensor x_87_pad_0 = const()[name = string("x_87_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_87_mode_0 = const()[name = string("x_87_mode_0"), val = string("constant")]; + fp16 const_118_to_fp16 = const()[name = string("const_118_to_fp16"), val = fp16(0x0p+0)]; + tensor x_87_cast_fp16 = pad(constant_val = const_118_to_fp16, mode = x_87_mode_0, pad = x_87_pad_0, x = x_85_cast_fp16)[name = string("x_87_cast_fp16")]; + tensor var_1163 = const()[name = string("op_1163"), val = tensor([1, 8, -1, 28])]; + tensor x_89_cast_fp16 = reshape(shape = var_1163, x = x_87_cast_fp16)[name = string("x_89_cast_fp16")]; + tensor var_1167_begin_0 = const()[name = string("op_1167_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_1167_end_0 = const()[name = string("op_1167_end_0"), val = tensor([1, 8, 140, 28])]; + tensor var_1167_end_mask_0 = const()[name = string("op_1167_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_1167_cast_fp16 = slice_by_index(begin = var_1167_begin_0, end = var_1167_end_0, end_mask = var_1167_end_mask_0, x = x_89_cast_fp16)[name = string("op_1167_cast_fp16")]; + tensor var_1168 = const()[name = string("op_1168"), val = tensor([1, 8, 28, 139])]; + tensor matrix_bd_13_cast_fp16 = reshape(shape = var_1168, x = var_1167_cast_fp16)[name = string("matrix_bd_13_cast_fp16")]; + bool matrix_ac_7_transpose_x_0 = const()[name = string("matrix_ac_7_transpose_x_0"), val = bool(false)]; + bool matrix_ac_7_transpose_y_0 = const()[name = string("matrix_ac_7_transpose_y_0"), val = bool(false)]; + tensor transpose_102_perm_0 = const()[name = string("transpose_102_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_103_perm_0 = const()[name = string("transpose_103_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_103 = transpose(perm = transpose_103_perm_0, x = k_13_cast_fp16)[name = string("transpose_333")]; + tensor transpose_102 = transpose(perm = transpose_102_perm_0, x = var_1151_cast_fp16)[name = string("transpose_334")]; + tensor matrix_ac_7_cast_fp16 = matmul(transpose_x = matrix_ac_7_transpose_x_0, transpose_y = matrix_ac_7_transpose_y_0, x = transpose_102, y = transpose_103)[name = string("matrix_ac_7_cast_fp16")]; + tensor matrix_bd_15_begin_0 = const()[name = string("matrix_bd_15_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_15_end_0 = const()[name = string("matrix_bd_15_end_0"), val = tensor([1, 8, 28, 70])]; + tensor matrix_bd_15_end_mask_0 = const()[name = string("matrix_bd_15_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_15_cast_fp16 = slice_by_index(begin = matrix_bd_15_begin_0, end = matrix_bd_15_end_0, end_mask = matrix_bd_15_end_mask_0, x = matrix_bd_13_cast_fp16)[name = string("matrix_bd_15_cast_fp16")]; + tensor var_1177_cast_fp16 = add(x = matrix_ac_7_cast_fp16, y = matrix_bd_15_cast_fp16)[name = string("op_1177_cast_fp16")]; + fp16 _inversed_scores_13_y_0_to_fp16 = const()[name = string("_inversed_scores_13_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_13_cast_fp16 = mul(x = var_1177_cast_fp16, y = _inversed_scores_13_y_0_to_fp16)[name = string("_inversed_scores_13_cast_fp16")]; + tensor scores_15_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_13_cast_fp16, cond = mask_11)[name = string("scores_15_cast_fp16")]; + tensor var_1183_cast_fp16 = softmax(axis = var_59, x = scores_15_cast_fp16)[name = string("op_1183_cast_fp16")]; + tensor input_197_cast_fp16 = select(a = var_44_to_fp16, b = var_1183_cast_fp16, cond = mask_11)[name = string("input_197_cast_fp16")]; + bool x_91_transpose_x_0 = const()[name = string("x_91_transpose_x_0"), val = bool(false)]; + bool x_91_transpose_y_0 = const()[name = string("x_91_transpose_y_0"), val = bool(false)]; + tensor value_15_cast_fp16 = transpose(perm = value_15_perm_0, x = v_7_cast_fp16)[name = string("transpose_332")]; + tensor x_91_cast_fp16 = matmul(transpose_x = x_91_transpose_x_0, transpose_y = x_91_transpose_y_0, x = input_197_cast_fp16, y = value_15_cast_fp16)[name = string("x_91_cast_fp16")]; + tensor var_1187_perm_0 = const()[name = string("op_1187_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1188 = const()[name = string("op_1188"), val = tensor([1, -1, 1024])]; + tensor var_1187_cast_fp16 = transpose(perm = var_1187_perm_0, x = x_91_cast_fp16)[name = string("transpose_331")]; + tensor input_199_cast_fp16 = reshape(shape = var_1188, x = var_1187_cast_fp16)[name = string("input_199_cast_fp16")]; + tensor encoder_layers_3_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84159424))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84945920))))[name = string("encoder_layers_3_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_3_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_3_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84946112)))]; + tensor linear_34_cast_fp16 = linear(bias = encoder_layers_3_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_3_self_attn_linear_out_weight_to_fp16_palettized, x = input_199_cast_fp16)[name = string("linear_34_cast_fp16")]; + tensor input_203_cast_fp16 = add(x = input_193_cast_fp16, y = linear_34_cast_fp16)[name = string("input_203_cast_fp16")]; + tensor x_95_axes_0 = const()[name = string("x_95_axes_0"), val = tensor([-1])]; + tensor encoder_layers_3_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_3_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84948224)))]; + tensor encoder_layers_3_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_3_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84950336)))]; + tensor x_95_cast_fp16 = layer_norm(axes = x_95_axes_0, beta = encoder_layers_3_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_3_norm_conv_weight_to_fp16, x = input_203_cast_fp16)[name = string("x_95_cast_fp16")]; + tensor input_205_perm_0 = const()[name = string("input_205_perm_0"), val = tensor([0, 2, 1])]; + string input_207_pad_type_0 = const()[name = string("input_207_pad_type_0"), val = string("valid")]; + tensor input_207_strides_0 = const()[name = string("input_207_strides_0"), val = tensor([1])]; + tensor input_207_pad_0 = const()[name = string("input_207_pad_0"), val = tensor([0, 0])]; + tensor input_207_dilations_0 = const()[name = string("input_207_dilations_0"), val = tensor([1])]; + int32 input_207_groups_0 = const()[name = string("input_207_groups_0"), val = int32(1)]; + tensor encoder_layers_3_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84952448))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87049664))))[name = string("encoder_layers_3_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_205_cast_fp16 = transpose(perm = input_205_perm_0, x = x_95_cast_fp16)[name = string("transpose_330")]; + tensor input_207_cast_fp16 = conv(dilations = input_207_dilations_0, groups = input_207_groups_0, pad = input_207_pad_0, pad_type = input_207_pad_type_0, strides = input_207_strides_0, weight = encoder_layers_3_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_205_cast_fp16)[name = string("input_207_cast_fp16")]; + int32 x_97_split_num_splits_0 = const()[name = string("x_97_split_num_splits_0"), val = int32(2)]; + int32 x_97_split_axis_0 = const()[name = string("x_97_split_axis_0"), val = int32(1)]; + tensor x_97_split_cast_fp16_0, tensor x_97_split_cast_fp16_1 = split(axis = x_97_split_axis_0, num_splits = x_97_split_num_splits_0, x = input_207_cast_fp16)[name = string("x_97_split_cast_fp16")]; + tensor x_97_split_1_sigmoid_cast_fp16 = sigmoid(x = x_97_split_cast_fp16_1)[name = string("x_97_split_1_sigmoid_cast_fp16")]; + tensor x_97_cast_fp16 = mul(x = x_97_split_cast_fp16_0, y = x_97_split_1_sigmoid_cast_fp16)[name = string("x_97_cast_fp16")]; + tensor input_209_cast_fp16 = select(a = var_44_to_fp16, b = x_97_cast_fp16, cond = var_575)[name = string("input_209_cast_fp16")]; + bool new_x_15_interleave_0 = const()[name = string("new_x_15_interleave_0"), val = bool(false)]; + tensor new_x_15_cast_fp16 = concat(axis = var_59, interleave = new_x_15_interleave_0, values = (cache_15_cast_fp16, input_209_cast_fp16))[name = string("new_x_15_cast_fp16")]; + tensor var_1227_begin_0 = const()[name = string("op_1227_begin_0"), val = tensor([0, 0, 28])]; + tensor var_1227_end_0 = const()[name = string("op_1227_end_0"), val = tensor([1, 1024, 36])]; + tensor var_1227_end_mask_0 = const()[name = string("op_1227_end_mask_0"), val = tensor([true, true, true])]; + tensor var_1227_cast_fp16 = slice_by_index(begin = var_1227_begin_0, end = var_1227_end_0, end_mask = var_1227_end_mask_0, x = new_x_15_cast_fp16)[name = string("op_1227_cast_fp16")]; + string x_99_pad_type_0 = const()[name = string("x_99_pad_type_0"), val = string("valid")]; + int32 x_99_groups_0 = const()[name = string("x_99_groups_0"), val = int32(1024)]; + tensor x_99_strides_0 = const()[name = string("x_99_strides_0"), val = tensor([1])]; + tensor x_99_pad_0 = const()[name = string("x_99_pad_0"), val = tensor([0, 0])]; + tensor x_99_dilations_0 = const()[name = string("x_99_dilations_0"), val = tensor([1])]; + tensor encoder_layers_3_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87053824))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87063104))))[name = string("encoder_layers_3_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_99_cast_fp16 = conv(dilations = x_99_dilations_0, groups = x_99_groups_0, pad = x_99_pad_0, pad_type = x_99_pad_type_0, strides = x_99_strides_0, weight = encoder_layers_3_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_15_cast_fp16)[name = string("x_99_cast_fp16")]; + tensor input_211_perm_0 = const()[name = string("input_211_perm_0"), val = tensor([0, 2, 1])]; + tensor x_101_axes_0 = const()[name = string("x_101_axes_0"), val = tensor([-1])]; + tensor encoder_layers_3_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_3_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87065216)))]; + tensor encoder_layers_3_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_3_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87067328)))]; + tensor input_211_cast_fp16 = transpose(perm = input_211_perm_0, x = x_99_cast_fp16)[name = string("transpose_329")]; + tensor x_101_cast_fp16 = layer_norm(axes = x_101_axes_0, beta = encoder_layers_3_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_3_conv_batch_norm_weight_to_fp16, x = input_211_cast_fp16)[name = string("x_101_cast_fp16")]; + tensor input_213_perm_0 = const()[name = string("input_213_perm_0"), val = tensor([0, 2, 1])]; + tensor input_213_cast_fp16 = transpose(perm = input_213_perm_0, x = x_101_cast_fp16)[name = string("transpose_328")]; + tensor input_215_cast_fp16 = silu(x = input_213_cast_fp16)[name = string("input_215_cast_fp16")]; + string x_103_pad_type_0 = const()[name = string("x_103_pad_type_0"), val = string("valid")]; + tensor x_103_strides_0 = const()[name = string("x_103_strides_0"), val = tensor([1])]; + tensor x_103_pad_0 = const()[name = string("x_103_pad_0"), val = tensor([0, 0])]; + tensor x_103_dilations_0 = const()[name = string("x_103_dilations_0"), val = tensor([1])]; + int32 x_103_groups_0 = const()[name = string("x_103_groups_0"), val = int32(1)]; + tensor encoder_layers_3_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87069440))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88118080))))[name = string("encoder_layers_3_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_103_cast_fp16 = conv(dilations = x_103_dilations_0, groups = x_103_groups_0, pad = x_103_pad_0, pad_type = x_103_pad_type_0, strides = x_103_strides_0, weight = encoder_layers_3_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_215_cast_fp16)[name = string("x_103_cast_fp16")]; + tensor input_217_perm_0 = const()[name = string("input_217_perm_0"), val = tensor([0, 2, 1])]; + tensor input_217_cast_fp16 = transpose(perm = input_217_perm_0, x = x_103_cast_fp16)[name = string("transpose_327")]; + tensor input_219_cast_fp16 = add(x = input_203_cast_fp16, y = input_217_cast_fp16)[name = string("input_219_cast_fp16")]; + tensor input_221_axes_0 = const()[name = string("input_221_axes_0"), val = tensor([-1])]; + tensor encoder_layers_3_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_3_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88120192)))]; + tensor encoder_layers_3_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_3_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88122304)))]; + tensor input_221_cast_fp16 = layer_norm(axes = input_221_axes_0, beta = encoder_layers_3_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_3_norm_feed_forward2_weight_to_fp16, x = input_219_cast_fp16)[name = string("input_221_cast_fp16")]; + tensor encoder_layers_3_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88124416))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(91270208))))[name = string("encoder_layers_3_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_3_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_3_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(91270400)))]; + tensor linear_35_cast_fp16 = linear(bias = encoder_layers_3_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_3_feed_forward2_linear1_weight_to_fp16_palettized, x = input_221_cast_fp16)[name = string("linear_35_cast_fp16")]; + tensor input_225_cast_fp16 = silu(x = linear_35_cast_fp16)[name = string("input_225_cast_fp16")]; + tensor encoder_layers_3_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(91278656))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94424448))))[name = string("encoder_layers_3_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_3_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_3_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94424640)))]; + tensor linear_36_cast_fp16 = linear(bias = encoder_layers_3_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_3_feed_forward2_linear2_weight_to_fp16_palettized, x = input_225_cast_fp16)[name = string("linear_36_cast_fp16")]; + fp16 var_1270_to_fp16 = const()[name = string("op_1270_to_fp16"), val = fp16(0x1p-1)]; + tensor var_1271_cast_fp16 = mul(x = linear_36_cast_fp16, y = var_1270_to_fp16)[name = string("op_1271_cast_fp16")]; + tensor input_231_cast_fp16 = add(x = input_219_cast_fp16, y = var_1271_cast_fp16)[name = string("input_231_cast_fp16")]; + tensor input_233_axes_0 = const()[name = string("input_233_axes_0"), val = tensor([-1])]; + tensor encoder_layers_3_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_3_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94426752)))]; + tensor encoder_layers_3_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_3_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94428864)))]; + tensor input_233_cast_fp16 = layer_norm(axes = input_233_axes_0, beta = encoder_layers_3_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_3_norm_out_weight_to_fp16, x = input_231_cast_fp16)[name = string("input_233_cast_fp16")]; + tensor cache_17_begin_0 = const()[name = string("cache_17_begin_0"), val = tensor([4, 0, 0, 0])]; + tensor cache_17_end_0 = const()[name = string("cache_17_end_0"), val = tensor([5, 1, 42, 1024])]; + tensor cache_17_end_mask_0 = const()[name = string("cache_17_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_17_squeeze_mask_0 = const()[name = string("cache_17_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_17_cast_fp16 = slice_by_index(begin = cache_17_begin_0, end = cache_17_end_0, end_mask = cache_17_end_mask_0, squeeze_mask = cache_17_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_17_cast_fp16")]; + tensor cache_19_begin_0 = const()[name = string("cache_19_begin_0"), val = tensor([4, 0, 0, 0])]; + tensor cache_19_end_0 = const()[name = string("cache_19_end_0"), val = tensor([5, 1, 1024, 8])]; + tensor cache_19_end_mask_0 = const()[name = string("cache_19_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_19_squeeze_mask_0 = const()[name = string("cache_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_19_cast_fp16 = slice_by_index(begin = cache_19_begin_0, end = cache_19_end_0, end_mask = cache_19_end_mask_0, squeeze_mask = cache_19_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_19_cast_fp16")]; + tensor input_235_axes_0 = const()[name = string("input_235_axes_0"), val = tensor([-1])]; + tensor encoder_layers_4_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_4_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94430976)))]; + tensor encoder_layers_4_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_4_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94433088)))]; + tensor input_235_cast_fp16 = layer_norm(axes = input_235_axes_0, beta = encoder_layers_4_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_4_norm_feed_forward1_weight_to_fp16, x = input_233_cast_fp16)[name = string("input_235_cast_fp16")]; + tensor encoder_layers_4_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94435200))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97580992))))[name = string("encoder_layers_4_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_4_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_4_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97581184)))]; + tensor linear_37_cast_fp16 = linear(bias = encoder_layers_4_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_4_feed_forward1_linear1_weight_to_fp16_palettized, x = input_235_cast_fp16)[name = string("linear_37_cast_fp16")]; + tensor input_239_cast_fp16 = silu(x = linear_37_cast_fp16)[name = string("input_239_cast_fp16")]; + tensor encoder_layers_4_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97589440))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100735232))))[name = string("encoder_layers_4_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_4_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_4_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100735424)))]; + tensor linear_38_cast_fp16 = linear(bias = encoder_layers_4_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_4_feed_forward1_linear2_weight_to_fp16_palettized, x = input_239_cast_fp16)[name = string("linear_38_cast_fp16")]; + fp16 var_1307_to_fp16 = const()[name = string("op_1307_to_fp16"), val = fp16(0x1p-1)]; + tensor var_1308_cast_fp16 = mul(x = linear_38_cast_fp16, y = var_1307_to_fp16)[name = string("op_1308_cast_fp16")]; + tensor input_245_cast_fp16 = add(x = input_233_cast_fp16, y = var_1308_cast_fp16)[name = string("input_245_cast_fp16")]; + tensor key_9_axes_0 = const()[name = string("key_9_axes_0"), val = tensor([-1])]; + tensor encoder_layers_4_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_4_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100737536)))]; + tensor encoder_layers_4_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_4_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100739648)))]; + tensor key_9_cast_fp16 = layer_norm(axes = key_9_axes_0, beta = encoder_layers_4_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_4_norm_self_att_weight_to_fp16, x = input_245_cast_fp16)[name = string("key_9_cast_fp16")]; + bool input_247_interleave_0 = const()[name = string("input_247_interleave_0"), val = bool(false)]; + tensor input_247_cast_fp16 = concat(axis = var_68, interleave = input_247_interleave_0, values = (cache_17_cast_fp16, key_9_cast_fp16))[name = string("input_247_cast_fp16")]; + tensor var_1330_begin_0 = const()[name = string("op_1330_begin_0"), val = tensor([0, 28, 0])]; + tensor var_1330_end_0 = const()[name = string("op_1330_end_0"), val = tensor([1, 42, 1024])]; + tensor var_1330_end_mask_0 = const()[name = string("op_1330_end_mask_0"), val = tensor([true, true, true])]; + tensor var_1330_cast_fp16 = slice_by_index(begin = var_1330_begin_0, end = var_1330_end_0, end_mask = var_1330_end_mask_0, x = cache_17_cast_fp16)[name = string("op_1330_cast_fp16")]; + bool var_1336_interleave_0 = const()[name = string("op_1336_interleave_0"), val = bool(false)]; + tensor var_1336_cast_fp16 = concat(axis = var_68, interleave = var_1336_interleave_0, values = (var_1330_cast_fp16, key_9_cast_fp16))[name = string("op_1336_cast_fp16")]; + tensor encoder_layers_4_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100741760))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(101528256))))[name = string("encoder_layers_4_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_4_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_4_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(101528448)))]; + tensor linear_39_cast_fp16 = linear(bias = encoder_layers_4_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_4_self_attn_linear_q_weight_to_fp16_palettized, x = key_9_cast_fp16)[name = string("linear_39_cast_fp16")]; + tensor var_1341 = const()[name = string("op_1341"), val = tensor([1, -1, 8, 128])]; + tensor q_25_cast_fp16 = reshape(shape = var_1341, x = linear_39_cast_fp16)[name = string("q_25_cast_fp16")]; + tensor encoder_layers_4_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(101530560))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102317056))))[name = string("encoder_layers_4_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_4_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_4_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102317248)))]; + tensor linear_40_cast_fp16 = linear(bias = encoder_layers_4_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_4_self_attn_linear_k_weight_to_fp16_palettized, x = input_247_cast_fp16)[name = string("linear_40_cast_fp16")]; + tensor var_1346 = const()[name = string("op_1346"), val = tensor([1, -1, 8, 128])]; + tensor k_17_cast_fp16 = reshape(shape = var_1346, x = linear_40_cast_fp16)[name = string("k_17_cast_fp16")]; + tensor encoder_layers_4_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102319360))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103105856))))[name = string("encoder_layers_4_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_4_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_4_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103106048)))]; + tensor linear_41_cast_fp16 = linear(bias = encoder_layers_4_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_4_self_attn_linear_v_weight_to_fp16_palettized, x = input_247_cast_fp16)[name = string("linear_41_cast_fp16")]; + tensor var_1351 = const()[name = string("op_1351"), val = tensor([1, -1, 8, 128])]; + tensor v_9_cast_fp16 = reshape(shape = var_1351, x = linear_41_cast_fp16)[name = string("v_9_cast_fp16")]; + tensor value_17_perm_0 = const()[name = string("value_17_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_4_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_4_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103108160)))]; + tensor var_1364_cast_fp16 = add(x = q_25_cast_fp16, y = encoder_layers_4_self_attn_pos_bias_u_to_fp16)[name = string("op_1364_cast_fp16")]; + tensor encoder_layers_4_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_4_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103110272)))]; + tensor var_1366_cast_fp16 = add(x = q_25_cast_fp16, y = encoder_layers_4_self_attn_pos_bias_v_to_fp16)[name = string("op_1366_cast_fp16")]; + tensor q_with_bias_v_9_perm_0 = const()[name = string("q_with_bias_v_9_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_111_transpose_x_0 = const()[name = string("x_111_transpose_x_0"), val = bool(false)]; + bool x_111_transpose_y_0 = const()[name = string("x_111_transpose_y_0"), val = bool(false)]; + tensor op_1368_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103112384))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103254784))))[name = string("op_1368_to_fp16_quantized")]; + tensor q_with_bias_v_9_cast_fp16 = transpose(perm = q_with_bias_v_9_perm_0, x = var_1366_cast_fp16)[name = string("transpose_326")]; + tensor x_111_cast_fp16 = matmul(transpose_x = x_111_transpose_x_0, transpose_y = x_111_transpose_y_0, x = q_with_bias_v_9_cast_fp16, y = op_1368_to_fp16_quantized)[name = string("x_111_cast_fp16")]; + tensor x_113_pad_0 = const()[name = string("x_113_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_113_mode_0 = const()[name = string("x_113_mode_0"), val = string("constant")]; + fp16 const_131_to_fp16 = const()[name = string("const_131_to_fp16"), val = fp16(0x0p+0)]; + tensor x_113_cast_fp16 = pad(constant_val = const_131_to_fp16, mode = x_113_mode_0, pad = x_113_pad_0, x = x_111_cast_fp16)[name = string("x_113_cast_fp16")]; + tensor var_1376 = const()[name = string("op_1376"), val = tensor([1, 8, -1, 28])]; + tensor x_115_cast_fp16 = reshape(shape = var_1376, x = x_113_cast_fp16)[name = string("x_115_cast_fp16")]; + tensor var_1380_begin_0 = const()[name = string("op_1380_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_1380_end_0 = const()[name = string("op_1380_end_0"), val = tensor([1, 8, 140, 28])]; + tensor var_1380_end_mask_0 = const()[name = string("op_1380_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_1380_cast_fp16 = slice_by_index(begin = var_1380_begin_0, end = var_1380_end_0, end_mask = var_1380_end_mask_0, x = x_115_cast_fp16)[name = string("op_1380_cast_fp16")]; + tensor var_1381 = const()[name = string("op_1381"), val = tensor([1, 8, 28, 139])]; + tensor matrix_bd_17_cast_fp16 = reshape(shape = var_1381, x = var_1380_cast_fp16)[name = string("matrix_bd_17_cast_fp16")]; + bool matrix_ac_9_transpose_x_0 = const()[name = string("matrix_ac_9_transpose_x_0"), val = bool(false)]; + bool matrix_ac_9_transpose_y_0 = const()[name = string("matrix_ac_9_transpose_y_0"), val = bool(false)]; + tensor transpose_104_perm_0 = const()[name = string("transpose_104_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_105_perm_0 = const()[name = string("transpose_105_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_105 = transpose(perm = transpose_105_perm_0, x = k_17_cast_fp16)[name = string("transpose_324")]; + tensor transpose_104 = transpose(perm = transpose_104_perm_0, x = var_1364_cast_fp16)[name = string("transpose_325")]; + tensor matrix_ac_9_cast_fp16 = matmul(transpose_x = matrix_ac_9_transpose_x_0, transpose_y = matrix_ac_9_transpose_y_0, x = transpose_104, y = transpose_105)[name = string("matrix_ac_9_cast_fp16")]; + tensor matrix_bd_19_begin_0 = const()[name = string("matrix_bd_19_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_19_end_0 = const()[name = string("matrix_bd_19_end_0"), val = tensor([1, 8, 28, 70])]; + tensor matrix_bd_19_end_mask_0 = const()[name = string("matrix_bd_19_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_19_cast_fp16 = slice_by_index(begin = matrix_bd_19_begin_0, end = matrix_bd_19_end_0, end_mask = matrix_bd_19_end_mask_0, x = matrix_bd_17_cast_fp16)[name = string("matrix_bd_19_cast_fp16")]; + tensor var_1390_cast_fp16 = add(x = matrix_ac_9_cast_fp16, y = matrix_bd_19_cast_fp16)[name = string("op_1390_cast_fp16")]; + fp16 _inversed_scores_17_y_0_to_fp16 = const()[name = string("_inversed_scores_17_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_17_cast_fp16 = mul(x = var_1390_cast_fp16, y = _inversed_scores_17_y_0_to_fp16)[name = string("_inversed_scores_17_cast_fp16")]; + tensor scores_19_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_17_cast_fp16, cond = mask_11)[name = string("scores_19_cast_fp16")]; + tensor var_1396_cast_fp16 = softmax(axis = var_59, x = scores_19_cast_fp16)[name = string("op_1396_cast_fp16")]; + tensor input_249_cast_fp16 = select(a = var_44_to_fp16, b = var_1396_cast_fp16, cond = mask_11)[name = string("input_249_cast_fp16")]; + bool x_117_transpose_x_0 = const()[name = string("x_117_transpose_x_0"), val = bool(false)]; + bool x_117_transpose_y_0 = const()[name = string("x_117_transpose_y_0"), val = bool(false)]; + tensor value_17_cast_fp16 = transpose(perm = value_17_perm_0, x = v_9_cast_fp16)[name = string("transpose_323")]; + tensor x_117_cast_fp16 = matmul(transpose_x = x_117_transpose_x_0, transpose_y = x_117_transpose_y_0, x = input_249_cast_fp16, y = value_17_cast_fp16)[name = string("x_117_cast_fp16")]; + tensor var_1400_perm_0 = const()[name = string("op_1400_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1401 = const()[name = string("op_1401"), val = tensor([1, -1, 1024])]; + tensor var_1400_cast_fp16 = transpose(perm = var_1400_perm_0, x = x_117_cast_fp16)[name = string("transpose_322")]; + tensor input_251_cast_fp16 = reshape(shape = var_1401, x = var_1400_cast_fp16)[name = string("input_251_cast_fp16")]; + tensor encoder_layers_4_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103255168))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104041664))))[name = string("encoder_layers_4_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_4_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_4_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104041856)))]; + tensor linear_43_cast_fp16 = linear(bias = encoder_layers_4_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_4_self_attn_linear_out_weight_to_fp16_palettized, x = input_251_cast_fp16)[name = string("linear_43_cast_fp16")]; + tensor input_255_cast_fp16 = add(x = input_245_cast_fp16, y = linear_43_cast_fp16)[name = string("input_255_cast_fp16")]; + tensor x_121_axes_0 = const()[name = string("x_121_axes_0"), val = tensor([-1])]; + tensor encoder_layers_4_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_4_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104043968)))]; + tensor encoder_layers_4_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_4_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104046080)))]; + tensor x_121_cast_fp16 = layer_norm(axes = x_121_axes_0, beta = encoder_layers_4_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_4_norm_conv_weight_to_fp16, x = input_255_cast_fp16)[name = string("x_121_cast_fp16")]; + tensor input_257_perm_0 = const()[name = string("input_257_perm_0"), val = tensor([0, 2, 1])]; + string input_259_pad_type_0 = const()[name = string("input_259_pad_type_0"), val = string("valid")]; + tensor input_259_strides_0 = const()[name = string("input_259_strides_0"), val = tensor([1])]; + tensor input_259_pad_0 = const()[name = string("input_259_pad_0"), val = tensor([0, 0])]; + tensor input_259_dilations_0 = const()[name = string("input_259_dilations_0"), val = tensor([1])]; + int32 input_259_groups_0 = const()[name = string("input_259_groups_0"), val = int32(1)]; + tensor encoder_layers_4_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104048192))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106145408))))[name = string("encoder_layers_4_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_257_cast_fp16 = transpose(perm = input_257_perm_0, x = x_121_cast_fp16)[name = string("transpose_321")]; + tensor input_259_cast_fp16 = conv(dilations = input_259_dilations_0, groups = input_259_groups_0, pad = input_259_pad_0, pad_type = input_259_pad_type_0, strides = input_259_strides_0, weight = encoder_layers_4_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_257_cast_fp16)[name = string("input_259_cast_fp16")]; + int32 x_123_split_num_splits_0 = const()[name = string("x_123_split_num_splits_0"), val = int32(2)]; + int32 x_123_split_axis_0 = const()[name = string("x_123_split_axis_0"), val = int32(1)]; + tensor x_123_split_cast_fp16_0, tensor x_123_split_cast_fp16_1 = split(axis = x_123_split_axis_0, num_splits = x_123_split_num_splits_0, x = input_259_cast_fp16)[name = string("x_123_split_cast_fp16")]; + tensor x_123_split_1_sigmoid_cast_fp16 = sigmoid(x = x_123_split_cast_fp16_1)[name = string("x_123_split_1_sigmoid_cast_fp16")]; + tensor x_123_cast_fp16 = mul(x = x_123_split_cast_fp16_0, y = x_123_split_1_sigmoid_cast_fp16)[name = string("x_123_cast_fp16")]; + tensor input_261_cast_fp16 = select(a = var_44_to_fp16, b = x_123_cast_fp16, cond = var_575)[name = string("input_261_cast_fp16")]; + bool new_x_19_interleave_0 = const()[name = string("new_x_19_interleave_0"), val = bool(false)]; + tensor new_x_19_cast_fp16 = concat(axis = var_59, interleave = new_x_19_interleave_0, values = (cache_19_cast_fp16, input_261_cast_fp16))[name = string("new_x_19_cast_fp16")]; + tensor var_1440_begin_0 = const()[name = string("op_1440_begin_0"), val = tensor([0, 0, 28])]; + tensor var_1440_end_0 = const()[name = string("op_1440_end_0"), val = tensor([1, 1024, 36])]; + tensor var_1440_end_mask_0 = const()[name = string("op_1440_end_mask_0"), val = tensor([true, true, true])]; + tensor var_1440_cast_fp16 = slice_by_index(begin = var_1440_begin_0, end = var_1440_end_0, end_mask = var_1440_end_mask_0, x = new_x_19_cast_fp16)[name = string("op_1440_cast_fp16")]; + string x_125_pad_type_0 = const()[name = string("x_125_pad_type_0"), val = string("valid")]; + int32 x_125_groups_0 = const()[name = string("x_125_groups_0"), val = int32(1024)]; + tensor x_125_strides_0 = const()[name = string("x_125_strides_0"), val = tensor([1])]; + tensor x_125_pad_0 = const()[name = string("x_125_pad_0"), val = tensor([0, 0])]; + tensor x_125_dilations_0 = const()[name = string("x_125_dilations_0"), val = tensor([1])]; + tensor encoder_layers_4_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106149568))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106158848))))[name = string("encoder_layers_4_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_125_cast_fp16 = conv(dilations = x_125_dilations_0, groups = x_125_groups_0, pad = x_125_pad_0, pad_type = x_125_pad_type_0, strides = x_125_strides_0, weight = encoder_layers_4_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_19_cast_fp16)[name = string("x_125_cast_fp16")]; + tensor input_263_perm_0 = const()[name = string("input_263_perm_0"), val = tensor([0, 2, 1])]; + tensor x_127_axes_0 = const()[name = string("x_127_axes_0"), val = tensor([-1])]; + tensor encoder_layers_4_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_4_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106160960)))]; + tensor encoder_layers_4_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_4_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106163072)))]; + tensor input_263_cast_fp16 = transpose(perm = input_263_perm_0, x = x_125_cast_fp16)[name = string("transpose_320")]; + tensor x_127_cast_fp16 = layer_norm(axes = x_127_axes_0, beta = encoder_layers_4_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_4_conv_batch_norm_weight_to_fp16, x = input_263_cast_fp16)[name = string("x_127_cast_fp16")]; + tensor input_265_perm_0 = const()[name = string("input_265_perm_0"), val = tensor([0, 2, 1])]; + tensor input_265_cast_fp16 = transpose(perm = input_265_perm_0, x = x_127_cast_fp16)[name = string("transpose_319")]; + tensor input_267_cast_fp16 = silu(x = input_265_cast_fp16)[name = string("input_267_cast_fp16")]; + string x_129_pad_type_0 = const()[name = string("x_129_pad_type_0"), val = string("valid")]; + tensor x_129_strides_0 = const()[name = string("x_129_strides_0"), val = tensor([1])]; + tensor x_129_pad_0 = const()[name = string("x_129_pad_0"), val = tensor([0, 0])]; + tensor x_129_dilations_0 = const()[name = string("x_129_dilations_0"), val = tensor([1])]; + int32 x_129_groups_0 = const()[name = string("x_129_groups_0"), val = int32(1)]; + tensor encoder_layers_4_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106165184))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107213824))))[name = string("encoder_layers_4_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_129_cast_fp16 = conv(dilations = x_129_dilations_0, groups = x_129_groups_0, pad = x_129_pad_0, pad_type = x_129_pad_type_0, strides = x_129_strides_0, weight = encoder_layers_4_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_267_cast_fp16)[name = string("x_129_cast_fp16")]; + tensor input_269_perm_0 = const()[name = string("input_269_perm_0"), val = tensor([0, 2, 1])]; + tensor input_269_cast_fp16 = transpose(perm = input_269_perm_0, x = x_129_cast_fp16)[name = string("transpose_318")]; + tensor input_271_cast_fp16 = add(x = input_255_cast_fp16, y = input_269_cast_fp16)[name = string("input_271_cast_fp16")]; + tensor input_273_axes_0 = const()[name = string("input_273_axes_0"), val = tensor([-1])]; + tensor encoder_layers_4_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_4_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107215936)))]; + tensor encoder_layers_4_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_4_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107218048)))]; + tensor input_273_cast_fp16 = layer_norm(axes = input_273_axes_0, beta = encoder_layers_4_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_4_norm_feed_forward2_weight_to_fp16, x = input_271_cast_fp16)[name = string("input_273_cast_fp16")]; + tensor encoder_layers_4_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107220160))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110365952))))[name = string("encoder_layers_4_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_4_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_4_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110366144)))]; + tensor linear_44_cast_fp16 = linear(bias = encoder_layers_4_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_4_feed_forward2_linear1_weight_to_fp16_palettized, x = input_273_cast_fp16)[name = string("linear_44_cast_fp16")]; + tensor input_277_cast_fp16 = silu(x = linear_44_cast_fp16)[name = string("input_277_cast_fp16")]; + tensor encoder_layers_4_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110374400))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113520192))))[name = string("encoder_layers_4_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_4_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_4_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113520384)))]; + tensor linear_45_cast_fp16 = linear(bias = encoder_layers_4_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_4_feed_forward2_linear2_weight_to_fp16_palettized, x = input_277_cast_fp16)[name = string("linear_45_cast_fp16")]; + fp16 var_1483_to_fp16 = const()[name = string("op_1483_to_fp16"), val = fp16(0x1p-1)]; + tensor var_1484_cast_fp16 = mul(x = linear_45_cast_fp16, y = var_1483_to_fp16)[name = string("op_1484_cast_fp16")]; + tensor input_283_cast_fp16 = add(x = input_271_cast_fp16, y = var_1484_cast_fp16)[name = string("input_283_cast_fp16")]; + tensor input_285_axes_0 = const()[name = string("input_285_axes_0"), val = tensor([-1])]; + tensor encoder_layers_4_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_4_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113522496)))]; + tensor encoder_layers_4_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_4_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113524608)))]; + tensor input_285_cast_fp16 = layer_norm(axes = input_285_axes_0, beta = encoder_layers_4_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_4_norm_out_weight_to_fp16, x = input_283_cast_fp16)[name = string("input_285_cast_fp16")]; + tensor cache_21_begin_0 = const()[name = string("cache_21_begin_0"), val = tensor([5, 0, 0, 0])]; + tensor cache_21_end_0 = const()[name = string("cache_21_end_0"), val = tensor([6, 1, 42, 1024])]; + tensor cache_21_end_mask_0 = const()[name = string("cache_21_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_21_squeeze_mask_0 = const()[name = string("cache_21_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_21_cast_fp16 = slice_by_index(begin = cache_21_begin_0, end = cache_21_end_0, end_mask = cache_21_end_mask_0, squeeze_mask = cache_21_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_21_cast_fp16")]; + tensor cache_23_begin_0 = const()[name = string("cache_23_begin_0"), val = tensor([5, 0, 0, 0])]; + tensor cache_23_end_0 = const()[name = string("cache_23_end_0"), val = tensor([6, 1, 1024, 8])]; + tensor cache_23_end_mask_0 = const()[name = string("cache_23_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_23_squeeze_mask_0 = const()[name = string("cache_23_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_23_cast_fp16 = slice_by_index(begin = cache_23_begin_0, end = cache_23_end_0, end_mask = cache_23_end_mask_0, squeeze_mask = cache_23_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_23_cast_fp16")]; + tensor input_287_axes_0 = const()[name = string("input_287_axes_0"), val = tensor([-1])]; + tensor encoder_layers_5_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_5_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113526720)))]; + tensor encoder_layers_5_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_5_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113528832)))]; + tensor input_287_cast_fp16 = layer_norm(axes = input_287_axes_0, beta = encoder_layers_5_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_5_norm_feed_forward1_weight_to_fp16, x = input_285_cast_fp16)[name = string("input_287_cast_fp16")]; + tensor encoder_layers_5_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113530944))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116676736))))[name = string("encoder_layers_5_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_5_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_5_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116676928)))]; + tensor linear_46_cast_fp16 = linear(bias = encoder_layers_5_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_5_feed_forward1_linear1_weight_to_fp16_palettized, x = input_287_cast_fp16)[name = string("linear_46_cast_fp16")]; + tensor input_291_cast_fp16 = silu(x = linear_46_cast_fp16)[name = string("input_291_cast_fp16")]; + tensor encoder_layers_5_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116685184))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119830976))))[name = string("encoder_layers_5_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_5_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_5_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119831168)))]; + tensor linear_47_cast_fp16 = linear(bias = encoder_layers_5_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_5_feed_forward1_linear2_weight_to_fp16_palettized, x = input_291_cast_fp16)[name = string("linear_47_cast_fp16")]; + fp16 var_1520_to_fp16 = const()[name = string("op_1520_to_fp16"), val = fp16(0x1p-1)]; + tensor var_1521_cast_fp16 = mul(x = linear_47_cast_fp16, y = var_1520_to_fp16)[name = string("op_1521_cast_fp16")]; + tensor input_297_cast_fp16 = add(x = input_285_cast_fp16, y = var_1521_cast_fp16)[name = string("input_297_cast_fp16")]; + tensor key_11_axes_0 = const()[name = string("key_11_axes_0"), val = tensor([-1])]; + tensor encoder_layers_5_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_5_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119833280)))]; + tensor encoder_layers_5_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_5_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119835392)))]; + tensor key_11_cast_fp16 = layer_norm(axes = key_11_axes_0, beta = encoder_layers_5_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_5_norm_self_att_weight_to_fp16, x = input_297_cast_fp16)[name = string("key_11_cast_fp16")]; + bool input_299_interleave_0 = const()[name = string("input_299_interleave_0"), val = bool(false)]; + tensor input_299_cast_fp16 = concat(axis = var_68, interleave = input_299_interleave_0, values = (cache_21_cast_fp16, key_11_cast_fp16))[name = string("input_299_cast_fp16")]; + tensor var_1543_begin_0 = const()[name = string("op_1543_begin_0"), val = tensor([0, 28, 0])]; + tensor var_1543_end_0 = const()[name = string("op_1543_end_0"), val = tensor([1, 42, 1024])]; + tensor var_1543_end_mask_0 = const()[name = string("op_1543_end_mask_0"), val = tensor([true, true, true])]; + tensor var_1543_cast_fp16 = slice_by_index(begin = var_1543_begin_0, end = var_1543_end_0, end_mask = var_1543_end_mask_0, x = cache_21_cast_fp16)[name = string("op_1543_cast_fp16")]; + bool var_1549_interleave_0 = const()[name = string("op_1549_interleave_0"), val = bool(false)]; + tensor var_1549_cast_fp16 = concat(axis = var_68, interleave = var_1549_interleave_0, values = (var_1543_cast_fp16, key_11_cast_fp16))[name = string("op_1549_cast_fp16")]; + tensor encoder_layers_5_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119837504))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120624000))))[name = string("encoder_layers_5_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_5_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_5_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120624192)))]; + tensor linear_48_cast_fp16 = linear(bias = encoder_layers_5_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_5_self_attn_linear_q_weight_to_fp16_palettized, x = key_11_cast_fp16)[name = string("linear_48_cast_fp16")]; + tensor var_1554 = const()[name = string("op_1554"), val = tensor([1, -1, 8, 128])]; + tensor q_31_cast_fp16 = reshape(shape = var_1554, x = linear_48_cast_fp16)[name = string("q_31_cast_fp16")]; + tensor encoder_layers_5_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120626304))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121412800))))[name = string("encoder_layers_5_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_5_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_5_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121412992)))]; + tensor linear_49_cast_fp16 = linear(bias = encoder_layers_5_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_5_self_attn_linear_k_weight_to_fp16_palettized, x = input_299_cast_fp16)[name = string("linear_49_cast_fp16")]; + tensor var_1559 = const()[name = string("op_1559"), val = tensor([1, -1, 8, 128])]; + tensor k_21_cast_fp16 = reshape(shape = var_1559, x = linear_49_cast_fp16)[name = string("k_21_cast_fp16")]; + tensor encoder_layers_5_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121415104))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122201600))))[name = string("encoder_layers_5_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_5_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_5_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122201792)))]; + tensor linear_50_cast_fp16 = linear(bias = encoder_layers_5_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_5_self_attn_linear_v_weight_to_fp16_palettized, x = input_299_cast_fp16)[name = string("linear_50_cast_fp16")]; + tensor var_1564 = const()[name = string("op_1564"), val = tensor([1, -1, 8, 128])]; + tensor v_11_cast_fp16 = reshape(shape = var_1564, x = linear_50_cast_fp16)[name = string("v_11_cast_fp16")]; + tensor value_19_perm_0 = const()[name = string("value_19_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_5_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_5_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122203904)))]; + tensor var_1577_cast_fp16 = add(x = q_31_cast_fp16, y = encoder_layers_5_self_attn_pos_bias_u_to_fp16)[name = string("op_1577_cast_fp16")]; + tensor encoder_layers_5_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_5_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122206016)))]; + tensor var_1579_cast_fp16 = add(x = q_31_cast_fp16, y = encoder_layers_5_self_attn_pos_bias_v_to_fp16)[name = string("op_1579_cast_fp16")]; + tensor q_with_bias_v_11_perm_0 = const()[name = string("q_with_bias_v_11_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_137_transpose_x_0 = const()[name = string("x_137_transpose_x_0"), val = bool(false)]; + bool x_137_transpose_y_0 = const()[name = string("x_137_transpose_y_0"), val = bool(false)]; + tensor op_1581_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122208128))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122350528))))[name = string("op_1581_to_fp16_quantized")]; + tensor q_with_bias_v_11_cast_fp16 = transpose(perm = q_with_bias_v_11_perm_0, x = var_1579_cast_fp16)[name = string("transpose_317")]; + tensor x_137_cast_fp16 = matmul(transpose_x = x_137_transpose_x_0, transpose_y = x_137_transpose_y_0, x = q_with_bias_v_11_cast_fp16, y = op_1581_to_fp16_quantized)[name = string("x_137_cast_fp16")]; + tensor x_139_pad_0 = const()[name = string("x_139_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_139_mode_0 = const()[name = string("x_139_mode_0"), val = string("constant")]; + fp16 const_144_to_fp16 = const()[name = string("const_144_to_fp16"), val = fp16(0x0p+0)]; + tensor x_139_cast_fp16 = pad(constant_val = const_144_to_fp16, mode = x_139_mode_0, pad = x_139_pad_0, x = x_137_cast_fp16)[name = string("x_139_cast_fp16")]; + tensor var_1589 = const()[name = string("op_1589"), val = tensor([1, 8, -1, 28])]; + tensor x_141_cast_fp16 = reshape(shape = var_1589, x = x_139_cast_fp16)[name = string("x_141_cast_fp16")]; + tensor var_1593_begin_0 = const()[name = string("op_1593_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_1593_end_0 = const()[name = string("op_1593_end_0"), val = tensor([1, 8, 140, 28])]; + tensor var_1593_end_mask_0 = const()[name = string("op_1593_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_1593_cast_fp16 = slice_by_index(begin = var_1593_begin_0, end = var_1593_end_0, end_mask = var_1593_end_mask_0, x = x_141_cast_fp16)[name = string("op_1593_cast_fp16")]; + tensor var_1594 = const()[name = string("op_1594"), val = tensor([1, 8, 28, 139])]; + tensor matrix_bd_21_cast_fp16 = reshape(shape = var_1594, x = var_1593_cast_fp16)[name = string("matrix_bd_21_cast_fp16")]; + bool matrix_ac_11_transpose_x_0 = const()[name = string("matrix_ac_11_transpose_x_0"), val = bool(false)]; + bool matrix_ac_11_transpose_y_0 = const()[name = string("matrix_ac_11_transpose_y_0"), val = bool(false)]; + tensor transpose_106_perm_0 = const()[name = string("transpose_106_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_107_perm_0 = const()[name = string("transpose_107_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_107 = transpose(perm = transpose_107_perm_0, x = k_21_cast_fp16)[name = string("transpose_315")]; + tensor transpose_106 = transpose(perm = transpose_106_perm_0, x = var_1577_cast_fp16)[name = string("transpose_316")]; + tensor matrix_ac_11_cast_fp16 = matmul(transpose_x = matrix_ac_11_transpose_x_0, transpose_y = matrix_ac_11_transpose_y_0, x = transpose_106, y = transpose_107)[name = string("matrix_ac_11_cast_fp16")]; + tensor matrix_bd_23_begin_0 = const()[name = string("matrix_bd_23_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_23_end_0 = const()[name = string("matrix_bd_23_end_0"), val = tensor([1, 8, 28, 70])]; + tensor matrix_bd_23_end_mask_0 = const()[name = string("matrix_bd_23_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_23_cast_fp16 = slice_by_index(begin = matrix_bd_23_begin_0, end = matrix_bd_23_end_0, end_mask = matrix_bd_23_end_mask_0, x = matrix_bd_21_cast_fp16)[name = string("matrix_bd_23_cast_fp16")]; + tensor var_1603_cast_fp16 = add(x = matrix_ac_11_cast_fp16, y = matrix_bd_23_cast_fp16)[name = string("op_1603_cast_fp16")]; + fp16 _inversed_scores_21_y_0_to_fp16 = const()[name = string("_inversed_scores_21_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_21_cast_fp16 = mul(x = var_1603_cast_fp16, y = _inversed_scores_21_y_0_to_fp16)[name = string("_inversed_scores_21_cast_fp16")]; + tensor scores_23_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_21_cast_fp16, cond = mask_11)[name = string("scores_23_cast_fp16")]; + tensor var_1609_cast_fp16 = softmax(axis = var_59, x = scores_23_cast_fp16)[name = string("op_1609_cast_fp16")]; + tensor input_301_cast_fp16 = select(a = var_44_to_fp16, b = var_1609_cast_fp16, cond = mask_11)[name = string("input_301_cast_fp16")]; + bool x_143_transpose_x_0 = const()[name = string("x_143_transpose_x_0"), val = bool(false)]; + bool x_143_transpose_y_0 = const()[name = string("x_143_transpose_y_0"), val = bool(false)]; + tensor value_19_cast_fp16 = transpose(perm = value_19_perm_0, x = v_11_cast_fp16)[name = string("transpose_314")]; + tensor x_143_cast_fp16 = matmul(transpose_x = x_143_transpose_x_0, transpose_y = x_143_transpose_y_0, x = input_301_cast_fp16, y = value_19_cast_fp16)[name = string("x_143_cast_fp16")]; + tensor var_1613_perm_0 = const()[name = string("op_1613_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1614 = const()[name = string("op_1614"), val = tensor([1, -1, 1024])]; + tensor var_1613_cast_fp16 = transpose(perm = var_1613_perm_0, x = x_143_cast_fp16)[name = string("transpose_313")]; + tensor input_303_cast_fp16 = reshape(shape = var_1614, x = var_1613_cast_fp16)[name = string("input_303_cast_fp16")]; + tensor encoder_layers_5_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122350912))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123137408))))[name = string("encoder_layers_5_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_5_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_5_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123137600)))]; + tensor linear_52_cast_fp16 = linear(bias = encoder_layers_5_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_5_self_attn_linear_out_weight_to_fp16_palettized, x = input_303_cast_fp16)[name = string("linear_52_cast_fp16")]; + tensor input_307_cast_fp16 = add(x = input_297_cast_fp16, y = linear_52_cast_fp16)[name = string("input_307_cast_fp16")]; + tensor x_147_axes_0 = const()[name = string("x_147_axes_0"), val = tensor([-1])]; + tensor encoder_layers_5_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_5_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123139712)))]; + tensor encoder_layers_5_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_5_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123141824)))]; + tensor x_147_cast_fp16 = layer_norm(axes = x_147_axes_0, beta = encoder_layers_5_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_5_norm_conv_weight_to_fp16, x = input_307_cast_fp16)[name = string("x_147_cast_fp16")]; + tensor input_309_perm_0 = const()[name = string("input_309_perm_0"), val = tensor([0, 2, 1])]; + string input_311_pad_type_0 = const()[name = string("input_311_pad_type_0"), val = string("valid")]; + tensor input_311_strides_0 = const()[name = string("input_311_strides_0"), val = tensor([1])]; + tensor input_311_pad_0 = const()[name = string("input_311_pad_0"), val = tensor([0, 0])]; + tensor input_311_dilations_0 = const()[name = string("input_311_dilations_0"), val = tensor([1])]; + int32 input_311_groups_0 = const()[name = string("input_311_groups_0"), val = int32(1)]; + tensor encoder_layers_5_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123143936))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125241152))))[name = string("encoder_layers_5_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_309_cast_fp16 = transpose(perm = input_309_perm_0, x = x_147_cast_fp16)[name = string("transpose_312")]; + tensor input_311_cast_fp16 = conv(dilations = input_311_dilations_0, groups = input_311_groups_0, pad = input_311_pad_0, pad_type = input_311_pad_type_0, strides = input_311_strides_0, weight = encoder_layers_5_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_309_cast_fp16)[name = string("input_311_cast_fp16")]; + int32 x_149_split_num_splits_0 = const()[name = string("x_149_split_num_splits_0"), val = int32(2)]; + int32 x_149_split_axis_0 = const()[name = string("x_149_split_axis_0"), val = int32(1)]; + tensor x_149_split_cast_fp16_0, tensor x_149_split_cast_fp16_1 = split(axis = x_149_split_axis_0, num_splits = x_149_split_num_splits_0, x = input_311_cast_fp16)[name = string("x_149_split_cast_fp16")]; + tensor x_149_split_1_sigmoid_cast_fp16 = sigmoid(x = x_149_split_cast_fp16_1)[name = string("x_149_split_1_sigmoid_cast_fp16")]; + tensor x_149_cast_fp16 = mul(x = x_149_split_cast_fp16_0, y = x_149_split_1_sigmoid_cast_fp16)[name = string("x_149_cast_fp16")]; + tensor input_313_cast_fp16 = select(a = var_44_to_fp16, b = x_149_cast_fp16, cond = var_575)[name = string("input_313_cast_fp16")]; + bool new_x_23_interleave_0 = const()[name = string("new_x_23_interleave_0"), val = bool(false)]; + tensor new_x_23_cast_fp16 = concat(axis = var_59, interleave = new_x_23_interleave_0, values = (cache_23_cast_fp16, input_313_cast_fp16))[name = string("new_x_23_cast_fp16")]; + tensor var_1653_begin_0 = const()[name = string("op_1653_begin_0"), val = tensor([0, 0, 28])]; + tensor var_1653_end_0 = const()[name = string("op_1653_end_0"), val = tensor([1, 1024, 36])]; + tensor var_1653_end_mask_0 = const()[name = string("op_1653_end_mask_0"), val = tensor([true, true, true])]; + tensor var_1653_cast_fp16 = slice_by_index(begin = var_1653_begin_0, end = var_1653_end_0, end_mask = var_1653_end_mask_0, x = new_x_23_cast_fp16)[name = string("op_1653_cast_fp16")]; + string x_151_pad_type_0 = const()[name = string("x_151_pad_type_0"), val = string("valid")]; + int32 x_151_groups_0 = const()[name = string("x_151_groups_0"), val = int32(1024)]; + tensor x_151_strides_0 = const()[name = string("x_151_strides_0"), val = tensor([1])]; + tensor x_151_pad_0 = const()[name = string("x_151_pad_0"), val = tensor([0, 0])]; + tensor x_151_dilations_0 = const()[name = string("x_151_dilations_0"), val = tensor([1])]; + tensor encoder_layers_5_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125245312))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125254592))))[name = string("encoder_layers_5_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_151_cast_fp16 = conv(dilations = x_151_dilations_0, groups = x_151_groups_0, pad = x_151_pad_0, pad_type = x_151_pad_type_0, strides = x_151_strides_0, weight = encoder_layers_5_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_23_cast_fp16)[name = string("x_151_cast_fp16")]; + tensor input_315_perm_0 = const()[name = string("input_315_perm_0"), val = tensor([0, 2, 1])]; + tensor x_153_axes_0 = const()[name = string("x_153_axes_0"), val = tensor([-1])]; + tensor encoder_layers_5_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_5_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125256704)))]; + tensor encoder_layers_5_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_5_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125258816)))]; + tensor input_315_cast_fp16 = transpose(perm = input_315_perm_0, x = x_151_cast_fp16)[name = string("transpose_311")]; + tensor x_153_cast_fp16 = layer_norm(axes = x_153_axes_0, beta = encoder_layers_5_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_5_conv_batch_norm_weight_to_fp16, x = input_315_cast_fp16)[name = string("x_153_cast_fp16")]; + tensor input_317_perm_0 = const()[name = string("input_317_perm_0"), val = tensor([0, 2, 1])]; + tensor input_317_cast_fp16 = transpose(perm = input_317_perm_0, x = x_153_cast_fp16)[name = string("transpose_310")]; + tensor input_319_cast_fp16 = silu(x = input_317_cast_fp16)[name = string("input_319_cast_fp16")]; + string x_155_pad_type_0 = const()[name = string("x_155_pad_type_0"), val = string("valid")]; + tensor x_155_strides_0 = const()[name = string("x_155_strides_0"), val = tensor([1])]; + tensor x_155_pad_0 = const()[name = string("x_155_pad_0"), val = tensor([0, 0])]; + tensor x_155_dilations_0 = const()[name = string("x_155_dilations_0"), val = tensor([1])]; + int32 x_155_groups_0 = const()[name = string("x_155_groups_0"), val = int32(1)]; + tensor encoder_layers_5_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125260928))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126309568))))[name = string("encoder_layers_5_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_155_cast_fp16 = conv(dilations = x_155_dilations_0, groups = x_155_groups_0, pad = x_155_pad_0, pad_type = x_155_pad_type_0, strides = x_155_strides_0, weight = encoder_layers_5_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_319_cast_fp16)[name = string("x_155_cast_fp16")]; + tensor input_321_perm_0 = const()[name = string("input_321_perm_0"), val = tensor([0, 2, 1])]; + tensor input_321_cast_fp16 = transpose(perm = input_321_perm_0, x = x_155_cast_fp16)[name = string("transpose_309")]; + tensor input_323_cast_fp16 = add(x = input_307_cast_fp16, y = input_321_cast_fp16)[name = string("input_323_cast_fp16")]; + tensor input_325_axes_0 = const()[name = string("input_325_axes_0"), val = tensor([-1])]; + tensor encoder_layers_5_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_5_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126311680)))]; + tensor encoder_layers_5_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_5_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126313792)))]; + tensor input_325_cast_fp16 = layer_norm(axes = input_325_axes_0, beta = encoder_layers_5_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_5_norm_feed_forward2_weight_to_fp16, x = input_323_cast_fp16)[name = string("input_325_cast_fp16")]; + tensor encoder_layers_5_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126315904))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(129461696))))[name = string("encoder_layers_5_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_5_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_5_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(129461888)))]; + tensor linear_53_cast_fp16 = linear(bias = encoder_layers_5_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_5_feed_forward2_linear1_weight_to_fp16_palettized, x = input_325_cast_fp16)[name = string("linear_53_cast_fp16")]; + tensor input_329_cast_fp16 = silu(x = linear_53_cast_fp16)[name = string("input_329_cast_fp16")]; + tensor encoder_layers_5_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(129470144))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132615936))))[name = string("encoder_layers_5_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_5_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_5_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132616128)))]; + tensor linear_54_cast_fp16 = linear(bias = encoder_layers_5_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_5_feed_forward2_linear2_weight_to_fp16_palettized, x = input_329_cast_fp16)[name = string("linear_54_cast_fp16")]; + fp16 var_1696_to_fp16 = const()[name = string("op_1696_to_fp16"), val = fp16(0x1p-1)]; + tensor var_1697_cast_fp16 = mul(x = linear_54_cast_fp16, y = var_1696_to_fp16)[name = string("op_1697_cast_fp16")]; + tensor input_335_cast_fp16 = add(x = input_323_cast_fp16, y = var_1697_cast_fp16)[name = string("input_335_cast_fp16")]; + tensor input_337_axes_0 = const()[name = string("input_337_axes_0"), val = tensor([-1])]; + tensor encoder_layers_5_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_5_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132618240)))]; + tensor encoder_layers_5_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_5_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132620352)))]; + tensor input_337_cast_fp16 = layer_norm(axes = input_337_axes_0, beta = encoder_layers_5_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_5_norm_out_weight_to_fp16, x = input_335_cast_fp16)[name = string("input_337_cast_fp16")]; + tensor cache_25_begin_0 = const()[name = string("cache_25_begin_0"), val = tensor([6, 0, 0, 0])]; + tensor cache_25_end_0 = const()[name = string("cache_25_end_0"), val = tensor([7, 1, 42, 1024])]; + tensor cache_25_end_mask_0 = const()[name = string("cache_25_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_25_squeeze_mask_0 = const()[name = string("cache_25_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_25_cast_fp16 = slice_by_index(begin = cache_25_begin_0, end = cache_25_end_0, end_mask = cache_25_end_mask_0, squeeze_mask = cache_25_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_25_cast_fp16")]; + tensor cache_27_begin_0 = const()[name = string("cache_27_begin_0"), val = tensor([6, 0, 0, 0])]; + tensor cache_27_end_0 = const()[name = string("cache_27_end_0"), val = tensor([7, 1, 1024, 8])]; + tensor cache_27_end_mask_0 = const()[name = string("cache_27_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_27_squeeze_mask_0 = const()[name = string("cache_27_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_27_cast_fp16 = slice_by_index(begin = cache_27_begin_0, end = cache_27_end_0, end_mask = cache_27_end_mask_0, squeeze_mask = cache_27_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_27_cast_fp16")]; + tensor input_339_axes_0 = const()[name = string("input_339_axes_0"), val = tensor([-1])]; + tensor encoder_layers_6_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_6_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132622464)))]; + tensor encoder_layers_6_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_6_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132624576)))]; + tensor input_339_cast_fp16 = layer_norm(axes = input_339_axes_0, beta = encoder_layers_6_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_6_norm_feed_forward1_weight_to_fp16, x = input_337_cast_fp16)[name = string("input_339_cast_fp16")]; + tensor encoder_layers_6_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132626688))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135772480))))[name = string("encoder_layers_6_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_6_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_6_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135772672)))]; + tensor linear_55_cast_fp16 = linear(bias = encoder_layers_6_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_6_feed_forward1_linear1_weight_to_fp16_palettized, x = input_339_cast_fp16)[name = string("linear_55_cast_fp16")]; + tensor input_343_cast_fp16 = silu(x = linear_55_cast_fp16)[name = string("input_343_cast_fp16")]; + tensor encoder_layers_6_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135780928))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138926720))))[name = string("encoder_layers_6_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_6_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_6_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138926912)))]; + tensor linear_56_cast_fp16 = linear(bias = encoder_layers_6_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_6_feed_forward1_linear2_weight_to_fp16_palettized, x = input_343_cast_fp16)[name = string("linear_56_cast_fp16")]; + fp16 var_1733_to_fp16 = const()[name = string("op_1733_to_fp16"), val = fp16(0x1p-1)]; + tensor var_1734_cast_fp16 = mul(x = linear_56_cast_fp16, y = var_1733_to_fp16)[name = string("op_1734_cast_fp16")]; + tensor input_349_cast_fp16 = add(x = input_337_cast_fp16, y = var_1734_cast_fp16)[name = string("input_349_cast_fp16")]; + tensor key_13_axes_0 = const()[name = string("key_13_axes_0"), val = tensor([-1])]; + tensor encoder_layers_6_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_6_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138929024)))]; + tensor encoder_layers_6_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_6_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138931136)))]; + tensor key_13_cast_fp16 = layer_norm(axes = key_13_axes_0, beta = encoder_layers_6_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_6_norm_self_att_weight_to_fp16, x = input_349_cast_fp16)[name = string("key_13_cast_fp16")]; + bool input_351_interleave_0 = const()[name = string("input_351_interleave_0"), val = bool(false)]; + tensor input_351_cast_fp16 = concat(axis = var_68, interleave = input_351_interleave_0, values = (cache_25_cast_fp16, key_13_cast_fp16))[name = string("input_351_cast_fp16")]; + tensor var_1756_begin_0 = const()[name = string("op_1756_begin_0"), val = tensor([0, 28, 0])]; + tensor var_1756_end_0 = const()[name = string("op_1756_end_0"), val = tensor([1, 42, 1024])]; + tensor var_1756_end_mask_0 = const()[name = string("op_1756_end_mask_0"), val = tensor([true, true, true])]; + tensor var_1756_cast_fp16 = slice_by_index(begin = var_1756_begin_0, end = var_1756_end_0, end_mask = var_1756_end_mask_0, x = cache_25_cast_fp16)[name = string("op_1756_cast_fp16")]; + bool var_1762_interleave_0 = const()[name = string("op_1762_interleave_0"), val = bool(false)]; + tensor var_1762_cast_fp16 = concat(axis = var_68, interleave = var_1762_interleave_0, values = (var_1756_cast_fp16, key_13_cast_fp16))[name = string("op_1762_cast_fp16")]; + tensor encoder_layers_6_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138933248))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(139719744))))[name = string("encoder_layers_6_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_6_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_6_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(139719936)))]; + tensor linear_57_cast_fp16 = linear(bias = encoder_layers_6_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_6_self_attn_linear_q_weight_to_fp16_palettized, x = key_13_cast_fp16)[name = string("linear_57_cast_fp16")]; + tensor var_1767 = const()[name = string("op_1767"), val = tensor([1, -1, 8, 128])]; + tensor q_37_cast_fp16 = reshape(shape = var_1767, x = linear_57_cast_fp16)[name = string("q_37_cast_fp16")]; + tensor encoder_layers_6_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(139722048))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140508544))))[name = string("encoder_layers_6_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_6_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_6_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140508736)))]; + tensor linear_58_cast_fp16 = linear(bias = encoder_layers_6_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_6_self_attn_linear_k_weight_to_fp16_palettized, x = input_351_cast_fp16)[name = string("linear_58_cast_fp16")]; + tensor var_1772 = const()[name = string("op_1772"), val = tensor([1, -1, 8, 128])]; + tensor k_25_cast_fp16 = reshape(shape = var_1772, x = linear_58_cast_fp16)[name = string("k_25_cast_fp16")]; + tensor encoder_layers_6_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140510848))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141297344))))[name = string("encoder_layers_6_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_6_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_6_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141297536)))]; + tensor linear_59_cast_fp16 = linear(bias = encoder_layers_6_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_6_self_attn_linear_v_weight_to_fp16_palettized, x = input_351_cast_fp16)[name = string("linear_59_cast_fp16")]; + tensor var_1777 = const()[name = string("op_1777"), val = tensor([1, -1, 8, 128])]; + tensor v_13_cast_fp16 = reshape(shape = var_1777, x = linear_59_cast_fp16)[name = string("v_13_cast_fp16")]; + tensor value_21_perm_0 = const()[name = string("value_21_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_6_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_6_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141299648)))]; + tensor var_1790_cast_fp16 = add(x = q_37_cast_fp16, y = encoder_layers_6_self_attn_pos_bias_u_to_fp16)[name = string("op_1790_cast_fp16")]; + tensor encoder_layers_6_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_6_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141301760)))]; + tensor var_1792_cast_fp16 = add(x = q_37_cast_fp16, y = encoder_layers_6_self_attn_pos_bias_v_to_fp16)[name = string("op_1792_cast_fp16")]; + tensor q_with_bias_v_13_perm_0 = const()[name = string("q_with_bias_v_13_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_163_transpose_x_0 = const()[name = string("x_163_transpose_x_0"), val = bool(false)]; + bool x_163_transpose_y_0 = const()[name = string("x_163_transpose_y_0"), val = bool(false)]; + tensor op_1794_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141303872))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141446272))))[name = string("op_1794_to_fp16_quantized")]; + tensor q_with_bias_v_13_cast_fp16 = transpose(perm = q_with_bias_v_13_perm_0, x = var_1792_cast_fp16)[name = string("transpose_308")]; + tensor x_163_cast_fp16 = matmul(transpose_x = x_163_transpose_x_0, transpose_y = x_163_transpose_y_0, x = q_with_bias_v_13_cast_fp16, y = op_1794_to_fp16_quantized)[name = string("x_163_cast_fp16")]; + tensor x_165_pad_0 = const()[name = string("x_165_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_165_mode_0 = const()[name = string("x_165_mode_0"), val = string("constant")]; + fp16 const_157_to_fp16 = const()[name = string("const_157_to_fp16"), val = fp16(0x0p+0)]; + tensor x_165_cast_fp16 = pad(constant_val = const_157_to_fp16, mode = x_165_mode_0, pad = x_165_pad_0, x = x_163_cast_fp16)[name = string("x_165_cast_fp16")]; + tensor var_1802 = const()[name = string("op_1802"), val = tensor([1, 8, -1, 28])]; + tensor x_167_cast_fp16 = reshape(shape = var_1802, x = x_165_cast_fp16)[name = string("x_167_cast_fp16")]; + tensor var_1806_begin_0 = const()[name = string("op_1806_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_1806_end_0 = const()[name = string("op_1806_end_0"), val = tensor([1, 8, 140, 28])]; + tensor var_1806_end_mask_0 = const()[name = string("op_1806_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_1806_cast_fp16 = slice_by_index(begin = var_1806_begin_0, end = var_1806_end_0, end_mask = var_1806_end_mask_0, x = x_167_cast_fp16)[name = string("op_1806_cast_fp16")]; + tensor var_1807 = const()[name = string("op_1807"), val = tensor([1, 8, 28, 139])]; + tensor matrix_bd_25_cast_fp16 = reshape(shape = var_1807, x = var_1806_cast_fp16)[name = string("matrix_bd_25_cast_fp16")]; + bool matrix_ac_13_transpose_x_0 = const()[name = string("matrix_ac_13_transpose_x_0"), val = bool(false)]; + bool matrix_ac_13_transpose_y_0 = const()[name = string("matrix_ac_13_transpose_y_0"), val = bool(false)]; + tensor transpose_108_perm_0 = const()[name = string("transpose_108_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_109_perm_0 = const()[name = string("transpose_109_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_109 = transpose(perm = transpose_109_perm_0, x = k_25_cast_fp16)[name = string("transpose_306")]; + tensor transpose_108 = transpose(perm = transpose_108_perm_0, x = var_1790_cast_fp16)[name = string("transpose_307")]; + tensor matrix_ac_13_cast_fp16 = matmul(transpose_x = matrix_ac_13_transpose_x_0, transpose_y = matrix_ac_13_transpose_y_0, x = transpose_108, y = transpose_109)[name = string("matrix_ac_13_cast_fp16")]; + tensor matrix_bd_27_begin_0 = const()[name = string("matrix_bd_27_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_27_end_0 = const()[name = string("matrix_bd_27_end_0"), val = tensor([1, 8, 28, 70])]; + tensor matrix_bd_27_end_mask_0 = const()[name = string("matrix_bd_27_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_27_cast_fp16 = slice_by_index(begin = matrix_bd_27_begin_0, end = matrix_bd_27_end_0, end_mask = matrix_bd_27_end_mask_0, x = matrix_bd_25_cast_fp16)[name = string("matrix_bd_27_cast_fp16")]; + tensor var_1816_cast_fp16 = add(x = matrix_ac_13_cast_fp16, y = matrix_bd_27_cast_fp16)[name = string("op_1816_cast_fp16")]; + fp16 _inversed_scores_25_y_0_to_fp16 = const()[name = string("_inversed_scores_25_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_25_cast_fp16 = mul(x = var_1816_cast_fp16, y = _inversed_scores_25_y_0_to_fp16)[name = string("_inversed_scores_25_cast_fp16")]; + tensor scores_27_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_25_cast_fp16, cond = mask_11)[name = string("scores_27_cast_fp16")]; + tensor var_1822_cast_fp16 = softmax(axis = var_59, x = scores_27_cast_fp16)[name = string("op_1822_cast_fp16")]; + tensor input_353_cast_fp16 = select(a = var_44_to_fp16, b = var_1822_cast_fp16, cond = mask_11)[name = string("input_353_cast_fp16")]; + bool x_169_transpose_x_0 = const()[name = string("x_169_transpose_x_0"), val = bool(false)]; + bool x_169_transpose_y_0 = const()[name = string("x_169_transpose_y_0"), val = bool(false)]; + tensor value_21_cast_fp16 = transpose(perm = value_21_perm_0, x = v_13_cast_fp16)[name = string("transpose_305")]; + tensor x_169_cast_fp16 = matmul(transpose_x = x_169_transpose_x_0, transpose_y = x_169_transpose_y_0, x = input_353_cast_fp16, y = value_21_cast_fp16)[name = string("x_169_cast_fp16")]; + tensor var_1826_perm_0 = const()[name = string("op_1826_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1827 = const()[name = string("op_1827"), val = tensor([1, -1, 1024])]; + tensor var_1826_cast_fp16 = transpose(perm = var_1826_perm_0, x = x_169_cast_fp16)[name = string("transpose_304")]; + tensor input_355_cast_fp16 = reshape(shape = var_1827, x = var_1826_cast_fp16)[name = string("input_355_cast_fp16")]; + tensor encoder_layers_6_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141446656))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142233152))))[name = string("encoder_layers_6_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_6_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_6_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142233344)))]; + tensor linear_61_cast_fp16 = linear(bias = encoder_layers_6_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_6_self_attn_linear_out_weight_to_fp16_palettized, x = input_355_cast_fp16)[name = string("linear_61_cast_fp16")]; + tensor input_359_cast_fp16 = add(x = input_349_cast_fp16, y = linear_61_cast_fp16)[name = string("input_359_cast_fp16")]; + tensor x_173_axes_0 = const()[name = string("x_173_axes_0"), val = tensor([-1])]; + tensor encoder_layers_6_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_6_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142235456)))]; + tensor encoder_layers_6_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_6_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142237568)))]; + tensor x_173_cast_fp16 = layer_norm(axes = x_173_axes_0, beta = encoder_layers_6_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_6_norm_conv_weight_to_fp16, x = input_359_cast_fp16)[name = string("x_173_cast_fp16")]; + tensor input_361_perm_0 = const()[name = string("input_361_perm_0"), val = tensor([0, 2, 1])]; + string input_363_pad_type_0 = const()[name = string("input_363_pad_type_0"), val = string("valid")]; + tensor input_363_strides_0 = const()[name = string("input_363_strides_0"), val = tensor([1])]; + tensor input_363_pad_0 = const()[name = string("input_363_pad_0"), val = tensor([0, 0])]; + tensor input_363_dilations_0 = const()[name = string("input_363_dilations_0"), val = tensor([1])]; + int32 input_363_groups_0 = const()[name = string("input_363_groups_0"), val = int32(1)]; + tensor encoder_layers_6_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142239680))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144336896))))[name = string("encoder_layers_6_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_361_cast_fp16 = transpose(perm = input_361_perm_0, x = x_173_cast_fp16)[name = string("transpose_303")]; + tensor input_363_cast_fp16 = conv(dilations = input_363_dilations_0, groups = input_363_groups_0, pad = input_363_pad_0, pad_type = input_363_pad_type_0, strides = input_363_strides_0, weight = encoder_layers_6_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_361_cast_fp16)[name = string("input_363_cast_fp16")]; + int32 x_175_split_num_splits_0 = const()[name = string("x_175_split_num_splits_0"), val = int32(2)]; + int32 x_175_split_axis_0 = const()[name = string("x_175_split_axis_0"), val = int32(1)]; + tensor x_175_split_cast_fp16_0, tensor x_175_split_cast_fp16_1 = split(axis = x_175_split_axis_0, num_splits = x_175_split_num_splits_0, x = input_363_cast_fp16)[name = string("x_175_split_cast_fp16")]; + tensor x_175_split_1_sigmoid_cast_fp16 = sigmoid(x = x_175_split_cast_fp16_1)[name = string("x_175_split_1_sigmoid_cast_fp16")]; + tensor x_175_cast_fp16 = mul(x = x_175_split_cast_fp16_0, y = x_175_split_1_sigmoid_cast_fp16)[name = string("x_175_cast_fp16")]; + tensor input_365_cast_fp16 = select(a = var_44_to_fp16, b = x_175_cast_fp16, cond = var_575)[name = string("input_365_cast_fp16")]; + bool new_x_27_interleave_0 = const()[name = string("new_x_27_interleave_0"), val = bool(false)]; + tensor new_x_27_cast_fp16 = concat(axis = var_59, interleave = new_x_27_interleave_0, values = (cache_27_cast_fp16, input_365_cast_fp16))[name = string("new_x_27_cast_fp16")]; + tensor var_1866_begin_0 = const()[name = string("op_1866_begin_0"), val = tensor([0, 0, 28])]; + tensor var_1866_end_0 = const()[name = string("op_1866_end_0"), val = tensor([1, 1024, 36])]; + tensor var_1866_end_mask_0 = const()[name = string("op_1866_end_mask_0"), val = tensor([true, true, true])]; + tensor var_1866_cast_fp16 = slice_by_index(begin = var_1866_begin_0, end = var_1866_end_0, end_mask = var_1866_end_mask_0, x = new_x_27_cast_fp16)[name = string("op_1866_cast_fp16")]; + string x_177_pad_type_0 = const()[name = string("x_177_pad_type_0"), val = string("valid")]; + int32 x_177_groups_0 = const()[name = string("x_177_groups_0"), val = int32(1024)]; + tensor x_177_strides_0 = const()[name = string("x_177_strides_0"), val = tensor([1])]; + tensor x_177_pad_0 = const()[name = string("x_177_pad_0"), val = tensor([0, 0])]; + tensor x_177_dilations_0 = const()[name = string("x_177_dilations_0"), val = tensor([1])]; + tensor encoder_layers_6_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144341056))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144350336))))[name = string("encoder_layers_6_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_177_cast_fp16 = conv(dilations = x_177_dilations_0, groups = x_177_groups_0, pad = x_177_pad_0, pad_type = x_177_pad_type_0, strides = x_177_strides_0, weight = encoder_layers_6_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_27_cast_fp16)[name = string("x_177_cast_fp16")]; + tensor input_367_perm_0 = const()[name = string("input_367_perm_0"), val = tensor([0, 2, 1])]; + tensor x_179_axes_0 = const()[name = string("x_179_axes_0"), val = tensor([-1])]; + tensor encoder_layers_6_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_6_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144352448)))]; + tensor encoder_layers_6_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_6_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144354560)))]; + tensor input_367_cast_fp16 = transpose(perm = input_367_perm_0, x = x_177_cast_fp16)[name = string("transpose_302")]; + tensor x_179_cast_fp16 = layer_norm(axes = x_179_axes_0, beta = encoder_layers_6_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_6_conv_batch_norm_weight_to_fp16, x = input_367_cast_fp16)[name = string("x_179_cast_fp16")]; + tensor input_369_perm_0 = const()[name = string("input_369_perm_0"), val = tensor([0, 2, 1])]; + tensor input_369_cast_fp16 = transpose(perm = input_369_perm_0, x = x_179_cast_fp16)[name = string("transpose_301")]; + tensor input_371_cast_fp16 = silu(x = input_369_cast_fp16)[name = string("input_371_cast_fp16")]; + string x_181_pad_type_0 = const()[name = string("x_181_pad_type_0"), val = string("valid")]; + tensor x_181_strides_0 = const()[name = string("x_181_strides_0"), val = tensor([1])]; + tensor x_181_pad_0 = const()[name = string("x_181_pad_0"), val = tensor([0, 0])]; + tensor x_181_dilations_0 = const()[name = string("x_181_dilations_0"), val = tensor([1])]; + int32 x_181_groups_0 = const()[name = string("x_181_groups_0"), val = int32(1)]; + tensor encoder_layers_6_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144356672))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(145405312))))[name = string("encoder_layers_6_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_181_cast_fp16 = conv(dilations = x_181_dilations_0, groups = x_181_groups_0, pad = x_181_pad_0, pad_type = x_181_pad_type_0, strides = x_181_strides_0, weight = encoder_layers_6_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_371_cast_fp16)[name = string("x_181_cast_fp16")]; + tensor input_373_perm_0 = const()[name = string("input_373_perm_0"), val = tensor([0, 2, 1])]; + tensor input_373_cast_fp16 = transpose(perm = input_373_perm_0, x = x_181_cast_fp16)[name = string("transpose_300")]; + tensor input_375_cast_fp16 = add(x = input_359_cast_fp16, y = input_373_cast_fp16)[name = string("input_375_cast_fp16")]; + tensor input_377_axes_0 = const()[name = string("input_377_axes_0"), val = tensor([-1])]; + tensor encoder_layers_6_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_6_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(145407424)))]; + tensor encoder_layers_6_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_6_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(145409536)))]; + tensor input_377_cast_fp16 = layer_norm(axes = input_377_axes_0, beta = encoder_layers_6_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_6_norm_feed_forward2_weight_to_fp16, x = input_375_cast_fp16)[name = string("input_377_cast_fp16")]; + tensor encoder_layers_6_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(145411648))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(148557440))))[name = string("encoder_layers_6_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_6_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_6_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(148557632)))]; + tensor linear_62_cast_fp16 = linear(bias = encoder_layers_6_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_6_feed_forward2_linear1_weight_to_fp16_palettized, x = input_377_cast_fp16)[name = string("linear_62_cast_fp16")]; + tensor input_381_cast_fp16 = silu(x = linear_62_cast_fp16)[name = string("input_381_cast_fp16")]; + tensor encoder_layers_6_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(148565888))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151711680))))[name = string("encoder_layers_6_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_6_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_6_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151711872)))]; + tensor linear_63_cast_fp16 = linear(bias = encoder_layers_6_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_6_feed_forward2_linear2_weight_to_fp16_palettized, x = input_381_cast_fp16)[name = string("linear_63_cast_fp16")]; + fp16 var_1909_to_fp16 = const()[name = string("op_1909_to_fp16"), val = fp16(0x1p-1)]; + tensor var_1910_cast_fp16 = mul(x = linear_63_cast_fp16, y = var_1909_to_fp16)[name = string("op_1910_cast_fp16")]; + tensor input_387_cast_fp16 = add(x = input_375_cast_fp16, y = var_1910_cast_fp16)[name = string("input_387_cast_fp16")]; + tensor input_389_axes_0 = const()[name = string("input_389_axes_0"), val = tensor([-1])]; + tensor encoder_layers_6_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_6_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151713984)))]; + tensor encoder_layers_6_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_6_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151716096)))]; + tensor input_389_cast_fp16 = layer_norm(axes = input_389_axes_0, beta = encoder_layers_6_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_6_norm_out_weight_to_fp16, x = input_387_cast_fp16)[name = string("input_389_cast_fp16")]; + tensor cache_29_begin_0 = const()[name = string("cache_29_begin_0"), val = tensor([7, 0, 0, 0])]; + tensor cache_29_end_0 = const()[name = string("cache_29_end_0"), val = tensor([8, 1, 42, 1024])]; + tensor cache_29_end_mask_0 = const()[name = string("cache_29_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_29_squeeze_mask_0 = const()[name = string("cache_29_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_29_cast_fp16 = slice_by_index(begin = cache_29_begin_0, end = cache_29_end_0, end_mask = cache_29_end_mask_0, squeeze_mask = cache_29_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_29_cast_fp16")]; + tensor cache_31_begin_0 = const()[name = string("cache_31_begin_0"), val = tensor([7, 0, 0, 0])]; + tensor cache_31_end_0 = const()[name = string("cache_31_end_0"), val = tensor([8, 1, 1024, 8])]; + tensor cache_31_end_mask_0 = const()[name = string("cache_31_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_31_squeeze_mask_0 = const()[name = string("cache_31_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_31_cast_fp16 = slice_by_index(begin = cache_31_begin_0, end = cache_31_end_0, end_mask = cache_31_end_mask_0, squeeze_mask = cache_31_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_31_cast_fp16")]; + tensor input_391_axes_0 = const()[name = string("input_391_axes_0"), val = tensor([-1])]; + tensor encoder_layers_7_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_7_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151718208)))]; + tensor encoder_layers_7_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_7_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151720320)))]; + tensor input_391_cast_fp16 = layer_norm(axes = input_391_axes_0, beta = encoder_layers_7_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_7_norm_feed_forward1_weight_to_fp16, x = input_389_cast_fp16)[name = string("input_391_cast_fp16")]; + tensor encoder_layers_7_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151722432))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(154868224))))[name = string("encoder_layers_7_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_7_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_7_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(154868416)))]; + tensor linear_64_cast_fp16 = linear(bias = encoder_layers_7_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_7_feed_forward1_linear1_weight_to_fp16_palettized, x = input_391_cast_fp16)[name = string("linear_64_cast_fp16")]; + tensor input_395_cast_fp16 = silu(x = linear_64_cast_fp16)[name = string("input_395_cast_fp16")]; + tensor encoder_layers_7_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(154876672))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158022464))))[name = string("encoder_layers_7_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_7_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_7_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158022656)))]; + tensor linear_65_cast_fp16 = linear(bias = encoder_layers_7_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_7_feed_forward1_linear2_weight_to_fp16_palettized, x = input_395_cast_fp16)[name = string("linear_65_cast_fp16")]; + fp16 var_1946_to_fp16 = const()[name = string("op_1946_to_fp16"), val = fp16(0x1p-1)]; + tensor var_1947_cast_fp16 = mul(x = linear_65_cast_fp16, y = var_1946_to_fp16)[name = string("op_1947_cast_fp16")]; + tensor input_401_cast_fp16 = add(x = input_389_cast_fp16, y = var_1947_cast_fp16)[name = string("input_401_cast_fp16")]; + tensor key_15_axes_0 = const()[name = string("key_15_axes_0"), val = tensor([-1])]; + tensor encoder_layers_7_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_7_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158024768)))]; + tensor encoder_layers_7_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_7_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158026880)))]; + tensor key_15_cast_fp16 = layer_norm(axes = key_15_axes_0, beta = encoder_layers_7_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_7_norm_self_att_weight_to_fp16, x = input_401_cast_fp16)[name = string("key_15_cast_fp16")]; + bool input_403_interleave_0 = const()[name = string("input_403_interleave_0"), val = bool(false)]; + tensor input_403_cast_fp16 = concat(axis = var_68, interleave = input_403_interleave_0, values = (cache_29_cast_fp16, key_15_cast_fp16))[name = string("input_403_cast_fp16")]; + tensor var_1969_begin_0 = const()[name = string("op_1969_begin_0"), val = tensor([0, 28, 0])]; + tensor var_1969_end_0 = const()[name = string("op_1969_end_0"), val = tensor([1, 42, 1024])]; + tensor var_1969_end_mask_0 = const()[name = string("op_1969_end_mask_0"), val = tensor([true, true, true])]; + tensor var_1969_cast_fp16 = slice_by_index(begin = var_1969_begin_0, end = var_1969_end_0, end_mask = var_1969_end_mask_0, x = cache_29_cast_fp16)[name = string("op_1969_cast_fp16")]; + bool var_1975_interleave_0 = const()[name = string("op_1975_interleave_0"), val = bool(false)]; + tensor var_1975_cast_fp16 = concat(axis = var_68, interleave = var_1975_interleave_0, values = (var_1969_cast_fp16, key_15_cast_fp16))[name = string("op_1975_cast_fp16")]; + tensor encoder_layers_7_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158028992))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158815488))))[name = string("encoder_layers_7_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_7_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_7_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158815680)))]; + tensor linear_66_cast_fp16 = linear(bias = encoder_layers_7_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_7_self_attn_linear_q_weight_to_fp16_palettized, x = key_15_cast_fp16)[name = string("linear_66_cast_fp16")]; + tensor var_1980 = const()[name = string("op_1980"), val = tensor([1, -1, 8, 128])]; + tensor q_43_cast_fp16 = reshape(shape = var_1980, x = linear_66_cast_fp16)[name = string("q_43_cast_fp16")]; + tensor encoder_layers_7_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158817792))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(159604288))))[name = string("encoder_layers_7_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_7_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_7_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(159604480)))]; + tensor linear_67_cast_fp16 = linear(bias = encoder_layers_7_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_7_self_attn_linear_k_weight_to_fp16_palettized, x = input_403_cast_fp16)[name = string("linear_67_cast_fp16")]; + tensor var_1985 = const()[name = string("op_1985"), val = tensor([1, -1, 8, 128])]; + tensor k_29_cast_fp16 = reshape(shape = var_1985, x = linear_67_cast_fp16)[name = string("k_29_cast_fp16")]; + tensor encoder_layers_7_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(159606592))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160393088))))[name = string("encoder_layers_7_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_7_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_7_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160393280)))]; + tensor linear_68_cast_fp16 = linear(bias = encoder_layers_7_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_7_self_attn_linear_v_weight_to_fp16_palettized, x = input_403_cast_fp16)[name = string("linear_68_cast_fp16")]; + tensor var_1990 = const()[name = string("op_1990"), val = tensor([1, -1, 8, 128])]; + tensor v_15_cast_fp16 = reshape(shape = var_1990, x = linear_68_cast_fp16)[name = string("v_15_cast_fp16")]; + tensor value_23_perm_0 = const()[name = string("value_23_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_7_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_7_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160395392)))]; + tensor var_2003_cast_fp16 = add(x = q_43_cast_fp16, y = encoder_layers_7_self_attn_pos_bias_u_to_fp16)[name = string("op_2003_cast_fp16")]; + tensor encoder_layers_7_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_7_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160397504)))]; + tensor var_2005_cast_fp16 = add(x = q_43_cast_fp16, y = encoder_layers_7_self_attn_pos_bias_v_to_fp16)[name = string("op_2005_cast_fp16")]; + tensor q_with_bias_v_15_perm_0 = const()[name = string("q_with_bias_v_15_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_189_transpose_x_0 = const()[name = string("x_189_transpose_x_0"), val = bool(false)]; + bool x_189_transpose_y_0 = const()[name = string("x_189_transpose_y_0"), val = bool(false)]; + tensor op_2007_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160399616))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160542016))))[name = string("op_2007_to_fp16_quantized")]; + tensor q_with_bias_v_15_cast_fp16 = transpose(perm = q_with_bias_v_15_perm_0, x = var_2005_cast_fp16)[name = string("transpose_299")]; + tensor x_189_cast_fp16 = matmul(transpose_x = x_189_transpose_x_0, transpose_y = x_189_transpose_y_0, x = q_with_bias_v_15_cast_fp16, y = op_2007_to_fp16_quantized)[name = string("x_189_cast_fp16")]; + tensor x_191_pad_0 = const()[name = string("x_191_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_191_mode_0 = const()[name = string("x_191_mode_0"), val = string("constant")]; + fp16 const_170_to_fp16 = const()[name = string("const_170_to_fp16"), val = fp16(0x0p+0)]; + tensor x_191_cast_fp16 = pad(constant_val = const_170_to_fp16, mode = x_191_mode_0, pad = x_191_pad_0, x = x_189_cast_fp16)[name = string("x_191_cast_fp16")]; + tensor var_2015 = const()[name = string("op_2015"), val = tensor([1, 8, -1, 28])]; + tensor x_193_cast_fp16 = reshape(shape = var_2015, x = x_191_cast_fp16)[name = string("x_193_cast_fp16")]; + tensor var_2019_begin_0 = const()[name = string("op_2019_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_2019_end_0 = const()[name = string("op_2019_end_0"), val = tensor([1, 8, 140, 28])]; + tensor var_2019_end_mask_0 = const()[name = string("op_2019_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_2019_cast_fp16 = slice_by_index(begin = var_2019_begin_0, end = var_2019_end_0, end_mask = var_2019_end_mask_0, x = x_193_cast_fp16)[name = string("op_2019_cast_fp16")]; + tensor var_2020 = const()[name = string("op_2020"), val = tensor([1, 8, 28, 139])]; + tensor matrix_bd_29_cast_fp16 = reshape(shape = var_2020, x = var_2019_cast_fp16)[name = string("matrix_bd_29_cast_fp16")]; + bool matrix_ac_15_transpose_x_0 = const()[name = string("matrix_ac_15_transpose_x_0"), val = bool(false)]; + bool matrix_ac_15_transpose_y_0 = const()[name = string("matrix_ac_15_transpose_y_0"), val = bool(false)]; + tensor transpose_110_perm_0 = const()[name = string("transpose_110_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_111_perm_0 = const()[name = string("transpose_111_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_111 = transpose(perm = transpose_111_perm_0, x = k_29_cast_fp16)[name = string("transpose_297")]; + tensor transpose_110 = transpose(perm = transpose_110_perm_0, x = var_2003_cast_fp16)[name = string("transpose_298")]; + tensor matrix_ac_15_cast_fp16 = matmul(transpose_x = matrix_ac_15_transpose_x_0, transpose_y = matrix_ac_15_transpose_y_0, x = transpose_110, y = transpose_111)[name = string("matrix_ac_15_cast_fp16")]; + tensor matrix_bd_31_begin_0 = const()[name = string("matrix_bd_31_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_31_end_0 = const()[name = string("matrix_bd_31_end_0"), val = tensor([1, 8, 28, 70])]; + tensor matrix_bd_31_end_mask_0 = const()[name = string("matrix_bd_31_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_31_cast_fp16 = slice_by_index(begin = matrix_bd_31_begin_0, end = matrix_bd_31_end_0, end_mask = matrix_bd_31_end_mask_0, x = matrix_bd_29_cast_fp16)[name = string("matrix_bd_31_cast_fp16")]; + tensor var_2029_cast_fp16 = add(x = matrix_ac_15_cast_fp16, y = matrix_bd_31_cast_fp16)[name = string("op_2029_cast_fp16")]; + fp16 _inversed_scores_29_y_0_to_fp16 = const()[name = string("_inversed_scores_29_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_29_cast_fp16 = mul(x = var_2029_cast_fp16, y = _inversed_scores_29_y_0_to_fp16)[name = string("_inversed_scores_29_cast_fp16")]; + tensor scores_31_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_29_cast_fp16, cond = mask_11)[name = string("scores_31_cast_fp16")]; + tensor var_2035_cast_fp16 = softmax(axis = var_59, x = scores_31_cast_fp16)[name = string("op_2035_cast_fp16")]; + tensor input_405_cast_fp16 = select(a = var_44_to_fp16, b = var_2035_cast_fp16, cond = mask_11)[name = string("input_405_cast_fp16")]; + bool x_195_transpose_x_0 = const()[name = string("x_195_transpose_x_0"), val = bool(false)]; + bool x_195_transpose_y_0 = const()[name = string("x_195_transpose_y_0"), val = bool(false)]; + tensor value_23_cast_fp16 = transpose(perm = value_23_perm_0, x = v_15_cast_fp16)[name = string("transpose_296")]; + tensor x_195_cast_fp16 = matmul(transpose_x = x_195_transpose_x_0, transpose_y = x_195_transpose_y_0, x = input_405_cast_fp16, y = value_23_cast_fp16)[name = string("x_195_cast_fp16")]; + tensor var_2039_perm_0 = const()[name = string("op_2039_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_2040 = const()[name = string("op_2040"), val = tensor([1, -1, 1024])]; + tensor var_2039_cast_fp16 = transpose(perm = var_2039_perm_0, x = x_195_cast_fp16)[name = string("transpose_295")]; + tensor input_407_cast_fp16 = reshape(shape = var_2040, x = var_2039_cast_fp16)[name = string("input_407_cast_fp16")]; + tensor encoder_layers_7_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160542400))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161328896))))[name = string("encoder_layers_7_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_7_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_7_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161329088)))]; + tensor linear_70_cast_fp16 = linear(bias = encoder_layers_7_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_7_self_attn_linear_out_weight_to_fp16_palettized, x = input_407_cast_fp16)[name = string("linear_70_cast_fp16")]; + tensor input_411_cast_fp16 = add(x = input_401_cast_fp16, y = linear_70_cast_fp16)[name = string("input_411_cast_fp16")]; + tensor x_199_axes_0 = const()[name = string("x_199_axes_0"), val = tensor([-1])]; + tensor encoder_layers_7_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_7_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161331200)))]; + tensor encoder_layers_7_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_7_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161333312)))]; + tensor x_199_cast_fp16 = layer_norm(axes = x_199_axes_0, beta = encoder_layers_7_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_7_norm_conv_weight_to_fp16, x = input_411_cast_fp16)[name = string("x_199_cast_fp16")]; + tensor input_413_perm_0 = const()[name = string("input_413_perm_0"), val = tensor([0, 2, 1])]; + string input_415_pad_type_0 = const()[name = string("input_415_pad_type_0"), val = string("valid")]; + tensor input_415_strides_0 = const()[name = string("input_415_strides_0"), val = tensor([1])]; + tensor input_415_pad_0 = const()[name = string("input_415_pad_0"), val = tensor([0, 0])]; + tensor input_415_dilations_0 = const()[name = string("input_415_dilations_0"), val = tensor([1])]; + int32 input_415_groups_0 = const()[name = string("input_415_groups_0"), val = int32(1)]; + tensor encoder_layers_7_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161335424))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163432640))))[name = string("encoder_layers_7_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_413_cast_fp16 = transpose(perm = input_413_perm_0, x = x_199_cast_fp16)[name = string("transpose_294")]; + tensor input_415_cast_fp16 = conv(dilations = input_415_dilations_0, groups = input_415_groups_0, pad = input_415_pad_0, pad_type = input_415_pad_type_0, strides = input_415_strides_0, weight = encoder_layers_7_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_413_cast_fp16)[name = string("input_415_cast_fp16")]; + int32 x_201_split_num_splits_0 = const()[name = string("x_201_split_num_splits_0"), val = int32(2)]; + int32 x_201_split_axis_0 = const()[name = string("x_201_split_axis_0"), val = int32(1)]; + tensor x_201_split_cast_fp16_0, tensor x_201_split_cast_fp16_1 = split(axis = x_201_split_axis_0, num_splits = x_201_split_num_splits_0, x = input_415_cast_fp16)[name = string("x_201_split_cast_fp16")]; + tensor x_201_split_1_sigmoid_cast_fp16 = sigmoid(x = x_201_split_cast_fp16_1)[name = string("x_201_split_1_sigmoid_cast_fp16")]; + tensor x_201_cast_fp16 = mul(x = x_201_split_cast_fp16_0, y = x_201_split_1_sigmoid_cast_fp16)[name = string("x_201_cast_fp16")]; + tensor input_417_cast_fp16 = select(a = var_44_to_fp16, b = x_201_cast_fp16, cond = var_575)[name = string("input_417_cast_fp16")]; + bool new_x_31_interleave_0 = const()[name = string("new_x_31_interleave_0"), val = bool(false)]; + tensor new_x_31_cast_fp16 = concat(axis = var_59, interleave = new_x_31_interleave_0, values = (cache_31_cast_fp16, input_417_cast_fp16))[name = string("new_x_31_cast_fp16")]; + tensor var_2079_begin_0 = const()[name = string("op_2079_begin_0"), val = tensor([0, 0, 28])]; + tensor var_2079_end_0 = const()[name = string("op_2079_end_0"), val = tensor([1, 1024, 36])]; + tensor var_2079_end_mask_0 = const()[name = string("op_2079_end_mask_0"), val = tensor([true, true, true])]; + tensor var_2079_cast_fp16 = slice_by_index(begin = var_2079_begin_0, end = var_2079_end_0, end_mask = var_2079_end_mask_0, x = new_x_31_cast_fp16)[name = string("op_2079_cast_fp16")]; + string x_203_pad_type_0 = const()[name = string("x_203_pad_type_0"), val = string("valid")]; + int32 x_203_groups_0 = const()[name = string("x_203_groups_0"), val = int32(1024)]; + tensor x_203_strides_0 = const()[name = string("x_203_strides_0"), val = tensor([1])]; + tensor x_203_pad_0 = const()[name = string("x_203_pad_0"), val = tensor([0, 0])]; + tensor x_203_dilations_0 = const()[name = string("x_203_dilations_0"), val = tensor([1])]; + tensor encoder_layers_7_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163436800))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163446080))))[name = string("encoder_layers_7_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_203_cast_fp16 = conv(dilations = x_203_dilations_0, groups = x_203_groups_0, pad = x_203_pad_0, pad_type = x_203_pad_type_0, strides = x_203_strides_0, weight = encoder_layers_7_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_31_cast_fp16)[name = string("x_203_cast_fp16")]; + tensor input_419_perm_0 = const()[name = string("input_419_perm_0"), val = tensor([0, 2, 1])]; + tensor x_205_axes_0 = const()[name = string("x_205_axes_0"), val = tensor([-1])]; + tensor encoder_layers_7_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_7_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163448192)))]; + tensor encoder_layers_7_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_7_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163450304)))]; + tensor input_419_cast_fp16 = transpose(perm = input_419_perm_0, x = x_203_cast_fp16)[name = string("transpose_293")]; + tensor x_205_cast_fp16 = layer_norm(axes = x_205_axes_0, beta = encoder_layers_7_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_7_conv_batch_norm_weight_to_fp16, x = input_419_cast_fp16)[name = string("x_205_cast_fp16")]; + tensor input_421_perm_0 = const()[name = string("input_421_perm_0"), val = tensor([0, 2, 1])]; + tensor input_421_cast_fp16 = transpose(perm = input_421_perm_0, x = x_205_cast_fp16)[name = string("transpose_292")]; + tensor input_423_cast_fp16 = silu(x = input_421_cast_fp16)[name = string("input_423_cast_fp16")]; + string x_207_pad_type_0 = const()[name = string("x_207_pad_type_0"), val = string("valid")]; + tensor x_207_strides_0 = const()[name = string("x_207_strides_0"), val = tensor([1])]; + tensor x_207_pad_0 = const()[name = string("x_207_pad_0"), val = tensor([0, 0])]; + tensor x_207_dilations_0 = const()[name = string("x_207_dilations_0"), val = tensor([1])]; + int32 x_207_groups_0 = const()[name = string("x_207_groups_0"), val = int32(1)]; + tensor encoder_layers_7_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163452416))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(164501056))))[name = string("encoder_layers_7_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_207_cast_fp16 = conv(dilations = x_207_dilations_0, groups = x_207_groups_0, pad = x_207_pad_0, pad_type = x_207_pad_type_0, strides = x_207_strides_0, weight = encoder_layers_7_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_423_cast_fp16)[name = string("x_207_cast_fp16")]; + tensor input_425_perm_0 = const()[name = string("input_425_perm_0"), val = tensor([0, 2, 1])]; + tensor input_425_cast_fp16 = transpose(perm = input_425_perm_0, x = x_207_cast_fp16)[name = string("transpose_291")]; + tensor input_427_cast_fp16 = add(x = input_411_cast_fp16, y = input_425_cast_fp16)[name = string("input_427_cast_fp16")]; + tensor input_429_axes_0 = const()[name = string("input_429_axes_0"), val = tensor([-1])]; + tensor encoder_layers_7_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_7_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(164503168)))]; + tensor encoder_layers_7_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_7_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(164505280)))]; + tensor input_429_cast_fp16 = layer_norm(axes = input_429_axes_0, beta = encoder_layers_7_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_7_norm_feed_forward2_weight_to_fp16, x = input_427_cast_fp16)[name = string("input_429_cast_fp16")]; + tensor encoder_layers_7_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(164507392))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167653184))))[name = string("encoder_layers_7_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_7_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_7_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167653376)))]; + tensor linear_71_cast_fp16 = linear(bias = encoder_layers_7_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_7_feed_forward2_linear1_weight_to_fp16_palettized, x = input_429_cast_fp16)[name = string("linear_71_cast_fp16")]; + tensor input_433_cast_fp16 = silu(x = linear_71_cast_fp16)[name = string("input_433_cast_fp16")]; + tensor encoder_layers_7_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167661632))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170807424))))[name = string("encoder_layers_7_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_7_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_7_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170807616)))]; + tensor linear_72_cast_fp16 = linear(bias = encoder_layers_7_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_7_feed_forward2_linear2_weight_to_fp16_palettized, x = input_433_cast_fp16)[name = string("linear_72_cast_fp16")]; + fp16 var_2122_to_fp16 = const()[name = string("op_2122_to_fp16"), val = fp16(0x1p-1)]; + tensor var_2123_cast_fp16 = mul(x = linear_72_cast_fp16, y = var_2122_to_fp16)[name = string("op_2123_cast_fp16")]; + tensor input_439_cast_fp16 = add(x = input_427_cast_fp16, y = var_2123_cast_fp16)[name = string("input_439_cast_fp16")]; + tensor input_441_axes_0 = const()[name = string("input_441_axes_0"), val = tensor([-1])]; + tensor encoder_layers_7_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_7_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170809728)))]; + tensor encoder_layers_7_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_7_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170811840)))]; + tensor input_441_cast_fp16 = layer_norm(axes = input_441_axes_0, beta = encoder_layers_7_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_7_norm_out_weight_to_fp16, x = input_439_cast_fp16)[name = string("input_441_cast_fp16")]; + tensor cache_33_begin_0 = const()[name = string("cache_33_begin_0"), val = tensor([8, 0, 0, 0])]; + tensor cache_33_end_0 = const()[name = string("cache_33_end_0"), val = tensor([9, 1, 42, 1024])]; + tensor cache_33_end_mask_0 = const()[name = string("cache_33_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_33_squeeze_mask_0 = const()[name = string("cache_33_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_33_cast_fp16 = slice_by_index(begin = cache_33_begin_0, end = cache_33_end_0, end_mask = cache_33_end_mask_0, squeeze_mask = cache_33_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_33_cast_fp16")]; + tensor cache_35_begin_0 = const()[name = string("cache_35_begin_0"), val = tensor([8, 0, 0, 0])]; + tensor cache_35_end_0 = const()[name = string("cache_35_end_0"), val = tensor([9, 1, 1024, 8])]; + tensor cache_35_end_mask_0 = const()[name = string("cache_35_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_35_squeeze_mask_0 = const()[name = string("cache_35_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_35_cast_fp16 = slice_by_index(begin = cache_35_begin_0, end = cache_35_end_0, end_mask = cache_35_end_mask_0, squeeze_mask = cache_35_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_35_cast_fp16")]; + tensor input_443_axes_0 = const()[name = string("input_443_axes_0"), val = tensor([-1])]; + tensor encoder_layers_8_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_8_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170813952)))]; + tensor encoder_layers_8_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_8_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170816064)))]; + tensor input_443_cast_fp16 = layer_norm(axes = input_443_axes_0, beta = encoder_layers_8_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_8_norm_feed_forward1_weight_to_fp16, x = input_441_cast_fp16)[name = string("input_443_cast_fp16")]; + tensor encoder_layers_8_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170818176))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(173963968))))[name = string("encoder_layers_8_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_8_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_8_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(173964160)))]; + tensor linear_73_cast_fp16 = linear(bias = encoder_layers_8_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_8_feed_forward1_linear1_weight_to_fp16_palettized, x = input_443_cast_fp16)[name = string("linear_73_cast_fp16")]; + tensor input_447_cast_fp16 = silu(x = linear_73_cast_fp16)[name = string("input_447_cast_fp16")]; + tensor encoder_layers_8_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(173972416))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(177118208))))[name = string("encoder_layers_8_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_8_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_8_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(177118400)))]; + tensor linear_74_cast_fp16 = linear(bias = encoder_layers_8_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_8_feed_forward1_linear2_weight_to_fp16_palettized, x = input_447_cast_fp16)[name = string("linear_74_cast_fp16")]; + fp16 var_2159_to_fp16 = const()[name = string("op_2159_to_fp16"), val = fp16(0x1p-1)]; + tensor var_2160_cast_fp16 = mul(x = linear_74_cast_fp16, y = var_2159_to_fp16)[name = string("op_2160_cast_fp16")]; + tensor input_453_cast_fp16 = add(x = input_441_cast_fp16, y = var_2160_cast_fp16)[name = string("input_453_cast_fp16")]; + tensor key_17_axes_0 = const()[name = string("key_17_axes_0"), val = tensor([-1])]; + tensor encoder_layers_8_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_8_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(177120512)))]; + tensor encoder_layers_8_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_8_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(177122624)))]; + tensor key_17_cast_fp16 = layer_norm(axes = key_17_axes_0, beta = encoder_layers_8_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_8_norm_self_att_weight_to_fp16, x = input_453_cast_fp16)[name = string("key_17_cast_fp16")]; + bool input_455_interleave_0 = const()[name = string("input_455_interleave_0"), val = bool(false)]; + tensor input_455_cast_fp16 = concat(axis = var_68, interleave = input_455_interleave_0, values = (cache_33_cast_fp16, key_17_cast_fp16))[name = string("input_455_cast_fp16")]; + tensor var_2182_begin_0 = const()[name = string("op_2182_begin_0"), val = tensor([0, 28, 0])]; + tensor var_2182_end_0 = const()[name = string("op_2182_end_0"), val = tensor([1, 42, 1024])]; + tensor var_2182_end_mask_0 = const()[name = string("op_2182_end_mask_0"), val = tensor([true, true, true])]; + tensor var_2182_cast_fp16 = slice_by_index(begin = var_2182_begin_0, end = var_2182_end_0, end_mask = var_2182_end_mask_0, x = cache_33_cast_fp16)[name = string("op_2182_cast_fp16")]; + bool var_2188_interleave_0 = const()[name = string("op_2188_interleave_0"), val = bool(false)]; + tensor var_2188_cast_fp16 = concat(axis = var_68, interleave = var_2188_interleave_0, values = (var_2182_cast_fp16, key_17_cast_fp16))[name = string("op_2188_cast_fp16")]; + tensor encoder_layers_8_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(177124736))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(177911232))))[name = string("encoder_layers_8_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_8_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_8_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(177911424)))]; + tensor linear_75_cast_fp16 = linear(bias = encoder_layers_8_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_8_self_attn_linear_q_weight_to_fp16_palettized, x = key_17_cast_fp16)[name = string("linear_75_cast_fp16")]; + tensor var_2193 = const()[name = string("op_2193"), val = tensor([1, -1, 8, 128])]; + tensor q_49_cast_fp16 = reshape(shape = var_2193, x = linear_75_cast_fp16)[name = string("q_49_cast_fp16")]; + tensor encoder_layers_8_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(177913536))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(178700032))))[name = string("encoder_layers_8_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_8_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_8_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(178700224)))]; + tensor linear_76_cast_fp16 = linear(bias = encoder_layers_8_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_8_self_attn_linear_k_weight_to_fp16_palettized, x = input_455_cast_fp16)[name = string("linear_76_cast_fp16")]; + tensor var_2198 = const()[name = string("op_2198"), val = tensor([1, -1, 8, 128])]; + tensor k_33_cast_fp16 = reshape(shape = var_2198, x = linear_76_cast_fp16)[name = string("k_33_cast_fp16")]; + tensor encoder_layers_8_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(178702336))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179488832))))[name = string("encoder_layers_8_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_8_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_8_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179489024)))]; + tensor linear_77_cast_fp16 = linear(bias = encoder_layers_8_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_8_self_attn_linear_v_weight_to_fp16_palettized, x = input_455_cast_fp16)[name = string("linear_77_cast_fp16")]; + tensor var_2203 = const()[name = string("op_2203"), val = tensor([1, -1, 8, 128])]; + tensor v_17_cast_fp16 = reshape(shape = var_2203, x = linear_77_cast_fp16)[name = string("v_17_cast_fp16")]; + tensor value_25_perm_0 = const()[name = string("value_25_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_8_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_8_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179491136)))]; + tensor var_2216_cast_fp16 = add(x = q_49_cast_fp16, y = encoder_layers_8_self_attn_pos_bias_u_to_fp16)[name = string("op_2216_cast_fp16")]; + tensor encoder_layers_8_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_8_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179493248)))]; + tensor var_2218_cast_fp16 = add(x = q_49_cast_fp16, y = encoder_layers_8_self_attn_pos_bias_v_to_fp16)[name = string("op_2218_cast_fp16")]; + tensor q_with_bias_v_17_perm_0 = const()[name = string("q_with_bias_v_17_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_215_transpose_x_0 = const()[name = string("x_215_transpose_x_0"), val = bool(false)]; + bool x_215_transpose_y_0 = const()[name = string("x_215_transpose_y_0"), val = bool(false)]; + tensor op_2220_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179495360))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179637760))))[name = string("op_2220_to_fp16_quantized")]; + tensor q_with_bias_v_17_cast_fp16 = transpose(perm = q_with_bias_v_17_perm_0, x = var_2218_cast_fp16)[name = string("transpose_290")]; + tensor x_215_cast_fp16 = matmul(transpose_x = x_215_transpose_x_0, transpose_y = x_215_transpose_y_0, x = q_with_bias_v_17_cast_fp16, y = op_2220_to_fp16_quantized)[name = string("x_215_cast_fp16")]; + tensor x_217_pad_0 = const()[name = string("x_217_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_217_mode_0 = const()[name = string("x_217_mode_0"), val = string("constant")]; + fp16 const_183_to_fp16 = const()[name = string("const_183_to_fp16"), val = fp16(0x0p+0)]; + tensor x_217_cast_fp16 = pad(constant_val = const_183_to_fp16, mode = x_217_mode_0, pad = x_217_pad_0, x = x_215_cast_fp16)[name = string("x_217_cast_fp16")]; + tensor var_2228 = const()[name = string("op_2228"), val = tensor([1, 8, -1, 28])]; + tensor x_219_cast_fp16 = reshape(shape = var_2228, x = x_217_cast_fp16)[name = string("x_219_cast_fp16")]; + tensor var_2232_begin_0 = const()[name = string("op_2232_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_2232_end_0 = const()[name = string("op_2232_end_0"), val = tensor([1, 8, 140, 28])]; + tensor var_2232_end_mask_0 = const()[name = string("op_2232_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_2232_cast_fp16 = slice_by_index(begin = var_2232_begin_0, end = var_2232_end_0, end_mask = var_2232_end_mask_0, x = x_219_cast_fp16)[name = string("op_2232_cast_fp16")]; + tensor var_2233 = const()[name = string("op_2233"), val = tensor([1, 8, 28, 139])]; + tensor matrix_bd_33_cast_fp16 = reshape(shape = var_2233, x = var_2232_cast_fp16)[name = string("matrix_bd_33_cast_fp16")]; + bool matrix_ac_17_transpose_x_0 = const()[name = string("matrix_ac_17_transpose_x_0"), val = bool(false)]; + bool matrix_ac_17_transpose_y_0 = const()[name = string("matrix_ac_17_transpose_y_0"), val = bool(false)]; + tensor transpose_112_perm_0 = const()[name = string("transpose_112_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_113_perm_0 = const()[name = string("transpose_113_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_113 = transpose(perm = transpose_113_perm_0, x = k_33_cast_fp16)[name = string("transpose_288")]; + tensor transpose_112 = transpose(perm = transpose_112_perm_0, x = var_2216_cast_fp16)[name = string("transpose_289")]; + tensor matrix_ac_17_cast_fp16 = matmul(transpose_x = matrix_ac_17_transpose_x_0, transpose_y = matrix_ac_17_transpose_y_0, x = transpose_112, y = transpose_113)[name = string("matrix_ac_17_cast_fp16")]; + tensor matrix_bd_35_begin_0 = const()[name = string("matrix_bd_35_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_35_end_0 = const()[name = string("matrix_bd_35_end_0"), val = tensor([1, 8, 28, 70])]; + tensor matrix_bd_35_end_mask_0 = const()[name = string("matrix_bd_35_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_35_cast_fp16 = slice_by_index(begin = matrix_bd_35_begin_0, end = matrix_bd_35_end_0, end_mask = matrix_bd_35_end_mask_0, x = matrix_bd_33_cast_fp16)[name = string("matrix_bd_35_cast_fp16")]; + tensor var_2242_cast_fp16 = add(x = matrix_ac_17_cast_fp16, y = matrix_bd_35_cast_fp16)[name = string("op_2242_cast_fp16")]; + fp16 _inversed_scores_33_y_0_to_fp16 = const()[name = string("_inversed_scores_33_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_33_cast_fp16 = mul(x = var_2242_cast_fp16, y = _inversed_scores_33_y_0_to_fp16)[name = string("_inversed_scores_33_cast_fp16")]; + tensor scores_35_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_33_cast_fp16, cond = mask_11)[name = string("scores_35_cast_fp16")]; + tensor var_2248_cast_fp16 = softmax(axis = var_59, x = scores_35_cast_fp16)[name = string("op_2248_cast_fp16")]; + tensor input_457_cast_fp16 = select(a = var_44_to_fp16, b = var_2248_cast_fp16, cond = mask_11)[name = string("input_457_cast_fp16")]; + bool x_221_transpose_x_0 = const()[name = string("x_221_transpose_x_0"), val = bool(false)]; + bool x_221_transpose_y_0 = const()[name = string("x_221_transpose_y_0"), val = bool(false)]; + tensor value_25_cast_fp16 = transpose(perm = value_25_perm_0, x = v_17_cast_fp16)[name = string("transpose_287")]; + tensor x_221_cast_fp16 = matmul(transpose_x = x_221_transpose_x_0, transpose_y = x_221_transpose_y_0, x = input_457_cast_fp16, y = value_25_cast_fp16)[name = string("x_221_cast_fp16")]; + tensor var_2252_perm_0 = const()[name = string("op_2252_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_2253 = const()[name = string("op_2253"), val = tensor([1, -1, 1024])]; + tensor var_2252_cast_fp16 = transpose(perm = var_2252_perm_0, x = x_221_cast_fp16)[name = string("transpose_286")]; + tensor input_459_cast_fp16 = reshape(shape = var_2253, x = var_2252_cast_fp16)[name = string("input_459_cast_fp16")]; + tensor encoder_layers_8_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179638144))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(180424640))))[name = string("encoder_layers_8_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_8_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_8_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(180424832)))]; + tensor linear_79_cast_fp16 = linear(bias = encoder_layers_8_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_8_self_attn_linear_out_weight_to_fp16_palettized, x = input_459_cast_fp16)[name = string("linear_79_cast_fp16")]; + tensor input_463_cast_fp16 = add(x = input_453_cast_fp16, y = linear_79_cast_fp16)[name = string("input_463_cast_fp16")]; + tensor x_225_axes_0 = const()[name = string("x_225_axes_0"), val = tensor([-1])]; + tensor encoder_layers_8_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_8_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(180426944)))]; + tensor encoder_layers_8_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_8_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(180429056)))]; + tensor x_225_cast_fp16 = layer_norm(axes = x_225_axes_0, beta = encoder_layers_8_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_8_norm_conv_weight_to_fp16, x = input_463_cast_fp16)[name = string("x_225_cast_fp16")]; + tensor input_465_perm_0 = const()[name = string("input_465_perm_0"), val = tensor([0, 2, 1])]; + string input_467_pad_type_0 = const()[name = string("input_467_pad_type_0"), val = string("valid")]; + tensor input_467_strides_0 = const()[name = string("input_467_strides_0"), val = tensor([1])]; + tensor input_467_pad_0 = const()[name = string("input_467_pad_0"), val = tensor([0, 0])]; + tensor input_467_dilations_0 = const()[name = string("input_467_dilations_0"), val = tensor([1])]; + int32 input_467_groups_0 = const()[name = string("input_467_groups_0"), val = int32(1)]; + tensor encoder_layers_8_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(180431168))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182528384))))[name = string("encoder_layers_8_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_465_cast_fp16 = transpose(perm = input_465_perm_0, x = x_225_cast_fp16)[name = string("transpose_285")]; + tensor input_467_cast_fp16 = conv(dilations = input_467_dilations_0, groups = input_467_groups_0, pad = input_467_pad_0, pad_type = input_467_pad_type_0, strides = input_467_strides_0, weight = encoder_layers_8_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_465_cast_fp16)[name = string("input_467_cast_fp16")]; + int32 x_227_split_num_splits_0 = const()[name = string("x_227_split_num_splits_0"), val = int32(2)]; + int32 x_227_split_axis_0 = const()[name = string("x_227_split_axis_0"), val = int32(1)]; + tensor x_227_split_cast_fp16_0, tensor x_227_split_cast_fp16_1 = split(axis = x_227_split_axis_0, num_splits = x_227_split_num_splits_0, x = input_467_cast_fp16)[name = string("x_227_split_cast_fp16")]; + tensor x_227_split_1_sigmoid_cast_fp16 = sigmoid(x = x_227_split_cast_fp16_1)[name = string("x_227_split_1_sigmoid_cast_fp16")]; + tensor x_227_cast_fp16 = mul(x = x_227_split_cast_fp16_0, y = x_227_split_1_sigmoid_cast_fp16)[name = string("x_227_cast_fp16")]; + tensor input_469_cast_fp16 = select(a = var_44_to_fp16, b = x_227_cast_fp16, cond = var_575)[name = string("input_469_cast_fp16")]; + bool new_x_35_interleave_0 = const()[name = string("new_x_35_interleave_0"), val = bool(false)]; + tensor new_x_35_cast_fp16 = concat(axis = var_59, interleave = new_x_35_interleave_0, values = (cache_35_cast_fp16, input_469_cast_fp16))[name = string("new_x_35_cast_fp16")]; + tensor var_2292_begin_0 = const()[name = string("op_2292_begin_0"), val = tensor([0, 0, 28])]; + tensor var_2292_end_0 = const()[name = string("op_2292_end_0"), val = tensor([1, 1024, 36])]; + tensor var_2292_end_mask_0 = const()[name = string("op_2292_end_mask_0"), val = tensor([true, true, true])]; + tensor var_2292_cast_fp16 = slice_by_index(begin = var_2292_begin_0, end = var_2292_end_0, end_mask = var_2292_end_mask_0, x = new_x_35_cast_fp16)[name = string("op_2292_cast_fp16")]; + string x_229_pad_type_0 = const()[name = string("x_229_pad_type_0"), val = string("valid")]; + int32 x_229_groups_0 = const()[name = string("x_229_groups_0"), val = int32(1024)]; + tensor x_229_strides_0 = const()[name = string("x_229_strides_0"), val = tensor([1])]; + tensor x_229_pad_0 = const()[name = string("x_229_pad_0"), val = tensor([0, 0])]; + tensor x_229_dilations_0 = const()[name = string("x_229_dilations_0"), val = tensor([1])]; + tensor encoder_layers_8_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182532544))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182541824))))[name = string("encoder_layers_8_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_229_cast_fp16 = conv(dilations = x_229_dilations_0, groups = x_229_groups_0, pad = x_229_pad_0, pad_type = x_229_pad_type_0, strides = x_229_strides_0, weight = encoder_layers_8_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_35_cast_fp16)[name = string("x_229_cast_fp16")]; + tensor input_471_perm_0 = const()[name = string("input_471_perm_0"), val = tensor([0, 2, 1])]; + tensor x_231_axes_0 = const()[name = string("x_231_axes_0"), val = tensor([-1])]; + tensor encoder_layers_8_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_8_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182543936)))]; + tensor encoder_layers_8_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_8_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182546048)))]; + tensor input_471_cast_fp16 = transpose(perm = input_471_perm_0, x = x_229_cast_fp16)[name = string("transpose_284")]; + tensor x_231_cast_fp16 = layer_norm(axes = x_231_axes_0, beta = encoder_layers_8_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_8_conv_batch_norm_weight_to_fp16, x = input_471_cast_fp16)[name = string("x_231_cast_fp16")]; + tensor input_473_perm_0 = const()[name = string("input_473_perm_0"), val = tensor([0, 2, 1])]; + tensor input_473_cast_fp16 = transpose(perm = input_473_perm_0, x = x_231_cast_fp16)[name = string("transpose_283")]; + tensor input_475_cast_fp16 = silu(x = input_473_cast_fp16)[name = string("input_475_cast_fp16")]; + string x_233_pad_type_0 = const()[name = string("x_233_pad_type_0"), val = string("valid")]; + tensor x_233_strides_0 = const()[name = string("x_233_strides_0"), val = tensor([1])]; + tensor x_233_pad_0 = const()[name = string("x_233_pad_0"), val = tensor([0, 0])]; + tensor x_233_dilations_0 = const()[name = string("x_233_dilations_0"), val = tensor([1])]; + int32 x_233_groups_0 = const()[name = string("x_233_groups_0"), val = int32(1)]; + tensor encoder_layers_8_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182548160))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183596800))))[name = string("encoder_layers_8_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_233_cast_fp16 = conv(dilations = x_233_dilations_0, groups = x_233_groups_0, pad = x_233_pad_0, pad_type = x_233_pad_type_0, strides = x_233_strides_0, weight = encoder_layers_8_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_475_cast_fp16)[name = string("x_233_cast_fp16")]; + tensor input_477_perm_0 = const()[name = string("input_477_perm_0"), val = tensor([0, 2, 1])]; + tensor input_477_cast_fp16 = transpose(perm = input_477_perm_0, x = x_233_cast_fp16)[name = string("transpose_282")]; + tensor input_479_cast_fp16 = add(x = input_463_cast_fp16, y = input_477_cast_fp16)[name = string("input_479_cast_fp16")]; + tensor input_481_axes_0 = const()[name = string("input_481_axes_0"), val = tensor([-1])]; + tensor encoder_layers_8_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_8_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183598912)))]; + tensor encoder_layers_8_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_8_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183601024)))]; + tensor input_481_cast_fp16 = layer_norm(axes = input_481_axes_0, beta = encoder_layers_8_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_8_norm_feed_forward2_weight_to_fp16, x = input_479_cast_fp16)[name = string("input_481_cast_fp16")]; + tensor encoder_layers_8_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183603136))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186748928))))[name = string("encoder_layers_8_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_8_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_8_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186749120)))]; + tensor linear_80_cast_fp16 = linear(bias = encoder_layers_8_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_8_feed_forward2_linear1_weight_to_fp16_palettized, x = input_481_cast_fp16)[name = string("linear_80_cast_fp16")]; + tensor input_485_cast_fp16 = silu(x = linear_80_cast_fp16)[name = string("input_485_cast_fp16")]; + tensor encoder_layers_8_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186757376))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(189903168))))[name = string("encoder_layers_8_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_8_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_8_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(189903360)))]; + tensor linear_81_cast_fp16 = linear(bias = encoder_layers_8_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_8_feed_forward2_linear2_weight_to_fp16_palettized, x = input_485_cast_fp16)[name = string("linear_81_cast_fp16")]; + fp16 var_2335_to_fp16 = const()[name = string("op_2335_to_fp16"), val = fp16(0x1p-1)]; + tensor var_2336_cast_fp16 = mul(x = linear_81_cast_fp16, y = var_2335_to_fp16)[name = string("op_2336_cast_fp16")]; + tensor input_491_cast_fp16 = add(x = input_479_cast_fp16, y = var_2336_cast_fp16)[name = string("input_491_cast_fp16")]; + tensor input_493_axes_0 = const()[name = string("input_493_axes_0"), val = tensor([-1])]; + tensor encoder_layers_8_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_8_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(189905472)))]; + tensor encoder_layers_8_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_8_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(189907584)))]; + tensor input_493_cast_fp16 = layer_norm(axes = input_493_axes_0, beta = encoder_layers_8_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_8_norm_out_weight_to_fp16, x = input_491_cast_fp16)[name = string("input_493_cast_fp16")]; + tensor cache_37_begin_0 = const()[name = string("cache_37_begin_0"), val = tensor([9, 0, 0, 0])]; + tensor cache_37_end_0 = const()[name = string("cache_37_end_0"), val = tensor([10, 1, 42, 1024])]; + tensor cache_37_end_mask_0 = const()[name = string("cache_37_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_37_squeeze_mask_0 = const()[name = string("cache_37_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_37_cast_fp16 = slice_by_index(begin = cache_37_begin_0, end = cache_37_end_0, end_mask = cache_37_end_mask_0, squeeze_mask = cache_37_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_37_cast_fp16")]; + tensor cache_39_begin_0 = const()[name = string("cache_39_begin_0"), val = tensor([9, 0, 0, 0])]; + tensor cache_39_end_0 = const()[name = string("cache_39_end_0"), val = tensor([10, 1, 1024, 8])]; + tensor cache_39_end_mask_0 = const()[name = string("cache_39_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_39_squeeze_mask_0 = const()[name = string("cache_39_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_39_cast_fp16 = slice_by_index(begin = cache_39_begin_0, end = cache_39_end_0, end_mask = cache_39_end_mask_0, squeeze_mask = cache_39_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_39_cast_fp16")]; + tensor input_495_axes_0 = const()[name = string("input_495_axes_0"), val = tensor([-1])]; + tensor encoder_layers_9_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_9_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(189909696)))]; + tensor encoder_layers_9_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_9_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(189911808)))]; + tensor input_495_cast_fp16 = layer_norm(axes = input_495_axes_0, beta = encoder_layers_9_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_9_norm_feed_forward1_weight_to_fp16, x = input_493_cast_fp16)[name = string("input_495_cast_fp16")]; + tensor encoder_layers_9_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(189913920))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(193059712))))[name = string("encoder_layers_9_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_9_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_9_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(193059904)))]; + tensor linear_82_cast_fp16 = linear(bias = encoder_layers_9_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_9_feed_forward1_linear1_weight_to_fp16_palettized, x = input_495_cast_fp16)[name = string("linear_82_cast_fp16")]; + tensor input_499_cast_fp16 = silu(x = linear_82_cast_fp16)[name = string("input_499_cast_fp16")]; + tensor encoder_layers_9_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(193068160))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196213952))))[name = string("encoder_layers_9_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_9_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_9_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196214144)))]; + tensor linear_83_cast_fp16 = linear(bias = encoder_layers_9_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_9_feed_forward1_linear2_weight_to_fp16_palettized, x = input_499_cast_fp16)[name = string("linear_83_cast_fp16")]; + fp16 var_2372_to_fp16 = const()[name = string("op_2372_to_fp16"), val = fp16(0x1p-1)]; + tensor var_2373_cast_fp16 = mul(x = linear_83_cast_fp16, y = var_2372_to_fp16)[name = string("op_2373_cast_fp16")]; + tensor input_505_cast_fp16 = add(x = input_493_cast_fp16, y = var_2373_cast_fp16)[name = string("input_505_cast_fp16")]; + tensor key_19_axes_0 = const()[name = string("key_19_axes_0"), val = tensor([-1])]; + tensor encoder_layers_9_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_9_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196216256)))]; + tensor encoder_layers_9_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_9_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196218368)))]; + tensor key_19_cast_fp16 = layer_norm(axes = key_19_axes_0, beta = encoder_layers_9_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_9_norm_self_att_weight_to_fp16, x = input_505_cast_fp16)[name = string("key_19_cast_fp16")]; + bool input_507_interleave_0 = const()[name = string("input_507_interleave_0"), val = bool(false)]; + tensor input_507_cast_fp16 = concat(axis = var_68, interleave = input_507_interleave_0, values = (cache_37_cast_fp16, key_19_cast_fp16))[name = string("input_507_cast_fp16")]; + tensor var_2395_begin_0 = const()[name = string("op_2395_begin_0"), val = tensor([0, 28, 0])]; + tensor var_2395_end_0 = const()[name = string("op_2395_end_0"), val = tensor([1, 42, 1024])]; + tensor var_2395_end_mask_0 = const()[name = string("op_2395_end_mask_0"), val = tensor([true, true, true])]; + tensor var_2395_cast_fp16 = slice_by_index(begin = var_2395_begin_0, end = var_2395_end_0, end_mask = var_2395_end_mask_0, x = cache_37_cast_fp16)[name = string("op_2395_cast_fp16")]; + bool var_2401_interleave_0 = const()[name = string("op_2401_interleave_0"), val = bool(false)]; + tensor var_2401_cast_fp16 = concat(axis = var_68, interleave = var_2401_interleave_0, values = (var_2395_cast_fp16, key_19_cast_fp16))[name = string("op_2401_cast_fp16")]; + tensor encoder_layers_9_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196220480))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197006976))))[name = string("encoder_layers_9_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_9_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_9_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197007168)))]; + tensor linear_84_cast_fp16 = linear(bias = encoder_layers_9_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_9_self_attn_linear_q_weight_to_fp16_palettized, x = key_19_cast_fp16)[name = string("linear_84_cast_fp16")]; + tensor var_2406 = const()[name = string("op_2406"), val = tensor([1, -1, 8, 128])]; + tensor q_55_cast_fp16 = reshape(shape = var_2406, x = linear_84_cast_fp16)[name = string("q_55_cast_fp16")]; + tensor encoder_layers_9_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197009280))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197795776))))[name = string("encoder_layers_9_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_9_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_9_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197795968)))]; + tensor linear_85_cast_fp16 = linear(bias = encoder_layers_9_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_9_self_attn_linear_k_weight_to_fp16_palettized, x = input_507_cast_fp16)[name = string("linear_85_cast_fp16")]; + tensor var_2411 = const()[name = string("op_2411"), val = tensor([1, -1, 8, 128])]; + tensor k_37_cast_fp16 = reshape(shape = var_2411, x = linear_85_cast_fp16)[name = string("k_37_cast_fp16")]; + tensor encoder_layers_9_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197798080))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198584576))))[name = string("encoder_layers_9_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_9_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_9_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198584768)))]; + tensor linear_86_cast_fp16 = linear(bias = encoder_layers_9_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_9_self_attn_linear_v_weight_to_fp16_palettized, x = input_507_cast_fp16)[name = string("linear_86_cast_fp16")]; + tensor var_2416 = const()[name = string("op_2416"), val = tensor([1, -1, 8, 128])]; + tensor v_19_cast_fp16 = reshape(shape = var_2416, x = linear_86_cast_fp16)[name = string("v_19_cast_fp16")]; + tensor value_27_perm_0 = const()[name = string("value_27_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_9_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_9_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198586880)))]; + tensor var_2429_cast_fp16 = add(x = q_55_cast_fp16, y = encoder_layers_9_self_attn_pos_bias_u_to_fp16)[name = string("op_2429_cast_fp16")]; + tensor encoder_layers_9_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_9_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198588992)))]; + tensor var_2431_cast_fp16 = add(x = q_55_cast_fp16, y = encoder_layers_9_self_attn_pos_bias_v_to_fp16)[name = string("op_2431_cast_fp16")]; + tensor q_with_bias_v_19_perm_0 = const()[name = string("q_with_bias_v_19_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_241_transpose_x_0 = const()[name = string("x_241_transpose_x_0"), val = bool(false)]; + bool x_241_transpose_y_0 = const()[name = string("x_241_transpose_y_0"), val = bool(false)]; + tensor op_2433_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198591104))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198733504))))[name = string("op_2433_to_fp16_quantized")]; + tensor q_with_bias_v_19_cast_fp16 = transpose(perm = q_with_bias_v_19_perm_0, x = var_2431_cast_fp16)[name = string("transpose_281")]; + tensor x_241_cast_fp16 = matmul(transpose_x = x_241_transpose_x_0, transpose_y = x_241_transpose_y_0, x = q_with_bias_v_19_cast_fp16, y = op_2433_to_fp16_quantized)[name = string("x_241_cast_fp16")]; + tensor x_243_pad_0 = const()[name = string("x_243_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_243_mode_0 = const()[name = string("x_243_mode_0"), val = string("constant")]; + fp16 const_196_to_fp16 = const()[name = string("const_196_to_fp16"), val = fp16(0x0p+0)]; + tensor x_243_cast_fp16 = pad(constant_val = const_196_to_fp16, mode = x_243_mode_0, pad = x_243_pad_0, x = x_241_cast_fp16)[name = string("x_243_cast_fp16")]; + tensor var_2441 = const()[name = string("op_2441"), val = tensor([1, 8, -1, 28])]; + tensor x_245_cast_fp16 = reshape(shape = var_2441, x = x_243_cast_fp16)[name = string("x_245_cast_fp16")]; + tensor var_2445_begin_0 = const()[name = string("op_2445_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_2445_end_0 = const()[name = string("op_2445_end_0"), val = tensor([1, 8, 140, 28])]; + tensor var_2445_end_mask_0 = const()[name = string("op_2445_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_2445_cast_fp16 = slice_by_index(begin = var_2445_begin_0, end = var_2445_end_0, end_mask = var_2445_end_mask_0, x = x_245_cast_fp16)[name = string("op_2445_cast_fp16")]; + tensor var_2446 = const()[name = string("op_2446"), val = tensor([1, 8, 28, 139])]; + tensor matrix_bd_37_cast_fp16 = reshape(shape = var_2446, x = var_2445_cast_fp16)[name = string("matrix_bd_37_cast_fp16")]; + bool matrix_ac_19_transpose_x_0 = const()[name = string("matrix_ac_19_transpose_x_0"), val = bool(false)]; + bool matrix_ac_19_transpose_y_0 = const()[name = string("matrix_ac_19_transpose_y_0"), val = bool(false)]; + tensor transpose_114_perm_0 = const()[name = string("transpose_114_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_115_perm_0 = const()[name = string("transpose_115_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_115 = transpose(perm = transpose_115_perm_0, x = k_37_cast_fp16)[name = string("transpose_279")]; + tensor transpose_114 = transpose(perm = transpose_114_perm_0, x = var_2429_cast_fp16)[name = string("transpose_280")]; + tensor matrix_ac_19_cast_fp16 = matmul(transpose_x = matrix_ac_19_transpose_x_0, transpose_y = matrix_ac_19_transpose_y_0, x = transpose_114, y = transpose_115)[name = string("matrix_ac_19_cast_fp16")]; + tensor matrix_bd_39_begin_0 = const()[name = string("matrix_bd_39_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_39_end_0 = const()[name = string("matrix_bd_39_end_0"), val = tensor([1, 8, 28, 70])]; + tensor matrix_bd_39_end_mask_0 = const()[name = string("matrix_bd_39_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_39_cast_fp16 = slice_by_index(begin = matrix_bd_39_begin_0, end = matrix_bd_39_end_0, end_mask = matrix_bd_39_end_mask_0, x = matrix_bd_37_cast_fp16)[name = string("matrix_bd_39_cast_fp16")]; + tensor var_2455_cast_fp16 = add(x = matrix_ac_19_cast_fp16, y = matrix_bd_39_cast_fp16)[name = string("op_2455_cast_fp16")]; + fp16 _inversed_scores_37_y_0_to_fp16 = const()[name = string("_inversed_scores_37_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_37_cast_fp16 = mul(x = var_2455_cast_fp16, y = _inversed_scores_37_y_0_to_fp16)[name = string("_inversed_scores_37_cast_fp16")]; + tensor scores_39_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_37_cast_fp16, cond = mask_11)[name = string("scores_39_cast_fp16")]; + tensor var_2461_cast_fp16 = softmax(axis = var_59, x = scores_39_cast_fp16)[name = string("op_2461_cast_fp16")]; + tensor input_509_cast_fp16 = select(a = var_44_to_fp16, b = var_2461_cast_fp16, cond = mask_11)[name = string("input_509_cast_fp16")]; + bool x_247_transpose_x_0 = const()[name = string("x_247_transpose_x_0"), val = bool(false)]; + bool x_247_transpose_y_0 = const()[name = string("x_247_transpose_y_0"), val = bool(false)]; + tensor value_27_cast_fp16 = transpose(perm = value_27_perm_0, x = v_19_cast_fp16)[name = string("transpose_278")]; + tensor x_247_cast_fp16 = matmul(transpose_x = x_247_transpose_x_0, transpose_y = x_247_transpose_y_0, x = input_509_cast_fp16, y = value_27_cast_fp16)[name = string("x_247_cast_fp16")]; + tensor var_2465_perm_0 = const()[name = string("op_2465_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_2466 = const()[name = string("op_2466"), val = tensor([1, -1, 1024])]; + tensor var_2465_cast_fp16 = transpose(perm = var_2465_perm_0, x = x_247_cast_fp16)[name = string("transpose_277")]; + tensor input_511_cast_fp16 = reshape(shape = var_2466, x = var_2465_cast_fp16)[name = string("input_511_cast_fp16")]; + tensor encoder_layers_9_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198733888))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199520384))))[name = string("encoder_layers_9_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_9_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_9_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199520576)))]; + tensor linear_88_cast_fp16 = linear(bias = encoder_layers_9_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_9_self_attn_linear_out_weight_to_fp16_palettized, x = input_511_cast_fp16)[name = string("linear_88_cast_fp16")]; + tensor input_515_cast_fp16 = add(x = input_505_cast_fp16, y = linear_88_cast_fp16)[name = string("input_515_cast_fp16")]; + tensor x_251_axes_0 = const()[name = string("x_251_axes_0"), val = tensor([-1])]; + tensor encoder_layers_9_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_9_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199522688)))]; + tensor encoder_layers_9_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_9_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199524800)))]; + tensor x_251_cast_fp16 = layer_norm(axes = x_251_axes_0, beta = encoder_layers_9_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_9_norm_conv_weight_to_fp16, x = input_515_cast_fp16)[name = string("x_251_cast_fp16")]; + tensor input_517_perm_0 = const()[name = string("input_517_perm_0"), val = tensor([0, 2, 1])]; + string input_519_pad_type_0 = const()[name = string("input_519_pad_type_0"), val = string("valid")]; + tensor input_519_strides_0 = const()[name = string("input_519_strides_0"), val = tensor([1])]; + tensor input_519_pad_0 = const()[name = string("input_519_pad_0"), val = tensor([0, 0])]; + tensor input_519_dilations_0 = const()[name = string("input_519_dilations_0"), val = tensor([1])]; + int32 input_519_groups_0 = const()[name = string("input_519_groups_0"), val = int32(1)]; + tensor encoder_layers_9_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199526912))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(201624128))))[name = string("encoder_layers_9_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_517_cast_fp16 = transpose(perm = input_517_perm_0, x = x_251_cast_fp16)[name = string("transpose_276")]; + tensor input_519_cast_fp16 = conv(dilations = input_519_dilations_0, groups = input_519_groups_0, pad = input_519_pad_0, pad_type = input_519_pad_type_0, strides = input_519_strides_0, weight = encoder_layers_9_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_517_cast_fp16)[name = string("input_519_cast_fp16")]; + int32 x_253_split_num_splits_0 = const()[name = string("x_253_split_num_splits_0"), val = int32(2)]; + int32 x_253_split_axis_0 = const()[name = string("x_253_split_axis_0"), val = int32(1)]; + tensor x_253_split_cast_fp16_0, tensor x_253_split_cast_fp16_1 = split(axis = x_253_split_axis_0, num_splits = x_253_split_num_splits_0, x = input_519_cast_fp16)[name = string("x_253_split_cast_fp16")]; + tensor x_253_split_1_sigmoid_cast_fp16 = sigmoid(x = x_253_split_cast_fp16_1)[name = string("x_253_split_1_sigmoid_cast_fp16")]; + tensor x_253_cast_fp16 = mul(x = x_253_split_cast_fp16_0, y = x_253_split_1_sigmoid_cast_fp16)[name = string("x_253_cast_fp16")]; + tensor input_521_cast_fp16 = select(a = var_44_to_fp16, b = x_253_cast_fp16, cond = var_575)[name = string("input_521_cast_fp16")]; + bool new_x_39_interleave_0 = const()[name = string("new_x_39_interleave_0"), val = bool(false)]; + tensor new_x_39_cast_fp16 = concat(axis = var_59, interleave = new_x_39_interleave_0, values = (cache_39_cast_fp16, input_521_cast_fp16))[name = string("new_x_39_cast_fp16")]; + tensor var_2505_begin_0 = const()[name = string("op_2505_begin_0"), val = tensor([0, 0, 28])]; + tensor var_2505_end_0 = const()[name = string("op_2505_end_0"), val = tensor([1, 1024, 36])]; + tensor var_2505_end_mask_0 = const()[name = string("op_2505_end_mask_0"), val = tensor([true, true, true])]; + tensor var_2505_cast_fp16 = slice_by_index(begin = var_2505_begin_0, end = var_2505_end_0, end_mask = var_2505_end_mask_0, x = new_x_39_cast_fp16)[name = string("op_2505_cast_fp16")]; + string x_255_pad_type_0 = const()[name = string("x_255_pad_type_0"), val = string("valid")]; + int32 x_255_groups_0 = const()[name = string("x_255_groups_0"), val = int32(1024)]; + tensor x_255_strides_0 = const()[name = string("x_255_strides_0"), val = tensor([1])]; + tensor x_255_pad_0 = const()[name = string("x_255_pad_0"), val = tensor([0, 0])]; + tensor x_255_dilations_0 = const()[name = string("x_255_dilations_0"), val = tensor([1])]; + tensor encoder_layers_9_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(201628288))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(201637568))))[name = string("encoder_layers_9_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_255_cast_fp16 = conv(dilations = x_255_dilations_0, groups = x_255_groups_0, pad = x_255_pad_0, pad_type = x_255_pad_type_0, strides = x_255_strides_0, weight = encoder_layers_9_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_39_cast_fp16)[name = string("x_255_cast_fp16")]; + tensor input_523_perm_0 = const()[name = string("input_523_perm_0"), val = tensor([0, 2, 1])]; + tensor x_257_axes_0 = const()[name = string("x_257_axes_0"), val = tensor([-1])]; + tensor encoder_layers_9_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_9_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(201639680)))]; + tensor encoder_layers_9_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_9_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(201641792)))]; + tensor input_523_cast_fp16 = transpose(perm = input_523_perm_0, x = x_255_cast_fp16)[name = string("transpose_275")]; + tensor x_257_cast_fp16 = layer_norm(axes = x_257_axes_0, beta = encoder_layers_9_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_9_conv_batch_norm_weight_to_fp16, x = input_523_cast_fp16)[name = string("x_257_cast_fp16")]; + tensor input_525_perm_0 = const()[name = string("input_525_perm_0"), val = tensor([0, 2, 1])]; + tensor input_525_cast_fp16 = transpose(perm = input_525_perm_0, x = x_257_cast_fp16)[name = string("transpose_274")]; + tensor input_527_cast_fp16 = silu(x = input_525_cast_fp16)[name = string("input_527_cast_fp16")]; + string x_259_pad_type_0 = const()[name = string("x_259_pad_type_0"), val = string("valid")]; + tensor x_259_strides_0 = const()[name = string("x_259_strides_0"), val = tensor([1])]; + tensor x_259_pad_0 = const()[name = string("x_259_pad_0"), val = tensor([0, 0])]; + tensor x_259_dilations_0 = const()[name = string("x_259_dilations_0"), val = tensor([1])]; + int32 x_259_groups_0 = const()[name = string("x_259_groups_0"), val = int32(1)]; + tensor encoder_layers_9_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(201643904))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202692544))))[name = string("encoder_layers_9_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_259_cast_fp16 = conv(dilations = x_259_dilations_0, groups = x_259_groups_0, pad = x_259_pad_0, pad_type = x_259_pad_type_0, strides = x_259_strides_0, weight = encoder_layers_9_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_527_cast_fp16)[name = string("x_259_cast_fp16")]; + tensor input_529_perm_0 = const()[name = string("input_529_perm_0"), val = tensor([0, 2, 1])]; + tensor input_529_cast_fp16 = transpose(perm = input_529_perm_0, x = x_259_cast_fp16)[name = string("transpose_273")]; + tensor input_531_cast_fp16 = add(x = input_515_cast_fp16, y = input_529_cast_fp16)[name = string("input_531_cast_fp16")]; + tensor input_533_axes_0 = const()[name = string("input_533_axes_0"), val = tensor([-1])]; + tensor encoder_layers_9_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_9_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202694656)))]; + tensor encoder_layers_9_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_9_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202696768)))]; + tensor input_533_cast_fp16 = layer_norm(axes = input_533_axes_0, beta = encoder_layers_9_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_9_norm_feed_forward2_weight_to_fp16, x = input_531_cast_fp16)[name = string("input_533_cast_fp16")]; + tensor encoder_layers_9_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202698880))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(205844672))))[name = string("encoder_layers_9_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_9_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_9_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(205844864)))]; + tensor linear_89_cast_fp16 = linear(bias = encoder_layers_9_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_9_feed_forward2_linear1_weight_to_fp16_palettized, x = input_533_cast_fp16)[name = string("linear_89_cast_fp16")]; + tensor input_537_cast_fp16 = silu(x = linear_89_cast_fp16)[name = string("input_537_cast_fp16")]; + tensor encoder_layers_9_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(205853120))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(208998912))))[name = string("encoder_layers_9_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_9_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_9_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(208999104)))]; + tensor linear_90_cast_fp16 = linear(bias = encoder_layers_9_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_9_feed_forward2_linear2_weight_to_fp16_palettized, x = input_537_cast_fp16)[name = string("linear_90_cast_fp16")]; + fp16 var_2548_to_fp16 = const()[name = string("op_2548_to_fp16"), val = fp16(0x1p-1)]; + tensor var_2549_cast_fp16 = mul(x = linear_90_cast_fp16, y = var_2548_to_fp16)[name = string("op_2549_cast_fp16")]; + tensor input_543_cast_fp16 = add(x = input_531_cast_fp16, y = var_2549_cast_fp16)[name = string("input_543_cast_fp16")]; + tensor input_545_axes_0 = const()[name = string("input_545_axes_0"), val = tensor([-1])]; + tensor encoder_layers_9_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_9_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(209001216)))]; + tensor encoder_layers_9_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_9_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(209003328)))]; + tensor input_545_cast_fp16 = layer_norm(axes = input_545_axes_0, beta = encoder_layers_9_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_9_norm_out_weight_to_fp16, x = input_543_cast_fp16)[name = string("input_545_cast_fp16")]; + tensor cache_41_begin_0 = const()[name = string("cache_41_begin_0"), val = tensor([10, 0, 0, 0])]; + tensor cache_41_end_0 = const()[name = string("cache_41_end_0"), val = tensor([11, 1, 42, 1024])]; + tensor cache_41_end_mask_0 = const()[name = string("cache_41_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_41_squeeze_mask_0 = const()[name = string("cache_41_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_41_cast_fp16 = slice_by_index(begin = cache_41_begin_0, end = cache_41_end_0, end_mask = cache_41_end_mask_0, squeeze_mask = cache_41_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_41_cast_fp16")]; + tensor cache_43_begin_0 = const()[name = string("cache_43_begin_0"), val = tensor([10, 0, 0, 0])]; + tensor cache_43_end_0 = const()[name = string("cache_43_end_0"), val = tensor([11, 1, 1024, 8])]; + tensor cache_43_end_mask_0 = const()[name = string("cache_43_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_43_squeeze_mask_0 = const()[name = string("cache_43_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_43_cast_fp16 = slice_by_index(begin = cache_43_begin_0, end = cache_43_end_0, end_mask = cache_43_end_mask_0, squeeze_mask = cache_43_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_43_cast_fp16")]; + tensor input_547_axes_0 = const()[name = string("input_547_axes_0"), val = tensor([-1])]; + tensor encoder_layers_10_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_10_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(209005440)))]; + tensor encoder_layers_10_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_10_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(209007552)))]; + tensor input_547_cast_fp16 = layer_norm(axes = input_547_axes_0, beta = encoder_layers_10_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_10_norm_feed_forward1_weight_to_fp16, x = input_545_cast_fp16)[name = string("input_547_cast_fp16")]; + tensor encoder_layers_10_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(209009664))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212155456))))[name = string("encoder_layers_10_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_10_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_10_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212155648)))]; + tensor linear_91_cast_fp16 = linear(bias = encoder_layers_10_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_10_feed_forward1_linear1_weight_to_fp16_palettized, x = input_547_cast_fp16)[name = string("linear_91_cast_fp16")]; + tensor input_551_cast_fp16 = silu(x = linear_91_cast_fp16)[name = string("input_551_cast_fp16")]; + tensor encoder_layers_10_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212163904))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215309696))))[name = string("encoder_layers_10_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_10_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_10_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215309888)))]; + tensor linear_92_cast_fp16 = linear(bias = encoder_layers_10_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_10_feed_forward1_linear2_weight_to_fp16_palettized, x = input_551_cast_fp16)[name = string("linear_92_cast_fp16")]; + fp16 var_2585_to_fp16 = const()[name = string("op_2585_to_fp16"), val = fp16(0x1p-1)]; + tensor var_2586_cast_fp16 = mul(x = linear_92_cast_fp16, y = var_2585_to_fp16)[name = string("op_2586_cast_fp16")]; + tensor input_557_cast_fp16 = add(x = input_545_cast_fp16, y = var_2586_cast_fp16)[name = string("input_557_cast_fp16")]; + tensor key_21_axes_0 = const()[name = string("key_21_axes_0"), val = tensor([-1])]; + tensor encoder_layers_10_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_10_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215312000)))]; + tensor encoder_layers_10_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_10_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215314112)))]; + tensor key_21_cast_fp16 = layer_norm(axes = key_21_axes_0, beta = encoder_layers_10_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_10_norm_self_att_weight_to_fp16, x = input_557_cast_fp16)[name = string("key_21_cast_fp16")]; + bool input_559_interleave_0 = const()[name = string("input_559_interleave_0"), val = bool(false)]; + tensor input_559_cast_fp16 = concat(axis = var_68, interleave = input_559_interleave_0, values = (cache_41_cast_fp16, key_21_cast_fp16))[name = string("input_559_cast_fp16")]; + tensor var_2608_begin_0 = const()[name = string("op_2608_begin_0"), val = tensor([0, 28, 0])]; + tensor var_2608_end_0 = const()[name = string("op_2608_end_0"), val = tensor([1, 42, 1024])]; + tensor var_2608_end_mask_0 = const()[name = string("op_2608_end_mask_0"), val = tensor([true, true, true])]; + tensor var_2608_cast_fp16 = slice_by_index(begin = var_2608_begin_0, end = var_2608_end_0, end_mask = var_2608_end_mask_0, x = cache_41_cast_fp16)[name = string("op_2608_cast_fp16")]; + bool var_2614_interleave_0 = const()[name = string("op_2614_interleave_0"), val = bool(false)]; + tensor var_2614_cast_fp16 = concat(axis = var_68, interleave = var_2614_interleave_0, values = (var_2608_cast_fp16, key_21_cast_fp16))[name = string("op_2614_cast_fp16")]; + tensor encoder_layers_10_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215316224))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216102720))))[name = string("encoder_layers_10_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_10_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_10_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216102912)))]; + tensor linear_93_cast_fp16 = linear(bias = encoder_layers_10_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_10_self_attn_linear_q_weight_to_fp16_palettized, x = key_21_cast_fp16)[name = string("linear_93_cast_fp16")]; + tensor var_2619 = const()[name = string("op_2619"), val = tensor([1, -1, 8, 128])]; + tensor q_61_cast_fp16 = reshape(shape = var_2619, x = linear_93_cast_fp16)[name = string("q_61_cast_fp16")]; + tensor encoder_layers_10_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216105024))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216891520))))[name = string("encoder_layers_10_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_10_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_10_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216891712)))]; + tensor linear_94_cast_fp16 = linear(bias = encoder_layers_10_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_10_self_attn_linear_k_weight_to_fp16_palettized, x = input_559_cast_fp16)[name = string("linear_94_cast_fp16")]; + tensor var_2624 = const()[name = string("op_2624"), val = tensor([1, -1, 8, 128])]; + tensor k_41_cast_fp16 = reshape(shape = var_2624, x = linear_94_cast_fp16)[name = string("k_41_cast_fp16")]; + tensor encoder_layers_10_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216893824))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217680320))))[name = string("encoder_layers_10_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_10_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_10_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217680512)))]; + tensor linear_95_cast_fp16 = linear(bias = encoder_layers_10_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_10_self_attn_linear_v_weight_to_fp16_palettized, x = input_559_cast_fp16)[name = string("linear_95_cast_fp16")]; + tensor var_2629 = const()[name = string("op_2629"), val = tensor([1, -1, 8, 128])]; + tensor v_21_cast_fp16 = reshape(shape = var_2629, x = linear_95_cast_fp16)[name = string("v_21_cast_fp16")]; + tensor value_29_perm_0 = const()[name = string("value_29_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_10_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_10_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217682624)))]; + tensor var_2642_cast_fp16 = add(x = q_61_cast_fp16, y = encoder_layers_10_self_attn_pos_bias_u_to_fp16)[name = string("op_2642_cast_fp16")]; + tensor encoder_layers_10_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_10_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217684736)))]; + tensor var_2644_cast_fp16 = add(x = q_61_cast_fp16, y = encoder_layers_10_self_attn_pos_bias_v_to_fp16)[name = string("op_2644_cast_fp16")]; + tensor q_with_bias_v_21_perm_0 = const()[name = string("q_with_bias_v_21_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_267_transpose_x_0 = const()[name = string("x_267_transpose_x_0"), val = bool(false)]; + bool x_267_transpose_y_0 = const()[name = string("x_267_transpose_y_0"), val = bool(false)]; + tensor op_2646_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217686848))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217829248))))[name = string("op_2646_to_fp16_quantized")]; + tensor q_with_bias_v_21_cast_fp16 = transpose(perm = q_with_bias_v_21_perm_0, x = var_2644_cast_fp16)[name = string("transpose_272")]; + tensor x_267_cast_fp16 = matmul(transpose_x = x_267_transpose_x_0, transpose_y = x_267_transpose_y_0, x = q_with_bias_v_21_cast_fp16, y = op_2646_to_fp16_quantized)[name = string("x_267_cast_fp16")]; + tensor x_269_pad_0 = const()[name = string("x_269_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_269_mode_0 = const()[name = string("x_269_mode_0"), val = string("constant")]; + fp16 const_209_to_fp16 = const()[name = string("const_209_to_fp16"), val = fp16(0x0p+0)]; + tensor x_269_cast_fp16 = pad(constant_val = const_209_to_fp16, mode = x_269_mode_0, pad = x_269_pad_0, x = x_267_cast_fp16)[name = string("x_269_cast_fp16")]; + tensor var_2654 = const()[name = string("op_2654"), val = tensor([1, 8, -1, 28])]; + tensor x_271_cast_fp16 = reshape(shape = var_2654, x = x_269_cast_fp16)[name = string("x_271_cast_fp16")]; + tensor var_2658_begin_0 = const()[name = string("op_2658_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_2658_end_0 = const()[name = string("op_2658_end_0"), val = tensor([1, 8, 140, 28])]; + tensor var_2658_end_mask_0 = const()[name = string("op_2658_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_2658_cast_fp16 = slice_by_index(begin = var_2658_begin_0, end = var_2658_end_0, end_mask = var_2658_end_mask_0, x = x_271_cast_fp16)[name = string("op_2658_cast_fp16")]; + tensor var_2659 = const()[name = string("op_2659"), val = tensor([1, 8, 28, 139])]; + tensor matrix_bd_41_cast_fp16 = reshape(shape = var_2659, x = var_2658_cast_fp16)[name = string("matrix_bd_41_cast_fp16")]; + bool matrix_ac_21_transpose_x_0 = const()[name = string("matrix_ac_21_transpose_x_0"), val = bool(false)]; + bool matrix_ac_21_transpose_y_0 = const()[name = string("matrix_ac_21_transpose_y_0"), val = bool(false)]; + tensor transpose_116_perm_0 = const()[name = string("transpose_116_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_117_perm_0 = const()[name = string("transpose_117_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_117 = transpose(perm = transpose_117_perm_0, x = k_41_cast_fp16)[name = string("transpose_270")]; + tensor transpose_116 = transpose(perm = transpose_116_perm_0, x = var_2642_cast_fp16)[name = string("transpose_271")]; + tensor matrix_ac_21_cast_fp16 = matmul(transpose_x = matrix_ac_21_transpose_x_0, transpose_y = matrix_ac_21_transpose_y_0, x = transpose_116, y = transpose_117)[name = string("matrix_ac_21_cast_fp16")]; + tensor matrix_bd_43_begin_0 = const()[name = string("matrix_bd_43_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_43_end_0 = const()[name = string("matrix_bd_43_end_0"), val = tensor([1, 8, 28, 70])]; + tensor matrix_bd_43_end_mask_0 = const()[name = string("matrix_bd_43_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_43_cast_fp16 = slice_by_index(begin = matrix_bd_43_begin_0, end = matrix_bd_43_end_0, end_mask = matrix_bd_43_end_mask_0, x = matrix_bd_41_cast_fp16)[name = string("matrix_bd_43_cast_fp16")]; + tensor var_2668_cast_fp16 = add(x = matrix_ac_21_cast_fp16, y = matrix_bd_43_cast_fp16)[name = string("op_2668_cast_fp16")]; + fp16 _inversed_scores_41_y_0_to_fp16 = const()[name = string("_inversed_scores_41_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_41_cast_fp16 = mul(x = var_2668_cast_fp16, y = _inversed_scores_41_y_0_to_fp16)[name = string("_inversed_scores_41_cast_fp16")]; + tensor scores_43_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_41_cast_fp16, cond = mask_11)[name = string("scores_43_cast_fp16")]; + tensor var_2674_cast_fp16 = softmax(axis = var_59, x = scores_43_cast_fp16)[name = string("op_2674_cast_fp16")]; + tensor input_561_cast_fp16 = select(a = var_44_to_fp16, b = var_2674_cast_fp16, cond = mask_11)[name = string("input_561_cast_fp16")]; + bool x_273_transpose_x_0 = const()[name = string("x_273_transpose_x_0"), val = bool(false)]; + bool x_273_transpose_y_0 = const()[name = string("x_273_transpose_y_0"), val = bool(false)]; + tensor value_29_cast_fp16 = transpose(perm = value_29_perm_0, x = v_21_cast_fp16)[name = string("transpose_269")]; + tensor x_273_cast_fp16 = matmul(transpose_x = x_273_transpose_x_0, transpose_y = x_273_transpose_y_0, x = input_561_cast_fp16, y = value_29_cast_fp16)[name = string("x_273_cast_fp16")]; + tensor var_2678_perm_0 = const()[name = string("op_2678_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_2679 = const()[name = string("op_2679"), val = tensor([1, -1, 1024])]; + tensor var_2678_cast_fp16 = transpose(perm = var_2678_perm_0, x = x_273_cast_fp16)[name = string("transpose_268")]; + tensor input_563_cast_fp16 = reshape(shape = var_2679, x = var_2678_cast_fp16)[name = string("input_563_cast_fp16")]; + tensor encoder_layers_10_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217829632))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218616128))))[name = string("encoder_layers_10_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_10_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_10_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218616320)))]; + tensor linear_97_cast_fp16 = linear(bias = encoder_layers_10_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_10_self_attn_linear_out_weight_to_fp16_palettized, x = input_563_cast_fp16)[name = string("linear_97_cast_fp16")]; + tensor input_567_cast_fp16 = add(x = input_557_cast_fp16, y = linear_97_cast_fp16)[name = string("input_567_cast_fp16")]; + tensor x_277_axes_0 = const()[name = string("x_277_axes_0"), val = tensor([-1])]; + tensor encoder_layers_10_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_10_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218618432)))]; + tensor encoder_layers_10_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_10_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218620544)))]; + tensor x_277_cast_fp16 = layer_norm(axes = x_277_axes_0, beta = encoder_layers_10_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_10_norm_conv_weight_to_fp16, x = input_567_cast_fp16)[name = string("x_277_cast_fp16")]; + tensor input_569_perm_0 = const()[name = string("input_569_perm_0"), val = tensor([0, 2, 1])]; + string input_571_pad_type_0 = const()[name = string("input_571_pad_type_0"), val = string("valid")]; + tensor input_571_strides_0 = const()[name = string("input_571_strides_0"), val = tensor([1])]; + tensor input_571_pad_0 = const()[name = string("input_571_pad_0"), val = tensor([0, 0])]; + tensor input_571_dilations_0 = const()[name = string("input_571_dilations_0"), val = tensor([1])]; + int32 input_571_groups_0 = const()[name = string("input_571_groups_0"), val = int32(1)]; + tensor encoder_layers_10_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218622656))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220719872))))[name = string("encoder_layers_10_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_569_cast_fp16 = transpose(perm = input_569_perm_0, x = x_277_cast_fp16)[name = string("transpose_267")]; + tensor input_571_cast_fp16 = conv(dilations = input_571_dilations_0, groups = input_571_groups_0, pad = input_571_pad_0, pad_type = input_571_pad_type_0, strides = input_571_strides_0, weight = encoder_layers_10_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_569_cast_fp16)[name = string("input_571_cast_fp16")]; + int32 x_279_split_num_splits_0 = const()[name = string("x_279_split_num_splits_0"), val = int32(2)]; + int32 x_279_split_axis_0 = const()[name = string("x_279_split_axis_0"), val = int32(1)]; + tensor x_279_split_cast_fp16_0, tensor x_279_split_cast_fp16_1 = split(axis = x_279_split_axis_0, num_splits = x_279_split_num_splits_0, x = input_571_cast_fp16)[name = string("x_279_split_cast_fp16")]; + tensor x_279_split_1_sigmoid_cast_fp16 = sigmoid(x = x_279_split_cast_fp16_1)[name = string("x_279_split_1_sigmoid_cast_fp16")]; + tensor x_279_cast_fp16 = mul(x = x_279_split_cast_fp16_0, y = x_279_split_1_sigmoid_cast_fp16)[name = string("x_279_cast_fp16")]; + tensor input_573_cast_fp16 = select(a = var_44_to_fp16, b = x_279_cast_fp16, cond = var_575)[name = string("input_573_cast_fp16")]; + bool new_x_43_interleave_0 = const()[name = string("new_x_43_interleave_0"), val = bool(false)]; + tensor new_x_43_cast_fp16 = concat(axis = var_59, interleave = new_x_43_interleave_0, values = (cache_43_cast_fp16, input_573_cast_fp16))[name = string("new_x_43_cast_fp16")]; + tensor var_2718_begin_0 = const()[name = string("op_2718_begin_0"), val = tensor([0, 0, 28])]; + tensor var_2718_end_0 = const()[name = string("op_2718_end_0"), val = tensor([1, 1024, 36])]; + tensor var_2718_end_mask_0 = const()[name = string("op_2718_end_mask_0"), val = tensor([true, true, true])]; + tensor var_2718_cast_fp16 = slice_by_index(begin = var_2718_begin_0, end = var_2718_end_0, end_mask = var_2718_end_mask_0, x = new_x_43_cast_fp16)[name = string("op_2718_cast_fp16")]; + string x_281_pad_type_0 = const()[name = string("x_281_pad_type_0"), val = string("valid")]; + int32 x_281_groups_0 = const()[name = string("x_281_groups_0"), val = int32(1024)]; + tensor x_281_strides_0 = const()[name = string("x_281_strides_0"), val = tensor([1])]; + tensor x_281_pad_0 = const()[name = string("x_281_pad_0"), val = tensor([0, 0])]; + tensor x_281_dilations_0 = const()[name = string("x_281_dilations_0"), val = tensor([1])]; + tensor encoder_layers_10_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220724032))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220733312))))[name = string("encoder_layers_10_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_281_cast_fp16 = conv(dilations = x_281_dilations_0, groups = x_281_groups_0, pad = x_281_pad_0, pad_type = x_281_pad_type_0, strides = x_281_strides_0, weight = encoder_layers_10_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_43_cast_fp16)[name = string("x_281_cast_fp16")]; + tensor input_575_perm_0 = const()[name = string("input_575_perm_0"), val = tensor([0, 2, 1])]; + tensor x_283_axes_0 = const()[name = string("x_283_axes_0"), val = tensor([-1])]; + tensor encoder_layers_10_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_10_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220735424)))]; + tensor encoder_layers_10_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_10_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220737536)))]; + tensor input_575_cast_fp16 = transpose(perm = input_575_perm_0, x = x_281_cast_fp16)[name = string("transpose_266")]; + tensor x_283_cast_fp16 = layer_norm(axes = x_283_axes_0, beta = encoder_layers_10_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_10_conv_batch_norm_weight_to_fp16, x = input_575_cast_fp16)[name = string("x_283_cast_fp16")]; + tensor input_577_perm_0 = const()[name = string("input_577_perm_0"), val = tensor([0, 2, 1])]; + tensor input_577_cast_fp16 = transpose(perm = input_577_perm_0, x = x_283_cast_fp16)[name = string("transpose_265")]; + tensor input_579_cast_fp16 = silu(x = input_577_cast_fp16)[name = string("input_579_cast_fp16")]; + string x_285_pad_type_0 = const()[name = string("x_285_pad_type_0"), val = string("valid")]; + tensor x_285_strides_0 = const()[name = string("x_285_strides_0"), val = tensor([1])]; + tensor x_285_pad_0 = const()[name = string("x_285_pad_0"), val = tensor([0, 0])]; + tensor x_285_dilations_0 = const()[name = string("x_285_dilations_0"), val = tensor([1])]; + int32 x_285_groups_0 = const()[name = string("x_285_groups_0"), val = int32(1)]; + tensor encoder_layers_10_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220739648))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221788288))))[name = string("encoder_layers_10_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_285_cast_fp16 = conv(dilations = x_285_dilations_0, groups = x_285_groups_0, pad = x_285_pad_0, pad_type = x_285_pad_type_0, strides = x_285_strides_0, weight = encoder_layers_10_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_579_cast_fp16)[name = string("x_285_cast_fp16")]; + tensor input_581_perm_0 = const()[name = string("input_581_perm_0"), val = tensor([0, 2, 1])]; + tensor input_581_cast_fp16 = transpose(perm = input_581_perm_0, x = x_285_cast_fp16)[name = string("transpose_264")]; + tensor input_583_cast_fp16 = add(x = input_567_cast_fp16, y = input_581_cast_fp16)[name = string("input_583_cast_fp16")]; + tensor input_585_axes_0 = const()[name = string("input_585_axes_0"), val = tensor([-1])]; + tensor encoder_layers_10_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_10_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221790400)))]; + tensor encoder_layers_10_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_10_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221792512)))]; + tensor input_585_cast_fp16 = layer_norm(axes = input_585_axes_0, beta = encoder_layers_10_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_10_norm_feed_forward2_weight_to_fp16, x = input_583_cast_fp16)[name = string("input_585_cast_fp16")]; + tensor encoder_layers_10_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221794624))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(224940416))))[name = string("encoder_layers_10_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_10_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_10_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(224940608)))]; + tensor linear_98_cast_fp16 = linear(bias = encoder_layers_10_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_10_feed_forward2_linear1_weight_to_fp16_palettized, x = input_585_cast_fp16)[name = string("linear_98_cast_fp16")]; + tensor input_589_cast_fp16 = silu(x = linear_98_cast_fp16)[name = string("input_589_cast_fp16")]; + tensor encoder_layers_10_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(224948864))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(228094656))))[name = string("encoder_layers_10_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_10_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_10_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(228094848)))]; + tensor linear_99_cast_fp16 = linear(bias = encoder_layers_10_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_10_feed_forward2_linear2_weight_to_fp16_palettized, x = input_589_cast_fp16)[name = string("linear_99_cast_fp16")]; + fp16 var_2761_to_fp16 = const()[name = string("op_2761_to_fp16"), val = fp16(0x1p-1)]; + tensor var_2762_cast_fp16 = mul(x = linear_99_cast_fp16, y = var_2761_to_fp16)[name = string("op_2762_cast_fp16")]; + tensor input_595_cast_fp16 = add(x = input_583_cast_fp16, y = var_2762_cast_fp16)[name = string("input_595_cast_fp16")]; + tensor input_597_axes_0 = const()[name = string("input_597_axes_0"), val = tensor([-1])]; + tensor encoder_layers_10_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_10_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(228096960)))]; + tensor encoder_layers_10_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_10_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(228099072)))]; + tensor input_597_cast_fp16 = layer_norm(axes = input_597_axes_0, beta = encoder_layers_10_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_10_norm_out_weight_to_fp16, x = input_595_cast_fp16)[name = string("input_597_cast_fp16")]; + tensor cache_45_begin_0 = const()[name = string("cache_45_begin_0"), val = tensor([11, 0, 0, 0])]; + tensor cache_45_end_0 = const()[name = string("cache_45_end_0"), val = tensor([12, 1, 42, 1024])]; + tensor cache_45_end_mask_0 = const()[name = string("cache_45_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_45_squeeze_mask_0 = const()[name = string("cache_45_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_45_cast_fp16 = slice_by_index(begin = cache_45_begin_0, end = cache_45_end_0, end_mask = cache_45_end_mask_0, squeeze_mask = cache_45_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_45_cast_fp16")]; + tensor cache_47_begin_0 = const()[name = string("cache_47_begin_0"), val = tensor([11, 0, 0, 0])]; + tensor cache_47_end_0 = const()[name = string("cache_47_end_0"), val = tensor([12, 1, 1024, 8])]; + tensor cache_47_end_mask_0 = const()[name = string("cache_47_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_47_squeeze_mask_0 = const()[name = string("cache_47_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_47_cast_fp16 = slice_by_index(begin = cache_47_begin_0, end = cache_47_end_0, end_mask = cache_47_end_mask_0, squeeze_mask = cache_47_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_47_cast_fp16")]; + tensor input_599_axes_0 = const()[name = string("input_599_axes_0"), val = tensor([-1])]; + tensor encoder_layers_11_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_11_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(228101184)))]; + tensor encoder_layers_11_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_11_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(228103296)))]; + tensor input_599_cast_fp16 = layer_norm(axes = input_599_axes_0, beta = encoder_layers_11_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_11_norm_feed_forward1_weight_to_fp16, x = input_597_cast_fp16)[name = string("input_599_cast_fp16")]; + tensor encoder_layers_11_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(228105408))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(231251200))))[name = string("encoder_layers_11_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_11_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_11_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(231251392)))]; + tensor linear_100_cast_fp16 = linear(bias = encoder_layers_11_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_11_feed_forward1_linear1_weight_to_fp16_palettized, x = input_599_cast_fp16)[name = string("linear_100_cast_fp16")]; + tensor input_603_cast_fp16 = silu(x = linear_100_cast_fp16)[name = string("input_603_cast_fp16")]; + tensor encoder_layers_11_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(231259648))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234405440))))[name = string("encoder_layers_11_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_11_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_11_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234405632)))]; + tensor linear_101_cast_fp16 = linear(bias = encoder_layers_11_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_11_feed_forward1_linear2_weight_to_fp16_palettized, x = input_603_cast_fp16)[name = string("linear_101_cast_fp16")]; + fp16 var_2798_to_fp16 = const()[name = string("op_2798_to_fp16"), val = fp16(0x1p-1)]; + tensor var_2799_cast_fp16 = mul(x = linear_101_cast_fp16, y = var_2798_to_fp16)[name = string("op_2799_cast_fp16")]; + tensor input_609_cast_fp16 = add(x = input_597_cast_fp16, y = var_2799_cast_fp16)[name = string("input_609_cast_fp16")]; + tensor key_23_axes_0 = const()[name = string("key_23_axes_0"), val = tensor([-1])]; + tensor encoder_layers_11_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_11_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234407744)))]; + tensor encoder_layers_11_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_11_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234409856)))]; + tensor key_23_cast_fp16 = layer_norm(axes = key_23_axes_0, beta = encoder_layers_11_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_11_norm_self_att_weight_to_fp16, x = input_609_cast_fp16)[name = string("key_23_cast_fp16")]; + bool input_611_interleave_0 = const()[name = string("input_611_interleave_0"), val = bool(false)]; + tensor input_611_cast_fp16 = concat(axis = var_68, interleave = input_611_interleave_0, values = (cache_45_cast_fp16, key_23_cast_fp16))[name = string("input_611_cast_fp16")]; + tensor var_2821_begin_0 = const()[name = string("op_2821_begin_0"), val = tensor([0, 28, 0])]; + tensor var_2821_end_0 = const()[name = string("op_2821_end_0"), val = tensor([1, 42, 1024])]; + tensor var_2821_end_mask_0 = const()[name = string("op_2821_end_mask_0"), val = tensor([true, true, true])]; + tensor var_2821_cast_fp16 = slice_by_index(begin = var_2821_begin_0, end = var_2821_end_0, end_mask = var_2821_end_mask_0, x = cache_45_cast_fp16)[name = string("op_2821_cast_fp16")]; + bool var_2827_interleave_0 = const()[name = string("op_2827_interleave_0"), val = bool(false)]; + tensor var_2827_cast_fp16 = concat(axis = var_68, interleave = var_2827_interleave_0, values = (var_2821_cast_fp16, key_23_cast_fp16))[name = string("op_2827_cast_fp16")]; + tensor encoder_layers_11_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234411968))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(235198464))))[name = string("encoder_layers_11_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_11_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_11_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(235198656)))]; + tensor linear_102_cast_fp16 = linear(bias = encoder_layers_11_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_11_self_attn_linear_q_weight_to_fp16_palettized, x = key_23_cast_fp16)[name = string("linear_102_cast_fp16")]; + tensor var_2832 = const()[name = string("op_2832"), val = tensor([1, -1, 8, 128])]; + tensor q_67_cast_fp16 = reshape(shape = var_2832, x = linear_102_cast_fp16)[name = string("q_67_cast_fp16")]; + tensor encoder_layers_11_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(235200768))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(235987264))))[name = string("encoder_layers_11_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_11_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_11_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(235987456)))]; + tensor linear_103_cast_fp16 = linear(bias = encoder_layers_11_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_11_self_attn_linear_k_weight_to_fp16_palettized, x = input_611_cast_fp16)[name = string("linear_103_cast_fp16")]; + tensor var_2837 = const()[name = string("op_2837"), val = tensor([1, -1, 8, 128])]; + tensor k_45_cast_fp16 = reshape(shape = var_2837, x = linear_103_cast_fp16)[name = string("k_45_cast_fp16")]; + tensor encoder_layers_11_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(235989568))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236776064))))[name = string("encoder_layers_11_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_11_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_11_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236776256)))]; + tensor linear_104_cast_fp16 = linear(bias = encoder_layers_11_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_11_self_attn_linear_v_weight_to_fp16_palettized, x = input_611_cast_fp16)[name = string("linear_104_cast_fp16")]; + tensor var_2842 = const()[name = string("op_2842"), val = tensor([1, -1, 8, 128])]; + tensor v_23_cast_fp16 = reshape(shape = var_2842, x = linear_104_cast_fp16)[name = string("v_23_cast_fp16")]; + tensor value_31_perm_0 = const()[name = string("value_31_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_11_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_11_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236778368)))]; + tensor var_2855_cast_fp16 = add(x = q_67_cast_fp16, y = encoder_layers_11_self_attn_pos_bias_u_to_fp16)[name = string("op_2855_cast_fp16")]; + tensor encoder_layers_11_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_11_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236780480)))]; + tensor var_2857_cast_fp16 = add(x = q_67_cast_fp16, y = encoder_layers_11_self_attn_pos_bias_v_to_fp16)[name = string("op_2857_cast_fp16")]; + tensor q_with_bias_v_23_perm_0 = const()[name = string("q_with_bias_v_23_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_293_transpose_x_0 = const()[name = string("x_293_transpose_x_0"), val = bool(false)]; + bool x_293_transpose_y_0 = const()[name = string("x_293_transpose_y_0"), val = bool(false)]; + tensor op_2859_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236782592))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236924992))))[name = string("op_2859_to_fp16_quantized")]; + tensor q_with_bias_v_23_cast_fp16 = transpose(perm = q_with_bias_v_23_perm_0, x = var_2857_cast_fp16)[name = string("transpose_263")]; + tensor x_293_cast_fp16 = matmul(transpose_x = x_293_transpose_x_0, transpose_y = x_293_transpose_y_0, x = q_with_bias_v_23_cast_fp16, y = op_2859_to_fp16_quantized)[name = string("x_293_cast_fp16")]; + tensor x_295_pad_0 = const()[name = string("x_295_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_295_mode_0 = const()[name = string("x_295_mode_0"), val = string("constant")]; + fp16 const_222_to_fp16 = const()[name = string("const_222_to_fp16"), val = fp16(0x0p+0)]; + tensor x_295_cast_fp16 = pad(constant_val = const_222_to_fp16, mode = x_295_mode_0, pad = x_295_pad_0, x = x_293_cast_fp16)[name = string("x_295_cast_fp16")]; + tensor var_2867 = const()[name = string("op_2867"), val = tensor([1, 8, -1, 28])]; + tensor x_297_cast_fp16 = reshape(shape = var_2867, x = x_295_cast_fp16)[name = string("x_297_cast_fp16")]; + tensor var_2871_begin_0 = const()[name = string("op_2871_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_2871_end_0 = const()[name = string("op_2871_end_0"), val = tensor([1, 8, 140, 28])]; + tensor var_2871_end_mask_0 = const()[name = string("op_2871_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_2871_cast_fp16 = slice_by_index(begin = var_2871_begin_0, end = var_2871_end_0, end_mask = var_2871_end_mask_0, x = x_297_cast_fp16)[name = string("op_2871_cast_fp16")]; + tensor var_2872 = const()[name = string("op_2872"), val = tensor([1, 8, 28, 139])]; + tensor matrix_bd_45_cast_fp16 = reshape(shape = var_2872, x = var_2871_cast_fp16)[name = string("matrix_bd_45_cast_fp16")]; + bool matrix_ac_23_transpose_x_0 = const()[name = string("matrix_ac_23_transpose_x_0"), val = bool(false)]; + bool matrix_ac_23_transpose_y_0 = const()[name = string("matrix_ac_23_transpose_y_0"), val = bool(false)]; + tensor transpose_118_perm_0 = const()[name = string("transpose_118_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_119_perm_0 = const()[name = string("transpose_119_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_119 = transpose(perm = transpose_119_perm_0, x = k_45_cast_fp16)[name = string("transpose_261")]; + tensor transpose_118 = transpose(perm = transpose_118_perm_0, x = var_2855_cast_fp16)[name = string("transpose_262")]; + tensor matrix_ac_23_cast_fp16 = matmul(transpose_x = matrix_ac_23_transpose_x_0, transpose_y = matrix_ac_23_transpose_y_0, x = transpose_118, y = transpose_119)[name = string("matrix_ac_23_cast_fp16")]; + tensor matrix_bd_47_begin_0 = const()[name = string("matrix_bd_47_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_47_end_0 = const()[name = string("matrix_bd_47_end_0"), val = tensor([1, 8, 28, 70])]; + tensor matrix_bd_47_end_mask_0 = const()[name = string("matrix_bd_47_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_47_cast_fp16 = slice_by_index(begin = matrix_bd_47_begin_0, end = matrix_bd_47_end_0, end_mask = matrix_bd_47_end_mask_0, x = matrix_bd_45_cast_fp16)[name = string("matrix_bd_47_cast_fp16")]; + tensor var_2881_cast_fp16 = add(x = matrix_ac_23_cast_fp16, y = matrix_bd_47_cast_fp16)[name = string("op_2881_cast_fp16")]; + fp16 _inversed_scores_45_y_0_to_fp16 = const()[name = string("_inversed_scores_45_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_45_cast_fp16 = mul(x = var_2881_cast_fp16, y = _inversed_scores_45_y_0_to_fp16)[name = string("_inversed_scores_45_cast_fp16")]; + tensor scores_47_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_45_cast_fp16, cond = mask_11)[name = string("scores_47_cast_fp16")]; + tensor var_2887_cast_fp16 = softmax(axis = var_59, x = scores_47_cast_fp16)[name = string("op_2887_cast_fp16")]; + tensor input_613_cast_fp16 = select(a = var_44_to_fp16, b = var_2887_cast_fp16, cond = mask_11)[name = string("input_613_cast_fp16")]; + bool x_299_transpose_x_0 = const()[name = string("x_299_transpose_x_0"), val = bool(false)]; + bool x_299_transpose_y_0 = const()[name = string("x_299_transpose_y_0"), val = bool(false)]; + tensor value_31_cast_fp16 = transpose(perm = value_31_perm_0, x = v_23_cast_fp16)[name = string("transpose_260")]; + tensor x_299_cast_fp16 = matmul(transpose_x = x_299_transpose_x_0, transpose_y = x_299_transpose_y_0, x = input_613_cast_fp16, y = value_31_cast_fp16)[name = string("x_299_cast_fp16")]; + tensor var_2891_perm_0 = const()[name = string("op_2891_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_2892 = const()[name = string("op_2892"), val = tensor([1, -1, 1024])]; + tensor var_2891_cast_fp16 = transpose(perm = var_2891_perm_0, x = x_299_cast_fp16)[name = string("transpose_259")]; + tensor input_615_cast_fp16 = reshape(shape = var_2892, x = var_2891_cast_fp16)[name = string("input_615_cast_fp16")]; + tensor encoder_layers_11_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236925376))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(237711872))))[name = string("encoder_layers_11_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_11_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_11_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(237712064)))]; + tensor linear_106_cast_fp16 = linear(bias = encoder_layers_11_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_11_self_attn_linear_out_weight_to_fp16_palettized, x = input_615_cast_fp16)[name = string("linear_106_cast_fp16")]; + tensor input_619_cast_fp16 = add(x = input_609_cast_fp16, y = linear_106_cast_fp16)[name = string("input_619_cast_fp16")]; + tensor x_303_axes_0 = const()[name = string("x_303_axes_0"), val = tensor([-1])]; + tensor encoder_layers_11_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_11_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(237714176)))]; + tensor encoder_layers_11_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_11_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(237716288)))]; + tensor x_303_cast_fp16 = layer_norm(axes = x_303_axes_0, beta = encoder_layers_11_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_11_norm_conv_weight_to_fp16, x = input_619_cast_fp16)[name = string("x_303_cast_fp16")]; + tensor input_621_perm_0 = const()[name = string("input_621_perm_0"), val = tensor([0, 2, 1])]; + string input_623_pad_type_0 = const()[name = string("input_623_pad_type_0"), val = string("valid")]; + tensor input_623_strides_0 = const()[name = string("input_623_strides_0"), val = tensor([1])]; + tensor input_623_pad_0 = const()[name = string("input_623_pad_0"), val = tensor([0, 0])]; + tensor input_623_dilations_0 = const()[name = string("input_623_dilations_0"), val = tensor([1])]; + int32 input_623_groups_0 = const()[name = string("input_623_groups_0"), val = int32(1)]; + tensor encoder_layers_11_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(237718400))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(239815616))))[name = string("encoder_layers_11_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_621_cast_fp16 = transpose(perm = input_621_perm_0, x = x_303_cast_fp16)[name = string("transpose_258")]; + tensor input_623_cast_fp16 = conv(dilations = input_623_dilations_0, groups = input_623_groups_0, pad = input_623_pad_0, pad_type = input_623_pad_type_0, strides = input_623_strides_0, weight = encoder_layers_11_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_621_cast_fp16)[name = string("input_623_cast_fp16")]; + int32 x_305_split_num_splits_0 = const()[name = string("x_305_split_num_splits_0"), val = int32(2)]; + int32 x_305_split_axis_0 = const()[name = string("x_305_split_axis_0"), val = int32(1)]; + tensor x_305_split_cast_fp16_0, tensor x_305_split_cast_fp16_1 = split(axis = x_305_split_axis_0, num_splits = x_305_split_num_splits_0, x = input_623_cast_fp16)[name = string("x_305_split_cast_fp16")]; + tensor x_305_split_1_sigmoid_cast_fp16 = sigmoid(x = x_305_split_cast_fp16_1)[name = string("x_305_split_1_sigmoid_cast_fp16")]; + tensor x_305_cast_fp16 = mul(x = x_305_split_cast_fp16_0, y = x_305_split_1_sigmoid_cast_fp16)[name = string("x_305_cast_fp16")]; + tensor input_625_cast_fp16 = select(a = var_44_to_fp16, b = x_305_cast_fp16, cond = var_575)[name = string("input_625_cast_fp16")]; + bool new_x_47_interleave_0 = const()[name = string("new_x_47_interleave_0"), val = bool(false)]; + tensor new_x_47_cast_fp16 = concat(axis = var_59, interleave = new_x_47_interleave_0, values = (cache_47_cast_fp16, input_625_cast_fp16))[name = string("new_x_47_cast_fp16")]; + tensor var_2931_begin_0 = const()[name = string("op_2931_begin_0"), val = tensor([0, 0, 28])]; + tensor var_2931_end_0 = const()[name = string("op_2931_end_0"), val = tensor([1, 1024, 36])]; + tensor var_2931_end_mask_0 = const()[name = string("op_2931_end_mask_0"), val = tensor([true, true, true])]; + tensor var_2931_cast_fp16 = slice_by_index(begin = var_2931_begin_0, end = var_2931_end_0, end_mask = var_2931_end_mask_0, x = new_x_47_cast_fp16)[name = string("op_2931_cast_fp16")]; + string x_307_pad_type_0 = const()[name = string("x_307_pad_type_0"), val = string("valid")]; + int32 x_307_groups_0 = const()[name = string("x_307_groups_0"), val = int32(1024)]; + tensor x_307_strides_0 = const()[name = string("x_307_strides_0"), val = tensor([1])]; + tensor x_307_pad_0 = const()[name = string("x_307_pad_0"), val = tensor([0, 0])]; + tensor x_307_dilations_0 = const()[name = string("x_307_dilations_0"), val = tensor([1])]; + tensor encoder_layers_11_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(239819776))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(239829056))))[name = string("encoder_layers_11_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_307_cast_fp16 = conv(dilations = x_307_dilations_0, groups = x_307_groups_0, pad = x_307_pad_0, pad_type = x_307_pad_type_0, strides = x_307_strides_0, weight = encoder_layers_11_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_47_cast_fp16)[name = string("x_307_cast_fp16")]; + tensor input_627_perm_0 = const()[name = string("input_627_perm_0"), val = tensor([0, 2, 1])]; + tensor x_309_axes_0 = const()[name = string("x_309_axes_0"), val = tensor([-1])]; + tensor encoder_layers_11_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_11_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(239831168)))]; + tensor encoder_layers_11_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_11_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(239833280)))]; + tensor input_627_cast_fp16 = transpose(perm = input_627_perm_0, x = x_307_cast_fp16)[name = string("transpose_257")]; + tensor x_309_cast_fp16 = layer_norm(axes = x_309_axes_0, beta = encoder_layers_11_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_11_conv_batch_norm_weight_to_fp16, x = input_627_cast_fp16)[name = string("x_309_cast_fp16")]; + tensor input_629_perm_0 = const()[name = string("input_629_perm_0"), val = tensor([0, 2, 1])]; + tensor input_629_cast_fp16 = transpose(perm = input_629_perm_0, x = x_309_cast_fp16)[name = string("transpose_256")]; + tensor input_631_cast_fp16 = silu(x = input_629_cast_fp16)[name = string("input_631_cast_fp16")]; + string x_311_pad_type_0 = const()[name = string("x_311_pad_type_0"), val = string("valid")]; + tensor x_311_strides_0 = const()[name = string("x_311_strides_0"), val = tensor([1])]; + tensor x_311_pad_0 = const()[name = string("x_311_pad_0"), val = tensor([0, 0])]; + tensor x_311_dilations_0 = const()[name = string("x_311_dilations_0"), val = tensor([1])]; + int32 x_311_groups_0 = const()[name = string("x_311_groups_0"), val = int32(1)]; + tensor encoder_layers_11_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(239835392))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240884032))))[name = string("encoder_layers_11_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_311_cast_fp16 = conv(dilations = x_311_dilations_0, groups = x_311_groups_0, pad = x_311_pad_0, pad_type = x_311_pad_type_0, strides = x_311_strides_0, weight = encoder_layers_11_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_631_cast_fp16)[name = string("x_311_cast_fp16")]; + tensor input_633_perm_0 = const()[name = string("input_633_perm_0"), val = tensor([0, 2, 1])]; + tensor input_633_cast_fp16 = transpose(perm = input_633_perm_0, x = x_311_cast_fp16)[name = string("transpose_255")]; + tensor input_635_cast_fp16 = add(x = input_619_cast_fp16, y = input_633_cast_fp16)[name = string("input_635_cast_fp16")]; + tensor input_637_axes_0 = const()[name = string("input_637_axes_0"), val = tensor([-1])]; + tensor encoder_layers_11_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_11_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240886144)))]; + tensor encoder_layers_11_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_11_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240888256)))]; + tensor input_637_cast_fp16 = layer_norm(axes = input_637_axes_0, beta = encoder_layers_11_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_11_norm_feed_forward2_weight_to_fp16, x = input_635_cast_fp16)[name = string("input_637_cast_fp16")]; + tensor encoder_layers_11_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240890368))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(244036160))))[name = string("encoder_layers_11_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_11_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_11_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(244036352)))]; + tensor linear_107_cast_fp16 = linear(bias = encoder_layers_11_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_11_feed_forward2_linear1_weight_to_fp16_palettized, x = input_637_cast_fp16)[name = string("linear_107_cast_fp16")]; + tensor input_641_cast_fp16 = silu(x = linear_107_cast_fp16)[name = string("input_641_cast_fp16")]; + tensor encoder_layers_11_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(244044608))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(247190400))))[name = string("encoder_layers_11_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_11_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_11_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(247190592)))]; + tensor linear_108_cast_fp16 = linear(bias = encoder_layers_11_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_11_feed_forward2_linear2_weight_to_fp16_palettized, x = input_641_cast_fp16)[name = string("linear_108_cast_fp16")]; + fp16 var_2974_to_fp16 = const()[name = string("op_2974_to_fp16"), val = fp16(0x1p-1)]; + tensor var_2975_cast_fp16 = mul(x = linear_108_cast_fp16, y = var_2974_to_fp16)[name = string("op_2975_cast_fp16")]; + tensor input_647_cast_fp16 = add(x = input_635_cast_fp16, y = var_2975_cast_fp16)[name = string("input_647_cast_fp16")]; + tensor input_649_axes_0 = const()[name = string("input_649_axes_0"), val = tensor([-1])]; + tensor encoder_layers_11_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_11_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(247192704)))]; + tensor encoder_layers_11_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_11_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(247194816)))]; + tensor input_649_cast_fp16 = layer_norm(axes = input_649_axes_0, beta = encoder_layers_11_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_11_norm_out_weight_to_fp16, x = input_647_cast_fp16)[name = string("input_649_cast_fp16")]; + tensor cache_49_begin_0 = const()[name = string("cache_49_begin_0"), val = tensor([12, 0, 0, 0])]; + tensor cache_49_end_0 = const()[name = string("cache_49_end_0"), val = tensor([13, 1, 42, 1024])]; + tensor cache_49_end_mask_0 = const()[name = string("cache_49_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_49_squeeze_mask_0 = const()[name = string("cache_49_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_49_cast_fp16 = slice_by_index(begin = cache_49_begin_0, end = cache_49_end_0, end_mask = cache_49_end_mask_0, squeeze_mask = cache_49_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_49_cast_fp16")]; + tensor cache_51_begin_0 = const()[name = string("cache_51_begin_0"), val = tensor([12, 0, 0, 0])]; + tensor cache_51_end_0 = const()[name = string("cache_51_end_0"), val = tensor([13, 1, 1024, 8])]; + tensor cache_51_end_mask_0 = const()[name = string("cache_51_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_51_squeeze_mask_0 = const()[name = string("cache_51_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_51_cast_fp16 = slice_by_index(begin = cache_51_begin_0, end = cache_51_end_0, end_mask = cache_51_end_mask_0, squeeze_mask = cache_51_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_51_cast_fp16")]; + tensor input_651_axes_0 = const()[name = string("input_651_axes_0"), val = tensor([-1])]; + tensor encoder_layers_12_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_12_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(247196928)))]; + tensor encoder_layers_12_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_12_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(247199040)))]; + tensor input_651_cast_fp16 = layer_norm(axes = input_651_axes_0, beta = encoder_layers_12_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_12_norm_feed_forward1_weight_to_fp16, x = input_649_cast_fp16)[name = string("input_651_cast_fp16")]; + tensor encoder_layers_12_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(247201152))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(250346944))))[name = string("encoder_layers_12_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_12_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_12_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(250347136)))]; + tensor linear_109_cast_fp16 = linear(bias = encoder_layers_12_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_12_feed_forward1_linear1_weight_to_fp16_palettized, x = input_651_cast_fp16)[name = string("linear_109_cast_fp16")]; + tensor input_655_cast_fp16 = silu(x = linear_109_cast_fp16)[name = string("input_655_cast_fp16")]; + tensor encoder_layers_12_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(250355392))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253501184))))[name = string("encoder_layers_12_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_12_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_12_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253501376)))]; + tensor linear_110_cast_fp16 = linear(bias = encoder_layers_12_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_12_feed_forward1_linear2_weight_to_fp16_palettized, x = input_655_cast_fp16)[name = string("linear_110_cast_fp16")]; + fp16 var_3011_to_fp16 = const()[name = string("op_3011_to_fp16"), val = fp16(0x1p-1)]; + tensor var_3012_cast_fp16 = mul(x = linear_110_cast_fp16, y = var_3011_to_fp16)[name = string("op_3012_cast_fp16")]; + tensor input_661_cast_fp16 = add(x = input_649_cast_fp16, y = var_3012_cast_fp16)[name = string("input_661_cast_fp16")]; + tensor key_25_axes_0 = const()[name = string("key_25_axes_0"), val = tensor([-1])]; + tensor encoder_layers_12_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_12_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253503488)))]; + tensor encoder_layers_12_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_12_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253505600)))]; + tensor key_25_cast_fp16 = layer_norm(axes = key_25_axes_0, beta = encoder_layers_12_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_12_norm_self_att_weight_to_fp16, x = input_661_cast_fp16)[name = string("key_25_cast_fp16")]; + bool input_663_interleave_0 = const()[name = string("input_663_interleave_0"), val = bool(false)]; + tensor input_663_cast_fp16 = concat(axis = var_68, interleave = input_663_interleave_0, values = (cache_49_cast_fp16, key_25_cast_fp16))[name = string("input_663_cast_fp16")]; + tensor var_3034_begin_0 = const()[name = string("op_3034_begin_0"), val = tensor([0, 28, 0])]; + tensor var_3034_end_0 = const()[name = string("op_3034_end_0"), val = tensor([1, 42, 1024])]; + tensor var_3034_end_mask_0 = const()[name = string("op_3034_end_mask_0"), val = tensor([true, true, true])]; + tensor var_3034_cast_fp16 = slice_by_index(begin = var_3034_begin_0, end = var_3034_end_0, end_mask = var_3034_end_mask_0, x = cache_49_cast_fp16)[name = string("op_3034_cast_fp16")]; + bool var_3040_interleave_0 = const()[name = string("op_3040_interleave_0"), val = bool(false)]; + tensor var_3040_cast_fp16 = concat(axis = var_68, interleave = var_3040_interleave_0, values = (var_3034_cast_fp16, key_25_cast_fp16))[name = string("op_3040_cast_fp16")]; + tensor encoder_layers_12_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253507712))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(254294208))))[name = string("encoder_layers_12_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_12_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_12_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(254294400)))]; + tensor linear_111_cast_fp16 = linear(bias = encoder_layers_12_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_12_self_attn_linear_q_weight_to_fp16_palettized, x = key_25_cast_fp16)[name = string("linear_111_cast_fp16")]; + tensor var_3045 = const()[name = string("op_3045"), val = tensor([1, -1, 8, 128])]; + tensor q_73_cast_fp16 = reshape(shape = var_3045, x = linear_111_cast_fp16)[name = string("q_73_cast_fp16")]; + tensor encoder_layers_12_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(254296512))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255083008))))[name = string("encoder_layers_12_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_12_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_12_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255083200)))]; + tensor linear_112_cast_fp16 = linear(bias = encoder_layers_12_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_12_self_attn_linear_k_weight_to_fp16_palettized, x = input_663_cast_fp16)[name = string("linear_112_cast_fp16")]; + tensor var_3050 = const()[name = string("op_3050"), val = tensor([1, -1, 8, 128])]; + tensor k_49_cast_fp16 = reshape(shape = var_3050, x = linear_112_cast_fp16)[name = string("k_49_cast_fp16")]; + tensor encoder_layers_12_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255085312))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255871808))))[name = string("encoder_layers_12_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_12_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_12_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255872000)))]; + tensor linear_113_cast_fp16 = linear(bias = encoder_layers_12_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_12_self_attn_linear_v_weight_to_fp16_palettized, x = input_663_cast_fp16)[name = string("linear_113_cast_fp16")]; + tensor var_3055 = const()[name = string("op_3055"), val = tensor([1, -1, 8, 128])]; + tensor v_25_cast_fp16 = reshape(shape = var_3055, x = linear_113_cast_fp16)[name = string("v_25_cast_fp16")]; + tensor value_33_perm_0 = const()[name = string("value_33_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_12_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_12_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255874112)))]; + tensor var_3068_cast_fp16 = add(x = q_73_cast_fp16, y = encoder_layers_12_self_attn_pos_bias_u_to_fp16)[name = string("op_3068_cast_fp16")]; + tensor encoder_layers_12_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_12_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255876224)))]; + tensor var_3070_cast_fp16 = add(x = q_73_cast_fp16, y = encoder_layers_12_self_attn_pos_bias_v_to_fp16)[name = string("op_3070_cast_fp16")]; + tensor q_with_bias_v_25_perm_0 = const()[name = string("q_with_bias_v_25_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_319_transpose_x_0 = const()[name = string("x_319_transpose_x_0"), val = bool(false)]; + bool x_319_transpose_y_0 = const()[name = string("x_319_transpose_y_0"), val = bool(false)]; + tensor op_3072_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255878336))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256020736))))[name = string("op_3072_to_fp16_quantized")]; + tensor q_with_bias_v_25_cast_fp16 = transpose(perm = q_with_bias_v_25_perm_0, x = var_3070_cast_fp16)[name = string("transpose_254")]; + tensor x_319_cast_fp16 = matmul(transpose_x = x_319_transpose_x_0, transpose_y = x_319_transpose_y_0, x = q_with_bias_v_25_cast_fp16, y = op_3072_to_fp16_quantized)[name = string("x_319_cast_fp16")]; + tensor x_321_pad_0 = const()[name = string("x_321_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_321_mode_0 = const()[name = string("x_321_mode_0"), val = string("constant")]; + fp16 const_235_to_fp16 = const()[name = string("const_235_to_fp16"), val = fp16(0x0p+0)]; + tensor x_321_cast_fp16 = pad(constant_val = const_235_to_fp16, mode = x_321_mode_0, pad = x_321_pad_0, x = x_319_cast_fp16)[name = string("x_321_cast_fp16")]; + tensor var_3080 = const()[name = string("op_3080"), val = tensor([1, 8, -1, 28])]; + tensor x_323_cast_fp16 = reshape(shape = var_3080, x = x_321_cast_fp16)[name = string("x_323_cast_fp16")]; + tensor var_3084_begin_0 = const()[name = string("op_3084_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_3084_end_0 = const()[name = string("op_3084_end_0"), val = tensor([1, 8, 140, 28])]; + tensor var_3084_end_mask_0 = const()[name = string("op_3084_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_3084_cast_fp16 = slice_by_index(begin = var_3084_begin_0, end = var_3084_end_0, end_mask = var_3084_end_mask_0, x = x_323_cast_fp16)[name = string("op_3084_cast_fp16")]; + tensor var_3085 = const()[name = string("op_3085"), val = tensor([1, 8, 28, 139])]; + tensor matrix_bd_49_cast_fp16 = reshape(shape = var_3085, x = var_3084_cast_fp16)[name = string("matrix_bd_49_cast_fp16")]; + bool matrix_ac_25_transpose_x_0 = const()[name = string("matrix_ac_25_transpose_x_0"), val = bool(false)]; + bool matrix_ac_25_transpose_y_0 = const()[name = string("matrix_ac_25_transpose_y_0"), val = bool(false)]; + tensor transpose_120_perm_0 = const()[name = string("transpose_120_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_121_perm_0 = const()[name = string("transpose_121_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_121 = transpose(perm = transpose_121_perm_0, x = k_49_cast_fp16)[name = string("transpose_252")]; + tensor transpose_120 = transpose(perm = transpose_120_perm_0, x = var_3068_cast_fp16)[name = string("transpose_253")]; + tensor matrix_ac_25_cast_fp16 = matmul(transpose_x = matrix_ac_25_transpose_x_0, transpose_y = matrix_ac_25_transpose_y_0, x = transpose_120, y = transpose_121)[name = string("matrix_ac_25_cast_fp16")]; + tensor matrix_bd_51_begin_0 = const()[name = string("matrix_bd_51_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_51_end_0 = const()[name = string("matrix_bd_51_end_0"), val = tensor([1, 8, 28, 70])]; + tensor matrix_bd_51_end_mask_0 = const()[name = string("matrix_bd_51_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_51_cast_fp16 = slice_by_index(begin = matrix_bd_51_begin_0, end = matrix_bd_51_end_0, end_mask = matrix_bd_51_end_mask_0, x = matrix_bd_49_cast_fp16)[name = string("matrix_bd_51_cast_fp16")]; + tensor var_3094_cast_fp16 = add(x = matrix_ac_25_cast_fp16, y = matrix_bd_51_cast_fp16)[name = string("op_3094_cast_fp16")]; + fp16 _inversed_scores_49_y_0_to_fp16 = const()[name = string("_inversed_scores_49_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_49_cast_fp16 = mul(x = var_3094_cast_fp16, y = _inversed_scores_49_y_0_to_fp16)[name = string("_inversed_scores_49_cast_fp16")]; + tensor scores_51_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_49_cast_fp16, cond = mask_11)[name = string("scores_51_cast_fp16")]; + tensor var_3100_cast_fp16 = softmax(axis = var_59, x = scores_51_cast_fp16)[name = string("op_3100_cast_fp16")]; + tensor input_665_cast_fp16 = select(a = var_44_to_fp16, b = var_3100_cast_fp16, cond = mask_11)[name = string("input_665_cast_fp16")]; + bool x_325_transpose_x_0 = const()[name = string("x_325_transpose_x_0"), val = bool(false)]; + bool x_325_transpose_y_0 = const()[name = string("x_325_transpose_y_0"), val = bool(false)]; + tensor value_33_cast_fp16 = transpose(perm = value_33_perm_0, x = v_25_cast_fp16)[name = string("transpose_251")]; + tensor x_325_cast_fp16 = matmul(transpose_x = x_325_transpose_x_0, transpose_y = x_325_transpose_y_0, x = input_665_cast_fp16, y = value_33_cast_fp16)[name = string("x_325_cast_fp16")]; + tensor var_3104_perm_0 = const()[name = string("op_3104_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_3105 = const()[name = string("op_3105"), val = tensor([1, -1, 1024])]; + tensor var_3104_cast_fp16 = transpose(perm = var_3104_perm_0, x = x_325_cast_fp16)[name = string("transpose_250")]; + tensor input_667_cast_fp16 = reshape(shape = var_3105, x = var_3104_cast_fp16)[name = string("input_667_cast_fp16")]; + tensor encoder_layers_12_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256021120))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256807616))))[name = string("encoder_layers_12_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_12_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_12_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256807808)))]; + tensor linear_115_cast_fp16 = linear(bias = encoder_layers_12_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_12_self_attn_linear_out_weight_to_fp16_palettized, x = input_667_cast_fp16)[name = string("linear_115_cast_fp16")]; + tensor input_671_cast_fp16 = add(x = input_661_cast_fp16, y = linear_115_cast_fp16)[name = string("input_671_cast_fp16")]; + tensor x_329_axes_0 = const()[name = string("x_329_axes_0"), val = tensor([-1])]; + tensor encoder_layers_12_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_12_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256809920)))]; + tensor encoder_layers_12_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_12_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256812032)))]; + tensor x_329_cast_fp16 = layer_norm(axes = x_329_axes_0, beta = encoder_layers_12_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_12_norm_conv_weight_to_fp16, x = input_671_cast_fp16)[name = string("x_329_cast_fp16")]; + tensor input_673_perm_0 = const()[name = string("input_673_perm_0"), val = tensor([0, 2, 1])]; + string input_675_pad_type_0 = const()[name = string("input_675_pad_type_0"), val = string("valid")]; + tensor input_675_strides_0 = const()[name = string("input_675_strides_0"), val = tensor([1])]; + tensor input_675_pad_0 = const()[name = string("input_675_pad_0"), val = tensor([0, 0])]; + tensor input_675_dilations_0 = const()[name = string("input_675_dilations_0"), val = tensor([1])]; + int32 input_675_groups_0 = const()[name = string("input_675_groups_0"), val = int32(1)]; + tensor encoder_layers_12_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256814144))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(258911360))))[name = string("encoder_layers_12_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_673_cast_fp16 = transpose(perm = input_673_perm_0, x = x_329_cast_fp16)[name = string("transpose_249")]; + tensor input_675_cast_fp16 = conv(dilations = input_675_dilations_0, groups = input_675_groups_0, pad = input_675_pad_0, pad_type = input_675_pad_type_0, strides = input_675_strides_0, weight = encoder_layers_12_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_673_cast_fp16)[name = string("input_675_cast_fp16")]; + int32 x_331_split_num_splits_0 = const()[name = string("x_331_split_num_splits_0"), val = int32(2)]; + int32 x_331_split_axis_0 = const()[name = string("x_331_split_axis_0"), val = int32(1)]; + tensor x_331_split_cast_fp16_0, tensor x_331_split_cast_fp16_1 = split(axis = x_331_split_axis_0, num_splits = x_331_split_num_splits_0, x = input_675_cast_fp16)[name = string("x_331_split_cast_fp16")]; + tensor x_331_split_1_sigmoid_cast_fp16 = sigmoid(x = x_331_split_cast_fp16_1)[name = string("x_331_split_1_sigmoid_cast_fp16")]; + tensor x_331_cast_fp16 = mul(x = x_331_split_cast_fp16_0, y = x_331_split_1_sigmoid_cast_fp16)[name = string("x_331_cast_fp16")]; + tensor input_677_cast_fp16 = select(a = var_44_to_fp16, b = x_331_cast_fp16, cond = var_575)[name = string("input_677_cast_fp16")]; + bool new_x_51_interleave_0 = const()[name = string("new_x_51_interleave_0"), val = bool(false)]; + tensor new_x_51_cast_fp16 = concat(axis = var_59, interleave = new_x_51_interleave_0, values = (cache_51_cast_fp16, input_677_cast_fp16))[name = string("new_x_51_cast_fp16")]; + tensor var_3144_begin_0 = const()[name = string("op_3144_begin_0"), val = tensor([0, 0, 28])]; + tensor var_3144_end_0 = const()[name = string("op_3144_end_0"), val = tensor([1, 1024, 36])]; + tensor var_3144_end_mask_0 = const()[name = string("op_3144_end_mask_0"), val = tensor([true, true, true])]; + tensor var_3144_cast_fp16 = slice_by_index(begin = var_3144_begin_0, end = var_3144_end_0, end_mask = var_3144_end_mask_0, x = new_x_51_cast_fp16)[name = string("op_3144_cast_fp16")]; + string x_333_pad_type_0 = const()[name = string("x_333_pad_type_0"), val = string("valid")]; + int32 x_333_groups_0 = const()[name = string("x_333_groups_0"), val = int32(1024)]; + tensor x_333_strides_0 = const()[name = string("x_333_strides_0"), val = tensor([1])]; + tensor x_333_pad_0 = const()[name = string("x_333_pad_0"), val = tensor([0, 0])]; + tensor x_333_dilations_0 = const()[name = string("x_333_dilations_0"), val = tensor([1])]; + tensor encoder_layers_12_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(258915520))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(258924800))))[name = string("encoder_layers_12_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_333_cast_fp16 = conv(dilations = x_333_dilations_0, groups = x_333_groups_0, pad = x_333_pad_0, pad_type = x_333_pad_type_0, strides = x_333_strides_0, weight = encoder_layers_12_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_51_cast_fp16)[name = string("x_333_cast_fp16")]; + tensor input_679_perm_0 = const()[name = string("input_679_perm_0"), val = tensor([0, 2, 1])]; + tensor x_335_axes_0 = const()[name = string("x_335_axes_0"), val = tensor([-1])]; + tensor encoder_layers_12_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_12_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(258926912)))]; + tensor encoder_layers_12_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_12_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(258929024)))]; + tensor input_679_cast_fp16 = transpose(perm = input_679_perm_0, x = x_333_cast_fp16)[name = string("transpose_248")]; + tensor x_335_cast_fp16 = layer_norm(axes = x_335_axes_0, beta = encoder_layers_12_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_12_conv_batch_norm_weight_to_fp16, x = input_679_cast_fp16)[name = string("x_335_cast_fp16")]; + tensor input_681_perm_0 = const()[name = string("input_681_perm_0"), val = tensor([0, 2, 1])]; + tensor input_681_cast_fp16 = transpose(perm = input_681_perm_0, x = x_335_cast_fp16)[name = string("transpose_247")]; + tensor input_683_cast_fp16 = silu(x = input_681_cast_fp16)[name = string("input_683_cast_fp16")]; + string x_337_pad_type_0 = const()[name = string("x_337_pad_type_0"), val = string("valid")]; + tensor x_337_strides_0 = const()[name = string("x_337_strides_0"), val = tensor([1])]; + tensor x_337_pad_0 = const()[name = string("x_337_pad_0"), val = tensor([0, 0])]; + tensor x_337_dilations_0 = const()[name = string("x_337_dilations_0"), val = tensor([1])]; + int32 x_337_groups_0 = const()[name = string("x_337_groups_0"), val = int32(1)]; + tensor encoder_layers_12_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(258931136))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259979776))))[name = string("encoder_layers_12_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_337_cast_fp16 = conv(dilations = x_337_dilations_0, groups = x_337_groups_0, pad = x_337_pad_0, pad_type = x_337_pad_type_0, strides = x_337_strides_0, weight = encoder_layers_12_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_683_cast_fp16)[name = string("x_337_cast_fp16")]; + tensor input_685_perm_0 = const()[name = string("input_685_perm_0"), val = tensor([0, 2, 1])]; + tensor input_685_cast_fp16 = transpose(perm = input_685_perm_0, x = x_337_cast_fp16)[name = string("transpose_246")]; + tensor input_687_cast_fp16 = add(x = input_671_cast_fp16, y = input_685_cast_fp16)[name = string("input_687_cast_fp16")]; + tensor input_689_axes_0 = const()[name = string("input_689_axes_0"), val = tensor([-1])]; + tensor encoder_layers_12_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_12_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259981888)))]; + tensor encoder_layers_12_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_12_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259984000)))]; + tensor input_689_cast_fp16 = layer_norm(axes = input_689_axes_0, beta = encoder_layers_12_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_12_norm_feed_forward2_weight_to_fp16, x = input_687_cast_fp16)[name = string("input_689_cast_fp16")]; + tensor encoder_layers_12_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259986112))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(263131904))))[name = string("encoder_layers_12_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_12_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_12_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(263132096)))]; + tensor linear_116_cast_fp16 = linear(bias = encoder_layers_12_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_12_feed_forward2_linear1_weight_to_fp16_palettized, x = input_689_cast_fp16)[name = string("linear_116_cast_fp16")]; + tensor input_693_cast_fp16 = silu(x = linear_116_cast_fp16)[name = string("input_693_cast_fp16")]; + tensor encoder_layers_12_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(263140352))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(266286144))))[name = string("encoder_layers_12_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_12_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_12_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(266286336)))]; + tensor linear_117_cast_fp16 = linear(bias = encoder_layers_12_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_12_feed_forward2_linear2_weight_to_fp16_palettized, x = input_693_cast_fp16)[name = string("linear_117_cast_fp16")]; + fp16 var_3187_to_fp16 = const()[name = string("op_3187_to_fp16"), val = fp16(0x1p-1)]; + tensor var_3188_cast_fp16 = mul(x = linear_117_cast_fp16, y = var_3187_to_fp16)[name = string("op_3188_cast_fp16")]; + tensor input_699_cast_fp16 = add(x = input_687_cast_fp16, y = var_3188_cast_fp16)[name = string("input_699_cast_fp16")]; + tensor input_701_axes_0 = const()[name = string("input_701_axes_0"), val = tensor([-1])]; + tensor encoder_layers_12_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_12_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(266288448)))]; + tensor encoder_layers_12_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_12_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(266290560)))]; + tensor input_701_cast_fp16 = layer_norm(axes = input_701_axes_0, beta = encoder_layers_12_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_12_norm_out_weight_to_fp16, x = input_699_cast_fp16)[name = string("input_701_cast_fp16")]; + tensor cache_53_begin_0 = const()[name = string("cache_53_begin_0"), val = tensor([13, 0, 0, 0])]; + tensor cache_53_end_0 = const()[name = string("cache_53_end_0"), val = tensor([14, 1, 42, 1024])]; + tensor cache_53_end_mask_0 = const()[name = string("cache_53_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_53_squeeze_mask_0 = const()[name = string("cache_53_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_53_cast_fp16 = slice_by_index(begin = cache_53_begin_0, end = cache_53_end_0, end_mask = cache_53_end_mask_0, squeeze_mask = cache_53_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_53_cast_fp16")]; + tensor cache_55_begin_0 = const()[name = string("cache_55_begin_0"), val = tensor([13, 0, 0, 0])]; + tensor cache_55_end_0 = const()[name = string("cache_55_end_0"), val = tensor([14, 1, 1024, 8])]; + tensor cache_55_end_mask_0 = const()[name = string("cache_55_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_55_squeeze_mask_0 = const()[name = string("cache_55_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_55_cast_fp16 = slice_by_index(begin = cache_55_begin_0, end = cache_55_end_0, end_mask = cache_55_end_mask_0, squeeze_mask = cache_55_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_55_cast_fp16")]; + tensor input_703_axes_0 = const()[name = string("input_703_axes_0"), val = tensor([-1])]; + tensor encoder_layers_13_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_13_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(266292672)))]; + tensor encoder_layers_13_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_13_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(266294784)))]; + tensor input_703_cast_fp16 = layer_norm(axes = input_703_axes_0, beta = encoder_layers_13_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_13_norm_feed_forward1_weight_to_fp16, x = input_701_cast_fp16)[name = string("input_703_cast_fp16")]; + tensor encoder_layers_13_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(266296896))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(269442688))))[name = string("encoder_layers_13_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_13_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_13_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(269442880)))]; + tensor linear_118_cast_fp16 = linear(bias = encoder_layers_13_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_13_feed_forward1_linear1_weight_to_fp16_palettized, x = input_703_cast_fp16)[name = string("linear_118_cast_fp16")]; + tensor input_707_cast_fp16 = silu(x = linear_118_cast_fp16)[name = string("input_707_cast_fp16")]; + tensor encoder_layers_13_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(269451136))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272596928))))[name = string("encoder_layers_13_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_13_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_13_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272597120)))]; + tensor linear_119_cast_fp16 = linear(bias = encoder_layers_13_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_13_feed_forward1_linear2_weight_to_fp16_palettized, x = input_707_cast_fp16)[name = string("linear_119_cast_fp16")]; + fp16 var_3224_to_fp16 = const()[name = string("op_3224_to_fp16"), val = fp16(0x1p-1)]; + tensor var_3225_cast_fp16 = mul(x = linear_119_cast_fp16, y = var_3224_to_fp16)[name = string("op_3225_cast_fp16")]; + tensor input_713_cast_fp16 = add(x = input_701_cast_fp16, y = var_3225_cast_fp16)[name = string("input_713_cast_fp16")]; + tensor key_27_axes_0 = const()[name = string("key_27_axes_0"), val = tensor([-1])]; + tensor encoder_layers_13_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_13_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272599232)))]; + tensor encoder_layers_13_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_13_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272601344)))]; + tensor key_27_cast_fp16 = layer_norm(axes = key_27_axes_0, beta = encoder_layers_13_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_13_norm_self_att_weight_to_fp16, x = input_713_cast_fp16)[name = string("key_27_cast_fp16")]; + bool input_715_interleave_0 = const()[name = string("input_715_interleave_0"), val = bool(false)]; + tensor input_715_cast_fp16 = concat(axis = var_68, interleave = input_715_interleave_0, values = (cache_53_cast_fp16, key_27_cast_fp16))[name = string("input_715_cast_fp16")]; + tensor var_3247_begin_0 = const()[name = string("op_3247_begin_0"), val = tensor([0, 28, 0])]; + tensor var_3247_end_0 = const()[name = string("op_3247_end_0"), val = tensor([1, 42, 1024])]; + tensor var_3247_end_mask_0 = const()[name = string("op_3247_end_mask_0"), val = tensor([true, true, true])]; + tensor var_3247_cast_fp16 = slice_by_index(begin = var_3247_begin_0, end = var_3247_end_0, end_mask = var_3247_end_mask_0, x = cache_53_cast_fp16)[name = string("op_3247_cast_fp16")]; + bool var_3253_interleave_0 = const()[name = string("op_3253_interleave_0"), val = bool(false)]; + tensor var_3253_cast_fp16 = concat(axis = var_68, interleave = var_3253_interleave_0, values = (var_3247_cast_fp16, key_27_cast_fp16))[name = string("op_3253_cast_fp16")]; + tensor encoder_layers_13_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272603456))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(273389952))))[name = string("encoder_layers_13_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_13_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_13_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(273390144)))]; + tensor linear_120_cast_fp16 = linear(bias = encoder_layers_13_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_13_self_attn_linear_q_weight_to_fp16_palettized, x = key_27_cast_fp16)[name = string("linear_120_cast_fp16")]; + tensor var_3258 = const()[name = string("op_3258"), val = tensor([1, -1, 8, 128])]; + tensor q_79_cast_fp16 = reshape(shape = var_3258, x = linear_120_cast_fp16)[name = string("q_79_cast_fp16")]; + tensor encoder_layers_13_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(273392256))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274178752))))[name = string("encoder_layers_13_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_13_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_13_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274178944)))]; + tensor linear_121_cast_fp16 = linear(bias = encoder_layers_13_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_13_self_attn_linear_k_weight_to_fp16_palettized, x = input_715_cast_fp16)[name = string("linear_121_cast_fp16")]; + tensor var_3263 = const()[name = string("op_3263"), val = tensor([1, -1, 8, 128])]; + tensor k_53_cast_fp16 = reshape(shape = var_3263, x = linear_121_cast_fp16)[name = string("k_53_cast_fp16")]; + tensor encoder_layers_13_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274181056))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274967552))))[name = string("encoder_layers_13_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_13_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_13_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274967744)))]; + tensor linear_122_cast_fp16 = linear(bias = encoder_layers_13_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_13_self_attn_linear_v_weight_to_fp16_palettized, x = input_715_cast_fp16)[name = string("linear_122_cast_fp16")]; + tensor var_3268 = const()[name = string("op_3268"), val = tensor([1, -1, 8, 128])]; + tensor v_27_cast_fp16 = reshape(shape = var_3268, x = linear_122_cast_fp16)[name = string("v_27_cast_fp16")]; + tensor value_35_perm_0 = const()[name = string("value_35_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_13_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_13_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274969856)))]; + tensor var_3281_cast_fp16 = add(x = q_79_cast_fp16, y = encoder_layers_13_self_attn_pos_bias_u_to_fp16)[name = string("op_3281_cast_fp16")]; + tensor encoder_layers_13_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_13_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274971968)))]; + tensor var_3283_cast_fp16 = add(x = q_79_cast_fp16, y = encoder_layers_13_self_attn_pos_bias_v_to_fp16)[name = string("op_3283_cast_fp16")]; + tensor q_with_bias_v_27_perm_0 = const()[name = string("q_with_bias_v_27_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_345_transpose_x_0 = const()[name = string("x_345_transpose_x_0"), val = bool(false)]; + bool x_345_transpose_y_0 = const()[name = string("x_345_transpose_y_0"), val = bool(false)]; + tensor op_3285_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274974080))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275116480))))[name = string("op_3285_to_fp16_quantized")]; + tensor q_with_bias_v_27_cast_fp16 = transpose(perm = q_with_bias_v_27_perm_0, x = var_3283_cast_fp16)[name = string("transpose_245")]; + tensor x_345_cast_fp16 = matmul(transpose_x = x_345_transpose_x_0, transpose_y = x_345_transpose_y_0, x = q_with_bias_v_27_cast_fp16, y = op_3285_to_fp16_quantized)[name = string("x_345_cast_fp16")]; + tensor x_347_pad_0 = const()[name = string("x_347_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_347_mode_0 = const()[name = string("x_347_mode_0"), val = string("constant")]; + fp16 const_248_to_fp16 = const()[name = string("const_248_to_fp16"), val = fp16(0x0p+0)]; + tensor x_347_cast_fp16 = pad(constant_val = const_248_to_fp16, mode = x_347_mode_0, pad = x_347_pad_0, x = x_345_cast_fp16)[name = string("x_347_cast_fp16")]; + tensor var_3293 = const()[name = string("op_3293"), val = tensor([1, 8, -1, 28])]; + tensor x_349_cast_fp16 = reshape(shape = var_3293, x = x_347_cast_fp16)[name = string("x_349_cast_fp16")]; + tensor var_3297_begin_0 = const()[name = string("op_3297_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_3297_end_0 = const()[name = string("op_3297_end_0"), val = tensor([1, 8, 140, 28])]; + tensor var_3297_end_mask_0 = const()[name = string("op_3297_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_3297_cast_fp16 = slice_by_index(begin = var_3297_begin_0, end = var_3297_end_0, end_mask = var_3297_end_mask_0, x = x_349_cast_fp16)[name = string("op_3297_cast_fp16")]; + tensor var_3298 = const()[name = string("op_3298"), val = tensor([1, 8, 28, 139])]; + tensor matrix_bd_53_cast_fp16 = reshape(shape = var_3298, x = var_3297_cast_fp16)[name = string("matrix_bd_53_cast_fp16")]; + bool matrix_ac_27_transpose_x_0 = const()[name = string("matrix_ac_27_transpose_x_0"), val = bool(false)]; + bool matrix_ac_27_transpose_y_0 = const()[name = string("matrix_ac_27_transpose_y_0"), val = bool(false)]; + tensor transpose_122_perm_0 = const()[name = string("transpose_122_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_123_perm_0 = const()[name = string("transpose_123_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_123 = transpose(perm = transpose_123_perm_0, x = k_53_cast_fp16)[name = string("transpose_243")]; + tensor transpose_122 = transpose(perm = transpose_122_perm_0, x = var_3281_cast_fp16)[name = string("transpose_244")]; + tensor matrix_ac_27_cast_fp16 = matmul(transpose_x = matrix_ac_27_transpose_x_0, transpose_y = matrix_ac_27_transpose_y_0, x = transpose_122, y = transpose_123)[name = string("matrix_ac_27_cast_fp16")]; + tensor matrix_bd_55_begin_0 = const()[name = string("matrix_bd_55_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_55_end_0 = const()[name = string("matrix_bd_55_end_0"), val = tensor([1, 8, 28, 70])]; + tensor matrix_bd_55_end_mask_0 = const()[name = string("matrix_bd_55_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_55_cast_fp16 = slice_by_index(begin = matrix_bd_55_begin_0, end = matrix_bd_55_end_0, end_mask = matrix_bd_55_end_mask_0, x = matrix_bd_53_cast_fp16)[name = string("matrix_bd_55_cast_fp16")]; + tensor var_3307_cast_fp16 = add(x = matrix_ac_27_cast_fp16, y = matrix_bd_55_cast_fp16)[name = string("op_3307_cast_fp16")]; + fp16 _inversed_scores_53_y_0_to_fp16 = const()[name = string("_inversed_scores_53_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_53_cast_fp16 = mul(x = var_3307_cast_fp16, y = _inversed_scores_53_y_0_to_fp16)[name = string("_inversed_scores_53_cast_fp16")]; + tensor scores_55_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_53_cast_fp16, cond = mask_11)[name = string("scores_55_cast_fp16")]; + tensor var_3313_cast_fp16 = softmax(axis = var_59, x = scores_55_cast_fp16)[name = string("op_3313_cast_fp16")]; + tensor input_717_cast_fp16 = select(a = var_44_to_fp16, b = var_3313_cast_fp16, cond = mask_11)[name = string("input_717_cast_fp16")]; + bool x_351_transpose_x_0 = const()[name = string("x_351_transpose_x_0"), val = bool(false)]; + bool x_351_transpose_y_0 = const()[name = string("x_351_transpose_y_0"), val = bool(false)]; + tensor value_35_cast_fp16 = transpose(perm = value_35_perm_0, x = v_27_cast_fp16)[name = string("transpose_242")]; + tensor x_351_cast_fp16 = matmul(transpose_x = x_351_transpose_x_0, transpose_y = x_351_transpose_y_0, x = input_717_cast_fp16, y = value_35_cast_fp16)[name = string("x_351_cast_fp16")]; + tensor var_3317_perm_0 = const()[name = string("op_3317_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_3318 = const()[name = string("op_3318"), val = tensor([1, -1, 1024])]; + tensor var_3317_cast_fp16 = transpose(perm = var_3317_perm_0, x = x_351_cast_fp16)[name = string("transpose_241")]; + tensor input_719_cast_fp16 = reshape(shape = var_3318, x = var_3317_cast_fp16)[name = string("input_719_cast_fp16")]; + tensor encoder_layers_13_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275116864))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275903360))))[name = string("encoder_layers_13_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_13_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_13_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275903552)))]; + tensor linear_124_cast_fp16 = linear(bias = encoder_layers_13_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_13_self_attn_linear_out_weight_to_fp16_palettized, x = input_719_cast_fp16)[name = string("linear_124_cast_fp16")]; + tensor input_723_cast_fp16 = add(x = input_713_cast_fp16, y = linear_124_cast_fp16)[name = string("input_723_cast_fp16")]; + tensor x_355_axes_0 = const()[name = string("x_355_axes_0"), val = tensor([-1])]; + tensor encoder_layers_13_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_13_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275905664)))]; + tensor encoder_layers_13_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_13_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275907776)))]; + tensor x_355_cast_fp16 = layer_norm(axes = x_355_axes_0, beta = encoder_layers_13_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_13_norm_conv_weight_to_fp16, x = input_723_cast_fp16)[name = string("x_355_cast_fp16")]; + tensor input_725_perm_0 = const()[name = string("input_725_perm_0"), val = tensor([0, 2, 1])]; + string input_727_pad_type_0 = const()[name = string("input_727_pad_type_0"), val = string("valid")]; + tensor input_727_strides_0 = const()[name = string("input_727_strides_0"), val = tensor([1])]; + tensor input_727_pad_0 = const()[name = string("input_727_pad_0"), val = tensor([0, 0])]; + tensor input_727_dilations_0 = const()[name = string("input_727_dilations_0"), val = tensor([1])]; + int32 input_727_groups_0 = const()[name = string("input_727_groups_0"), val = int32(1)]; + tensor encoder_layers_13_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275909888))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278007104))))[name = string("encoder_layers_13_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_725_cast_fp16 = transpose(perm = input_725_perm_0, x = x_355_cast_fp16)[name = string("transpose_240")]; + tensor input_727_cast_fp16 = conv(dilations = input_727_dilations_0, groups = input_727_groups_0, pad = input_727_pad_0, pad_type = input_727_pad_type_0, strides = input_727_strides_0, weight = encoder_layers_13_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_725_cast_fp16)[name = string("input_727_cast_fp16")]; + int32 x_357_split_num_splits_0 = const()[name = string("x_357_split_num_splits_0"), val = int32(2)]; + int32 x_357_split_axis_0 = const()[name = string("x_357_split_axis_0"), val = int32(1)]; + tensor x_357_split_cast_fp16_0, tensor x_357_split_cast_fp16_1 = split(axis = x_357_split_axis_0, num_splits = x_357_split_num_splits_0, x = input_727_cast_fp16)[name = string("x_357_split_cast_fp16")]; + tensor x_357_split_1_sigmoid_cast_fp16 = sigmoid(x = x_357_split_cast_fp16_1)[name = string("x_357_split_1_sigmoid_cast_fp16")]; + tensor x_357_cast_fp16 = mul(x = x_357_split_cast_fp16_0, y = x_357_split_1_sigmoid_cast_fp16)[name = string("x_357_cast_fp16")]; + tensor input_729_cast_fp16 = select(a = var_44_to_fp16, b = x_357_cast_fp16, cond = var_575)[name = string("input_729_cast_fp16")]; + bool new_x_55_interleave_0 = const()[name = string("new_x_55_interleave_0"), val = bool(false)]; + tensor new_x_55_cast_fp16 = concat(axis = var_59, interleave = new_x_55_interleave_0, values = (cache_55_cast_fp16, input_729_cast_fp16))[name = string("new_x_55_cast_fp16")]; + tensor var_3357_begin_0 = const()[name = string("op_3357_begin_0"), val = tensor([0, 0, 28])]; + tensor var_3357_end_0 = const()[name = string("op_3357_end_0"), val = tensor([1, 1024, 36])]; + tensor var_3357_end_mask_0 = const()[name = string("op_3357_end_mask_0"), val = tensor([true, true, true])]; + tensor var_3357_cast_fp16 = slice_by_index(begin = var_3357_begin_0, end = var_3357_end_0, end_mask = var_3357_end_mask_0, x = new_x_55_cast_fp16)[name = string("op_3357_cast_fp16")]; + string x_359_pad_type_0 = const()[name = string("x_359_pad_type_0"), val = string("valid")]; + int32 x_359_groups_0 = const()[name = string("x_359_groups_0"), val = int32(1024)]; + tensor x_359_strides_0 = const()[name = string("x_359_strides_0"), val = tensor([1])]; + tensor x_359_pad_0 = const()[name = string("x_359_pad_0"), val = tensor([0, 0])]; + tensor x_359_dilations_0 = const()[name = string("x_359_dilations_0"), val = tensor([1])]; + tensor encoder_layers_13_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278011264))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278020544))))[name = string("encoder_layers_13_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_359_cast_fp16 = conv(dilations = x_359_dilations_0, groups = x_359_groups_0, pad = x_359_pad_0, pad_type = x_359_pad_type_0, strides = x_359_strides_0, weight = encoder_layers_13_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_55_cast_fp16)[name = string("x_359_cast_fp16")]; + tensor input_731_perm_0 = const()[name = string("input_731_perm_0"), val = tensor([0, 2, 1])]; + tensor x_361_axes_0 = const()[name = string("x_361_axes_0"), val = tensor([-1])]; + tensor encoder_layers_13_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_13_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278022656)))]; + tensor encoder_layers_13_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_13_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278024768)))]; + tensor input_731_cast_fp16 = transpose(perm = input_731_perm_0, x = x_359_cast_fp16)[name = string("transpose_239")]; + tensor x_361_cast_fp16 = layer_norm(axes = x_361_axes_0, beta = encoder_layers_13_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_13_conv_batch_norm_weight_to_fp16, x = input_731_cast_fp16)[name = string("x_361_cast_fp16")]; + tensor input_733_perm_0 = const()[name = string("input_733_perm_0"), val = tensor([0, 2, 1])]; + tensor input_733_cast_fp16 = transpose(perm = input_733_perm_0, x = x_361_cast_fp16)[name = string("transpose_238")]; + tensor input_735_cast_fp16 = silu(x = input_733_cast_fp16)[name = string("input_735_cast_fp16")]; + string x_363_pad_type_0 = const()[name = string("x_363_pad_type_0"), val = string("valid")]; + tensor x_363_strides_0 = const()[name = string("x_363_strides_0"), val = tensor([1])]; + tensor x_363_pad_0 = const()[name = string("x_363_pad_0"), val = tensor([0, 0])]; + tensor x_363_dilations_0 = const()[name = string("x_363_dilations_0"), val = tensor([1])]; + int32 x_363_groups_0 = const()[name = string("x_363_groups_0"), val = int32(1)]; + tensor encoder_layers_13_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278026880))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(279075520))))[name = string("encoder_layers_13_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_363_cast_fp16 = conv(dilations = x_363_dilations_0, groups = x_363_groups_0, pad = x_363_pad_0, pad_type = x_363_pad_type_0, strides = x_363_strides_0, weight = encoder_layers_13_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_735_cast_fp16)[name = string("x_363_cast_fp16")]; + tensor input_737_perm_0 = const()[name = string("input_737_perm_0"), val = tensor([0, 2, 1])]; + tensor input_737_cast_fp16 = transpose(perm = input_737_perm_0, x = x_363_cast_fp16)[name = string("transpose_237")]; + tensor input_739_cast_fp16 = add(x = input_723_cast_fp16, y = input_737_cast_fp16)[name = string("input_739_cast_fp16")]; + tensor input_741_axes_0 = const()[name = string("input_741_axes_0"), val = tensor([-1])]; + tensor encoder_layers_13_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_13_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(279077632)))]; + tensor encoder_layers_13_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_13_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(279079744)))]; + tensor input_741_cast_fp16 = layer_norm(axes = input_741_axes_0, beta = encoder_layers_13_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_13_norm_feed_forward2_weight_to_fp16, x = input_739_cast_fp16)[name = string("input_741_cast_fp16")]; + tensor encoder_layers_13_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(279081856))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(282227648))))[name = string("encoder_layers_13_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_13_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_13_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(282227840)))]; + tensor linear_125_cast_fp16 = linear(bias = encoder_layers_13_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_13_feed_forward2_linear1_weight_to_fp16_palettized, x = input_741_cast_fp16)[name = string("linear_125_cast_fp16")]; + tensor input_745_cast_fp16 = silu(x = linear_125_cast_fp16)[name = string("input_745_cast_fp16")]; + tensor encoder_layers_13_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(282236096))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(285381888))))[name = string("encoder_layers_13_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_13_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_13_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(285382080)))]; + tensor linear_126_cast_fp16 = linear(bias = encoder_layers_13_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_13_feed_forward2_linear2_weight_to_fp16_palettized, x = input_745_cast_fp16)[name = string("linear_126_cast_fp16")]; + fp16 var_3400_to_fp16 = const()[name = string("op_3400_to_fp16"), val = fp16(0x1p-1)]; + tensor var_3401_cast_fp16 = mul(x = linear_126_cast_fp16, y = var_3400_to_fp16)[name = string("op_3401_cast_fp16")]; + tensor input_751_cast_fp16 = add(x = input_739_cast_fp16, y = var_3401_cast_fp16)[name = string("input_751_cast_fp16")]; + tensor input_753_axes_0 = const()[name = string("input_753_axes_0"), val = tensor([-1])]; + tensor encoder_layers_13_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_13_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(285384192)))]; + tensor encoder_layers_13_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_13_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(285386304)))]; + tensor input_753_cast_fp16 = layer_norm(axes = input_753_axes_0, beta = encoder_layers_13_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_13_norm_out_weight_to_fp16, x = input_751_cast_fp16)[name = string("input_753_cast_fp16")]; + tensor cache_57_begin_0 = const()[name = string("cache_57_begin_0"), val = tensor([14, 0, 0, 0])]; + tensor cache_57_end_0 = const()[name = string("cache_57_end_0"), val = tensor([15, 1, 42, 1024])]; + tensor cache_57_end_mask_0 = const()[name = string("cache_57_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_57_squeeze_mask_0 = const()[name = string("cache_57_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_57_cast_fp16 = slice_by_index(begin = cache_57_begin_0, end = cache_57_end_0, end_mask = cache_57_end_mask_0, squeeze_mask = cache_57_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_57_cast_fp16")]; + tensor cache_59_begin_0 = const()[name = string("cache_59_begin_0"), val = tensor([14, 0, 0, 0])]; + tensor cache_59_end_0 = const()[name = string("cache_59_end_0"), val = tensor([15, 1, 1024, 8])]; + tensor cache_59_end_mask_0 = const()[name = string("cache_59_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_59_squeeze_mask_0 = const()[name = string("cache_59_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_59_cast_fp16 = slice_by_index(begin = cache_59_begin_0, end = cache_59_end_0, end_mask = cache_59_end_mask_0, squeeze_mask = cache_59_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_59_cast_fp16")]; + tensor input_755_axes_0 = const()[name = string("input_755_axes_0"), val = tensor([-1])]; + tensor encoder_layers_14_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_14_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(285388416)))]; + tensor encoder_layers_14_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_14_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(285390528)))]; + tensor input_755_cast_fp16 = layer_norm(axes = input_755_axes_0, beta = encoder_layers_14_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_14_norm_feed_forward1_weight_to_fp16, x = input_753_cast_fp16)[name = string("input_755_cast_fp16")]; + tensor encoder_layers_14_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(285392640))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(288538432))))[name = string("encoder_layers_14_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_14_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_14_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(288538624)))]; + tensor linear_127_cast_fp16 = linear(bias = encoder_layers_14_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_14_feed_forward1_linear1_weight_to_fp16_palettized, x = input_755_cast_fp16)[name = string("linear_127_cast_fp16")]; + tensor input_759_cast_fp16 = silu(x = linear_127_cast_fp16)[name = string("input_759_cast_fp16")]; + tensor encoder_layers_14_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(288546880))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291692672))))[name = string("encoder_layers_14_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_14_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_14_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291692864)))]; + tensor linear_128_cast_fp16 = linear(bias = encoder_layers_14_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_14_feed_forward1_linear2_weight_to_fp16_palettized, x = input_759_cast_fp16)[name = string("linear_128_cast_fp16")]; + fp16 var_3437_to_fp16 = const()[name = string("op_3437_to_fp16"), val = fp16(0x1p-1)]; + tensor var_3438_cast_fp16 = mul(x = linear_128_cast_fp16, y = var_3437_to_fp16)[name = string("op_3438_cast_fp16")]; + tensor input_765_cast_fp16 = add(x = input_753_cast_fp16, y = var_3438_cast_fp16)[name = string("input_765_cast_fp16")]; + tensor key_29_axes_0 = const()[name = string("key_29_axes_0"), val = tensor([-1])]; + tensor encoder_layers_14_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_14_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291694976)))]; + tensor encoder_layers_14_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_14_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291697088)))]; + tensor key_29_cast_fp16 = layer_norm(axes = key_29_axes_0, beta = encoder_layers_14_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_14_norm_self_att_weight_to_fp16, x = input_765_cast_fp16)[name = string("key_29_cast_fp16")]; + bool input_767_interleave_0 = const()[name = string("input_767_interleave_0"), val = bool(false)]; + tensor input_767_cast_fp16 = concat(axis = var_68, interleave = input_767_interleave_0, values = (cache_57_cast_fp16, key_29_cast_fp16))[name = string("input_767_cast_fp16")]; + tensor var_3460_begin_0 = const()[name = string("op_3460_begin_0"), val = tensor([0, 28, 0])]; + tensor var_3460_end_0 = const()[name = string("op_3460_end_0"), val = tensor([1, 42, 1024])]; + tensor var_3460_end_mask_0 = const()[name = string("op_3460_end_mask_0"), val = tensor([true, true, true])]; + tensor var_3460_cast_fp16 = slice_by_index(begin = var_3460_begin_0, end = var_3460_end_0, end_mask = var_3460_end_mask_0, x = cache_57_cast_fp16)[name = string("op_3460_cast_fp16")]; + bool var_3466_interleave_0 = const()[name = string("op_3466_interleave_0"), val = bool(false)]; + tensor var_3466_cast_fp16 = concat(axis = var_68, interleave = var_3466_interleave_0, values = (var_3460_cast_fp16, key_29_cast_fp16))[name = string("op_3466_cast_fp16")]; + tensor encoder_layers_14_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291699200))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(292485696))))[name = string("encoder_layers_14_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_14_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_14_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(292485888)))]; + tensor linear_129_cast_fp16 = linear(bias = encoder_layers_14_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_14_self_attn_linear_q_weight_to_fp16_palettized, x = key_29_cast_fp16)[name = string("linear_129_cast_fp16")]; + tensor var_3471 = const()[name = string("op_3471"), val = tensor([1, -1, 8, 128])]; + tensor q_85_cast_fp16 = reshape(shape = var_3471, x = linear_129_cast_fp16)[name = string("q_85_cast_fp16")]; + tensor encoder_layers_14_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(292488000))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293274496))))[name = string("encoder_layers_14_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_14_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_14_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293274688)))]; + tensor linear_130_cast_fp16 = linear(bias = encoder_layers_14_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_14_self_attn_linear_k_weight_to_fp16_palettized, x = input_767_cast_fp16)[name = string("linear_130_cast_fp16")]; + tensor var_3476 = const()[name = string("op_3476"), val = tensor([1, -1, 8, 128])]; + tensor k_57_cast_fp16 = reshape(shape = var_3476, x = linear_130_cast_fp16)[name = string("k_57_cast_fp16")]; + tensor encoder_layers_14_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293276800))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294063296))))[name = string("encoder_layers_14_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_14_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_14_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294063488)))]; + tensor linear_131_cast_fp16 = linear(bias = encoder_layers_14_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_14_self_attn_linear_v_weight_to_fp16_palettized, x = input_767_cast_fp16)[name = string("linear_131_cast_fp16")]; + tensor var_3481 = const()[name = string("op_3481"), val = tensor([1, -1, 8, 128])]; + tensor v_29_cast_fp16 = reshape(shape = var_3481, x = linear_131_cast_fp16)[name = string("v_29_cast_fp16")]; + tensor value_37_perm_0 = const()[name = string("value_37_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_14_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_14_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294065600)))]; + tensor var_3494_cast_fp16 = add(x = q_85_cast_fp16, y = encoder_layers_14_self_attn_pos_bias_u_to_fp16)[name = string("op_3494_cast_fp16")]; + tensor encoder_layers_14_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_14_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294067712)))]; + tensor var_3496_cast_fp16 = add(x = q_85_cast_fp16, y = encoder_layers_14_self_attn_pos_bias_v_to_fp16)[name = string("op_3496_cast_fp16")]; + tensor q_with_bias_v_29_perm_0 = const()[name = string("q_with_bias_v_29_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_371_transpose_x_0 = const()[name = string("x_371_transpose_x_0"), val = bool(false)]; + bool x_371_transpose_y_0 = const()[name = string("x_371_transpose_y_0"), val = bool(false)]; + tensor op_3498_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294069824))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294212224))))[name = string("op_3498_to_fp16_quantized")]; + tensor q_with_bias_v_29_cast_fp16 = transpose(perm = q_with_bias_v_29_perm_0, x = var_3496_cast_fp16)[name = string("transpose_236")]; + tensor x_371_cast_fp16 = matmul(transpose_x = x_371_transpose_x_0, transpose_y = x_371_transpose_y_0, x = q_with_bias_v_29_cast_fp16, y = op_3498_to_fp16_quantized)[name = string("x_371_cast_fp16")]; + tensor x_373_pad_0 = const()[name = string("x_373_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_373_mode_0 = const()[name = string("x_373_mode_0"), val = string("constant")]; + fp16 const_261_to_fp16 = const()[name = string("const_261_to_fp16"), val = fp16(0x0p+0)]; + tensor x_373_cast_fp16 = pad(constant_val = const_261_to_fp16, mode = x_373_mode_0, pad = x_373_pad_0, x = x_371_cast_fp16)[name = string("x_373_cast_fp16")]; + tensor var_3506 = const()[name = string("op_3506"), val = tensor([1, 8, -1, 28])]; + tensor x_375_cast_fp16 = reshape(shape = var_3506, x = x_373_cast_fp16)[name = string("x_375_cast_fp16")]; + tensor var_3510_begin_0 = const()[name = string("op_3510_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_3510_end_0 = const()[name = string("op_3510_end_0"), val = tensor([1, 8, 140, 28])]; + tensor var_3510_end_mask_0 = const()[name = string("op_3510_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_3510_cast_fp16 = slice_by_index(begin = var_3510_begin_0, end = var_3510_end_0, end_mask = var_3510_end_mask_0, x = x_375_cast_fp16)[name = string("op_3510_cast_fp16")]; + tensor var_3511 = const()[name = string("op_3511"), val = tensor([1, 8, 28, 139])]; + tensor matrix_bd_57_cast_fp16 = reshape(shape = var_3511, x = var_3510_cast_fp16)[name = string("matrix_bd_57_cast_fp16")]; + bool matrix_ac_29_transpose_x_0 = const()[name = string("matrix_ac_29_transpose_x_0"), val = bool(false)]; + bool matrix_ac_29_transpose_y_0 = const()[name = string("matrix_ac_29_transpose_y_0"), val = bool(false)]; + tensor transpose_124_perm_0 = const()[name = string("transpose_124_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_125_perm_0 = const()[name = string("transpose_125_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_125 = transpose(perm = transpose_125_perm_0, x = k_57_cast_fp16)[name = string("transpose_234")]; + tensor transpose_124 = transpose(perm = transpose_124_perm_0, x = var_3494_cast_fp16)[name = string("transpose_235")]; + tensor matrix_ac_29_cast_fp16 = matmul(transpose_x = matrix_ac_29_transpose_x_0, transpose_y = matrix_ac_29_transpose_y_0, x = transpose_124, y = transpose_125)[name = string("matrix_ac_29_cast_fp16")]; + tensor matrix_bd_59_begin_0 = const()[name = string("matrix_bd_59_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_59_end_0 = const()[name = string("matrix_bd_59_end_0"), val = tensor([1, 8, 28, 70])]; + tensor matrix_bd_59_end_mask_0 = const()[name = string("matrix_bd_59_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_59_cast_fp16 = slice_by_index(begin = matrix_bd_59_begin_0, end = matrix_bd_59_end_0, end_mask = matrix_bd_59_end_mask_0, x = matrix_bd_57_cast_fp16)[name = string("matrix_bd_59_cast_fp16")]; + tensor var_3520_cast_fp16 = add(x = matrix_ac_29_cast_fp16, y = matrix_bd_59_cast_fp16)[name = string("op_3520_cast_fp16")]; + fp16 _inversed_scores_57_y_0_to_fp16 = const()[name = string("_inversed_scores_57_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_57_cast_fp16 = mul(x = var_3520_cast_fp16, y = _inversed_scores_57_y_0_to_fp16)[name = string("_inversed_scores_57_cast_fp16")]; + tensor scores_59_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_57_cast_fp16, cond = mask_11)[name = string("scores_59_cast_fp16")]; + tensor var_3526_cast_fp16 = softmax(axis = var_59, x = scores_59_cast_fp16)[name = string("op_3526_cast_fp16")]; + tensor input_769_cast_fp16 = select(a = var_44_to_fp16, b = var_3526_cast_fp16, cond = mask_11)[name = string("input_769_cast_fp16")]; + bool x_377_transpose_x_0 = const()[name = string("x_377_transpose_x_0"), val = bool(false)]; + bool x_377_transpose_y_0 = const()[name = string("x_377_transpose_y_0"), val = bool(false)]; + tensor value_37_cast_fp16 = transpose(perm = value_37_perm_0, x = v_29_cast_fp16)[name = string("transpose_233")]; + tensor x_377_cast_fp16 = matmul(transpose_x = x_377_transpose_x_0, transpose_y = x_377_transpose_y_0, x = input_769_cast_fp16, y = value_37_cast_fp16)[name = string("x_377_cast_fp16")]; + tensor var_3530_perm_0 = const()[name = string("op_3530_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_3531 = const()[name = string("op_3531"), val = tensor([1, -1, 1024])]; + tensor var_3530_cast_fp16 = transpose(perm = var_3530_perm_0, x = x_377_cast_fp16)[name = string("transpose_232")]; + tensor input_771_cast_fp16 = reshape(shape = var_3531, x = var_3530_cast_fp16)[name = string("input_771_cast_fp16")]; + tensor encoder_layers_14_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294212608))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294999104))))[name = string("encoder_layers_14_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_14_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_14_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294999296)))]; + tensor linear_133_cast_fp16 = linear(bias = encoder_layers_14_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_14_self_attn_linear_out_weight_to_fp16_palettized, x = input_771_cast_fp16)[name = string("linear_133_cast_fp16")]; + tensor input_775_cast_fp16 = add(x = input_765_cast_fp16, y = linear_133_cast_fp16)[name = string("input_775_cast_fp16")]; + tensor x_381_axes_0 = const()[name = string("x_381_axes_0"), val = tensor([-1])]; + tensor encoder_layers_14_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_14_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(295001408)))]; + tensor encoder_layers_14_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_14_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(295003520)))]; + tensor x_381_cast_fp16 = layer_norm(axes = x_381_axes_0, beta = encoder_layers_14_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_14_norm_conv_weight_to_fp16, x = input_775_cast_fp16)[name = string("x_381_cast_fp16")]; + tensor input_777_perm_0 = const()[name = string("input_777_perm_0"), val = tensor([0, 2, 1])]; + string input_779_pad_type_0 = const()[name = string("input_779_pad_type_0"), val = string("valid")]; + tensor input_779_strides_0 = const()[name = string("input_779_strides_0"), val = tensor([1])]; + tensor input_779_pad_0 = const()[name = string("input_779_pad_0"), val = tensor([0, 0])]; + tensor input_779_dilations_0 = const()[name = string("input_779_dilations_0"), val = tensor([1])]; + int32 input_779_groups_0 = const()[name = string("input_779_groups_0"), val = int32(1)]; + tensor encoder_layers_14_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(295005632))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297102848))))[name = string("encoder_layers_14_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_777_cast_fp16 = transpose(perm = input_777_perm_0, x = x_381_cast_fp16)[name = string("transpose_231")]; + tensor input_779_cast_fp16 = conv(dilations = input_779_dilations_0, groups = input_779_groups_0, pad = input_779_pad_0, pad_type = input_779_pad_type_0, strides = input_779_strides_0, weight = encoder_layers_14_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_777_cast_fp16)[name = string("input_779_cast_fp16")]; + int32 x_383_split_num_splits_0 = const()[name = string("x_383_split_num_splits_0"), val = int32(2)]; + int32 x_383_split_axis_0 = const()[name = string("x_383_split_axis_0"), val = int32(1)]; + tensor x_383_split_cast_fp16_0, tensor x_383_split_cast_fp16_1 = split(axis = x_383_split_axis_0, num_splits = x_383_split_num_splits_0, x = input_779_cast_fp16)[name = string("x_383_split_cast_fp16")]; + tensor x_383_split_1_sigmoid_cast_fp16 = sigmoid(x = x_383_split_cast_fp16_1)[name = string("x_383_split_1_sigmoid_cast_fp16")]; + tensor x_383_cast_fp16 = mul(x = x_383_split_cast_fp16_0, y = x_383_split_1_sigmoid_cast_fp16)[name = string("x_383_cast_fp16")]; + tensor input_781_cast_fp16 = select(a = var_44_to_fp16, b = x_383_cast_fp16, cond = var_575)[name = string("input_781_cast_fp16")]; + bool new_x_59_interleave_0 = const()[name = string("new_x_59_interleave_0"), val = bool(false)]; + tensor new_x_59_cast_fp16 = concat(axis = var_59, interleave = new_x_59_interleave_0, values = (cache_59_cast_fp16, input_781_cast_fp16))[name = string("new_x_59_cast_fp16")]; + tensor var_3570_begin_0 = const()[name = string("op_3570_begin_0"), val = tensor([0, 0, 28])]; + tensor var_3570_end_0 = const()[name = string("op_3570_end_0"), val = tensor([1, 1024, 36])]; + tensor var_3570_end_mask_0 = const()[name = string("op_3570_end_mask_0"), val = tensor([true, true, true])]; + tensor var_3570_cast_fp16 = slice_by_index(begin = var_3570_begin_0, end = var_3570_end_0, end_mask = var_3570_end_mask_0, x = new_x_59_cast_fp16)[name = string("op_3570_cast_fp16")]; + string x_385_pad_type_0 = const()[name = string("x_385_pad_type_0"), val = string("valid")]; + int32 x_385_groups_0 = const()[name = string("x_385_groups_0"), val = int32(1024)]; + tensor x_385_strides_0 = const()[name = string("x_385_strides_0"), val = tensor([1])]; + tensor x_385_pad_0 = const()[name = string("x_385_pad_0"), val = tensor([0, 0])]; + tensor x_385_dilations_0 = const()[name = string("x_385_dilations_0"), val = tensor([1])]; + tensor encoder_layers_14_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297107008))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297116288))))[name = string("encoder_layers_14_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_385_cast_fp16 = conv(dilations = x_385_dilations_0, groups = x_385_groups_0, pad = x_385_pad_0, pad_type = x_385_pad_type_0, strides = x_385_strides_0, weight = encoder_layers_14_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_59_cast_fp16)[name = string("x_385_cast_fp16")]; + tensor input_783_perm_0 = const()[name = string("input_783_perm_0"), val = tensor([0, 2, 1])]; + tensor x_387_axes_0 = const()[name = string("x_387_axes_0"), val = tensor([-1])]; + tensor encoder_layers_14_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_14_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297118400)))]; + tensor encoder_layers_14_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_14_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297120512)))]; + tensor input_783_cast_fp16 = transpose(perm = input_783_perm_0, x = x_385_cast_fp16)[name = string("transpose_230")]; + tensor x_387_cast_fp16 = layer_norm(axes = x_387_axes_0, beta = encoder_layers_14_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_14_conv_batch_norm_weight_to_fp16, x = input_783_cast_fp16)[name = string("x_387_cast_fp16")]; + tensor input_785_perm_0 = const()[name = string("input_785_perm_0"), val = tensor([0, 2, 1])]; + tensor input_785_cast_fp16 = transpose(perm = input_785_perm_0, x = x_387_cast_fp16)[name = string("transpose_229")]; + tensor input_787_cast_fp16 = silu(x = input_785_cast_fp16)[name = string("input_787_cast_fp16")]; + string x_389_pad_type_0 = const()[name = string("x_389_pad_type_0"), val = string("valid")]; + tensor x_389_strides_0 = const()[name = string("x_389_strides_0"), val = tensor([1])]; + tensor x_389_pad_0 = const()[name = string("x_389_pad_0"), val = tensor([0, 0])]; + tensor x_389_dilations_0 = const()[name = string("x_389_dilations_0"), val = tensor([1])]; + int32 x_389_groups_0 = const()[name = string("x_389_groups_0"), val = int32(1)]; + tensor encoder_layers_14_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297122624))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(298171264))))[name = string("encoder_layers_14_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_389_cast_fp16 = conv(dilations = x_389_dilations_0, groups = x_389_groups_0, pad = x_389_pad_0, pad_type = x_389_pad_type_0, strides = x_389_strides_0, weight = encoder_layers_14_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_787_cast_fp16)[name = string("x_389_cast_fp16")]; + tensor input_789_perm_0 = const()[name = string("input_789_perm_0"), val = tensor([0, 2, 1])]; + tensor input_789_cast_fp16 = transpose(perm = input_789_perm_0, x = x_389_cast_fp16)[name = string("transpose_228")]; + tensor input_791_cast_fp16 = add(x = input_775_cast_fp16, y = input_789_cast_fp16)[name = string("input_791_cast_fp16")]; + tensor input_793_axes_0 = const()[name = string("input_793_axes_0"), val = tensor([-1])]; + tensor encoder_layers_14_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_14_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(298173376)))]; + tensor encoder_layers_14_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_14_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(298175488)))]; + tensor input_793_cast_fp16 = layer_norm(axes = input_793_axes_0, beta = encoder_layers_14_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_14_norm_feed_forward2_weight_to_fp16, x = input_791_cast_fp16)[name = string("input_793_cast_fp16")]; + tensor encoder_layers_14_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(298177600))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301323392))))[name = string("encoder_layers_14_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_14_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_14_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301323584)))]; + tensor linear_134_cast_fp16 = linear(bias = encoder_layers_14_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_14_feed_forward2_linear1_weight_to_fp16_palettized, x = input_793_cast_fp16)[name = string("linear_134_cast_fp16")]; + tensor input_797_cast_fp16 = silu(x = linear_134_cast_fp16)[name = string("input_797_cast_fp16")]; + tensor encoder_layers_14_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301331840))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(304477632))))[name = string("encoder_layers_14_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_14_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_14_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(304477824)))]; + tensor linear_135_cast_fp16 = linear(bias = encoder_layers_14_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_14_feed_forward2_linear2_weight_to_fp16_palettized, x = input_797_cast_fp16)[name = string("linear_135_cast_fp16")]; + fp16 var_3613_to_fp16 = const()[name = string("op_3613_to_fp16"), val = fp16(0x1p-1)]; + tensor var_3614_cast_fp16 = mul(x = linear_135_cast_fp16, y = var_3613_to_fp16)[name = string("op_3614_cast_fp16")]; + tensor input_803_cast_fp16 = add(x = input_791_cast_fp16, y = var_3614_cast_fp16)[name = string("input_803_cast_fp16")]; + tensor input_805_axes_0 = const()[name = string("input_805_axes_0"), val = tensor([-1])]; + tensor encoder_layers_14_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_14_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(304479936)))]; + tensor encoder_layers_14_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_14_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(304482048)))]; + tensor input_805_cast_fp16 = layer_norm(axes = input_805_axes_0, beta = encoder_layers_14_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_14_norm_out_weight_to_fp16, x = input_803_cast_fp16)[name = string("input_805_cast_fp16")]; + tensor cache_61_begin_0 = const()[name = string("cache_61_begin_0"), val = tensor([15, 0, 0, 0])]; + tensor cache_61_end_0 = const()[name = string("cache_61_end_0"), val = tensor([16, 1, 42, 1024])]; + tensor cache_61_end_mask_0 = const()[name = string("cache_61_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_61_squeeze_mask_0 = const()[name = string("cache_61_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_61_cast_fp16 = slice_by_index(begin = cache_61_begin_0, end = cache_61_end_0, end_mask = cache_61_end_mask_0, squeeze_mask = cache_61_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_61_cast_fp16")]; + tensor cache_63_begin_0 = const()[name = string("cache_63_begin_0"), val = tensor([15, 0, 0, 0])]; + tensor cache_63_end_0 = const()[name = string("cache_63_end_0"), val = tensor([16, 1, 1024, 8])]; + tensor cache_63_end_mask_0 = const()[name = string("cache_63_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_63_squeeze_mask_0 = const()[name = string("cache_63_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_63_cast_fp16 = slice_by_index(begin = cache_63_begin_0, end = cache_63_end_0, end_mask = cache_63_end_mask_0, squeeze_mask = cache_63_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_63_cast_fp16")]; + tensor input_807_axes_0 = const()[name = string("input_807_axes_0"), val = tensor([-1])]; + tensor encoder_layers_15_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_15_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(304484160)))]; + tensor encoder_layers_15_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_15_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(304486272)))]; + tensor input_807_cast_fp16 = layer_norm(axes = input_807_axes_0, beta = encoder_layers_15_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_15_norm_feed_forward1_weight_to_fp16, x = input_805_cast_fp16)[name = string("input_807_cast_fp16")]; + tensor encoder_layers_15_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(304488384))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307634176))))[name = string("encoder_layers_15_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_15_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_15_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307634368)))]; + tensor linear_136_cast_fp16 = linear(bias = encoder_layers_15_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_15_feed_forward1_linear1_weight_to_fp16_palettized, x = input_807_cast_fp16)[name = string("linear_136_cast_fp16")]; + tensor input_811_cast_fp16 = silu(x = linear_136_cast_fp16)[name = string("input_811_cast_fp16")]; + tensor encoder_layers_15_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307642624))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310788416))))[name = string("encoder_layers_15_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_15_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_15_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310788608)))]; + tensor linear_137_cast_fp16 = linear(bias = encoder_layers_15_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_15_feed_forward1_linear2_weight_to_fp16_palettized, x = input_811_cast_fp16)[name = string("linear_137_cast_fp16")]; + fp16 var_3650_to_fp16 = const()[name = string("op_3650_to_fp16"), val = fp16(0x1p-1)]; + tensor var_3651_cast_fp16 = mul(x = linear_137_cast_fp16, y = var_3650_to_fp16)[name = string("op_3651_cast_fp16")]; + tensor input_817_cast_fp16 = add(x = input_805_cast_fp16, y = var_3651_cast_fp16)[name = string("input_817_cast_fp16")]; + tensor key_31_axes_0 = const()[name = string("key_31_axes_0"), val = tensor([-1])]; + tensor encoder_layers_15_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_15_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310790720)))]; + tensor encoder_layers_15_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_15_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310792832)))]; + tensor key_31_cast_fp16 = layer_norm(axes = key_31_axes_0, beta = encoder_layers_15_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_15_norm_self_att_weight_to_fp16, x = input_817_cast_fp16)[name = string("key_31_cast_fp16")]; + bool input_819_interleave_0 = const()[name = string("input_819_interleave_0"), val = bool(false)]; + tensor input_819_cast_fp16 = concat(axis = var_68, interleave = input_819_interleave_0, values = (cache_61_cast_fp16, key_31_cast_fp16))[name = string("input_819_cast_fp16")]; + tensor var_3673_begin_0 = const()[name = string("op_3673_begin_0"), val = tensor([0, 28, 0])]; + tensor var_3673_end_0 = const()[name = string("op_3673_end_0"), val = tensor([1, 42, 1024])]; + tensor var_3673_end_mask_0 = const()[name = string("op_3673_end_mask_0"), val = tensor([true, true, true])]; + tensor var_3673_cast_fp16 = slice_by_index(begin = var_3673_begin_0, end = var_3673_end_0, end_mask = var_3673_end_mask_0, x = cache_61_cast_fp16)[name = string("op_3673_cast_fp16")]; + bool var_3679_interleave_0 = const()[name = string("op_3679_interleave_0"), val = bool(false)]; + tensor var_3679_cast_fp16 = concat(axis = var_68, interleave = var_3679_interleave_0, values = (var_3673_cast_fp16, key_31_cast_fp16))[name = string("op_3679_cast_fp16")]; + tensor encoder_layers_15_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310794944))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311581440))))[name = string("encoder_layers_15_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_15_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_15_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311581632)))]; + tensor linear_138_cast_fp16 = linear(bias = encoder_layers_15_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_15_self_attn_linear_q_weight_to_fp16_palettized, x = key_31_cast_fp16)[name = string("linear_138_cast_fp16")]; + tensor var_3684 = const()[name = string("op_3684"), val = tensor([1, -1, 8, 128])]; + tensor q_91_cast_fp16 = reshape(shape = var_3684, x = linear_138_cast_fp16)[name = string("q_91_cast_fp16")]; + tensor encoder_layers_15_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311583744))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(312370240))))[name = string("encoder_layers_15_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_15_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_15_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(312370432)))]; + tensor linear_139_cast_fp16 = linear(bias = encoder_layers_15_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_15_self_attn_linear_k_weight_to_fp16_palettized, x = input_819_cast_fp16)[name = string("linear_139_cast_fp16")]; + tensor var_3689 = const()[name = string("op_3689"), val = tensor([1, -1, 8, 128])]; + tensor k_61_cast_fp16 = reshape(shape = var_3689, x = linear_139_cast_fp16)[name = string("k_61_cast_fp16")]; + tensor encoder_layers_15_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(312372544))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(313159040))))[name = string("encoder_layers_15_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_15_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_15_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(313159232)))]; + tensor linear_140_cast_fp16 = linear(bias = encoder_layers_15_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_15_self_attn_linear_v_weight_to_fp16_palettized, x = input_819_cast_fp16)[name = string("linear_140_cast_fp16")]; + tensor var_3694 = const()[name = string("op_3694"), val = tensor([1, -1, 8, 128])]; + tensor v_31_cast_fp16 = reshape(shape = var_3694, x = linear_140_cast_fp16)[name = string("v_31_cast_fp16")]; + tensor value_39_perm_0 = const()[name = string("value_39_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_15_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_15_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(313161344)))]; + tensor var_3707_cast_fp16 = add(x = q_91_cast_fp16, y = encoder_layers_15_self_attn_pos_bias_u_to_fp16)[name = string("op_3707_cast_fp16")]; + tensor encoder_layers_15_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_15_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(313163456)))]; + tensor var_3709_cast_fp16 = add(x = q_91_cast_fp16, y = encoder_layers_15_self_attn_pos_bias_v_to_fp16)[name = string("op_3709_cast_fp16")]; + tensor q_with_bias_v_31_perm_0 = const()[name = string("q_with_bias_v_31_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_397_transpose_x_0 = const()[name = string("x_397_transpose_x_0"), val = bool(false)]; + bool x_397_transpose_y_0 = const()[name = string("x_397_transpose_y_0"), val = bool(false)]; + tensor op_3711_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(313165568))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(313307968))))[name = string("op_3711_to_fp16_quantized")]; + tensor q_with_bias_v_31_cast_fp16 = transpose(perm = q_with_bias_v_31_perm_0, x = var_3709_cast_fp16)[name = string("transpose_227")]; + tensor x_397_cast_fp16 = matmul(transpose_x = x_397_transpose_x_0, transpose_y = x_397_transpose_y_0, x = q_with_bias_v_31_cast_fp16, y = op_3711_to_fp16_quantized)[name = string("x_397_cast_fp16")]; + tensor x_399_pad_0 = const()[name = string("x_399_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_399_mode_0 = const()[name = string("x_399_mode_0"), val = string("constant")]; + fp16 const_274_to_fp16 = const()[name = string("const_274_to_fp16"), val = fp16(0x0p+0)]; + tensor x_399_cast_fp16 = pad(constant_val = const_274_to_fp16, mode = x_399_mode_0, pad = x_399_pad_0, x = x_397_cast_fp16)[name = string("x_399_cast_fp16")]; + tensor var_3719 = const()[name = string("op_3719"), val = tensor([1, 8, -1, 28])]; + tensor x_401_cast_fp16 = reshape(shape = var_3719, x = x_399_cast_fp16)[name = string("x_401_cast_fp16")]; + tensor var_3723_begin_0 = const()[name = string("op_3723_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_3723_end_0 = const()[name = string("op_3723_end_0"), val = tensor([1, 8, 140, 28])]; + tensor var_3723_end_mask_0 = const()[name = string("op_3723_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_3723_cast_fp16 = slice_by_index(begin = var_3723_begin_0, end = var_3723_end_0, end_mask = var_3723_end_mask_0, x = x_401_cast_fp16)[name = string("op_3723_cast_fp16")]; + tensor var_3724 = const()[name = string("op_3724"), val = tensor([1, 8, 28, 139])]; + tensor matrix_bd_61_cast_fp16 = reshape(shape = var_3724, x = var_3723_cast_fp16)[name = string("matrix_bd_61_cast_fp16")]; + bool matrix_ac_31_transpose_x_0 = const()[name = string("matrix_ac_31_transpose_x_0"), val = bool(false)]; + bool matrix_ac_31_transpose_y_0 = const()[name = string("matrix_ac_31_transpose_y_0"), val = bool(false)]; + tensor transpose_126_perm_0 = const()[name = string("transpose_126_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_127_perm_0 = const()[name = string("transpose_127_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_127 = transpose(perm = transpose_127_perm_0, x = k_61_cast_fp16)[name = string("transpose_225")]; + tensor transpose_126 = transpose(perm = transpose_126_perm_0, x = var_3707_cast_fp16)[name = string("transpose_226")]; + tensor matrix_ac_31_cast_fp16 = matmul(transpose_x = matrix_ac_31_transpose_x_0, transpose_y = matrix_ac_31_transpose_y_0, x = transpose_126, y = transpose_127)[name = string("matrix_ac_31_cast_fp16")]; + tensor matrix_bd_63_begin_0 = const()[name = string("matrix_bd_63_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_63_end_0 = const()[name = string("matrix_bd_63_end_0"), val = tensor([1, 8, 28, 70])]; + tensor matrix_bd_63_end_mask_0 = const()[name = string("matrix_bd_63_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_63_cast_fp16 = slice_by_index(begin = matrix_bd_63_begin_0, end = matrix_bd_63_end_0, end_mask = matrix_bd_63_end_mask_0, x = matrix_bd_61_cast_fp16)[name = string("matrix_bd_63_cast_fp16")]; + tensor var_3733_cast_fp16 = add(x = matrix_ac_31_cast_fp16, y = matrix_bd_63_cast_fp16)[name = string("op_3733_cast_fp16")]; + fp16 _inversed_scores_61_y_0_to_fp16 = const()[name = string("_inversed_scores_61_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_61_cast_fp16 = mul(x = var_3733_cast_fp16, y = _inversed_scores_61_y_0_to_fp16)[name = string("_inversed_scores_61_cast_fp16")]; + tensor scores_63_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_61_cast_fp16, cond = mask_11)[name = string("scores_63_cast_fp16")]; + tensor var_3739_cast_fp16 = softmax(axis = var_59, x = scores_63_cast_fp16)[name = string("op_3739_cast_fp16")]; + tensor input_821_cast_fp16 = select(a = var_44_to_fp16, b = var_3739_cast_fp16, cond = mask_11)[name = string("input_821_cast_fp16")]; + bool x_403_transpose_x_0 = const()[name = string("x_403_transpose_x_0"), val = bool(false)]; + bool x_403_transpose_y_0 = const()[name = string("x_403_transpose_y_0"), val = bool(false)]; + tensor value_39_cast_fp16 = transpose(perm = value_39_perm_0, x = v_31_cast_fp16)[name = string("transpose_224")]; + tensor x_403_cast_fp16 = matmul(transpose_x = x_403_transpose_x_0, transpose_y = x_403_transpose_y_0, x = input_821_cast_fp16, y = value_39_cast_fp16)[name = string("x_403_cast_fp16")]; + tensor var_3743_perm_0 = const()[name = string("op_3743_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_3744 = const()[name = string("op_3744"), val = tensor([1, -1, 1024])]; + tensor var_3743_cast_fp16 = transpose(perm = var_3743_perm_0, x = x_403_cast_fp16)[name = string("transpose_223")]; + tensor input_823_cast_fp16 = reshape(shape = var_3744, x = var_3743_cast_fp16)[name = string("input_823_cast_fp16")]; + tensor encoder_layers_15_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(313308352))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(314094848))))[name = string("encoder_layers_15_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_15_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_15_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(314095040)))]; + tensor linear_142_cast_fp16 = linear(bias = encoder_layers_15_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_15_self_attn_linear_out_weight_to_fp16_palettized, x = input_823_cast_fp16)[name = string("linear_142_cast_fp16")]; + tensor input_827_cast_fp16 = add(x = input_817_cast_fp16, y = linear_142_cast_fp16)[name = string("input_827_cast_fp16")]; + tensor x_407_axes_0 = const()[name = string("x_407_axes_0"), val = tensor([-1])]; + tensor encoder_layers_15_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_15_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(314097152)))]; + tensor encoder_layers_15_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_15_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(314099264)))]; + tensor x_407_cast_fp16 = layer_norm(axes = x_407_axes_0, beta = encoder_layers_15_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_15_norm_conv_weight_to_fp16, x = input_827_cast_fp16)[name = string("x_407_cast_fp16")]; + tensor input_829_perm_0 = const()[name = string("input_829_perm_0"), val = tensor([0, 2, 1])]; + string input_831_pad_type_0 = const()[name = string("input_831_pad_type_0"), val = string("valid")]; + tensor input_831_strides_0 = const()[name = string("input_831_strides_0"), val = tensor([1])]; + tensor input_831_pad_0 = const()[name = string("input_831_pad_0"), val = tensor([0, 0])]; + tensor input_831_dilations_0 = const()[name = string("input_831_dilations_0"), val = tensor([1])]; + int32 input_831_groups_0 = const()[name = string("input_831_groups_0"), val = int32(1)]; + tensor encoder_layers_15_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(314101376))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316198592))))[name = string("encoder_layers_15_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_829_cast_fp16 = transpose(perm = input_829_perm_0, x = x_407_cast_fp16)[name = string("transpose_222")]; + tensor input_831_cast_fp16 = conv(dilations = input_831_dilations_0, groups = input_831_groups_0, pad = input_831_pad_0, pad_type = input_831_pad_type_0, strides = input_831_strides_0, weight = encoder_layers_15_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_829_cast_fp16)[name = string("input_831_cast_fp16")]; + int32 x_409_split_num_splits_0 = const()[name = string("x_409_split_num_splits_0"), val = int32(2)]; + int32 x_409_split_axis_0 = const()[name = string("x_409_split_axis_0"), val = int32(1)]; + tensor x_409_split_cast_fp16_0, tensor x_409_split_cast_fp16_1 = split(axis = x_409_split_axis_0, num_splits = x_409_split_num_splits_0, x = input_831_cast_fp16)[name = string("x_409_split_cast_fp16")]; + tensor x_409_split_1_sigmoid_cast_fp16 = sigmoid(x = x_409_split_cast_fp16_1)[name = string("x_409_split_1_sigmoid_cast_fp16")]; + tensor x_409_cast_fp16 = mul(x = x_409_split_cast_fp16_0, y = x_409_split_1_sigmoid_cast_fp16)[name = string("x_409_cast_fp16")]; + tensor input_833_cast_fp16 = select(a = var_44_to_fp16, b = x_409_cast_fp16, cond = var_575)[name = string("input_833_cast_fp16")]; + bool new_x_63_interleave_0 = const()[name = string("new_x_63_interleave_0"), val = bool(false)]; + tensor new_x_63_cast_fp16 = concat(axis = var_59, interleave = new_x_63_interleave_0, values = (cache_63_cast_fp16, input_833_cast_fp16))[name = string("new_x_63_cast_fp16")]; + tensor var_3783_begin_0 = const()[name = string("op_3783_begin_0"), val = tensor([0, 0, 28])]; + tensor var_3783_end_0 = const()[name = string("op_3783_end_0"), val = tensor([1, 1024, 36])]; + tensor var_3783_end_mask_0 = const()[name = string("op_3783_end_mask_0"), val = tensor([true, true, true])]; + tensor var_3783_cast_fp16 = slice_by_index(begin = var_3783_begin_0, end = var_3783_end_0, end_mask = var_3783_end_mask_0, x = new_x_63_cast_fp16)[name = string("op_3783_cast_fp16")]; + string x_411_pad_type_0 = const()[name = string("x_411_pad_type_0"), val = string("valid")]; + int32 x_411_groups_0 = const()[name = string("x_411_groups_0"), val = int32(1024)]; + tensor x_411_strides_0 = const()[name = string("x_411_strides_0"), val = tensor([1])]; + tensor x_411_pad_0 = const()[name = string("x_411_pad_0"), val = tensor([0, 0])]; + tensor x_411_dilations_0 = const()[name = string("x_411_dilations_0"), val = tensor([1])]; + tensor encoder_layers_15_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316202752))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316212032))))[name = string("encoder_layers_15_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_411_cast_fp16 = conv(dilations = x_411_dilations_0, groups = x_411_groups_0, pad = x_411_pad_0, pad_type = x_411_pad_type_0, strides = x_411_strides_0, weight = encoder_layers_15_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_63_cast_fp16)[name = string("x_411_cast_fp16")]; + tensor input_835_perm_0 = const()[name = string("input_835_perm_0"), val = tensor([0, 2, 1])]; + tensor x_413_axes_0 = const()[name = string("x_413_axes_0"), val = tensor([-1])]; + tensor encoder_layers_15_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_15_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316214144)))]; + tensor encoder_layers_15_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_15_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316216256)))]; + tensor input_835_cast_fp16 = transpose(perm = input_835_perm_0, x = x_411_cast_fp16)[name = string("transpose_221")]; + tensor x_413_cast_fp16 = layer_norm(axes = x_413_axes_0, beta = encoder_layers_15_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_15_conv_batch_norm_weight_to_fp16, x = input_835_cast_fp16)[name = string("x_413_cast_fp16")]; + tensor input_837_perm_0 = const()[name = string("input_837_perm_0"), val = tensor([0, 2, 1])]; + tensor input_837_cast_fp16 = transpose(perm = input_837_perm_0, x = x_413_cast_fp16)[name = string("transpose_220")]; + tensor input_839_cast_fp16 = silu(x = input_837_cast_fp16)[name = string("input_839_cast_fp16")]; + string x_415_pad_type_0 = const()[name = string("x_415_pad_type_0"), val = string("valid")]; + tensor x_415_strides_0 = const()[name = string("x_415_strides_0"), val = tensor([1])]; + tensor x_415_pad_0 = const()[name = string("x_415_pad_0"), val = tensor([0, 0])]; + tensor x_415_dilations_0 = const()[name = string("x_415_dilations_0"), val = tensor([1])]; + int32 x_415_groups_0 = const()[name = string("x_415_groups_0"), val = int32(1)]; + tensor encoder_layers_15_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316218368))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(317267008))))[name = string("encoder_layers_15_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_415_cast_fp16 = conv(dilations = x_415_dilations_0, groups = x_415_groups_0, pad = x_415_pad_0, pad_type = x_415_pad_type_0, strides = x_415_strides_0, weight = encoder_layers_15_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_839_cast_fp16)[name = string("x_415_cast_fp16")]; + tensor input_841_perm_0 = const()[name = string("input_841_perm_0"), val = tensor([0, 2, 1])]; + tensor input_841_cast_fp16 = transpose(perm = input_841_perm_0, x = x_415_cast_fp16)[name = string("transpose_219")]; + tensor input_843_cast_fp16 = add(x = input_827_cast_fp16, y = input_841_cast_fp16)[name = string("input_843_cast_fp16")]; + tensor input_845_axes_0 = const()[name = string("input_845_axes_0"), val = tensor([-1])]; + tensor encoder_layers_15_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_15_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(317269120)))]; + tensor encoder_layers_15_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_15_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(317271232)))]; + tensor input_845_cast_fp16 = layer_norm(axes = input_845_axes_0, beta = encoder_layers_15_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_15_norm_feed_forward2_weight_to_fp16, x = input_843_cast_fp16)[name = string("input_845_cast_fp16")]; + tensor encoder_layers_15_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(317273344))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(320419136))))[name = string("encoder_layers_15_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_15_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_15_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(320419328)))]; + tensor linear_143_cast_fp16 = linear(bias = encoder_layers_15_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_15_feed_forward2_linear1_weight_to_fp16_palettized, x = input_845_cast_fp16)[name = string("linear_143_cast_fp16")]; + tensor input_849_cast_fp16 = silu(x = linear_143_cast_fp16)[name = string("input_849_cast_fp16")]; + tensor encoder_layers_15_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(320427584))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(323573376))))[name = string("encoder_layers_15_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_15_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_15_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(323573568)))]; + tensor linear_144_cast_fp16 = linear(bias = encoder_layers_15_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_15_feed_forward2_linear2_weight_to_fp16_palettized, x = input_849_cast_fp16)[name = string("linear_144_cast_fp16")]; + fp16 var_3826_to_fp16 = const()[name = string("op_3826_to_fp16"), val = fp16(0x1p-1)]; + tensor var_3827_cast_fp16 = mul(x = linear_144_cast_fp16, y = var_3826_to_fp16)[name = string("op_3827_cast_fp16")]; + tensor input_855_cast_fp16 = add(x = input_843_cast_fp16, y = var_3827_cast_fp16)[name = string("input_855_cast_fp16")]; + tensor input_857_axes_0 = const()[name = string("input_857_axes_0"), val = tensor([-1])]; + tensor encoder_layers_15_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_15_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(323575680)))]; + tensor encoder_layers_15_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_15_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(323577792)))]; + tensor input_857_cast_fp16 = layer_norm(axes = input_857_axes_0, beta = encoder_layers_15_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_15_norm_out_weight_to_fp16, x = input_855_cast_fp16)[name = string("input_857_cast_fp16")]; + tensor cache_65_begin_0 = const()[name = string("cache_65_begin_0"), val = tensor([16, 0, 0, 0])]; + tensor cache_65_end_0 = const()[name = string("cache_65_end_0"), val = tensor([17, 1, 42, 1024])]; + tensor cache_65_end_mask_0 = const()[name = string("cache_65_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_65_squeeze_mask_0 = const()[name = string("cache_65_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_65_cast_fp16 = slice_by_index(begin = cache_65_begin_0, end = cache_65_end_0, end_mask = cache_65_end_mask_0, squeeze_mask = cache_65_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_65_cast_fp16")]; + tensor cache_67_begin_0 = const()[name = string("cache_67_begin_0"), val = tensor([16, 0, 0, 0])]; + tensor cache_67_end_0 = const()[name = string("cache_67_end_0"), val = tensor([17, 1, 1024, 8])]; + tensor cache_67_end_mask_0 = const()[name = string("cache_67_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_67_squeeze_mask_0 = const()[name = string("cache_67_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_67_cast_fp16 = slice_by_index(begin = cache_67_begin_0, end = cache_67_end_0, end_mask = cache_67_end_mask_0, squeeze_mask = cache_67_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_67_cast_fp16")]; + tensor input_859_axes_0 = const()[name = string("input_859_axes_0"), val = tensor([-1])]; + tensor encoder_layers_16_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_16_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(323579904)))]; + tensor encoder_layers_16_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_16_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(323582016)))]; + tensor input_859_cast_fp16 = layer_norm(axes = input_859_axes_0, beta = encoder_layers_16_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_16_norm_feed_forward1_weight_to_fp16, x = input_857_cast_fp16)[name = string("input_859_cast_fp16")]; + tensor encoder_layers_16_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(323584128))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(326729920))))[name = string("encoder_layers_16_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_16_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_16_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(326730112)))]; + tensor linear_145_cast_fp16 = linear(bias = encoder_layers_16_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_16_feed_forward1_linear1_weight_to_fp16_palettized, x = input_859_cast_fp16)[name = string("linear_145_cast_fp16")]; + tensor input_863_cast_fp16 = silu(x = linear_145_cast_fp16)[name = string("input_863_cast_fp16")]; + tensor encoder_layers_16_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(326738368))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(329884160))))[name = string("encoder_layers_16_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_16_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_16_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(329884352)))]; + tensor linear_146_cast_fp16 = linear(bias = encoder_layers_16_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_16_feed_forward1_linear2_weight_to_fp16_palettized, x = input_863_cast_fp16)[name = string("linear_146_cast_fp16")]; + fp16 var_3863_to_fp16 = const()[name = string("op_3863_to_fp16"), val = fp16(0x1p-1)]; + tensor var_3864_cast_fp16 = mul(x = linear_146_cast_fp16, y = var_3863_to_fp16)[name = string("op_3864_cast_fp16")]; + tensor input_869_cast_fp16 = add(x = input_857_cast_fp16, y = var_3864_cast_fp16)[name = string("input_869_cast_fp16")]; + tensor key_33_axes_0 = const()[name = string("key_33_axes_0"), val = tensor([-1])]; + tensor encoder_layers_16_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_16_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(329886464)))]; + tensor encoder_layers_16_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_16_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(329888576)))]; + tensor key_33_cast_fp16 = layer_norm(axes = key_33_axes_0, beta = encoder_layers_16_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_16_norm_self_att_weight_to_fp16, x = input_869_cast_fp16)[name = string("key_33_cast_fp16")]; + bool input_871_interleave_0 = const()[name = string("input_871_interleave_0"), val = bool(false)]; + tensor input_871_cast_fp16 = concat(axis = var_68, interleave = input_871_interleave_0, values = (cache_65_cast_fp16, key_33_cast_fp16))[name = string("input_871_cast_fp16")]; + tensor var_3886_begin_0 = const()[name = string("op_3886_begin_0"), val = tensor([0, 28, 0])]; + tensor var_3886_end_0 = const()[name = string("op_3886_end_0"), val = tensor([1, 42, 1024])]; + tensor var_3886_end_mask_0 = const()[name = string("op_3886_end_mask_0"), val = tensor([true, true, true])]; + tensor var_3886_cast_fp16 = slice_by_index(begin = var_3886_begin_0, end = var_3886_end_0, end_mask = var_3886_end_mask_0, x = cache_65_cast_fp16)[name = string("op_3886_cast_fp16")]; + bool var_3892_interleave_0 = const()[name = string("op_3892_interleave_0"), val = bool(false)]; + tensor var_3892_cast_fp16 = concat(axis = var_68, interleave = var_3892_interleave_0, values = (var_3886_cast_fp16, key_33_cast_fp16))[name = string("op_3892_cast_fp16")]; + tensor encoder_layers_16_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(329890688))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(330677184))))[name = string("encoder_layers_16_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_16_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_16_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(330677376)))]; + tensor linear_147_cast_fp16 = linear(bias = encoder_layers_16_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_16_self_attn_linear_q_weight_to_fp16_palettized, x = key_33_cast_fp16)[name = string("linear_147_cast_fp16")]; + tensor var_3897 = const()[name = string("op_3897"), val = tensor([1, -1, 8, 128])]; + tensor q_97_cast_fp16 = reshape(shape = var_3897, x = linear_147_cast_fp16)[name = string("q_97_cast_fp16")]; + tensor encoder_layers_16_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(330679488))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331465984))))[name = string("encoder_layers_16_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_16_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_16_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331466176)))]; + tensor linear_148_cast_fp16 = linear(bias = encoder_layers_16_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_16_self_attn_linear_k_weight_to_fp16_palettized, x = input_871_cast_fp16)[name = string("linear_148_cast_fp16")]; + tensor var_3902 = const()[name = string("op_3902"), val = tensor([1, -1, 8, 128])]; + tensor k_65_cast_fp16 = reshape(shape = var_3902, x = linear_148_cast_fp16)[name = string("k_65_cast_fp16")]; + tensor encoder_layers_16_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331468288))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332254784))))[name = string("encoder_layers_16_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_16_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_16_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332254976)))]; + tensor linear_149_cast_fp16 = linear(bias = encoder_layers_16_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_16_self_attn_linear_v_weight_to_fp16_palettized, x = input_871_cast_fp16)[name = string("linear_149_cast_fp16")]; + tensor var_3907 = const()[name = string("op_3907"), val = tensor([1, -1, 8, 128])]; + tensor v_33_cast_fp16 = reshape(shape = var_3907, x = linear_149_cast_fp16)[name = string("v_33_cast_fp16")]; + tensor value_41_perm_0 = const()[name = string("value_41_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_16_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_16_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332257088)))]; + tensor var_3920_cast_fp16 = add(x = q_97_cast_fp16, y = encoder_layers_16_self_attn_pos_bias_u_to_fp16)[name = string("op_3920_cast_fp16")]; + tensor encoder_layers_16_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_16_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332259200)))]; + tensor var_3922_cast_fp16 = add(x = q_97_cast_fp16, y = encoder_layers_16_self_attn_pos_bias_v_to_fp16)[name = string("op_3922_cast_fp16")]; + tensor q_with_bias_v_33_perm_0 = const()[name = string("q_with_bias_v_33_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_423_transpose_x_0 = const()[name = string("x_423_transpose_x_0"), val = bool(false)]; + bool x_423_transpose_y_0 = const()[name = string("x_423_transpose_y_0"), val = bool(false)]; + tensor op_3924_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332261312))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332403712))))[name = string("op_3924_to_fp16_quantized")]; + tensor q_with_bias_v_33_cast_fp16 = transpose(perm = q_with_bias_v_33_perm_0, x = var_3922_cast_fp16)[name = string("transpose_218")]; + tensor x_423_cast_fp16 = matmul(transpose_x = x_423_transpose_x_0, transpose_y = x_423_transpose_y_0, x = q_with_bias_v_33_cast_fp16, y = op_3924_to_fp16_quantized)[name = string("x_423_cast_fp16")]; + tensor x_425_pad_0 = const()[name = string("x_425_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_425_mode_0 = const()[name = string("x_425_mode_0"), val = string("constant")]; + fp16 const_287_to_fp16 = const()[name = string("const_287_to_fp16"), val = fp16(0x0p+0)]; + tensor x_425_cast_fp16 = pad(constant_val = const_287_to_fp16, mode = x_425_mode_0, pad = x_425_pad_0, x = x_423_cast_fp16)[name = string("x_425_cast_fp16")]; + tensor var_3932 = const()[name = string("op_3932"), val = tensor([1, 8, -1, 28])]; + tensor x_427_cast_fp16 = reshape(shape = var_3932, x = x_425_cast_fp16)[name = string("x_427_cast_fp16")]; + tensor var_3936_begin_0 = const()[name = string("op_3936_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_3936_end_0 = const()[name = string("op_3936_end_0"), val = tensor([1, 8, 140, 28])]; + tensor var_3936_end_mask_0 = const()[name = string("op_3936_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_3936_cast_fp16 = slice_by_index(begin = var_3936_begin_0, end = var_3936_end_0, end_mask = var_3936_end_mask_0, x = x_427_cast_fp16)[name = string("op_3936_cast_fp16")]; + tensor var_3937 = const()[name = string("op_3937"), val = tensor([1, 8, 28, 139])]; + tensor matrix_bd_65_cast_fp16 = reshape(shape = var_3937, x = var_3936_cast_fp16)[name = string("matrix_bd_65_cast_fp16")]; + bool matrix_ac_33_transpose_x_0 = const()[name = string("matrix_ac_33_transpose_x_0"), val = bool(false)]; + bool matrix_ac_33_transpose_y_0 = const()[name = string("matrix_ac_33_transpose_y_0"), val = bool(false)]; + tensor transpose_128_perm_0 = const()[name = string("transpose_128_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_129_perm_0 = const()[name = string("transpose_129_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_129 = transpose(perm = transpose_129_perm_0, x = k_65_cast_fp16)[name = string("transpose_216")]; + tensor transpose_128 = transpose(perm = transpose_128_perm_0, x = var_3920_cast_fp16)[name = string("transpose_217")]; + tensor matrix_ac_33_cast_fp16 = matmul(transpose_x = matrix_ac_33_transpose_x_0, transpose_y = matrix_ac_33_transpose_y_0, x = transpose_128, y = transpose_129)[name = string("matrix_ac_33_cast_fp16")]; + tensor matrix_bd_67_begin_0 = const()[name = string("matrix_bd_67_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_67_end_0 = const()[name = string("matrix_bd_67_end_0"), val = tensor([1, 8, 28, 70])]; + tensor matrix_bd_67_end_mask_0 = const()[name = string("matrix_bd_67_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_67_cast_fp16 = slice_by_index(begin = matrix_bd_67_begin_0, end = matrix_bd_67_end_0, end_mask = matrix_bd_67_end_mask_0, x = matrix_bd_65_cast_fp16)[name = string("matrix_bd_67_cast_fp16")]; + tensor var_3946_cast_fp16 = add(x = matrix_ac_33_cast_fp16, y = matrix_bd_67_cast_fp16)[name = string("op_3946_cast_fp16")]; + fp16 _inversed_scores_65_y_0_to_fp16 = const()[name = string("_inversed_scores_65_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_65_cast_fp16 = mul(x = var_3946_cast_fp16, y = _inversed_scores_65_y_0_to_fp16)[name = string("_inversed_scores_65_cast_fp16")]; + tensor scores_67_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_65_cast_fp16, cond = mask_11)[name = string("scores_67_cast_fp16")]; + tensor var_3952_cast_fp16 = softmax(axis = var_59, x = scores_67_cast_fp16)[name = string("op_3952_cast_fp16")]; + tensor input_873_cast_fp16 = select(a = var_44_to_fp16, b = var_3952_cast_fp16, cond = mask_11)[name = string("input_873_cast_fp16")]; + bool x_429_transpose_x_0 = const()[name = string("x_429_transpose_x_0"), val = bool(false)]; + bool x_429_transpose_y_0 = const()[name = string("x_429_transpose_y_0"), val = bool(false)]; + tensor value_41_cast_fp16 = transpose(perm = value_41_perm_0, x = v_33_cast_fp16)[name = string("transpose_215")]; + tensor x_429_cast_fp16 = matmul(transpose_x = x_429_transpose_x_0, transpose_y = x_429_transpose_y_0, x = input_873_cast_fp16, y = value_41_cast_fp16)[name = string("x_429_cast_fp16")]; + tensor var_3956_perm_0 = const()[name = string("op_3956_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_3957 = const()[name = string("op_3957"), val = tensor([1, -1, 1024])]; + tensor var_3956_cast_fp16 = transpose(perm = var_3956_perm_0, x = x_429_cast_fp16)[name = string("transpose_214")]; + tensor input_875_cast_fp16 = reshape(shape = var_3957, x = var_3956_cast_fp16)[name = string("input_875_cast_fp16")]; + tensor encoder_layers_16_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332404096))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(333190592))))[name = string("encoder_layers_16_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_16_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_16_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(333190784)))]; + tensor linear_151_cast_fp16 = linear(bias = encoder_layers_16_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_16_self_attn_linear_out_weight_to_fp16_palettized, x = input_875_cast_fp16)[name = string("linear_151_cast_fp16")]; + tensor input_879_cast_fp16 = add(x = input_869_cast_fp16, y = linear_151_cast_fp16)[name = string("input_879_cast_fp16")]; + tensor x_433_axes_0 = const()[name = string("x_433_axes_0"), val = tensor([-1])]; + tensor encoder_layers_16_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_16_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(333192896)))]; + tensor encoder_layers_16_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_16_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(333195008)))]; + tensor x_433_cast_fp16 = layer_norm(axes = x_433_axes_0, beta = encoder_layers_16_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_16_norm_conv_weight_to_fp16, x = input_879_cast_fp16)[name = string("x_433_cast_fp16")]; + tensor input_881_perm_0 = const()[name = string("input_881_perm_0"), val = tensor([0, 2, 1])]; + string input_883_pad_type_0 = const()[name = string("input_883_pad_type_0"), val = string("valid")]; + tensor input_883_strides_0 = const()[name = string("input_883_strides_0"), val = tensor([1])]; + tensor input_883_pad_0 = const()[name = string("input_883_pad_0"), val = tensor([0, 0])]; + tensor input_883_dilations_0 = const()[name = string("input_883_dilations_0"), val = tensor([1])]; + int32 input_883_groups_0 = const()[name = string("input_883_groups_0"), val = int32(1)]; + tensor encoder_layers_16_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(333197120))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(335294336))))[name = string("encoder_layers_16_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_881_cast_fp16 = transpose(perm = input_881_perm_0, x = x_433_cast_fp16)[name = string("transpose_213")]; + tensor input_883_cast_fp16 = conv(dilations = input_883_dilations_0, groups = input_883_groups_0, pad = input_883_pad_0, pad_type = input_883_pad_type_0, strides = input_883_strides_0, weight = encoder_layers_16_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_881_cast_fp16)[name = string("input_883_cast_fp16")]; + int32 x_435_split_num_splits_0 = const()[name = string("x_435_split_num_splits_0"), val = int32(2)]; + int32 x_435_split_axis_0 = const()[name = string("x_435_split_axis_0"), val = int32(1)]; + tensor x_435_split_cast_fp16_0, tensor x_435_split_cast_fp16_1 = split(axis = x_435_split_axis_0, num_splits = x_435_split_num_splits_0, x = input_883_cast_fp16)[name = string("x_435_split_cast_fp16")]; + tensor x_435_split_1_sigmoid_cast_fp16 = sigmoid(x = x_435_split_cast_fp16_1)[name = string("x_435_split_1_sigmoid_cast_fp16")]; + tensor x_435_cast_fp16 = mul(x = x_435_split_cast_fp16_0, y = x_435_split_1_sigmoid_cast_fp16)[name = string("x_435_cast_fp16")]; + tensor input_885_cast_fp16 = select(a = var_44_to_fp16, b = x_435_cast_fp16, cond = var_575)[name = string("input_885_cast_fp16")]; + bool new_x_67_interleave_0 = const()[name = string("new_x_67_interleave_0"), val = bool(false)]; + tensor new_x_67_cast_fp16 = concat(axis = var_59, interleave = new_x_67_interleave_0, values = (cache_67_cast_fp16, input_885_cast_fp16))[name = string("new_x_67_cast_fp16")]; + tensor var_3996_begin_0 = const()[name = string("op_3996_begin_0"), val = tensor([0, 0, 28])]; + tensor var_3996_end_0 = const()[name = string("op_3996_end_0"), val = tensor([1, 1024, 36])]; + tensor var_3996_end_mask_0 = const()[name = string("op_3996_end_mask_0"), val = tensor([true, true, true])]; + tensor var_3996_cast_fp16 = slice_by_index(begin = var_3996_begin_0, end = var_3996_end_0, end_mask = var_3996_end_mask_0, x = new_x_67_cast_fp16)[name = string("op_3996_cast_fp16")]; + string x_437_pad_type_0 = const()[name = string("x_437_pad_type_0"), val = string("valid")]; + int32 x_437_groups_0 = const()[name = string("x_437_groups_0"), val = int32(1024)]; + tensor x_437_strides_0 = const()[name = string("x_437_strides_0"), val = tensor([1])]; + tensor x_437_pad_0 = const()[name = string("x_437_pad_0"), val = tensor([0, 0])]; + tensor x_437_dilations_0 = const()[name = string("x_437_dilations_0"), val = tensor([1])]; + tensor encoder_layers_16_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(335298496))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(335307776))))[name = string("encoder_layers_16_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_437_cast_fp16 = conv(dilations = x_437_dilations_0, groups = x_437_groups_0, pad = x_437_pad_0, pad_type = x_437_pad_type_0, strides = x_437_strides_0, weight = encoder_layers_16_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_67_cast_fp16)[name = string("x_437_cast_fp16")]; + tensor input_887_perm_0 = const()[name = string("input_887_perm_0"), val = tensor([0, 2, 1])]; + tensor x_439_axes_0 = const()[name = string("x_439_axes_0"), val = tensor([-1])]; + tensor encoder_layers_16_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_16_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(335309888)))]; + tensor encoder_layers_16_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_16_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(335312000)))]; + tensor input_887_cast_fp16 = transpose(perm = input_887_perm_0, x = x_437_cast_fp16)[name = string("transpose_212")]; + tensor x_439_cast_fp16 = layer_norm(axes = x_439_axes_0, beta = encoder_layers_16_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_16_conv_batch_norm_weight_to_fp16, x = input_887_cast_fp16)[name = string("x_439_cast_fp16")]; + tensor input_889_perm_0 = const()[name = string("input_889_perm_0"), val = tensor([0, 2, 1])]; + tensor input_889_cast_fp16 = transpose(perm = input_889_perm_0, x = x_439_cast_fp16)[name = string("transpose_211")]; + tensor input_891_cast_fp16 = silu(x = input_889_cast_fp16)[name = string("input_891_cast_fp16")]; + string x_441_pad_type_0 = const()[name = string("x_441_pad_type_0"), val = string("valid")]; + tensor x_441_strides_0 = const()[name = string("x_441_strides_0"), val = tensor([1])]; + tensor x_441_pad_0 = const()[name = string("x_441_pad_0"), val = tensor([0, 0])]; + tensor x_441_dilations_0 = const()[name = string("x_441_dilations_0"), val = tensor([1])]; + int32 x_441_groups_0 = const()[name = string("x_441_groups_0"), val = int32(1)]; + tensor encoder_layers_16_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(335314112))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(336362752))))[name = string("encoder_layers_16_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_441_cast_fp16 = conv(dilations = x_441_dilations_0, groups = x_441_groups_0, pad = x_441_pad_0, pad_type = x_441_pad_type_0, strides = x_441_strides_0, weight = encoder_layers_16_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_891_cast_fp16)[name = string("x_441_cast_fp16")]; + tensor input_893_perm_0 = const()[name = string("input_893_perm_0"), val = tensor([0, 2, 1])]; + tensor input_893_cast_fp16 = transpose(perm = input_893_perm_0, x = x_441_cast_fp16)[name = string("transpose_210")]; + tensor input_895_cast_fp16 = add(x = input_879_cast_fp16, y = input_893_cast_fp16)[name = string("input_895_cast_fp16")]; + tensor input_897_axes_0 = const()[name = string("input_897_axes_0"), val = tensor([-1])]; + tensor encoder_layers_16_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_16_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(336364864)))]; + tensor encoder_layers_16_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_16_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(336366976)))]; + tensor input_897_cast_fp16 = layer_norm(axes = input_897_axes_0, beta = encoder_layers_16_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_16_norm_feed_forward2_weight_to_fp16, x = input_895_cast_fp16)[name = string("input_897_cast_fp16")]; + tensor encoder_layers_16_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(336369088))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(339514880))))[name = string("encoder_layers_16_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_16_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_16_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(339515072)))]; + tensor linear_152_cast_fp16 = linear(bias = encoder_layers_16_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_16_feed_forward2_linear1_weight_to_fp16_palettized, x = input_897_cast_fp16)[name = string("linear_152_cast_fp16")]; + tensor input_901_cast_fp16 = silu(x = linear_152_cast_fp16)[name = string("input_901_cast_fp16")]; + tensor encoder_layers_16_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(339523328))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(342669120))))[name = string("encoder_layers_16_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_16_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_16_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(342669312)))]; + tensor linear_153_cast_fp16 = linear(bias = encoder_layers_16_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_16_feed_forward2_linear2_weight_to_fp16_palettized, x = input_901_cast_fp16)[name = string("linear_153_cast_fp16")]; + fp16 var_4039_to_fp16 = const()[name = string("op_4039_to_fp16"), val = fp16(0x1p-1)]; + tensor var_4040_cast_fp16 = mul(x = linear_153_cast_fp16, y = var_4039_to_fp16)[name = string("op_4040_cast_fp16")]; + tensor input_907_cast_fp16 = add(x = input_895_cast_fp16, y = var_4040_cast_fp16)[name = string("input_907_cast_fp16")]; + tensor input_909_axes_0 = const()[name = string("input_909_axes_0"), val = tensor([-1])]; + tensor encoder_layers_16_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_16_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(342671424)))]; + tensor encoder_layers_16_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_16_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(342673536)))]; + tensor input_909_cast_fp16 = layer_norm(axes = input_909_axes_0, beta = encoder_layers_16_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_16_norm_out_weight_to_fp16, x = input_907_cast_fp16)[name = string("input_909_cast_fp16")]; + tensor cache_69_begin_0 = const()[name = string("cache_69_begin_0"), val = tensor([17, 0, 0, 0])]; + tensor cache_69_end_0 = const()[name = string("cache_69_end_0"), val = tensor([18, 1, 42, 1024])]; + tensor cache_69_end_mask_0 = const()[name = string("cache_69_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_69_squeeze_mask_0 = const()[name = string("cache_69_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_69_cast_fp16 = slice_by_index(begin = cache_69_begin_0, end = cache_69_end_0, end_mask = cache_69_end_mask_0, squeeze_mask = cache_69_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_69_cast_fp16")]; + tensor cache_71_begin_0 = const()[name = string("cache_71_begin_0"), val = tensor([17, 0, 0, 0])]; + tensor cache_71_end_0 = const()[name = string("cache_71_end_0"), val = tensor([18, 1, 1024, 8])]; + tensor cache_71_end_mask_0 = const()[name = string("cache_71_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_71_squeeze_mask_0 = const()[name = string("cache_71_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_71_cast_fp16 = slice_by_index(begin = cache_71_begin_0, end = cache_71_end_0, end_mask = cache_71_end_mask_0, squeeze_mask = cache_71_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_71_cast_fp16")]; + tensor input_911_axes_0 = const()[name = string("input_911_axes_0"), val = tensor([-1])]; + tensor encoder_layers_17_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_17_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(342675648)))]; + tensor encoder_layers_17_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_17_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(342677760)))]; + tensor input_911_cast_fp16 = layer_norm(axes = input_911_axes_0, beta = encoder_layers_17_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_17_norm_feed_forward1_weight_to_fp16, x = input_909_cast_fp16)[name = string("input_911_cast_fp16")]; + tensor encoder_layers_17_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(342679872))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(345825664))))[name = string("encoder_layers_17_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_17_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_17_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(345825856)))]; + tensor linear_154_cast_fp16 = linear(bias = encoder_layers_17_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_17_feed_forward1_linear1_weight_to_fp16_palettized, x = input_911_cast_fp16)[name = string("linear_154_cast_fp16")]; + tensor input_915_cast_fp16 = silu(x = linear_154_cast_fp16)[name = string("input_915_cast_fp16")]; + tensor encoder_layers_17_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(345834112))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(348979904))))[name = string("encoder_layers_17_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_17_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_17_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(348980096)))]; + tensor linear_155_cast_fp16 = linear(bias = encoder_layers_17_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_17_feed_forward1_linear2_weight_to_fp16_palettized, x = input_915_cast_fp16)[name = string("linear_155_cast_fp16")]; + fp16 var_4076_to_fp16 = const()[name = string("op_4076_to_fp16"), val = fp16(0x1p-1)]; + tensor var_4077_cast_fp16 = mul(x = linear_155_cast_fp16, y = var_4076_to_fp16)[name = string("op_4077_cast_fp16")]; + tensor input_921_cast_fp16 = add(x = input_909_cast_fp16, y = var_4077_cast_fp16)[name = string("input_921_cast_fp16")]; + tensor key_35_axes_0 = const()[name = string("key_35_axes_0"), val = tensor([-1])]; + tensor encoder_layers_17_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_17_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(348982208)))]; + tensor encoder_layers_17_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_17_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(348984320)))]; + tensor key_35_cast_fp16 = layer_norm(axes = key_35_axes_0, beta = encoder_layers_17_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_17_norm_self_att_weight_to_fp16, x = input_921_cast_fp16)[name = string("key_35_cast_fp16")]; + bool input_923_interleave_0 = const()[name = string("input_923_interleave_0"), val = bool(false)]; + tensor input_923_cast_fp16 = concat(axis = var_68, interleave = input_923_interleave_0, values = (cache_69_cast_fp16, key_35_cast_fp16))[name = string("input_923_cast_fp16")]; + tensor var_4099_begin_0 = const()[name = string("op_4099_begin_0"), val = tensor([0, 28, 0])]; + tensor var_4099_end_0 = const()[name = string("op_4099_end_0"), val = tensor([1, 42, 1024])]; + tensor var_4099_end_mask_0 = const()[name = string("op_4099_end_mask_0"), val = tensor([true, true, true])]; + tensor var_4099_cast_fp16 = slice_by_index(begin = var_4099_begin_0, end = var_4099_end_0, end_mask = var_4099_end_mask_0, x = cache_69_cast_fp16)[name = string("op_4099_cast_fp16")]; + bool var_4105_interleave_0 = const()[name = string("op_4105_interleave_0"), val = bool(false)]; + tensor var_4105_cast_fp16 = concat(axis = var_68, interleave = var_4105_interleave_0, values = (var_4099_cast_fp16, key_35_cast_fp16))[name = string("op_4105_cast_fp16")]; + tensor encoder_layers_17_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(348986432))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(349772928))))[name = string("encoder_layers_17_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_17_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_17_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(349773120)))]; + tensor linear_156_cast_fp16 = linear(bias = encoder_layers_17_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_17_self_attn_linear_q_weight_to_fp16_palettized, x = key_35_cast_fp16)[name = string("linear_156_cast_fp16")]; + tensor var_4110 = const()[name = string("op_4110"), val = tensor([1, -1, 8, 128])]; + tensor q_103_cast_fp16 = reshape(shape = var_4110, x = linear_156_cast_fp16)[name = string("q_103_cast_fp16")]; + tensor encoder_layers_17_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(349775232))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350561728))))[name = string("encoder_layers_17_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_17_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_17_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350561920)))]; + tensor linear_157_cast_fp16 = linear(bias = encoder_layers_17_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_17_self_attn_linear_k_weight_to_fp16_palettized, x = input_923_cast_fp16)[name = string("linear_157_cast_fp16")]; + tensor var_4115 = const()[name = string("op_4115"), val = tensor([1, -1, 8, 128])]; + tensor k_69_cast_fp16 = reshape(shape = var_4115, x = linear_157_cast_fp16)[name = string("k_69_cast_fp16")]; + tensor encoder_layers_17_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350564032))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351350528))))[name = string("encoder_layers_17_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_17_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_17_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351350720)))]; + tensor linear_158_cast_fp16 = linear(bias = encoder_layers_17_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_17_self_attn_linear_v_weight_to_fp16_palettized, x = input_923_cast_fp16)[name = string("linear_158_cast_fp16")]; + tensor var_4120 = const()[name = string("op_4120"), val = tensor([1, -1, 8, 128])]; + tensor v_35_cast_fp16 = reshape(shape = var_4120, x = linear_158_cast_fp16)[name = string("v_35_cast_fp16")]; + tensor value_43_perm_0 = const()[name = string("value_43_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_17_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_17_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351352832)))]; + tensor var_4133_cast_fp16 = add(x = q_103_cast_fp16, y = encoder_layers_17_self_attn_pos_bias_u_to_fp16)[name = string("op_4133_cast_fp16")]; + tensor encoder_layers_17_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_17_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351354944)))]; + tensor var_4135_cast_fp16 = add(x = q_103_cast_fp16, y = encoder_layers_17_self_attn_pos_bias_v_to_fp16)[name = string("op_4135_cast_fp16")]; + tensor q_with_bias_v_35_perm_0 = const()[name = string("q_with_bias_v_35_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_449_transpose_x_0 = const()[name = string("x_449_transpose_x_0"), val = bool(false)]; + bool x_449_transpose_y_0 = const()[name = string("x_449_transpose_y_0"), val = bool(false)]; + tensor op_4137_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351357056))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351499456))))[name = string("op_4137_to_fp16_quantized")]; + tensor q_with_bias_v_35_cast_fp16 = transpose(perm = q_with_bias_v_35_perm_0, x = var_4135_cast_fp16)[name = string("transpose_209")]; + tensor x_449_cast_fp16 = matmul(transpose_x = x_449_transpose_x_0, transpose_y = x_449_transpose_y_0, x = q_with_bias_v_35_cast_fp16, y = op_4137_to_fp16_quantized)[name = string("x_449_cast_fp16")]; + tensor x_451_pad_0 = const()[name = string("x_451_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_451_mode_0 = const()[name = string("x_451_mode_0"), val = string("constant")]; + fp16 const_300_to_fp16 = const()[name = string("const_300_to_fp16"), val = fp16(0x0p+0)]; + tensor x_451_cast_fp16 = pad(constant_val = const_300_to_fp16, mode = x_451_mode_0, pad = x_451_pad_0, x = x_449_cast_fp16)[name = string("x_451_cast_fp16")]; + tensor var_4145 = const()[name = string("op_4145"), val = tensor([1, 8, -1, 28])]; + tensor x_453_cast_fp16 = reshape(shape = var_4145, x = x_451_cast_fp16)[name = string("x_453_cast_fp16")]; + tensor var_4149_begin_0 = const()[name = string("op_4149_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_4149_end_0 = const()[name = string("op_4149_end_0"), val = tensor([1, 8, 140, 28])]; + tensor var_4149_end_mask_0 = const()[name = string("op_4149_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_4149_cast_fp16 = slice_by_index(begin = var_4149_begin_0, end = var_4149_end_0, end_mask = var_4149_end_mask_0, x = x_453_cast_fp16)[name = string("op_4149_cast_fp16")]; + tensor var_4150 = const()[name = string("op_4150"), val = tensor([1, 8, 28, 139])]; + tensor matrix_bd_69_cast_fp16 = reshape(shape = var_4150, x = var_4149_cast_fp16)[name = string("matrix_bd_69_cast_fp16")]; + bool matrix_ac_35_transpose_x_0 = const()[name = string("matrix_ac_35_transpose_x_0"), val = bool(false)]; + bool matrix_ac_35_transpose_y_0 = const()[name = string("matrix_ac_35_transpose_y_0"), val = bool(false)]; + tensor transpose_130_perm_0 = const()[name = string("transpose_130_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_131_perm_0 = const()[name = string("transpose_131_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_131 = transpose(perm = transpose_131_perm_0, x = k_69_cast_fp16)[name = string("transpose_207")]; + tensor transpose_130 = transpose(perm = transpose_130_perm_0, x = var_4133_cast_fp16)[name = string("transpose_208")]; + tensor matrix_ac_35_cast_fp16 = matmul(transpose_x = matrix_ac_35_transpose_x_0, transpose_y = matrix_ac_35_transpose_y_0, x = transpose_130, y = transpose_131)[name = string("matrix_ac_35_cast_fp16")]; + tensor matrix_bd_71_begin_0 = const()[name = string("matrix_bd_71_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_71_end_0 = const()[name = string("matrix_bd_71_end_0"), val = tensor([1, 8, 28, 70])]; + tensor matrix_bd_71_end_mask_0 = const()[name = string("matrix_bd_71_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_71_cast_fp16 = slice_by_index(begin = matrix_bd_71_begin_0, end = matrix_bd_71_end_0, end_mask = matrix_bd_71_end_mask_0, x = matrix_bd_69_cast_fp16)[name = string("matrix_bd_71_cast_fp16")]; + tensor var_4159_cast_fp16 = add(x = matrix_ac_35_cast_fp16, y = matrix_bd_71_cast_fp16)[name = string("op_4159_cast_fp16")]; + fp16 _inversed_scores_69_y_0_to_fp16 = const()[name = string("_inversed_scores_69_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_69_cast_fp16 = mul(x = var_4159_cast_fp16, y = _inversed_scores_69_y_0_to_fp16)[name = string("_inversed_scores_69_cast_fp16")]; + tensor scores_71_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_69_cast_fp16, cond = mask_11)[name = string("scores_71_cast_fp16")]; + tensor var_4165_cast_fp16 = softmax(axis = var_59, x = scores_71_cast_fp16)[name = string("op_4165_cast_fp16")]; + tensor input_925_cast_fp16 = select(a = var_44_to_fp16, b = var_4165_cast_fp16, cond = mask_11)[name = string("input_925_cast_fp16")]; + bool x_455_transpose_x_0 = const()[name = string("x_455_transpose_x_0"), val = bool(false)]; + bool x_455_transpose_y_0 = const()[name = string("x_455_transpose_y_0"), val = bool(false)]; + tensor value_43_cast_fp16 = transpose(perm = value_43_perm_0, x = v_35_cast_fp16)[name = string("transpose_206")]; + tensor x_455_cast_fp16 = matmul(transpose_x = x_455_transpose_x_0, transpose_y = x_455_transpose_y_0, x = input_925_cast_fp16, y = value_43_cast_fp16)[name = string("x_455_cast_fp16")]; + tensor var_4169_perm_0 = const()[name = string("op_4169_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_4170 = const()[name = string("op_4170"), val = tensor([1, -1, 1024])]; + tensor var_4169_cast_fp16 = transpose(perm = var_4169_perm_0, x = x_455_cast_fp16)[name = string("transpose_205")]; + tensor input_927_cast_fp16 = reshape(shape = var_4170, x = var_4169_cast_fp16)[name = string("input_927_cast_fp16")]; + tensor encoder_layers_17_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351499840))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(352286336))))[name = string("encoder_layers_17_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_17_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_17_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(352286528)))]; + tensor linear_160_cast_fp16 = linear(bias = encoder_layers_17_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_17_self_attn_linear_out_weight_to_fp16_palettized, x = input_927_cast_fp16)[name = string("linear_160_cast_fp16")]; + tensor input_931_cast_fp16 = add(x = input_921_cast_fp16, y = linear_160_cast_fp16)[name = string("input_931_cast_fp16")]; + tensor x_459_axes_0 = const()[name = string("x_459_axes_0"), val = tensor([-1])]; + tensor encoder_layers_17_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_17_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(352288640)))]; + tensor encoder_layers_17_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_17_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(352290752)))]; + tensor x_459_cast_fp16 = layer_norm(axes = x_459_axes_0, beta = encoder_layers_17_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_17_norm_conv_weight_to_fp16, x = input_931_cast_fp16)[name = string("x_459_cast_fp16")]; + tensor input_933_perm_0 = const()[name = string("input_933_perm_0"), val = tensor([0, 2, 1])]; + string input_935_pad_type_0 = const()[name = string("input_935_pad_type_0"), val = string("valid")]; + tensor input_935_strides_0 = const()[name = string("input_935_strides_0"), val = tensor([1])]; + tensor input_935_pad_0 = const()[name = string("input_935_pad_0"), val = tensor([0, 0])]; + tensor input_935_dilations_0 = const()[name = string("input_935_dilations_0"), val = tensor([1])]; + int32 input_935_groups_0 = const()[name = string("input_935_groups_0"), val = int32(1)]; + tensor encoder_layers_17_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(352292864))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(354390080))))[name = string("encoder_layers_17_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_933_cast_fp16 = transpose(perm = input_933_perm_0, x = x_459_cast_fp16)[name = string("transpose_204")]; + tensor input_935_cast_fp16 = conv(dilations = input_935_dilations_0, groups = input_935_groups_0, pad = input_935_pad_0, pad_type = input_935_pad_type_0, strides = input_935_strides_0, weight = encoder_layers_17_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_933_cast_fp16)[name = string("input_935_cast_fp16")]; + int32 x_461_split_num_splits_0 = const()[name = string("x_461_split_num_splits_0"), val = int32(2)]; + int32 x_461_split_axis_0 = const()[name = string("x_461_split_axis_0"), val = int32(1)]; + tensor x_461_split_cast_fp16_0, tensor x_461_split_cast_fp16_1 = split(axis = x_461_split_axis_0, num_splits = x_461_split_num_splits_0, x = input_935_cast_fp16)[name = string("x_461_split_cast_fp16")]; + tensor x_461_split_1_sigmoid_cast_fp16 = sigmoid(x = x_461_split_cast_fp16_1)[name = string("x_461_split_1_sigmoid_cast_fp16")]; + tensor x_461_cast_fp16 = mul(x = x_461_split_cast_fp16_0, y = x_461_split_1_sigmoid_cast_fp16)[name = string("x_461_cast_fp16")]; + tensor input_937_cast_fp16 = select(a = var_44_to_fp16, b = x_461_cast_fp16, cond = var_575)[name = string("input_937_cast_fp16")]; + bool new_x_71_interleave_0 = const()[name = string("new_x_71_interleave_0"), val = bool(false)]; + tensor new_x_71_cast_fp16 = concat(axis = var_59, interleave = new_x_71_interleave_0, values = (cache_71_cast_fp16, input_937_cast_fp16))[name = string("new_x_71_cast_fp16")]; + tensor var_4209_begin_0 = const()[name = string("op_4209_begin_0"), val = tensor([0, 0, 28])]; + tensor var_4209_end_0 = const()[name = string("op_4209_end_0"), val = tensor([1, 1024, 36])]; + tensor var_4209_end_mask_0 = const()[name = string("op_4209_end_mask_0"), val = tensor([true, true, true])]; + tensor var_4209_cast_fp16 = slice_by_index(begin = var_4209_begin_0, end = var_4209_end_0, end_mask = var_4209_end_mask_0, x = new_x_71_cast_fp16)[name = string("op_4209_cast_fp16")]; + string x_463_pad_type_0 = const()[name = string("x_463_pad_type_0"), val = string("valid")]; + int32 x_463_groups_0 = const()[name = string("x_463_groups_0"), val = int32(1024)]; + tensor x_463_strides_0 = const()[name = string("x_463_strides_0"), val = tensor([1])]; + tensor x_463_pad_0 = const()[name = string("x_463_pad_0"), val = tensor([0, 0])]; + tensor x_463_dilations_0 = const()[name = string("x_463_dilations_0"), val = tensor([1])]; + tensor encoder_layers_17_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(354394240))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(354403520))))[name = string("encoder_layers_17_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_463_cast_fp16 = conv(dilations = x_463_dilations_0, groups = x_463_groups_0, pad = x_463_pad_0, pad_type = x_463_pad_type_0, strides = x_463_strides_0, weight = encoder_layers_17_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_71_cast_fp16)[name = string("x_463_cast_fp16")]; + tensor input_939_perm_0 = const()[name = string("input_939_perm_0"), val = tensor([0, 2, 1])]; + tensor x_465_axes_0 = const()[name = string("x_465_axes_0"), val = tensor([-1])]; + tensor encoder_layers_17_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_17_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(354405632)))]; + tensor encoder_layers_17_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_17_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(354407744)))]; + tensor input_939_cast_fp16 = transpose(perm = input_939_perm_0, x = x_463_cast_fp16)[name = string("transpose_203")]; + tensor x_465_cast_fp16 = layer_norm(axes = x_465_axes_0, beta = encoder_layers_17_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_17_conv_batch_norm_weight_to_fp16, x = input_939_cast_fp16)[name = string("x_465_cast_fp16")]; + tensor input_941_perm_0 = const()[name = string("input_941_perm_0"), val = tensor([0, 2, 1])]; + tensor input_941_cast_fp16 = transpose(perm = input_941_perm_0, x = x_465_cast_fp16)[name = string("transpose_202")]; + tensor input_943_cast_fp16 = silu(x = input_941_cast_fp16)[name = string("input_943_cast_fp16")]; + string x_467_pad_type_0 = const()[name = string("x_467_pad_type_0"), val = string("valid")]; + tensor x_467_strides_0 = const()[name = string("x_467_strides_0"), val = tensor([1])]; + tensor x_467_pad_0 = const()[name = string("x_467_pad_0"), val = tensor([0, 0])]; + tensor x_467_dilations_0 = const()[name = string("x_467_dilations_0"), val = tensor([1])]; + int32 x_467_groups_0 = const()[name = string("x_467_groups_0"), val = int32(1)]; + tensor encoder_layers_17_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(354409856))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(355458496))))[name = string("encoder_layers_17_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_467_cast_fp16 = conv(dilations = x_467_dilations_0, groups = x_467_groups_0, pad = x_467_pad_0, pad_type = x_467_pad_type_0, strides = x_467_strides_0, weight = encoder_layers_17_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_943_cast_fp16)[name = string("x_467_cast_fp16")]; + tensor input_945_perm_0 = const()[name = string("input_945_perm_0"), val = tensor([0, 2, 1])]; + tensor input_945_cast_fp16 = transpose(perm = input_945_perm_0, x = x_467_cast_fp16)[name = string("transpose_201")]; + tensor input_947_cast_fp16 = add(x = input_931_cast_fp16, y = input_945_cast_fp16)[name = string("input_947_cast_fp16")]; + tensor input_949_axes_0 = const()[name = string("input_949_axes_0"), val = tensor([-1])]; + tensor encoder_layers_17_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_17_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(355460608)))]; + tensor encoder_layers_17_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_17_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(355462720)))]; + tensor input_949_cast_fp16 = layer_norm(axes = input_949_axes_0, beta = encoder_layers_17_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_17_norm_feed_forward2_weight_to_fp16, x = input_947_cast_fp16)[name = string("input_949_cast_fp16")]; + tensor encoder_layers_17_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(355464832))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(358610624))))[name = string("encoder_layers_17_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_17_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_17_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(358610816)))]; + tensor linear_161_cast_fp16 = linear(bias = encoder_layers_17_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_17_feed_forward2_linear1_weight_to_fp16_palettized, x = input_949_cast_fp16)[name = string("linear_161_cast_fp16")]; + tensor input_953_cast_fp16 = silu(x = linear_161_cast_fp16)[name = string("input_953_cast_fp16")]; + tensor encoder_layers_17_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(358619072))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361764864))))[name = string("encoder_layers_17_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_17_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_17_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361765056)))]; + tensor linear_162_cast_fp16 = linear(bias = encoder_layers_17_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_17_feed_forward2_linear2_weight_to_fp16_palettized, x = input_953_cast_fp16)[name = string("linear_162_cast_fp16")]; + fp16 var_4252_to_fp16 = const()[name = string("op_4252_to_fp16"), val = fp16(0x1p-1)]; + tensor var_4253_cast_fp16 = mul(x = linear_162_cast_fp16, y = var_4252_to_fp16)[name = string("op_4253_cast_fp16")]; + tensor input_959_cast_fp16 = add(x = input_947_cast_fp16, y = var_4253_cast_fp16)[name = string("input_959_cast_fp16")]; + tensor input_961_axes_0 = const()[name = string("input_961_axes_0"), val = tensor([-1])]; + tensor encoder_layers_17_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_17_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361767168)))]; + tensor encoder_layers_17_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_17_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361769280)))]; + tensor input_961_cast_fp16 = layer_norm(axes = input_961_axes_0, beta = encoder_layers_17_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_17_norm_out_weight_to_fp16, x = input_959_cast_fp16)[name = string("input_961_cast_fp16")]; + tensor cache_73_begin_0 = const()[name = string("cache_73_begin_0"), val = tensor([18, 0, 0, 0])]; + tensor cache_73_end_0 = const()[name = string("cache_73_end_0"), val = tensor([19, 1, 42, 1024])]; + tensor cache_73_end_mask_0 = const()[name = string("cache_73_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_73_squeeze_mask_0 = const()[name = string("cache_73_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_73_cast_fp16 = slice_by_index(begin = cache_73_begin_0, end = cache_73_end_0, end_mask = cache_73_end_mask_0, squeeze_mask = cache_73_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_73_cast_fp16")]; + tensor cache_75_begin_0 = const()[name = string("cache_75_begin_0"), val = tensor([18, 0, 0, 0])]; + tensor cache_75_end_0 = const()[name = string("cache_75_end_0"), val = tensor([19, 1, 1024, 8])]; + tensor cache_75_end_mask_0 = const()[name = string("cache_75_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_75_squeeze_mask_0 = const()[name = string("cache_75_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_75_cast_fp16 = slice_by_index(begin = cache_75_begin_0, end = cache_75_end_0, end_mask = cache_75_end_mask_0, squeeze_mask = cache_75_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_75_cast_fp16")]; + tensor input_963_axes_0 = const()[name = string("input_963_axes_0"), val = tensor([-1])]; + tensor encoder_layers_18_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_18_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361771392)))]; + tensor encoder_layers_18_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_18_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361773504)))]; + tensor input_963_cast_fp16 = layer_norm(axes = input_963_axes_0, beta = encoder_layers_18_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_18_norm_feed_forward1_weight_to_fp16, x = input_961_cast_fp16)[name = string("input_963_cast_fp16")]; + tensor encoder_layers_18_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361775616))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(364921408))))[name = string("encoder_layers_18_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_18_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_18_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(364921600)))]; + tensor linear_163_cast_fp16 = linear(bias = encoder_layers_18_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_18_feed_forward1_linear1_weight_to_fp16_palettized, x = input_963_cast_fp16)[name = string("linear_163_cast_fp16")]; + tensor input_967_cast_fp16 = silu(x = linear_163_cast_fp16)[name = string("input_967_cast_fp16")]; + tensor encoder_layers_18_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(364929856))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368075648))))[name = string("encoder_layers_18_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_18_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_18_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368075840)))]; + tensor linear_164_cast_fp16 = linear(bias = encoder_layers_18_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_18_feed_forward1_linear2_weight_to_fp16_palettized, x = input_967_cast_fp16)[name = string("linear_164_cast_fp16")]; + fp16 var_4289_to_fp16 = const()[name = string("op_4289_to_fp16"), val = fp16(0x1p-1)]; + tensor var_4290_cast_fp16 = mul(x = linear_164_cast_fp16, y = var_4289_to_fp16)[name = string("op_4290_cast_fp16")]; + tensor input_973_cast_fp16 = add(x = input_961_cast_fp16, y = var_4290_cast_fp16)[name = string("input_973_cast_fp16")]; + tensor key_37_axes_0 = const()[name = string("key_37_axes_0"), val = tensor([-1])]; + tensor encoder_layers_18_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_18_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368077952)))]; + tensor encoder_layers_18_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_18_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368080064)))]; + tensor key_37_cast_fp16 = layer_norm(axes = key_37_axes_0, beta = encoder_layers_18_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_18_norm_self_att_weight_to_fp16, x = input_973_cast_fp16)[name = string("key_37_cast_fp16")]; + bool input_975_interleave_0 = const()[name = string("input_975_interleave_0"), val = bool(false)]; + tensor input_975_cast_fp16 = concat(axis = var_68, interleave = input_975_interleave_0, values = (cache_73_cast_fp16, key_37_cast_fp16))[name = string("input_975_cast_fp16")]; + tensor var_4312_begin_0 = const()[name = string("op_4312_begin_0"), val = tensor([0, 28, 0])]; + tensor var_4312_end_0 = const()[name = string("op_4312_end_0"), val = tensor([1, 42, 1024])]; + tensor var_4312_end_mask_0 = const()[name = string("op_4312_end_mask_0"), val = tensor([true, true, true])]; + tensor var_4312_cast_fp16 = slice_by_index(begin = var_4312_begin_0, end = var_4312_end_0, end_mask = var_4312_end_mask_0, x = cache_73_cast_fp16)[name = string("op_4312_cast_fp16")]; + bool var_4318_interleave_0 = const()[name = string("op_4318_interleave_0"), val = bool(false)]; + tensor var_4318_cast_fp16 = concat(axis = var_68, interleave = var_4318_interleave_0, values = (var_4312_cast_fp16, key_37_cast_fp16))[name = string("op_4318_cast_fp16")]; + tensor encoder_layers_18_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368082176))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368868672))))[name = string("encoder_layers_18_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_18_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_18_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368868864)))]; + tensor linear_165_cast_fp16 = linear(bias = encoder_layers_18_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_18_self_attn_linear_q_weight_to_fp16_palettized, x = key_37_cast_fp16)[name = string("linear_165_cast_fp16")]; + tensor var_4323 = const()[name = string("op_4323"), val = tensor([1, -1, 8, 128])]; + tensor q_109_cast_fp16 = reshape(shape = var_4323, x = linear_165_cast_fp16)[name = string("q_109_cast_fp16")]; + tensor encoder_layers_18_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368870976))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369657472))))[name = string("encoder_layers_18_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_18_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_18_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369657664)))]; + tensor linear_166_cast_fp16 = linear(bias = encoder_layers_18_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_18_self_attn_linear_k_weight_to_fp16_palettized, x = input_975_cast_fp16)[name = string("linear_166_cast_fp16")]; + tensor var_4328 = const()[name = string("op_4328"), val = tensor([1, -1, 8, 128])]; + tensor k_73_cast_fp16 = reshape(shape = var_4328, x = linear_166_cast_fp16)[name = string("k_73_cast_fp16")]; + tensor encoder_layers_18_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369659776))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(370446272))))[name = string("encoder_layers_18_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_18_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_18_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(370446464)))]; + tensor linear_167_cast_fp16 = linear(bias = encoder_layers_18_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_18_self_attn_linear_v_weight_to_fp16_palettized, x = input_975_cast_fp16)[name = string("linear_167_cast_fp16")]; + tensor var_4333 = const()[name = string("op_4333"), val = tensor([1, -1, 8, 128])]; + tensor v_37_cast_fp16 = reshape(shape = var_4333, x = linear_167_cast_fp16)[name = string("v_37_cast_fp16")]; + tensor value_45_perm_0 = const()[name = string("value_45_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_18_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_18_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(370448576)))]; + tensor var_4346_cast_fp16 = add(x = q_109_cast_fp16, y = encoder_layers_18_self_attn_pos_bias_u_to_fp16)[name = string("op_4346_cast_fp16")]; + tensor encoder_layers_18_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_18_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(370450688)))]; + tensor var_4348_cast_fp16 = add(x = q_109_cast_fp16, y = encoder_layers_18_self_attn_pos_bias_v_to_fp16)[name = string("op_4348_cast_fp16")]; + tensor q_with_bias_v_37_perm_0 = const()[name = string("q_with_bias_v_37_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_475_transpose_x_0 = const()[name = string("x_475_transpose_x_0"), val = bool(false)]; + bool x_475_transpose_y_0 = const()[name = string("x_475_transpose_y_0"), val = bool(false)]; + tensor op_4350_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(370452800))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(370595200))))[name = string("op_4350_to_fp16_quantized")]; + tensor q_with_bias_v_37_cast_fp16 = transpose(perm = q_with_bias_v_37_perm_0, x = var_4348_cast_fp16)[name = string("transpose_200")]; + tensor x_475_cast_fp16 = matmul(transpose_x = x_475_transpose_x_0, transpose_y = x_475_transpose_y_0, x = q_with_bias_v_37_cast_fp16, y = op_4350_to_fp16_quantized)[name = string("x_475_cast_fp16")]; + tensor x_477_pad_0 = const()[name = string("x_477_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_477_mode_0 = const()[name = string("x_477_mode_0"), val = string("constant")]; + fp16 const_313_to_fp16 = const()[name = string("const_313_to_fp16"), val = fp16(0x0p+0)]; + tensor x_477_cast_fp16 = pad(constant_val = const_313_to_fp16, mode = x_477_mode_0, pad = x_477_pad_0, x = x_475_cast_fp16)[name = string("x_477_cast_fp16")]; + tensor var_4358 = const()[name = string("op_4358"), val = tensor([1, 8, -1, 28])]; + tensor x_479_cast_fp16 = reshape(shape = var_4358, x = x_477_cast_fp16)[name = string("x_479_cast_fp16")]; + tensor var_4362_begin_0 = const()[name = string("op_4362_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_4362_end_0 = const()[name = string("op_4362_end_0"), val = tensor([1, 8, 140, 28])]; + tensor var_4362_end_mask_0 = const()[name = string("op_4362_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_4362_cast_fp16 = slice_by_index(begin = var_4362_begin_0, end = var_4362_end_0, end_mask = var_4362_end_mask_0, x = x_479_cast_fp16)[name = string("op_4362_cast_fp16")]; + tensor var_4363 = const()[name = string("op_4363"), val = tensor([1, 8, 28, 139])]; + tensor matrix_bd_73_cast_fp16 = reshape(shape = var_4363, x = var_4362_cast_fp16)[name = string("matrix_bd_73_cast_fp16")]; + bool matrix_ac_37_transpose_x_0 = const()[name = string("matrix_ac_37_transpose_x_0"), val = bool(false)]; + bool matrix_ac_37_transpose_y_0 = const()[name = string("matrix_ac_37_transpose_y_0"), val = bool(false)]; + tensor transpose_132_perm_0 = const()[name = string("transpose_132_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_133_perm_0 = const()[name = string("transpose_133_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_133 = transpose(perm = transpose_133_perm_0, x = k_73_cast_fp16)[name = string("transpose_198")]; + tensor transpose_132 = transpose(perm = transpose_132_perm_0, x = var_4346_cast_fp16)[name = string("transpose_199")]; + tensor matrix_ac_37_cast_fp16 = matmul(transpose_x = matrix_ac_37_transpose_x_0, transpose_y = matrix_ac_37_transpose_y_0, x = transpose_132, y = transpose_133)[name = string("matrix_ac_37_cast_fp16")]; + tensor matrix_bd_75_begin_0 = const()[name = string("matrix_bd_75_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_75_end_0 = const()[name = string("matrix_bd_75_end_0"), val = tensor([1, 8, 28, 70])]; + tensor matrix_bd_75_end_mask_0 = const()[name = string("matrix_bd_75_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_75_cast_fp16 = slice_by_index(begin = matrix_bd_75_begin_0, end = matrix_bd_75_end_0, end_mask = matrix_bd_75_end_mask_0, x = matrix_bd_73_cast_fp16)[name = string("matrix_bd_75_cast_fp16")]; + tensor var_4372_cast_fp16 = add(x = matrix_ac_37_cast_fp16, y = matrix_bd_75_cast_fp16)[name = string("op_4372_cast_fp16")]; + fp16 _inversed_scores_73_y_0_to_fp16 = const()[name = string("_inversed_scores_73_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_73_cast_fp16 = mul(x = var_4372_cast_fp16, y = _inversed_scores_73_y_0_to_fp16)[name = string("_inversed_scores_73_cast_fp16")]; + tensor scores_75_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_73_cast_fp16, cond = mask_11)[name = string("scores_75_cast_fp16")]; + tensor var_4378_cast_fp16 = softmax(axis = var_59, x = scores_75_cast_fp16)[name = string("op_4378_cast_fp16")]; + tensor input_977_cast_fp16 = select(a = var_44_to_fp16, b = var_4378_cast_fp16, cond = mask_11)[name = string("input_977_cast_fp16")]; + bool x_481_transpose_x_0 = const()[name = string("x_481_transpose_x_0"), val = bool(false)]; + bool x_481_transpose_y_0 = const()[name = string("x_481_transpose_y_0"), val = bool(false)]; + tensor value_45_cast_fp16 = transpose(perm = value_45_perm_0, x = v_37_cast_fp16)[name = string("transpose_197")]; + tensor x_481_cast_fp16 = matmul(transpose_x = x_481_transpose_x_0, transpose_y = x_481_transpose_y_0, x = input_977_cast_fp16, y = value_45_cast_fp16)[name = string("x_481_cast_fp16")]; + tensor var_4382_perm_0 = const()[name = string("op_4382_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_4383 = const()[name = string("op_4383"), val = tensor([1, -1, 1024])]; + tensor var_4382_cast_fp16 = transpose(perm = var_4382_perm_0, x = x_481_cast_fp16)[name = string("transpose_196")]; + tensor input_979_cast_fp16 = reshape(shape = var_4383, x = var_4382_cast_fp16)[name = string("input_979_cast_fp16")]; + tensor encoder_layers_18_self_attn_linear_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(370595584))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371644224))))[name = string("encoder_layers_18_self_attn_linear_out_weight_to_fp16_quantized")]; + tensor encoder_layers_18_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_18_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371646336)))]; + tensor linear_169_cast_fp16 = linear(bias = encoder_layers_18_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_18_self_attn_linear_out_weight_to_fp16_quantized, x = input_979_cast_fp16)[name = string("linear_169_cast_fp16")]; + tensor input_983_cast_fp16 = add(x = input_973_cast_fp16, y = linear_169_cast_fp16)[name = string("input_983_cast_fp16")]; + tensor x_485_axes_0 = const()[name = string("x_485_axes_0"), val = tensor([-1])]; + tensor encoder_layers_18_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_18_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371648448)))]; + tensor encoder_layers_18_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_18_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371650560)))]; + tensor x_485_cast_fp16 = layer_norm(axes = x_485_axes_0, beta = encoder_layers_18_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_18_norm_conv_weight_to_fp16, x = input_983_cast_fp16)[name = string("x_485_cast_fp16")]; + tensor input_985_perm_0 = const()[name = string("input_985_perm_0"), val = tensor([0, 2, 1])]; + string input_987_pad_type_0 = const()[name = string("input_987_pad_type_0"), val = string("valid")]; + tensor input_987_strides_0 = const()[name = string("input_987_strides_0"), val = tensor([1])]; + tensor input_987_pad_0 = const()[name = string("input_987_pad_0"), val = tensor([0, 0])]; + tensor input_987_dilations_0 = const()[name = string("input_987_dilations_0"), val = tensor([1])]; + int32 input_987_groups_0 = const()[name = string("input_987_groups_0"), val = int32(1)]; + tensor encoder_layers_18_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371652672))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(373749888))))[name = string("encoder_layers_18_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_985_cast_fp16 = transpose(perm = input_985_perm_0, x = x_485_cast_fp16)[name = string("transpose_195")]; + tensor input_987_cast_fp16 = conv(dilations = input_987_dilations_0, groups = input_987_groups_0, pad = input_987_pad_0, pad_type = input_987_pad_type_0, strides = input_987_strides_0, weight = encoder_layers_18_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_985_cast_fp16)[name = string("input_987_cast_fp16")]; + int32 x_487_split_num_splits_0 = const()[name = string("x_487_split_num_splits_0"), val = int32(2)]; + int32 x_487_split_axis_0 = const()[name = string("x_487_split_axis_0"), val = int32(1)]; + tensor x_487_split_cast_fp16_0, tensor x_487_split_cast_fp16_1 = split(axis = x_487_split_axis_0, num_splits = x_487_split_num_splits_0, x = input_987_cast_fp16)[name = string("x_487_split_cast_fp16")]; + tensor x_487_split_1_sigmoid_cast_fp16 = sigmoid(x = x_487_split_cast_fp16_1)[name = string("x_487_split_1_sigmoid_cast_fp16")]; + tensor x_487_cast_fp16 = mul(x = x_487_split_cast_fp16_0, y = x_487_split_1_sigmoid_cast_fp16)[name = string("x_487_cast_fp16")]; + tensor input_989_cast_fp16 = select(a = var_44_to_fp16, b = x_487_cast_fp16, cond = var_575)[name = string("input_989_cast_fp16")]; + bool new_x_75_interleave_0 = const()[name = string("new_x_75_interleave_0"), val = bool(false)]; + tensor new_x_75_cast_fp16 = concat(axis = var_59, interleave = new_x_75_interleave_0, values = (cache_75_cast_fp16, input_989_cast_fp16))[name = string("new_x_75_cast_fp16")]; + tensor var_4422_begin_0 = const()[name = string("op_4422_begin_0"), val = tensor([0, 0, 28])]; + tensor var_4422_end_0 = const()[name = string("op_4422_end_0"), val = tensor([1, 1024, 36])]; + tensor var_4422_end_mask_0 = const()[name = string("op_4422_end_mask_0"), val = tensor([true, true, true])]; + tensor var_4422_cast_fp16 = slice_by_index(begin = var_4422_begin_0, end = var_4422_end_0, end_mask = var_4422_end_mask_0, x = new_x_75_cast_fp16)[name = string("op_4422_cast_fp16")]; + string x_489_pad_type_0 = const()[name = string("x_489_pad_type_0"), val = string("valid")]; + int32 x_489_groups_0 = const()[name = string("x_489_groups_0"), val = int32(1024)]; + tensor x_489_strides_0 = const()[name = string("x_489_strides_0"), val = tensor([1])]; + tensor x_489_pad_0 = const()[name = string("x_489_pad_0"), val = tensor([0, 0])]; + tensor x_489_dilations_0 = const()[name = string("x_489_dilations_0"), val = tensor([1])]; + tensor encoder_layers_18_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(373754048))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(373763328))))[name = string("encoder_layers_18_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_489_cast_fp16 = conv(dilations = x_489_dilations_0, groups = x_489_groups_0, pad = x_489_pad_0, pad_type = x_489_pad_type_0, strides = x_489_strides_0, weight = encoder_layers_18_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_75_cast_fp16)[name = string("x_489_cast_fp16")]; + tensor input_991_perm_0 = const()[name = string("input_991_perm_0"), val = tensor([0, 2, 1])]; + tensor x_491_axes_0 = const()[name = string("x_491_axes_0"), val = tensor([-1])]; + tensor encoder_layers_18_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_18_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(373765440)))]; + tensor encoder_layers_18_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_18_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(373767552)))]; + tensor input_991_cast_fp16 = transpose(perm = input_991_perm_0, x = x_489_cast_fp16)[name = string("transpose_194")]; + tensor x_491_cast_fp16 = layer_norm(axes = x_491_axes_0, beta = encoder_layers_18_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_18_conv_batch_norm_weight_to_fp16, x = input_991_cast_fp16)[name = string("x_491_cast_fp16")]; + tensor input_993_perm_0 = const()[name = string("input_993_perm_0"), val = tensor([0, 2, 1])]; + tensor input_993_cast_fp16 = transpose(perm = input_993_perm_0, x = x_491_cast_fp16)[name = string("transpose_193")]; + tensor input_995_cast_fp16 = silu(x = input_993_cast_fp16)[name = string("input_995_cast_fp16")]; + string x_493_pad_type_0 = const()[name = string("x_493_pad_type_0"), val = string("valid")]; + tensor x_493_strides_0 = const()[name = string("x_493_strides_0"), val = tensor([1])]; + tensor x_493_pad_0 = const()[name = string("x_493_pad_0"), val = tensor([0, 0])]; + tensor x_493_dilations_0 = const()[name = string("x_493_dilations_0"), val = tensor([1])]; + int32 x_493_groups_0 = const()[name = string("x_493_groups_0"), val = int32(1)]; + tensor encoder_layers_18_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(373769664))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(374818304))))[name = string("encoder_layers_18_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_493_cast_fp16 = conv(dilations = x_493_dilations_0, groups = x_493_groups_0, pad = x_493_pad_0, pad_type = x_493_pad_type_0, strides = x_493_strides_0, weight = encoder_layers_18_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_995_cast_fp16)[name = string("x_493_cast_fp16")]; + tensor input_997_perm_0 = const()[name = string("input_997_perm_0"), val = tensor([0, 2, 1])]; + tensor input_997_cast_fp16 = transpose(perm = input_997_perm_0, x = x_493_cast_fp16)[name = string("transpose_192")]; + tensor input_999_cast_fp16 = add(x = input_983_cast_fp16, y = input_997_cast_fp16)[name = string("input_999_cast_fp16")]; + tensor input_1001_axes_0 = const()[name = string("input_1001_axes_0"), val = tensor([-1])]; + tensor encoder_layers_18_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_18_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(374820416)))]; + tensor encoder_layers_18_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_18_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(374822528)))]; + tensor input_1001_cast_fp16 = layer_norm(axes = input_1001_axes_0, beta = encoder_layers_18_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_18_norm_feed_forward2_weight_to_fp16, x = input_999_cast_fp16)[name = string("input_1001_cast_fp16")]; + tensor encoder_layers_18_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(374824640))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(379019008))))[name = string("encoder_layers_18_feed_forward2_linear1_weight_to_fp16_quantized")]; + tensor encoder_layers_18_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_18_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(379027264)))]; + tensor linear_170_cast_fp16 = linear(bias = encoder_layers_18_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_18_feed_forward2_linear1_weight_to_fp16_quantized, x = input_1001_cast_fp16)[name = string("linear_170_cast_fp16")]; + tensor input_1005_cast_fp16 = silu(x = linear_170_cast_fp16)[name = string("input_1005_cast_fp16")]; + tensor encoder_layers_18_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(379035520))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(383229888))))[name = string("encoder_layers_18_feed_forward2_linear2_weight_to_fp16_quantized")]; + tensor encoder_layers_18_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_18_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(383232000)))]; + tensor linear_171_cast_fp16 = linear(bias = encoder_layers_18_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_18_feed_forward2_linear2_weight_to_fp16_quantized, x = input_1005_cast_fp16)[name = string("linear_171_cast_fp16")]; + fp16 var_4465_to_fp16 = const()[name = string("op_4465_to_fp16"), val = fp16(0x1p-1)]; + tensor var_4466_cast_fp16 = mul(x = linear_171_cast_fp16, y = var_4465_to_fp16)[name = string("op_4466_cast_fp16")]; + tensor input_1011_cast_fp16 = add(x = input_999_cast_fp16, y = var_4466_cast_fp16)[name = string("input_1011_cast_fp16")]; + tensor input_1013_axes_0 = const()[name = string("input_1013_axes_0"), val = tensor([-1])]; + tensor encoder_layers_18_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_18_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(383234112)))]; + tensor encoder_layers_18_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_18_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(383236224)))]; + tensor input_1013_cast_fp16 = layer_norm(axes = input_1013_axes_0, beta = encoder_layers_18_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_18_norm_out_weight_to_fp16, x = input_1011_cast_fp16)[name = string("input_1013_cast_fp16")]; + tensor cache_77_begin_0 = const()[name = string("cache_77_begin_0"), val = tensor([19, 0, 0, 0])]; + tensor cache_77_end_0 = const()[name = string("cache_77_end_0"), val = tensor([20, 1, 42, 1024])]; + tensor cache_77_end_mask_0 = const()[name = string("cache_77_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_77_squeeze_mask_0 = const()[name = string("cache_77_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_77_cast_fp16 = slice_by_index(begin = cache_77_begin_0, end = cache_77_end_0, end_mask = cache_77_end_mask_0, squeeze_mask = cache_77_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_77_cast_fp16")]; + tensor cache_79_begin_0 = const()[name = string("cache_79_begin_0"), val = tensor([19, 0, 0, 0])]; + tensor cache_79_end_0 = const()[name = string("cache_79_end_0"), val = tensor([20, 1, 1024, 8])]; + tensor cache_79_end_mask_0 = const()[name = string("cache_79_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_79_squeeze_mask_0 = const()[name = string("cache_79_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_79_cast_fp16 = slice_by_index(begin = cache_79_begin_0, end = cache_79_end_0, end_mask = cache_79_end_mask_0, squeeze_mask = cache_79_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_79_cast_fp16")]; + tensor input_1015_axes_0 = const()[name = string("input_1015_axes_0"), val = tensor([-1])]; + tensor encoder_layers_19_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_19_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(383238336)))]; + tensor encoder_layers_19_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_19_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(383240448)))]; + tensor input_1015_cast_fp16 = layer_norm(axes = input_1015_axes_0, beta = encoder_layers_19_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_19_norm_feed_forward1_weight_to_fp16, x = input_1013_cast_fp16)[name = string("input_1015_cast_fp16")]; + tensor encoder_layers_19_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(383242560))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(387436928))))[name = string("encoder_layers_19_feed_forward1_linear1_weight_to_fp16_quantized")]; + tensor encoder_layers_19_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_19_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(387445184)))]; + tensor linear_172_cast_fp16 = linear(bias = encoder_layers_19_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_19_feed_forward1_linear1_weight_to_fp16_quantized, x = input_1015_cast_fp16)[name = string("linear_172_cast_fp16")]; + tensor input_1019_cast_fp16 = silu(x = linear_172_cast_fp16)[name = string("input_1019_cast_fp16")]; + tensor encoder_layers_19_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(387453440))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391647808))))[name = string("encoder_layers_19_feed_forward1_linear2_weight_to_fp16_quantized")]; + tensor encoder_layers_19_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_19_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391649920)))]; + tensor linear_173_cast_fp16 = linear(bias = encoder_layers_19_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_19_feed_forward1_linear2_weight_to_fp16_quantized, x = input_1019_cast_fp16)[name = string("linear_173_cast_fp16")]; + fp16 var_4502_to_fp16 = const()[name = string("op_4502_to_fp16"), val = fp16(0x1p-1)]; + tensor var_4503_cast_fp16 = mul(x = linear_173_cast_fp16, y = var_4502_to_fp16)[name = string("op_4503_cast_fp16")]; + tensor input_1025_cast_fp16 = add(x = input_1013_cast_fp16, y = var_4503_cast_fp16)[name = string("input_1025_cast_fp16")]; + tensor key_39_axes_0 = const()[name = string("key_39_axes_0"), val = tensor([-1])]; + tensor encoder_layers_19_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_19_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391652032)))]; + tensor encoder_layers_19_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_19_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391654144)))]; + tensor key_39_cast_fp16 = layer_norm(axes = key_39_axes_0, beta = encoder_layers_19_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_19_norm_self_att_weight_to_fp16, x = input_1025_cast_fp16)[name = string("key_39_cast_fp16")]; + bool input_1027_interleave_0 = const()[name = string("input_1027_interleave_0"), val = bool(false)]; + tensor input_1027_cast_fp16 = concat(axis = var_68, interleave = input_1027_interleave_0, values = (cache_77_cast_fp16, key_39_cast_fp16))[name = string("input_1027_cast_fp16")]; + tensor var_4525_begin_0 = const()[name = string("op_4525_begin_0"), val = tensor([0, 28, 0])]; + tensor var_4525_end_0 = const()[name = string("op_4525_end_0"), val = tensor([1, 42, 1024])]; + tensor var_4525_end_mask_0 = const()[name = string("op_4525_end_mask_0"), val = tensor([true, true, true])]; + tensor var_4525_cast_fp16 = slice_by_index(begin = var_4525_begin_0, end = var_4525_end_0, end_mask = var_4525_end_mask_0, x = cache_77_cast_fp16)[name = string("op_4525_cast_fp16")]; + bool var_4531_interleave_0 = const()[name = string("op_4531_interleave_0"), val = bool(false)]; + tensor var_4531_cast_fp16 = concat(axis = var_68, interleave = var_4531_interleave_0, values = (var_4525_cast_fp16, key_39_cast_fp16))[name = string("op_4531_cast_fp16")]; + tensor encoder_layers_19_self_attn_linear_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391656256))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(392704896))))[name = string("encoder_layers_19_self_attn_linear_q_weight_to_fp16_quantized")]; + tensor encoder_layers_19_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_19_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(392707008)))]; + tensor linear_174_cast_fp16 = linear(bias = encoder_layers_19_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_19_self_attn_linear_q_weight_to_fp16_quantized, x = key_39_cast_fp16)[name = string("linear_174_cast_fp16")]; + tensor var_4536 = const()[name = string("op_4536"), val = tensor([1, -1, 8, 128])]; + tensor q_115_cast_fp16 = reshape(shape = var_4536, x = linear_174_cast_fp16)[name = string("q_115_cast_fp16")]; + tensor encoder_layers_19_self_attn_linear_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(392709120))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(393757760))))[name = string("encoder_layers_19_self_attn_linear_k_weight_to_fp16_quantized")]; + tensor encoder_layers_19_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_19_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(393759872)))]; + tensor linear_175_cast_fp16 = linear(bias = encoder_layers_19_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_19_self_attn_linear_k_weight_to_fp16_quantized, x = input_1027_cast_fp16)[name = string("linear_175_cast_fp16")]; + tensor var_4541 = const()[name = string("op_4541"), val = tensor([1, -1, 8, 128])]; + tensor k_77_cast_fp16 = reshape(shape = var_4541, x = linear_175_cast_fp16)[name = string("k_77_cast_fp16")]; + tensor encoder_layers_19_self_attn_linear_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(393761984))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(394810624))))[name = string("encoder_layers_19_self_attn_linear_v_weight_to_fp16_quantized")]; + tensor encoder_layers_19_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_19_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(394812736)))]; + tensor linear_176_cast_fp16 = linear(bias = encoder_layers_19_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_19_self_attn_linear_v_weight_to_fp16_quantized, x = input_1027_cast_fp16)[name = string("linear_176_cast_fp16")]; + tensor var_4546 = const()[name = string("op_4546"), val = tensor([1, -1, 8, 128])]; + tensor v_39_cast_fp16 = reshape(shape = var_4546, x = linear_176_cast_fp16)[name = string("v_39_cast_fp16")]; + tensor value_47_perm_0 = const()[name = string("value_47_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_19_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_19_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(394814848)))]; + tensor var_4559_cast_fp16 = add(x = q_115_cast_fp16, y = encoder_layers_19_self_attn_pos_bias_u_to_fp16)[name = string("op_4559_cast_fp16")]; + tensor encoder_layers_19_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_19_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(394816960)))]; + tensor var_4561_cast_fp16 = add(x = q_115_cast_fp16, y = encoder_layers_19_self_attn_pos_bias_v_to_fp16)[name = string("op_4561_cast_fp16")]; + tensor q_with_bias_v_39_perm_0 = const()[name = string("q_with_bias_v_39_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_501_transpose_x_0 = const()[name = string("x_501_transpose_x_0"), val = bool(false)]; + bool x_501_transpose_y_0 = const()[name = string("x_501_transpose_y_0"), val = bool(false)]; + tensor op_4563_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(394819072))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(394961472))))[name = string("op_4563_to_fp16_quantized")]; + tensor q_with_bias_v_39_cast_fp16 = transpose(perm = q_with_bias_v_39_perm_0, x = var_4561_cast_fp16)[name = string("transpose_191")]; + tensor x_501_cast_fp16 = matmul(transpose_x = x_501_transpose_x_0, transpose_y = x_501_transpose_y_0, x = q_with_bias_v_39_cast_fp16, y = op_4563_to_fp16_quantized)[name = string("x_501_cast_fp16")]; + tensor x_503_pad_0 = const()[name = string("x_503_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_503_mode_0 = const()[name = string("x_503_mode_0"), val = string("constant")]; + fp16 const_326_to_fp16 = const()[name = string("const_326_to_fp16"), val = fp16(0x0p+0)]; + tensor x_503_cast_fp16 = pad(constant_val = const_326_to_fp16, mode = x_503_mode_0, pad = x_503_pad_0, x = x_501_cast_fp16)[name = string("x_503_cast_fp16")]; + tensor var_4571 = const()[name = string("op_4571"), val = tensor([1, 8, -1, 28])]; + tensor x_505_cast_fp16 = reshape(shape = var_4571, x = x_503_cast_fp16)[name = string("x_505_cast_fp16")]; + tensor var_4575_begin_0 = const()[name = string("op_4575_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_4575_end_0 = const()[name = string("op_4575_end_0"), val = tensor([1, 8, 140, 28])]; + tensor var_4575_end_mask_0 = const()[name = string("op_4575_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_4575_cast_fp16 = slice_by_index(begin = var_4575_begin_0, end = var_4575_end_0, end_mask = var_4575_end_mask_0, x = x_505_cast_fp16)[name = string("op_4575_cast_fp16")]; + tensor var_4576 = const()[name = string("op_4576"), val = tensor([1, 8, 28, 139])]; + tensor matrix_bd_77_cast_fp16 = reshape(shape = var_4576, x = var_4575_cast_fp16)[name = string("matrix_bd_77_cast_fp16")]; + bool matrix_ac_39_transpose_x_0 = const()[name = string("matrix_ac_39_transpose_x_0"), val = bool(false)]; + bool matrix_ac_39_transpose_y_0 = const()[name = string("matrix_ac_39_transpose_y_0"), val = bool(false)]; + tensor transpose_134_perm_0 = const()[name = string("transpose_134_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_135_perm_0 = const()[name = string("transpose_135_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_135 = transpose(perm = transpose_135_perm_0, x = k_77_cast_fp16)[name = string("transpose_189")]; + tensor transpose_134 = transpose(perm = transpose_134_perm_0, x = var_4559_cast_fp16)[name = string("transpose_190")]; + tensor matrix_ac_39_cast_fp16 = matmul(transpose_x = matrix_ac_39_transpose_x_0, transpose_y = matrix_ac_39_transpose_y_0, x = transpose_134, y = transpose_135)[name = string("matrix_ac_39_cast_fp16")]; + tensor matrix_bd_79_begin_0 = const()[name = string("matrix_bd_79_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_79_end_0 = const()[name = string("matrix_bd_79_end_0"), val = tensor([1, 8, 28, 70])]; + tensor matrix_bd_79_end_mask_0 = const()[name = string("matrix_bd_79_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_79_cast_fp16 = slice_by_index(begin = matrix_bd_79_begin_0, end = matrix_bd_79_end_0, end_mask = matrix_bd_79_end_mask_0, x = matrix_bd_77_cast_fp16)[name = string("matrix_bd_79_cast_fp16")]; + tensor var_4585_cast_fp16 = add(x = matrix_ac_39_cast_fp16, y = matrix_bd_79_cast_fp16)[name = string("op_4585_cast_fp16")]; + fp16 _inversed_scores_77_y_0_to_fp16 = const()[name = string("_inversed_scores_77_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_77_cast_fp16 = mul(x = var_4585_cast_fp16, y = _inversed_scores_77_y_0_to_fp16)[name = string("_inversed_scores_77_cast_fp16")]; + tensor scores_79_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_77_cast_fp16, cond = mask_11)[name = string("scores_79_cast_fp16")]; + tensor var_4591_cast_fp16 = softmax(axis = var_59, x = scores_79_cast_fp16)[name = string("op_4591_cast_fp16")]; + tensor input_1029_cast_fp16 = select(a = var_44_to_fp16, b = var_4591_cast_fp16, cond = mask_11)[name = string("input_1029_cast_fp16")]; + bool x_507_transpose_x_0 = const()[name = string("x_507_transpose_x_0"), val = bool(false)]; + bool x_507_transpose_y_0 = const()[name = string("x_507_transpose_y_0"), val = bool(false)]; + tensor value_47_cast_fp16 = transpose(perm = value_47_perm_0, x = v_39_cast_fp16)[name = string("transpose_188")]; + tensor x_507_cast_fp16 = matmul(transpose_x = x_507_transpose_x_0, transpose_y = x_507_transpose_y_0, x = input_1029_cast_fp16, y = value_47_cast_fp16)[name = string("x_507_cast_fp16")]; + tensor var_4595_perm_0 = const()[name = string("op_4595_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_4596 = const()[name = string("op_4596"), val = tensor([1, -1, 1024])]; + tensor var_4595_cast_fp16 = transpose(perm = var_4595_perm_0, x = x_507_cast_fp16)[name = string("transpose_187")]; + tensor input_1031_cast_fp16 = reshape(shape = var_4596, x = var_4595_cast_fp16)[name = string("input_1031_cast_fp16")]; + tensor encoder_layers_19_self_attn_linear_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(394961856))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(396010496))))[name = string("encoder_layers_19_self_attn_linear_out_weight_to_fp16_quantized")]; + tensor encoder_layers_19_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_19_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(396012608)))]; + tensor linear_178_cast_fp16 = linear(bias = encoder_layers_19_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_19_self_attn_linear_out_weight_to_fp16_quantized, x = input_1031_cast_fp16)[name = string("linear_178_cast_fp16")]; + tensor input_1035_cast_fp16 = add(x = input_1025_cast_fp16, y = linear_178_cast_fp16)[name = string("input_1035_cast_fp16")]; + tensor x_511_axes_0 = const()[name = string("x_511_axes_0"), val = tensor([-1])]; + tensor encoder_layers_19_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_19_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(396014720)))]; + tensor encoder_layers_19_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_19_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(396016832)))]; + tensor x_511_cast_fp16 = layer_norm(axes = x_511_axes_0, beta = encoder_layers_19_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_19_norm_conv_weight_to_fp16, x = input_1035_cast_fp16)[name = string("x_511_cast_fp16")]; + tensor input_1037_perm_0 = const()[name = string("input_1037_perm_0"), val = tensor([0, 2, 1])]; + string input_1039_pad_type_0 = const()[name = string("input_1039_pad_type_0"), val = string("valid")]; + tensor input_1039_strides_0 = const()[name = string("input_1039_strides_0"), val = tensor([1])]; + tensor input_1039_pad_0 = const()[name = string("input_1039_pad_0"), val = tensor([0, 0])]; + tensor input_1039_dilations_0 = const()[name = string("input_1039_dilations_0"), val = tensor([1])]; + int32 input_1039_groups_0 = const()[name = string("input_1039_groups_0"), val = int32(1)]; + tensor encoder_layers_19_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(396018944))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(398116160))))[name = string("encoder_layers_19_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_1037_cast_fp16 = transpose(perm = input_1037_perm_0, x = x_511_cast_fp16)[name = string("transpose_186")]; + tensor input_1039_cast_fp16 = conv(dilations = input_1039_dilations_0, groups = input_1039_groups_0, pad = input_1039_pad_0, pad_type = input_1039_pad_type_0, strides = input_1039_strides_0, weight = encoder_layers_19_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_1037_cast_fp16)[name = string("input_1039_cast_fp16")]; + int32 x_513_split_num_splits_0 = const()[name = string("x_513_split_num_splits_0"), val = int32(2)]; + int32 x_513_split_axis_0 = const()[name = string("x_513_split_axis_0"), val = int32(1)]; + tensor x_513_split_cast_fp16_0, tensor x_513_split_cast_fp16_1 = split(axis = x_513_split_axis_0, num_splits = x_513_split_num_splits_0, x = input_1039_cast_fp16)[name = string("x_513_split_cast_fp16")]; + tensor x_513_split_1_sigmoid_cast_fp16 = sigmoid(x = x_513_split_cast_fp16_1)[name = string("x_513_split_1_sigmoid_cast_fp16")]; + tensor x_513_cast_fp16 = mul(x = x_513_split_cast_fp16_0, y = x_513_split_1_sigmoid_cast_fp16)[name = string("x_513_cast_fp16")]; + tensor input_1041_cast_fp16 = select(a = var_44_to_fp16, b = x_513_cast_fp16, cond = var_575)[name = string("input_1041_cast_fp16")]; + bool new_x_79_interleave_0 = const()[name = string("new_x_79_interleave_0"), val = bool(false)]; + tensor new_x_79_cast_fp16 = concat(axis = var_59, interleave = new_x_79_interleave_0, values = (cache_79_cast_fp16, input_1041_cast_fp16))[name = string("new_x_79_cast_fp16")]; + tensor var_4635_begin_0 = const()[name = string("op_4635_begin_0"), val = tensor([0, 0, 28])]; + tensor var_4635_end_0 = const()[name = string("op_4635_end_0"), val = tensor([1, 1024, 36])]; + tensor var_4635_end_mask_0 = const()[name = string("op_4635_end_mask_0"), val = tensor([true, true, true])]; + tensor var_4635_cast_fp16 = slice_by_index(begin = var_4635_begin_0, end = var_4635_end_0, end_mask = var_4635_end_mask_0, x = new_x_79_cast_fp16)[name = string("op_4635_cast_fp16")]; + string x_515_pad_type_0 = const()[name = string("x_515_pad_type_0"), val = string("valid")]; + int32 x_515_groups_0 = const()[name = string("x_515_groups_0"), val = int32(1024)]; + tensor x_515_strides_0 = const()[name = string("x_515_strides_0"), val = tensor([1])]; + tensor x_515_pad_0 = const()[name = string("x_515_pad_0"), val = tensor([0, 0])]; + tensor x_515_dilations_0 = const()[name = string("x_515_dilations_0"), val = tensor([1])]; + tensor encoder_layers_19_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(398120320))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(398129600))))[name = string("encoder_layers_19_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_515_cast_fp16 = conv(dilations = x_515_dilations_0, groups = x_515_groups_0, pad = x_515_pad_0, pad_type = x_515_pad_type_0, strides = x_515_strides_0, weight = encoder_layers_19_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_79_cast_fp16)[name = string("x_515_cast_fp16")]; + tensor input_1043_perm_0 = const()[name = string("input_1043_perm_0"), val = tensor([0, 2, 1])]; + tensor x_517_axes_0 = const()[name = string("x_517_axes_0"), val = tensor([-1])]; + tensor encoder_layers_19_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_19_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(398131712)))]; + tensor encoder_layers_19_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_19_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(398133824)))]; + tensor input_1043_cast_fp16 = transpose(perm = input_1043_perm_0, x = x_515_cast_fp16)[name = string("transpose_185")]; + tensor x_517_cast_fp16 = layer_norm(axes = x_517_axes_0, beta = encoder_layers_19_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_19_conv_batch_norm_weight_to_fp16, x = input_1043_cast_fp16)[name = string("x_517_cast_fp16")]; + tensor input_1045_perm_0 = const()[name = string("input_1045_perm_0"), val = tensor([0, 2, 1])]; + tensor input_1045_cast_fp16 = transpose(perm = input_1045_perm_0, x = x_517_cast_fp16)[name = string("transpose_184")]; + tensor input_1047_cast_fp16 = silu(x = input_1045_cast_fp16)[name = string("input_1047_cast_fp16")]; + string x_519_pad_type_0 = const()[name = string("x_519_pad_type_0"), val = string("valid")]; + tensor x_519_strides_0 = const()[name = string("x_519_strides_0"), val = tensor([1])]; + tensor x_519_pad_0 = const()[name = string("x_519_pad_0"), val = tensor([0, 0])]; + tensor x_519_dilations_0 = const()[name = string("x_519_dilations_0"), val = tensor([1])]; + int32 x_519_groups_0 = const()[name = string("x_519_groups_0"), val = int32(1)]; + tensor encoder_layers_19_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(398135936))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(399184576))))[name = string("encoder_layers_19_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_519_cast_fp16 = conv(dilations = x_519_dilations_0, groups = x_519_groups_0, pad = x_519_pad_0, pad_type = x_519_pad_type_0, strides = x_519_strides_0, weight = encoder_layers_19_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_1047_cast_fp16)[name = string("x_519_cast_fp16")]; + tensor input_1049_perm_0 = const()[name = string("input_1049_perm_0"), val = tensor([0, 2, 1])]; + tensor input_1049_cast_fp16 = transpose(perm = input_1049_perm_0, x = x_519_cast_fp16)[name = string("transpose_183")]; + tensor input_1051_cast_fp16 = add(x = input_1035_cast_fp16, y = input_1049_cast_fp16)[name = string("input_1051_cast_fp16")]; + tensor input_1053_axes_0 = const()[name = string("input_1053_axes_0"), val = tensor([-1])]; + tensor encoder_layers_19_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_19_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(399186688)))]; + tensor encoder_layers_19_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_19_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(399188800)))]; + tensor input_1053_cast_fp16 = layer_norm(axes = input_1053_axes_0, beta = encoder_layers_19_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_19_norm_feed_forward2_weight_to_fp16, x = input_1051_cast_fp16)[name = string("input_1053_cast_fp16")]; + tensor encoder_layers_19_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(399190912))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(403385280))))[name = string("encoder_layers_19_feed_forward2_linear1_weight_to_fp16_quantized")]; + tensor encoder_layers_19_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_19_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(403393536)))]; + tensor linear_179_cast_fp16 = linear(bias = encoder_layers_19_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_19_feed_forward2_linear1_weight_to_fp16_quantized, x = input_1053_cast_fp16)[name = string("linear_179_cast_fp16")]; + tensor input_1057_cast_fp16 = silu(x = linear_179_cast_fp16)[name = string("input_1057_cast_fp16")]; + tensor encoder_layers_19_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(403401792))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407596160))))[name = string("encoder_layers_19_feed_forward2_linear2_weight_to_fp16_quantized")]; + tensor encoder_layers_19_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_19_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407598272)))]; + tensor linear_180_cast_fp16 = linear(bias = encoder_layers_19_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_19_feed_forward2_linear2_weight_to_fp16_quantized, x = input_1057_cast_fp16)[name = string("linear_180_cast_fp16")]; + fp16 var_4678_to_fp16 = const()[name = string("op_4678_to_fp16"), val = fp16(0x1p-1)]; + tensor var_4679_cast_fp16 = mul(x = linear_180_cast_fp16, y = var_4678_to_fp16)[name = string("op_4679_cast_fp16")]; + tensor input_1063_cast_fp16 = add(x = input_1051_cast_fp16, y = var_4679_cast_fp16)[name = string("input_1063_cast_fp16")]; + tensor input_1065_axes_0 = const()[name = string("input_1065_axes_0"), val = tensor([-1])]; + tensor encoder_layers_19_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_19_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407600384)))]; + tensor encoder_layers_19_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_19_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407602496)))]; + tensor input_1065_cast_fp16 = layer_norm(axes = input_1065_axes_0, beta = encoder_layers_19_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_19_norm_out_weight_to_fp16, x = input_1063_cast_fp16)[name = string("input_1065_cast_fp16")]; + tensor cache_81_begin_0 = const()[name = string("cache_81_begin_0"), val = tensor([20, 0, 0, 0])]; + tensor cache_81_end_0 = const()[name = string("cache_81_end_0"), val = tensor([21, 1, 42, 1024])]; + tensor cache_81_end_mask_0 = const()[name = string("cache_81_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_81_squeeze_mask_0 = const()[name = string("cache_81_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_81_cast_fp16 = slice_by_index(begin = cache_81_begin_0, end = cache_81_end_0, end_mask = cache_81_end_mask_0, squeeze_mask = cache_81_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_81_cast_fp16")]; + tensor cache_83_begin_0 = const()[name = string("cache_83_begin_0"), val = tensor([20, 0, 0, 0])]; + tensor cache_83_end_0 = const()[name = string("cache_83_end_0"), val = tensor([21, 1, 1024, 8])]; + tensor cache_83_end_mask_0 = const()[name = string("cache_83_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_83_squeeze_mask_0 = const()[name = string("cache_83_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_83_cast_fp16 = slice_by_index(begin = cache_83_begin_0, end = cache_83_end_0, end_mask = cache_83_end_mask_0, squeeze_mask = cache_83_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_83_cast_fp16")]; + tensor input_1067_axes_0 = const()[name = string("input_1067_axes_0"), val = tensor([-1])]; + tensor encoder_layers_20_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_20_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407604608)))]; + tensor encoder_layers_20_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_20_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407606720)))]; + tensor input_1067_cast_fp16 = layer_norm(axes = input_1067_axes_0, beta = encoder_layers_20_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_20_norm_feed_forward1_weight_to_fp16, x = input_1065_cast_fp16)[name = string("input_1067_cast_fp16")]; + tensor encoder_layers_20_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407608832))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(411803200))))[name = string("encoder_layers_20_feed_forward1_linear1_weight_to_fp16_quantized")]; + tensor encoder_layers_20_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_20_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(411811456)))]; + tensor linear_181_cast_fp16 = linear(bias = encoder_layers_20_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_20_feed_forward1_linear1_weight_to_fp16_quantized, x = input_1067_cast_fp16)[name = string("linear_181_cast_fp16")]; + tensor input_1071_cast_fp16 = silu(x = linear_181_cast_fp16)[name = string("input_1071_cast_fp16")]; + tensor encoder_layers_20_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(411819712))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(416014080))))[name = string("encoder_layers_20_feed_forward1_linear2_weight_to_fp16_quantized")]; + tensor encoder_layers_20_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_20_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(416016192)))]; + tensor linear_182_cast_fp16 = linear(bias = encoder_layers_20_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_20_feed_forward1_linear2_weight_to_fp16_quantized, x = input_1071_cast_fp16)[name = string("linear_182_cast_fp16")]; + fp16 var_4715_to_fp16 = const()[name = string("op_4715_to_fp16"), val = fp16(0x1p-1)]; + tensor var_4716_cast_fp16 = mul(x = linear_182_cast_fp16, y = var_4715_to_fp16)[name = string("op_4716_cast_fp16")]; + tensor input_1077_cast_fp16 = add(x = input_1065_cast_fp16, y = var_4716_cast_fp16)[name = string("input_1077_cast_fp16")]; + tensor key_41_axes_0 = const()[name = string("key_41_axes_0"), val = tensor([-1])]; + tensor encoder_layers_20_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_20_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(416018304)))]; + tensor encoder_layers_20_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_20_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(416020416)))]; + tensor key_41_cast_fp16 = layer_norm(axes = key_41_axes_0, beta = encoder_layers_20_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_20_norm_self_att_weight_to_fp16, x = input_1077_cast_fp16)[name = string("key_41_cast_fp16")]; + bool input_1079_interleave_0 = const()[name = string("input_1079_interleave_0"), val = bool(false)]; + tensor input_1079_cast_fp16 = concat(axis = var_68, interleave = input_1079_interleave_0, values = (cache_81_cast_fp16, key_41_cast_fp16))[name = string("input_1079_cast_fp16")]; + tensor var_4738_begin_0 = const()[name = string("op_4738_begin_0"), val = tensor([0, 28, 0])]; + tensor var_4738_end_0 = const()[name = string("op_4738_end_0"), val = tensor([1, 42, 1024])]; + tensor var_4738_end_mask_0 = const()[name = string("op_4738_end_mask_0"), val = tensor([true, true, true])]; + tensor var_4738_cast_fp16 = slice_by_index(begin = var_4738_begin_0, end = var_4738_end_0, end_mask = var_4738_end_mask_0, x = cache_81_cast_fp16)[name = string("op_4738_cast_fp16")]; + bool var_4744_interleave_0 = const()[name = string("op_4744_interleave_0"), val = bool(false)]; + tensor var_4744_cast_fp16 = concat(axis = var_68, interleave = var_4744_interleave_0, values = (var_4738_cast_fp16, key_41_cast_fp16))[name = string("op_4744_cast_fp16")]; + tensor encoder_layers_20_self_attn_linear_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(416022528))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(417071168))))[name = string("encoder_layers_20_self_attn_linear_q_weight_to_fp16_quantized")]; + tensor encoder_layers_20_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_20_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(417073280)))]; + tensor linear_183_cast_fp16 = linear(bias = encoder_layers_20_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_20_self_attn_linear_q_weight_to_fp16_quantized, x = key_41_cast_fp16)[name = string("linear_183_cast_fp16")]; + tensor var_4749 = const()[name = string("op_4749"), val = tensor([1, -1, 8, 128])]; + tensor q_121_cast_fp16 = reshape(shape = var_4749, x = linear_183_cast_fp16)[name = string("q_121_cast_fp16")]; + tensor encoder_layers_20_self_attn_linear_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(417075392))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(418124032))))[name = string("encoder_layers_20_self_attn_linear_k_weight_to_fp16_quantized")]; + tensor encoder_layers_20_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_20_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(418126144)))]; + tensor linear_184_cast_fp16 = linear(bias = encoder_layers_20_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_20_self_attn_linear_k_weight_to_fp16_quantized, x = input_1079_cast_fp16)[name = string("linear_184_cast_fp16")]; + tensor var_4754 = const()[name = string("op_4754"), val = tensor([1, -1, 8, 128])]; + tensor k_81_cast_fp16 = reshape(shape = var_4754, x = linear_184_cast_fp16)[name = string("k_81_cast_fp16")]; + tensor encoder_layers_20_self_attn_linear_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(418128256))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(419176896))))[name = string("encoder_layers_20_self_attn_linear_v_weight_to_fp16_quantized")]; + tensor encoder_layers_20_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_20_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(419179008)))]; + tensor linear_185_cast_fp16 = linear(bias = encoder_layers_20_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_20_self_attn_linear_v_weight_to_fp16_quantized, x = input_1079_cast_fp16)[name = string("linear_185_cast_fp16")]; + tensor var_4759 = const()[name = string("op_4759"), val = tensor([1, -1, 8, 128])]; + tensor v_41_cast_fp16 = reshape(shape = var_4759, x = linear_185_cast_fp16)[name = string("v_41_cast_fp16")]; + tensor value_49_perm_0 = const()[name = string("value_49_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_20_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_20_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(419181120)))]; + tensor var_4772_cast_fp16 = add(x = q_121_cast_fp16, y = encoder_layers_20_self_attn_pos_bias_u_to_fp16)[name = string("op_4772_cast_fp16")]; + tensor encoder_layers_20_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_20_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(419183232)))]; + tensor var_4774_cast_fp16 = add(x = q_121_cast_fp16, y = encoder_layers_20_self_attn_pos_bias_v_to_fp16)[name = string("op_4774_cast_fp16")]; + tensor q_with_bias_v_41_perm_0 = const()[name = string("q_with_bias_v_41_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_527_transpose_x_0 = const()[name = string("x_527_transpose_x_0"), val = bool(false)]; + bool x_527_transpose_y_0 = const()[name = string("x_527_transpose_y_0"), val = bool(false)]; + tensor op_4776_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(419185344))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(419327744))))[name = string("op_4776_to_fp16_quantized")]; + tensor q_with_bias_v_41_cast_fp16 = transpose(perm = q_with_bias_v_41_perm_0, x = var_4774_cast_fp16)[name = string("transpose_182")]; + tensor x_527_cast_fp16 = matmul(transpose_x = x_527_transpose_x_0, transpose_y = x_527_transpose_y_0, x = q_with_bias_v_41_cast_fp16, y = op_4776_to_fp16_quantized)[name = string("x_527_cast_fp16")]; + tensor x_529_pad_0 = const()[name = string("x_529_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_529_mode_0 = const()[name = string("x_529_mode_0"), val = string("constant")]; + fp16 const_339_to_fp16 = const()[name = string("const_339_to_fp16"), val = fp16(0x0p+0)]; + tensor x_529_cast_fp16 = pad(constant_val = const_339_to_fp16, mode = x_529_mode_0, pad = x_529_pad_0, x = x_527_cast_fp16)[name = string("x_529_cast_fp16")]; + tensor var_4784 = const()[name = string("op_4784"), val = tensor([1, 8, -1, 28])]; + tensor x_531_cast_fp16 = reshape(shape = var_4784, x = x_529_cast_fp16)[name = string("x_531_cast_fp16")]; + tensor var_4788_begin_0 = const()[name = string("op_4788_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_4788_end_0 = const()[name = string("op_4788_end_0"), val = tensor([1, 8, 140, 28])]; + tensor var_4788_end_mask_0 = const()[name = string("op_4788_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_4788_cast_fp16 = slice_by_index(begin = var_4788_begin_0, end = var_4788_end_0, end_mask = var_4788_end_mask_0, x = x_531_cast_fp16)[name = string("op_4788_cast_fp16")]; + tensor var_4789 = const()[name = string("op_4789"), val = tensor([1, 8, 28, 139])]; + tensor matrix_bd_81_cast_fp16 = reshape(shape = var_4789, x = var_4788_cast_fp16)[name = string("matrix_bd_81_cast_fp16")]; + bool matrix_ac_41_transpose_x_0 = const()[name = string("matrix_ac_41_transpose_x_0"), val = bool(false)]; + bool matrix_ac_41_transpose_y_0 = const()[name = string("matrix_ac_41_transpose_y_0"), val = bool(false)]; + tensor transpose_136_perm_0 = const()[name = string("transpose_136_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_137_perm_0 = const()[name = string("transpose_137_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_137 = transpose(perm = transpose_137_perm_0, x = k_81_cast_fp16)[name = string("transpose_180")]; + tensor transpose_136 = transpose(perm = transpose_136_perm_0, x = var_4772_cast_fp16)[name = string("transpose_181")]; + tensor matrix_ac_41_cast_fp16 = matmul(transpose_x = matrix_ac_41_transpose_x_0, transpose_y = matrix_ac_41_transpose_y_0, x = transpose_136, y = transpose_137)[name = string("matrix_ac_41_cast_fp16")]; + tensor matrix_bd_83_begin_0 = const()[name = string("matrix_bd_83_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_83_end_0 = const()[name = string("matrix_bd_83_end_0"), val = tensor([1, 8, 28, 70])]; + tensor matrix_bd_83_end_mask_0 = const()[name = string("matrix_bd_83_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_83_cast_fp16 = slice_by_index(begin = matrix_bd_83_begin_0, end = matrix_bd_83_end_0, end_mask = matrix_bd_83_end_mask_0, x = matrix_bd_81_cast_fp16)[name = string("matrix_bd_83_cast_fp16")]; + tensor var_4798_cast_fp16 = add(x = matrix_ac_41_cast_fp16, y = matrix_bd_83_cast_fp16)[name = string("op_4798_cast_fp16")]; + fp16 _inversed_scores_81_y_0_to_fp16 = const()[name = string("_inversed_scores_81_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_81_cast_fp16 = mul(x = var_4798_cast_fp16, y = _inversed_scores_81_y_0_to_fp16)[name = string("_inversed_scores_81_cast_fp16")]; + tensor scores_83_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_81_cast_fp16, cond = mask_11)[name = string("scores_83_cast_fp16")]; + tensor var_4804_cast_fp16 = softmax(axis = var_59, x = scores_83_cast_fp16)[name = string("op_4804_cast_fp16")]; + tensor input_1081_cast_fp16 = select(a = var_44_to_fp16, b = var_4804_cast_fp16, cond = mask_11)[name = string("input_1081_cast_fp16")]; + bool x_533_transpose_x_0 = const()[name = string("x_533_transpose_x_0"), val = bool(false)]; + bool x_533_transpose_y_0 = const()[name = string("x_533_transpose_y_0"), val = bool(false)]; + tensor value_49_cast_fp16 = transpose(perm = value_49_perm_0, x = v_41_cast_fp16)[name = string("transpose_179")]; + tensor x_533_cast_fp16 = matmul(transpose_x = x_533_transpose_x_0, transpose_y = x_533_transpose_y_0, x = input_1081_cast_fp16, y = value_49_cast_fp16)[name = string("x_533_cast_fp16")]; + tensor var_4808_perm_0 = const()[name = string("op_4808_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_4809 = const()[name = string("op_4809"), val = tensor([1, -1, 1024])]; + tensor var_4808_cast_fp16 = transpose(perm = var_4808_perm_0, x = x_533_cast_fp16)[name = string("transpose_178")]; + tensor input_1083_cast_fp16 = reshape(shape = var_4809, x = var_4808_cast_fp16)[name = string("input_1083_cast_fp16")]; + tensor encoder_layers_20_self_attn_linear_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(419328128))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(420376768))))[name = string("encoder_layers_20_self_attn_linear_out_weight_to_fp16_quantized")]; + tensor encoder_layers_20_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_20_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(420378880)))]; + tensor linear_187_cast_fp16 = linear(bias = encoder_layers_20_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_20_self_attn_linear_out_weight_to_fp16_quantized, x = input_1083_cast_fp16)[name = string("linear_187_cast_fp16")]; + tensor input_1087_cast_fp16 = add(x = input_1077_cast_fp16, y = linear_187_cast_fp16)[name = string("input_1087_cast_fp16")]; + tensor x_537_axes_0 = const()[name = string("x_537_axes_0"), val = tensor([-1])]; + tensor encoder_layers_20_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_20_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(420380992)))]; + tensor encoder_layers_20_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_20_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(420383104)))]; + tensor x_537_cast_fp16 = layer_norm(axes = x_537_axes_0, beta = encoder_layers_20_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_20_norm_conv_weight_to_fp16, x = input_1087_cast_fp16)[name = string("x_537_cast_fp16")]; + tensor input_1089_perm_0 = const()[name = string("input_1089_perm_0"), val = tensor([0, 2, 1])]; + string input_1091_pad_type_0 = const()[name = string("input_1091_pad_type_0"), val = string("valid")]; + tensor input_1091_strides_0 = const()[name = string("input_1091_strides_0"), val = tensor([1])]; + tensor input_1091_pad_0 = const()[name = string("input_1091_pad_0"), val = tensor([0, 0])]; + tensor input_1091_dilations_0 = const()[name = string("input_1091_dilations_0"), val = tensor([1])]; + int32 input_1091_groups_0 = const()[name = string("input_1091_groups_0"), val = int32(1)]; + tensor encoder_layers_20_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(420385216))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(422482432))))[name = string("encoder_layers_20_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_1089_cast_fp16 = transpose(perm = input_1089_perm_0, x = x_537_cast_fp16)[name = string("transpose_177")]; + tensor input_1091_cast_fp16 = conv(dilations = input_1091_dilations_0, groups = input_1091_groups_0, pad = input_1091_pad_0, pad_type = input_1091_pad_type_0, strides = input_1091_strides_0, weight = encoder_layers_20_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_1089_cast_fp16)[name = string("input_1091_cast_fp16")]; + int32 x_539_split_num_splits_0 = const()[name = string("x_539_split_num_splits_0"), val = int32(2)]; + int32 x_539_split_axis_0 = const()[name = string("x_539_split_axis_0"), val = int32(1)]; + tensor x_539_split_cast_fp16_0, tensor x_539_split_cast_fp16_1 = split(axis = x_539_split_axis_0, num_splits = x_539_split_num_splits_0, x = input_1091_cast_fp16)[name = string("x_539_split_cast_fp16")]; + tensor x_539_split_1_sigmoid_cast_fp16 = sigmoid(x = x_539_split_cast_fp16_1)[name = string("x_539_split_1_sigmoid_cast_fp16")]; + tensor x_539_cast_fp16 = mul(x = x_539_split_cast_fp16_0, y = x_539_split_1_sigmoid_cast_fp16)[name = string("x_539_cast_fp16")]; + tensor input_1093_cast_fp16 = select(a = var_44_to_fp16, b = x_539_cast_fp16, cond = var_575)[name = string("input_1093_cast_fp16")]; + bool new_x_83_interleave_0 = const()[name = string("new_x_83_interleave_0"), val = bool(false)]; + tensor new_x_83_cast_fp16 = concat(axis = var_59, interleave = new_x_83_interleave_0, values = (cache_83_cast_fp16, input_1093_cast_fp16))[name = string("new_x_83_cast_fp16")]; + tensor var_4848_begin_0 = const()[name = string("op_4848_begin_0"), val = tensor([0, 0, 28])]; + tensor var_4848_end_0 = const()[name = string("op_4848_end_0"), val = tensor([1, 1024, 36])]; + tensor var_4848_end_mask_0 = const()[name = string("op_4848_end_mask_0"), val = tensor([true, true, true])]; + tensor var_4848_cast_fp16 = slice_by_index(begin = var_4848_begin_0, end = var_4848_end_0, end_mask = var_4848_end_mask_0, x = new_x_83_cast_fp16)[name = string("op_4848_cast_fp16")]; + string x_541_pad_type_0 = const()[name = string("x_541_pad_type_0"), val = string("valid")]; + int32 x_541_groups_0 = const()[name = string("x_541_groups_0"), val = int32(1024)]; + tensor x_541_strides_0 = const()[name = string("x_541_strides_0"), val = tensor([1])]; + tensor x_541_pad_0 = const()[name = string("x_541_pad_0"), val = tensor([0, 0])]; + tensor x_541_dilations_0 = const()[name = string("x_541_dilations_0"), val = tensor([1])]; + tensor encoder_layers_20_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(422486592))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(422495872))))[name = string("encoder_layers_20_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_541_cast_fp16 = conv(dilations = x_541_dilations_0, groups = x_541_groups_0, pad = x_541_pad_0, pad_type = x_541_pad_type_0, strides = x_541_strides_0, weight = encoder_layers_20_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_83_cast_fp16)[name = string("x_541_cast_fp16")]; + tensor input_1095_perm_0 = const()[name = string("input_1095_perm_0"), val = tensor([0, 2, 1])]; + tensor x_543_axes_0 = const()[name = string("x_543_axes_0"), val = tensor([-1])]; + tensor encoder_layers_20_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_20_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(422497984)))]; + tensor encoder_layers_20_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_20_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(422500096)))]; + tensor input_1095_cast_fp16 = transpose(perm = input_1095_perm_0, x = x_541_cast_fp16)[name = string("transpose_176")]; + tensor x_543_cast_fp16 = layer_norm(axes = x_543_axes_0, beta = encoder_layers_20_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_20_conv_batch_norm_weight_to_fp16, x = input_1095_cast_fp16)[name = string("x_543_cast_fp16")]; + tensor input_1097_perm_0 = const()[name = string("input_1097_perm_0"), val = tensor([0, 2, 1])]; + tensor input_1097_cast_fp16 = transpose(perm = input_1097_perm_0, x = x_543_cast_fp16)[name = string("transpose_175")]; + tensor input_1099_cast_fp16 = silu(x = input_1097_cast_fp16)[name = string("input_1099_cast_fp16")]; + string x_545_pad_type_0 = const()[name = string("x_545_pad_type_0"), val = string("valid")]; + tensor x_545_strides_0 = const()[name = string("x_545_strides_0"), val = tensor([1])]; + tensor x_545_pad_0 = const()[name = string("x_545_pad_0"), val = tensor([0, 0])]; + tensor x_545_dilations_0 = const()[name = string("x_545_dilations_0"), val = tensor([1])]; + int32 x_545_groups_0 = const()[name = string("x_545_groups_0"), val = int32(1)]; + tensor encoder_layers_20_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(422502208))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(423550848))))[name = string("encoder_layers_20_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_545_cast_fp16 = conv(dilations = x_545_dilations_0, groups = x_545_groups_0, pad = x_545_pad_0, pad_type = x_545_pad_type_0, strides = x_545_strides_0, weight = encoder_layers_20_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_1099_cast_fp16)[name = string("x_545_cast_fp16")]; + tensor input_1101_perm_0 = const()[name = string("input_1101_perm_0"), val = tensor([0, 2, 1])]; + tensor input_1101_cast_fp16 = transpose(perm = input_1101_perm_0, x = x_545_cast_fp16)[name = string("transpose_174")]; + tensor input_1103_cast_fp16 = add(x = input_1087_cast_fp16, y = input_1101_cast_fp16)[name = string("input_1103_cast_fp16")]; + tensor input_1105_axes_0 = const()[name = string("input_1105_axes_0"), val = tensor([-1])]; + tensor encoder_layers_20_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_20_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(423552960)))]; + tensor encoder_layers_20_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_20_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(423555072)))]; + tensor input_1105_cast_fp16 = layer_norm(axes = input_1105_axes_0, beta = encoder_layers_20_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_20_norm_feed_forward2_weight_to_fp16, x = input_1103_cast_fp16)[name = string("input_1105_cast_fp16")]; + tensor encoder_layers_20_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(423557184))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427751552))))[name = string("encoder_layers_20_feed_forward2_linear1_weight_to_fp16_quantized")]; + tensor encoder_layers_20_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_20_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427759808)))]; + tensor linear_188_cast_fp16 = linear(bias = encoder_layers_20_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_20_feed_forward2_linear1_weight_to_fp16_quantized, x = input_1105_cast_fp16)[name = string("linear_188_cast_fp16")]; + tensor input_1109_cast_fp16 = silu(x = linear_188_cast_fp16)[name = string("input_1109_cast_fp16")]; + tensor encoder_layers_20_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427768064))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(431962432))))[name = string("encoder_layers_20_feed_forward2_linear2_weight_to_fp16_quantized")]; + tensor encoder_layers_20_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_20_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(431964544)))]; + tensor linear_189_cast_fp16 = linear(bias = encoder_layers_20_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_20_feed_forward2_linear2_weight_to_fp16_quantized, x = input_1109_cast_fp16)[name = string("linear_189_cast_fp16")]; + fp16 var_4891_to_fp16 = const()[name = string("op_4891_to_fp16"), val = fp16(0x1p-1)]; + tensor var_4892_cast_fp16 = mul(x = linear_189_cast_fp16, y = var_4891_to_fp16)[name = string("op_4892_cast_fp16")]; + tensor input_1115_cast_fp16 = add(x = input_1103_cast_fp16, y = var_4892_cast_fp16)[name = string("input_1115_cast_fp16")]; + tensor input_1117_axes_0 = const()[name = string("input_1117_axes_0"), val = tensor([-1])]; + tensor encoder_layers_20_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_20_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(431966656)))]; + tensor encoder_layers_20_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_20_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(431968768)))]; + tensor input_1117_cast_fp16 = layer_norm(axes = input_1117_axes_0, beta = encoder_layers_20_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_20_norm_out_weight_to_fp16, x = input_1115_cast_fp16)[name = string("input_1117_cast_fp16")]; + tensor cache_85_begin_0 = const()[name = string("cache_85_begin_0"), val = tensor([21, 0, 0, 0])]; + tensor cache_85_end_0 = const()[name = string("cache_85_end_0"), val = tensor([22, 1, 42, 1024])]; + tensor cache_85_end_mask_0 = const()[name = string("cache_85_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_85_squeeze_mask_0 = const()[name = string("cache_85_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_85_cast_fp16 = slice_by_index(begin = cache_85_begin_0, end = cache_85_end_0, end_mask = cache_85_end_mask_0, squeeze_mask = cache_85_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_85_cast_fp16")]; + tensor cache_87_begin_0 = const()[name = string("cache_87_begin_0"), val = tensor([21, 0, 0, 0])]; + tensor cache_87_end_0 = const()[name = string("cache_87_end_0"), val = tensor([22, 1, 1024, 8])]; + tensor cache_87_end_mask_0 = const()[name = string("cache_87_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_87_squeeze_mask_0 = const()[name = string("cache_87_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_87_cast_fp16 = slice_by_index(begin = cache_87_begin_0, end = cache_87_end_0, end_mask = cache_87_end_mask_0, squeeze_mask = cache_87_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_87_cast_fp16")]; + tensor input_1119_axes_0 = const()[name = string("input_1119_axes_0"), val = tensor([-1])]; + tensor encoder_layers_21_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_21_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(431970880)))]; + tensor encoder_layers_21_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_21_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(431972992)))]; + tensor input_1119_cast_fp16 = layer_norm(axes = input_1119_axes_0, beta = encoder_layers_21_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_21_norm_feed_forward1_weight_to_fp16, x = input_1117_cast_fp16)[name = string("input_1119_cast_fp16")]; + tensor encoder_layers_21_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(431975104))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436169472))))[name = string("encoder_layers_21_feed_forward1_linear1_weight_to_fp16_quantized")]; + tensor encoder_layers_21_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_21_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436177728)))]; + tensor linear_190_cast_fp16 = linear(bias = encoder_layers_21_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_21_feed_forward1_linear1_weight_to_fp16_quantized, x = input_1119_cast_fp16)[name = string("linear_190_cast_fp16")]; + tensor input_1123_cast_fp16 = silu(x = linear_190_cast_fp16)[name = string("input_1123_cast_fp16")]; + tensor encoder_layers_21_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436185984))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(440380352))))[name = string("encoder_layers_21_feed_forward1_linear2_weight_to_fp16_quantized")]; + tensor encoder_layers_21_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_21_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(440382464)))]; + tensor linear_191_cast_fp16 = linear(bias = encoder_layers_21_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_21_feed_forward1_linear2_weight_to_fp16_quantized, x = input_1123_cast_fp16)[name = string("linear_191_cast_fp16")]; + fp16 var_4928_to_fp16 = const()[name = string("op_4928_to_fp16"), val = fp16(0x1p-1)]; + tensor var_4929_cast_fp16 = mul(x = linear_191_cast_fp16, y = var_4928_to_fp16)[name = string("op_4929_cast_fp16")]; + tensor input_1129_cast_fp16 = add(x = input_1117_cast_fp16, y = var_4929_cast_fp16)[name = string("input_1129_cast_fp16")]; + tensor key_43_axes_0 = const()[name = string("key_43_axes_0"), val = tensor([-1])]; + tensor encoder_layers_21_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_21_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(440384576)))]; + tensor encoder_layers_21_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_21_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(440386688)))]; + tensor key_43_cast_fp16 = layer_norm(axes = key_43_axes_0, beta = encoder_layers_21_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_21_norm_self_att_weight_to_fp16, x = input_1129_cast_fp16)[name = string("key_43_cast_fp16")]; + bool input_1131_interleave_0 = const()[name = string("input_1131_interleave_0"), val = bool(false)]; + tensor input_1131_cast_fp16 = concat(axis = var_68, interleave = input_1131_interleave_0, values = (cache_85_cast_fp16, key_43_cast_fp16))[name = string("input_1131_cast_fp16")]; + tensor var_4951_begin_0 = const()[name = string("op_4951_begin_0"), val = tensor([0, 28, 0])]; + tensor var_4951_end_0 = const()[name = string("op_4951_end_0"), val = tensor([1, 42, 1024])]; + tensor var_4951_end_mask_0 = const()[name = string("op_4951_end_mask_0"), val = tensor([true, true, true])]; + tensor var_4951_cast_fp16 = slice_by_index(begin = var_4951_begin_0, end = var_4951_end_0, end_mask = var_4951_end_mask_0, x = cache_85_cast_fp16)[name = string("op_4951_cast_fp16")]; + bool var_4957_interleave_0 = const()[name = string("op_4957_interleave_0"), val = bool(false)]; + tensor var_4957_cast_fp16 = concat(axis = var_68, interleave = var_4957_interleave_0, values = (var_4951_cast_fp16, key_43_cast_fp16))[name = string("op_4957_cast_fp16")]; + tensor encoder_layers_21_self_attn_linear_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(440388800))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(441437440))))[name = string("encoder_layers_21_self_attn_linear_q_weight_to_fp16_quantized")]; + tensor encoder_layers_21_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_21_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(441439552)))]; + tensor linear_192_cast_fp16 = linear(bias = encoder_layers_21_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_21_self_attn_linear_q_weight_to_fp16_quantized, x = key_43_cast_fp16)[name = string("linear_192_cast_fp16")]; + tensor var_4962 = const()[name = string("op_4962"), val = tensor([1, -1, 8, 128])]; + tensor q_127_cast_fp16 = reshape(shape = var_4962, x = linear_192_cast_fp16)[name = string("q_127_cast_fp16")]; + tensor encoder_layers_21_self_attn_linear_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(441441664))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(442490304))))[name = string("encoder_layers_21_self_attn_linear_k_weight_to_fp16_quantized")]; + tensor encoder_layers_21_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_21_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(442492416)))]; + tensor linear_193_cast_fp16 = linear(bias = encoder_layers_21_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_21_self_attn_linear_k_weight_to_fp16_quantized, x = input_1131_cast_fp16)[name = string("linear_193_cast_fp16")]; + tensor var_4967 = const()[name = string("op_4967"), val = tensor([1, -1, 8, 128])]; + tensor k_85_cast_fp16 = reshape(shape = var_4967, x = linear_193_cast_fp16)[name = string("k_85_cast_fp16")]; + tensor encoder_layers_21_self_attn_linear_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(442494528))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(443543168))))[name = string("encoder_layers_21_self_attn_linear_v_weight_to_fp16_quantized")]; + tensor encoder_layers_21_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_21_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(443545280)))]; + tensor linear_194_cast_fp16 = linear(bias = encoder_layers_21_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_21_self_attn_linear_v_weight_to_fp16_quantized, x = input_1131_cast_fp16)[name = string("linear_194_cast_fp16")]; + tensor var_4972 = const()[name = string("op_4972"), val = tensor([1, -1, 8, 128])]; + tensor v_43_cast_fp16 = reshape(shape = var_4972, x = linear_194_cast_fp16)[name = string("v_43_cast_fp16")]; + tensor value_51_perm_0 = const()[name = string("value_51_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_21_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_21_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(443547392)))]; + tensor var_4985_cast_fp16 = add(x = q_127_cast_fp16, y = encoder_layers_21_self_attn_pos_bias_u_to_fp16)[name = string("op_4985_cast_fp16")]; + tensor encoder_layers_21_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_21_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(443549504)))]; + tensor var_4987_cast_fp16 = add(x = q_127_cast_fp16, y = encoder_layers_21_self_attn_pos_bias_v_to_fp16)[name = string("op_4987_cast_fp16")]; + tensor q_with_bias_v_43_perm_0 = const()[name = string("q_with_bias_v_43_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_553_transpose_x_0 = const()[name = string("x_553_transpose_x_0"), val = bool(false)]; + bool x_553_transpose_y_0 = const()[name = string("x_553_transpose_y_0"), val = bool(false)]; + tensor op_4989_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(443551616))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(443694016))))[name = string("op_4989_to_fp16_quantized")]; + tensor q_with_bias_v_43_cast_fp16 = transpose(perm = q_with_bias_v_43_perm_0, x = var_4987_cast_fp16)[name = string("transpose_173")]; + tensor x_553_cast_fp16 = matmul(transpose_x = x_553_transpose_x_0, transpose_y = x_553_transpose_y_0, x = q_with_bias_v_43_cast_fp16, y = op_4989_to_fp16_quantized)[name = string("x_553_cast_fp16")]; + tensor x_555_pad_0 = const()[name = string("x_555_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_555_mode_0 = const()[name = string("x_555_mode_0"), val = string("constant")]; + fp16 const_352_to_fp16 = const()[name = string("const_352_to_fp16"), val = fp16(0x0p+0)]; + tensor x_555_cast_fp16 = pad(constant_val = const_352_to_fp16, mode = x_555_mode_0, pad = x_555_pad_0, x = x_553_cast_fp16)[name = string("x_555_cast_fp16")]; + tensor var_4997 = const()[name = string("op_4997"), val = tensor([1, 8, -1, 28])]; + tensor x_557_cast_fp16 = reshape(shape = var_4997, x = x_555_cast_fp16)[name = string("x_557_cast_fp16")]; + tensor var_5001_begin_0 = const()[name = string("op_5001_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_5001_end_0 = const()[name = string("op_5001_end_0"), val = tensor([1, 8, 140, 28])]; + tensor var_5001_end_mask_0 = const()[name = string("op_5001_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_5001_cast_fp16 = slice_by_index(begin = var_5001_begin_0, end = var_5001_end_0, end_mask = var_5001_end_mask_0, x = x_557_cast_fp16)[name = string("op_5001_cast_fp16")]; + tensor var_5002 = const()[name = string("op_5002"), val = tensor([1, 8, 28, 139])]; + tensor matrix_bd_85_cast_fp16 = reshape(shape = var_5002, x = var_5001_cast_fp16)[name = string("matrix_bd_85_cast_fp16")]; + bool matrix_ac_43_transpose_x_0 = const()[name = string("matrix_ac_43_transpose_x_0"), val = bool(false)]; + bool matrix_ac_43_transpose_y_0 = const()[name = string("matrix_ac_43_transpose_y_0"), val = bool(false)]; + tensor transpose_138_perm_0 = const()[name = string("transpose_138_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_139_perm_0 = const()[name = string("transpose_139_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_139 = transpose(perm = transpose_139_perm_0, x = k_85_cast_fp16)[name = string("transpose_171")]; + tensor transpose_138 = transpose(perm = transpose_138_perm_0, x = var_4985_cast_fp16)[name = string("transpose_172")]; + tensor matrix_ac_43_cast_fp16 = matmul(transpose_x = matrix_ac_43_transpose_x_0, transpose_y = matrix_ac_43_transpose_y_0, x = transpose_138, y = transpose_139)[name = string("matrix_ac_43_cast_fp16")]; + tensor matrix_bd_87_begin_0 = const()[name = string("matrix_bd_87_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_87_end_0 = const()[name = string("matrix_bd_87_end_0"), val = tensor([1, 8, 28, 70])]; + tensor matrix_bd_87_end_mask_0 = const()[name = string("matrix_bd_87_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_87_cast_fp16 = slice_by_index(begin = matrix_bd_87_begin_0, end = matrix_bd_87_end_0, end_mask = matrix_bd_87_end_mask_0, x = matrix_bd_85_cast_fp16)[name = string("matrix_bd_87_cast_fp16")]; + tensor var_5011_cast_fp16 = add(x = matrix_ac_43_cast_fp16, y = matrix_bd_87_cast_fp16)[name = string("op_5011_cast_fp16")]; + fp16 _inversed_scores_85_y_0_to_fp16 = const()[name = string("_inversed_scores_85_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_85_cast_fp16 = mul(x = var_5011_cast_fp16, y = _inversed_scores_85_y_0_to_fp16)[name = string("_inversed_scores_85_cast_fp16")]; + tensor scores_87_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_85_cast_fp16, cond = mask_11)[name = string("scores_87_cast_fp16")]; + tensor var_5017_cast_fp16 = softmax(axis = var_59, x = scores_87_cast_fp16)[name = string("op_5017_cast_fp16")]; + tensor input_1133_cast_fp16 = select(a = var_44_to_fp16, b = var_5017_cast_fp16, cond = mask_11)[name = string("input_1133_cast_fp16")]; + bool x_559_transpose_x_0 = const()[name = string("x_559_transpose_x_0"), val = bool(false)]; + bool x_559_transpose_y_0 = const()[name = string("x_559_transpose_y_0"), val = bool(false)]; + tensor value_51_cast_fp16 = transpose(perm = value_51_perm_0, x = v_43_cast_fp16)[name = string("transpose_170")]; + tensor x_559_cast_fp16 = matmul(transpose_x = x_559_transpose_x_0, transpose_y = x_559_transpose_y_0, x = input_1133_cast_fp16, y = value_51_cast_fp16)[name = string("x_559_cast_fp16")]; + tensor var_5021_perm_0 = const()[name = string("op_5021_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_5022 = const()[name = string("op_5022"), val = tensor([1, -1, 1024])]; + tensor var_5021_cast_fp16 = transpose(perm = var_5021_perm_0, x = x_559_cast_fp16)[name = string("transpose_169")]; + tensor input_1135_cast_fp16 = reshape(shape = var_5022, x = var_5021_cast_fp16)[name = string("input_1135_cast_fp16")]; + tensor encoder_layers_21_self_attn_linear_out_weight_to_fp16 = const()[name = string("encoder_layers_21_self_attn_linear_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(443694400)))]; + tensor encoder_layers_21_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_21_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(445791616)))]; + tensor linear_196_cast_fp16 = linear(bias = encoder_layers_21_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_21_self_attn_linear_out_weight_to_fp16, x = input_1135_cast_fp16)[name = string("linear_196_cast_fp16")]; + tensor input_1139_cast_fp16 = add(x = input_1129_cast_fp16, y = linear_196_cast_fp16)[name = string("input_1139_cast_fp16")]; + tensor x_563_axes_0 = const()[name = string("x_563_axes_0"), val = tensor([-1])]; + tensor encoder_layers_21_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_21_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(445793728)))]; + tensor encoder_layers_21_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_21_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(445795840)))]; + tensor x_563_cast_fp16 = layer_norm(axes = x_563_axes_0, beta = encoder_layers_21_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_21_norm_conv_weight_to_fp16, x = input_1139_cast_fp16)[name = string("x_563_cast_fp16")]; + tensor input_1141_perm_0 = const()[name = string("input_1141_perm_0"), val = tensor([0, 2, 1])]; + string input_1143_pad_type_0 = const()[name = string("input_1143_pad_type_0"), val = string("valid")]; + tensor input_1143_strides_0 = const()[name = string("input_1143_strides_0"), val = tensor([1])]; + tensor input_1143_pad_0 = const()[name = string("input_1143_pad_0"), val = tensor([0, 0])]; + tensor input_1143_dilations_0 = const()[name = string("input_1143_dilations_0"), val = tensor([1])]; + int32 input_1143_groups_0 = const()[name = string("input_1143_groups_0"), val = int32(1)]; + tensor encoder_layers_21_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(445797952))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(447895168))))[name = string("encoder_layers_21_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_1141_cast_fp16 = transpose(perm = input_1141_perm_0, x = x_563_cast_fp16)[name = string("transpose_168")]; + tensor input_1143_cast_fp16 = conv(dilations = input_1143_dilations_0, groups = input_1143_groups_0, pad = input_1143_pad_0, pad_type = input_1143_pad_type_0, strides = input_1143_strides_0, weight = encoder_layers_21_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_1141_cast_fp16)[name = string("input_1143_cast_fp16")]; + int32 x_565_split_num_splits_0 = const()[name = string("x_565_split_num_splits_0"), val = int32(2)]; + int32 x_565_split_axis_0 = const()[name = string("x_565_split_axis_0"), val = int32(1)]; + tensor x_565_split_cast_fp16_0, tensor x_565_split_cast_fp16_1 = split(axis = x_565_split_axis_0, num_splits = x_565_split_num_splits_0, x = input_1143_cast_fp16)[name = string("x_565_split_cast_fp16")]; + tensor x_565_split_1_sigmoid_cast_fp16 = sigmoid(x = x_565_split_cast_fp16_1)[name = string("x_565_split_1_sigmoid_cast_fp16")]; + tensor x_565_cast_fp16 = mul(x = x_565_split_cast_fp16_0, y = x_565_split_1_sigmoid_cast_fp16)[name = string("x_565_cast_fp16")]; + tensor input_1145_cast_fp16 = select(a = var_44_to_fp16, b = x_565_cast_fp16, cond = var_575)[name = string("input_1145_cast_fp16")]; + bool new_x_87_interleave_0 = const()[name = string("new_x_87_interleave_0"), val = bool(false)]; + tensor new_x_87_cast_fp16 = concat(axis = var_59, interleave = new_x_87_interleave_0, values = (cache_87_cast_fp16, input_1145_cast_fp16))[name = string("new_x_87_cast_fp16")]; + tensor var_5061_begin_0 = const()[name = string("op_5061_begin_0"), val = tensor([0, 0, 28])]; + tensor var_5061_end_0 = const()[name = string("op_5061_end_0"), val = tensor([1, 1024, 36])]; + tensor var_5061_end_mask_0 = const()[name = string("op_5061_end_mask_0"), val = tensor([true, true, true])]; + tensor var_5061_cast_fp16 = slice_by_index(begin = var_5061_begin_0, end = var_5061_end_0, end_mask = var_5061_end_mask_0, x = new_x_87_cast_fp16)[name = string("op_5061_cast_fp16")]; + string x_567_pad_type_0 = const()[name = string("x_567_pad_type_0"), val = string("valid")]; + int32 x_567_groups_0 = const()[name = string("x_567_groups_0"), val = int32(1024)]; + tensor x_567_strides_0 = const()[name = string("x_567_strides_0"), val = tensor([1])]; + tensor x_567_pad_0 = const()[name = string("x_567_pad_0"), val = tensor([0, 0])]; + tensor x_567_dilations_0 = const()[name = string("x_567_dilations_0"), val = tensor([1])]; + tensor encoder_layers_21_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(447899328))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(447908608))))[name = string("encoder_layers_21_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_567_cast_fp16 = conv(dilations = x_567_dilations_0, groups = x_567_groups_0, pad = x_567_pad_0, pad_type = x_567_pad_type_0, strides = x_567_strides_0, weight = encoder_layers_21_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_87_cast_fp16)[name = string("x_567_cast_fp16")]; + tensor input_1147_perm_0 = const()[name = string("input_1147_perm_0"), val = tensor([0, 2, 1])]; + tensor x_569_axes_0 = const()[name = string("x_569_axes_0"), val = tensor([-1])]; + tensor encoder_layers_21_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_21_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(447910720)))]; + tensor encoder_layers_21_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_21_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(447912832)))]; + tensor input_1147_cast_fp16 = transpose(perm = input_1147_perm_0, x = x_567_cast_fp16)[name = string("transpose_167")]; + tensor x_569_cast_fp16 = layer_norm(axes = x_569_axes_0, beta = encoder_layers_21_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_21_conv_batch_norm_weight_to_fp16, x = input_1147_cast_fp16)[name = string("x_569_cast_fp16")]; + tensor input_1149_perm_0 = const()[name = string("input_1149_perm_0"), val = tensor([0, 2, 1])]; + tensor input_1149_cast_fp16 = transpose(perm = input_1149_perm_0, x = x_569_cast_fp16)[name = string("transpose_166")]; + tensor input_1151_cast_fp16 = silu(x = input_1149_cast_fp16)[name = string("input_1151_cast_fp16")]; + string x_571_pad_type_0 = const()[name = string("x_571_pad_type_0"), val = string("valid")]; + tensor x_571_strides_0 = const()[name = string("x_571_strides_0"), val = tensor([1])]; + tensor x_571_pad_0 = const()[name = string("x_571_pad_0"), val = tensor([0, 0])]; + tensor x_571_dilations_0 = const()[name = string("x_571_dilations_0"), val = tensor([1])]; + int32 x_571_groups_0 = const()[name = string("x_571_groups_0"), val = int32(1)]; + tensor encoder_layers_21_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(447914944))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(448963584))))[name = string("encoder_layers_21_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_571_cast_fp16 = conv(dilations = x_571_dilations_0, groups = x_571_groups_0, pad = x_571_pad_0, pad_type = x_571_pad_type_0, strides = x_571_strides_0, weight = encoder_layers_21_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_1151_cast_fp16)[name = string("x_571_cast_fp16")]; + tensor input_1153_perm_0 = const()[name = string("input_1153_perm_0"), val = tensor([0, 2, 1])]; + tensor input_1153_cast_fp16 = transpose(perm = input_1153_perm_0, x = x_571_cast_fp16)[name = string("transpose_165")]; + tensor input_1155_cast_fp16 = add(x = input_1139_cast_fp16, y = input_1153_cast_fp16)[name = string("input_1155_cast_fp16")]; + tensor input_1157_axes_0 = const()[name = string("input_1157_axes_0"), val = tensor([-1])]; + tensor encoder_layers_21_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_21_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(448965696)))]; + tensor encoder_layers_21_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_21_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(448967808)))]; + tensor input_1157_cast_fp16 = layer_norm(axes = input_1157_axes_0, beta = encoder_layers_21_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_21_norm_feed_forward2_weight_to_fp16, x = input_1155_cast_fp16)[name = string("input_1157_cast_fp16")]; + tensor encoder_layers_21_feed_forward2_linear1_weight_to_fp16 = const()[name = string("encoder_layers_21_feed_forward2_linear1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(448969920)))]; + tensor encoder_layers_21_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_21_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(457358592)))]; + tensor linear_197_cast_fp16 = linear(bias = encoder_layers_21_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_21_feed_forward2_linear1_weight_to_fp16, x = input_1157_cast_fp16)[name = string("linear_197_cast_fp16")]; + tensor input_1161_cast_fp16 = silu(x = linear_197_cast_fp16)[name = string("input_1161_cast_fp16")]; + tensor encoder_layers_21_feed_forward2_linear2_weight_to_fp16 = const()[name = string("encoder_layers_21_feed_forward2_linear2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(457366848)))]; + tensor encoder_layers_21_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_21_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465755520)))]; + tensor linear_198_cast_fp16 = linear(bias = encoder_layers_21_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_21_feed_forward2_linear2_weight_to_fp16, x = input_1161_cast_fp16)[name = string("linear_198_cast_fp16")]; + fp16 var_5104_to_fp16 = const()[name = string("op_5104_to_fp16"), val = fp16(0x1p-1)]; + tensor var_5105_cast_fp16 = mul(x = linear_198_cast_fp16, y = var_5104_to_fp16)[name = string("op_5105_cast_fp16")]; + tensor input_1167_cast_fp16 = add(x = input_1155_cast_fp16, y = var_5105_cast_fp16)[name = string("input_1167_cast_fp16")]; + tensor input_1169_axes_0 = const()[name = string("input_1169_axes_0"), val = tensor([-1])]; + tensor encoder_layers_21_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_21_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465757632)))]; + tensor encoder_layers_21_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_21_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465759744)))]; + tensor input_1169_cast_fp16 = layer_norm(axes = input_1169_axes_0, beta = encoder_layers_21_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_21_norm_out_weight_to_fp16, x = input_1167_cast_fp16)[name = string("input_1169_cast_fp16")]; + tensor cache_89_begin_0 = const()[name = string("cache_89_begin_0"), val = tensor([22, 0, 0, 0])]; + tensor cache_89_end_0 = const()[name = string("cache_89_end_0"), val = tensor([23, 1, 42, 1024])]; + tensor cache_89_end_mask_0 = const()[name = string("cache_89_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_89_squeeze_mask_0 = const()[name = string("cache_89_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_89_cast_fp16 = slice_by_index(begin = cache_89_begin_0, end = cache_89_end_0, end_mask = cache_89_end_mask_0, squeeze_mask = cache_89_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_89_cast_fp16")]; + tensor cache_91_begin_0 = const()[name = string("cache_91_begin_0"), val = tensor([22, 0, 0, 0])]; + tensor cache_91_end_0 = const()[name = string("cache_91_end_0"), val = tensor([23, 1, 1024, 8])]; + tensor cache_91_end_mask_0 = const()[name = string("cache_91_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_91_squeeze_mask_0 = const()[name = string("cache_91_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_91_cast_fp16 = slice_by_index(begin = cache_91_begin_0, end = cache_91_end_0, end_mask = cache_91_end_mask_0, squeeze_mask = cache_91_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_91_cast_fp16")]; + tensor input_1171_axes_0 = const()[name = string("input_1171_axes_0"), val = tensor([-1])]; + tensor encoder_layers_22_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_22_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465761856)))]; + tensor encoder_layers_22_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_22_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465763968)))]; + tensor input_1171_cast_fp16 = layer_norm(axes = input_1171_axes_0, beta = encoder_layers_22_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_22_norm_feed_forward1_weight_to_fp16, x = input_1169_cast_fp16)[name = string("input_1171_cast_fp16")]; + tensor encoder_layers_22_feed_forward1_linear1_weight_to_fp16 = const()[name = string("encoder_layers_22_feed_forward1_linear1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465766080)))]; + tensor encoder_layers_22_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_22_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(474154752)))]; + tensor linear_199_cast_fp16 = linear(bias = encoder_layers_22_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_22_feed_forward1_linear1_weight_to_fp16, x = input_1171_cast_fp16)[name = string("linear_199_cast_fp16")]; + tensor input_1175_cast_fp16 = silu(x = linear_199_cast_fp16)[name = string("input_1175_cast_fp16")]; + tensor encoder_layers_22_feed_forward1_linear2_weight_to_fp16 = const()[name = string("encoder_layers_22_feed_forward1_linear2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(474163008)))]; + tensor encoder_layers_22_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_22_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(482551680)))]; + tensor linear_200_cast_fp16 = linear(bias = encoder_layers_22_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_22_feed_forward1_linear2_weight_to_fp16, x = input_1175_cast_fp16)[name = string("linear_200_cast_fp16")]; + fp16 var_5141_to_fp16 = const()[name = string("op_5141_to_fp16"), val = fp16(0x1p-1)]; + tensor var_5142_cast_fp16 = mul(x = linear_200_cast_fp16, y = var_5141_to_fp16)[name = string("op_5142_cast_fp16")]; + tensor input_1181_cast_fp16 = add(x = input_1169_cast_fp16, y = var_5142_cast_fp16)[name = string("input_1181_cast_fp16")]; + tensor key_45_axes_0 = const()[name = string("key_45_axes_0"), val = tensor([-1])]; + tensor encoder_layers_22_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_22_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(482553792)))]; + tensor encoder_layers_22_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_22_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(482555904)))]; + tensor key_45_cast_fp16 = layer_norm(axes = key_45_axes_0, beta = encoder_layers_22_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_22_norm_self_att_weight_to_fp16, x = input_1181_cast_fp16)[name = string("key_45_cast_fp16")]; + bool input_1183_interleave_0 = const()[name = string("input_1183_interleave_0"), val = bool(false)]; + tensor input_1183_cast_fp16 = concat(axis = var_68, interleave = input_1183_interleave_0, values = (cache_89_cast_fp16, key_45_cast_fp16))[name = string("input_1183_cast_fp16")]; + tensor var_5164_begin_0 = const()[name = string("op_5164_begin_0"), val = tensor([0, 28, 0])]; + tensor var_5164_end_0 = const()[name = string("op_5164_end_0"), val = tensor([1, 42, 1024])]; + tensor var_5164_end_mask_0 = const()[name = string("op_5164_end_mask_0"), val = tensor([true, true, true])]; + tensor var_5164_cast_fp16 = slice_by_index(begin = var_5164_begin_0, end = var_5164_end_0, end_mask = var_5164_end_mask_0, x = cache_89_cast_fp16)[name = string("op_5164_cast_fp16")]; + bool var_5170_interleave_0 = const()[name = string("op_5170_interleave_0"), val = bool(false)]; + tensor var_5170_cast_fp16 = concat(axis = var_68, interleave = var_5170_interleave_0, values = (var_5164_cast_fp16, key_45_cast_fp16))[name = string("op_5170_cast_fp16")]; + tensor encoder_layers_22_self_attn_linear_q_weight_to_fp16 = const()[name = string("encoder_layers_22_self_attn_linear_q_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(482558016)))]; + tensor encoder_layers_22_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_22_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(484655232)))]; + tensor linear_201_cast_fp16 = linear(bias = encoder_layers_22_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_22_self_attn_linear_q_weight_to_fp16, x = key_45_cast_fp16)[name = string("linear_201_cast_fp16")]; + tensor var_5175 = const()[name = string("op_5175"), val = tensor([1, -1, 8, 128])]; + tensor q_133_cast_fp16 = reshape(shape = var_5175, x = linear_201_cast_fp16)[name = string("q_133_cast_fp16")]; + tensor encoder_layers_22_self_attn_linear_k_weight_to_fp16 = const()[name = string("encoder_layers_22_self_attn_linear_k_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(484657344)))]; + tensor encoder_layers_22_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_22_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(486754560)))]; + tensor linear_202_cast_fp16 = linear(bias = encoder_layers_22_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_22_self_attn_linear_k_weight_to_fp16, x = input_1183_cast_fp16)[name = string("linear_202_cast_fp16")]; + tensor var_5180 = const()[name = string("op_5180"), val = tensor([1, -1, 8, 128])]; + tensor k_89_cast_fp16 = reshape(shape = var_5180, x = linear_202_cast_fp16)[name = string("k_89_cast_fp16")]; + tensor encoder_layers_22_self_attn_linear_v_weight_to_fp16 = const()[name = string("encoder_layers_22_self_attn_linear_v_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(486756672)))]; + tensor encoder_layers_22_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_22_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(488853888)))]; + tensor linear_203_cast_fp16 = linear(bias = encoder_layers_22_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_22_self_attn_linear_v_weight_to_fp16, x = input_1183_cast_fp16)[name = string("linear_203_cast_fp16")]; + tensor var_5185 = const()[name = string("op_5185"), val = tensor([1, -1, 8, 128])]; + tensor v_45_cast_fp16 = reshape(shape = var_5185, x = linear_203_cast_fp16)[name = string("v_45_cast_fp16")]; + tensor value_53_perm_0 = const()[name = string("value_53_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_22_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_22_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(488856000)))]; + tensor var_5198_cast_fp16 = add(x = q_133_cast_fp16, y = encoder_layers_22_self_attn_pos_bias_u_to_fp16)[name = string("op_5198_cast_fp16")]; + tensor encoder_layers_22_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_22_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(488858112)))]; + tensor var_5200_cast_fp16 = add(x = q_133_cast_fp16, y = encoder_layers_22_self_attn_pos_bias_v_to_fp16)[name = string("op_5200_cast_fp16")]; + tensor q_with_bias_v_45_perm_0 = const()[name = string("q_with_bias_v_45_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_579_transpose_x_0 = const()[name = string("x_579_transpose_x_0"), val = bool(false)]; + bool x_579_transpose_y_0 = const()[name = string("x_579_transpose_y_0"), val = bool(false)]; + tensor op_5202_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(488860224))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(489002624))))[name = string("op_5202_to_fp16_quantized")]; + tensor q_with_bias_v_45_cast_fp16 = transpose(perm = q_with_bias_v_45_perm_0, x = var_5200_cast_fp16)[name = string("transpose_164")]; + tensor x_579_cast_fp16 = matmul(transpose_x = x_579_transpose_x_0, transpose_y = x_579_transpose_y_0, x = q_with_bias_v_45_cast_fp16, y = op_5202_to_fp16_quantized)[name = string("x_579_cast_fp16")]; + tensor x_581_pad_0 = const()[name = string("x_581_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_581_mode_0 = const()[name = string("x_581_mode_0"), val = string("constant")]; + fp16 const_365_to_fp16 = const()[name = string("const_365_to_fp16"), val = fp16(0x0p+0)]; + tensor x_581_cast_fp16 = pad(constant_val = const_365_to_fp16, mode = x_581_mode_0, pad = x_581_pad_0, x = x_579_cast_fp16)[name = string("x_581_cast_fp16")]; + tensor var_5210 = const()[name = string("op_5210"), val = tensor([1, 8, -1, 28])]; + tensor x_583_cast_fp16 = reshape(shape = var_5210, x = x_581_cast_fp16)[name = string("x_583_cast_fp16")]; + tensor var_5214_begin_0 = const()[name = string("op_5214_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_5214_end_0 = const()[name = string("op_5214_end_0"), val = tensor([1, 8, 140, 28])]; + tensor var_5214_end_mask_0 = const()[name = string("op_5214_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_5214_cast_fp16 = slice_by_index(begin = var_5214_begin_0, end = var_5214_end_0, end_mask = var_5214_end_mask_0, x = x_583_cast_fp16)[name = string("op_5214_cast_fp16")]; + tensor var_5215 = const()[name = string("op_5215"), val = tensor([1, 8, 28, 139])]; + tensor matrix_bd_89_cast_fp16 = reshape(shape = var_5215, x = var_5214_cast_fp16)[name = string("matrix_bd_89_cast_fp16")]; + bool matrix_ac_45_transpose_x_0 = const()[name = string("matrix_ac_45_transpose_x_0"), val = bool(false)]; + bool matrix_ac_45_transpose_y_0 = const()[name = string("matrix_ac_45_transpose_y_0"), val = bool(false)]; + tensor transpose_140_perm_0 = const()[name = string("transpose_140_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_141_perm_0 = const()[name = string("transpose_141_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_141 = transpose(perm = transpose_141_perm_0, x = k_89_cast_fp16)[name = string("transpose_162")]; + tensor transpose_140 = transpose(perm = transpose_140_perm_0, x = var_5198_cast_fp16)[name = string("transpose_163")]; + tensor matrix_ac_45_cast_fp16 = matmul(transpose_x = matrix_ac_45_transpose_x_0, transpose_y = matrix_ac_45_transpose_y_0, x = transpose_140, y = transpose_141)[name = string("matrix_ac_45_cast_fp16")]; + tensor matrix_bd_91_begin_0 = const()[name = string("matrix_bd_91_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_91_end_0 = const()[name = string("matrix_bd_91_end_0"), val = tensor([1, 8, 28, 70])]; + tensor matrix_bd_91_end_mask_0 = const()[name = string("matrix_bd_91_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_91_cast_fp16 = slice_by_index(begin = matrix_bd_91_begin_0, end = matrix_bd_91_end_0, end_mask = matrix_bd_91_end_mask_0, x = matrix_bd_89_cast_fp16)[name = string("matrix_bd_91_cast_fp16")]; + tensor var_5224_cast_fp16 = add(x = matrix_ac_45_cast_fp16, y = matrix_bd_91_cast_fp16)[name = string("op_5224_cast_fp16")]; + fp16 _inversed_scores_89_y_0_to_fp16 = const()[name = string("_inversed_scores_89_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_89_cast_fp16 = mul(x = var_5224_cast_fp16, y = _inversed_scores_89_y_0_to_fp16)[name = string("_inversed_scores_89_cast_fp16")]; + tensor scores_91_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_89_cast_fp16, cond = mask_11)[name = string("scores_91_cast_fp16")]; + tensor var_5230_cast_fp16 = softmax(axis = var_59, x = scores_91_cast_fp16)[name = string("op_5230_cast_fp16")]; + tensor input_1185_cast_fp16 = select(a = var_44_to_fp16, b = var_5230_cast_fp16, cond = mask_11)[name = string("input_1185_cast_fp16")]; + bool x_585_transpose_x_0 = const()[name = string("x_585_transpose_x_0"), val = bool(false)]; + bool x_585_transpose_y_0 = const()[name = string("x_585_transpose_y_0"), val = bool(false)]; + tensor value_53_cast_fp16 = transpose(perm = value_53_perm_0, x = v_45_cast_fp16)[name = string("transpose_161")]; + tensor x_585_cast_fp16 = matmul(transpose_x = x_585_transpose_x_0, transpose_y = x_585_transpose_y_0, x = input_1185_cast_fp16, y = value_53_cast_fp16)[name = string("x_585_cast_fp16")]; + tensor var_5234_perm_0 = const()[name = string("op_5234_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_5235 = const()[name = string("op_5235"), val = tensor([1, -1, 1024])]; + tensor var_5234_cast_fp16 = transpose(perm = var_5234_perm_0, x = x_585_cast_fp16)[name = string("transpose_160")]; + tensor input_1187_cast_fp16 = reshape(shape = var_5235, x = var_5234_cast_fp16)[name = string("input_1187_cast_fp16")]; + tensor encoder_layers_22_self_attn_linear_out_weight_to_fp16 = const()[name = string("encoder_layers_22_self_attn_linear_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(489003008)))]; + tensor encoder_layers_22_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_22_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(491100224)))]; + tensor linear_205_cast_fp16 = linear(bias = encoder_layers_22_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_22_self_attn_linear_out_weight_to_fp16, x = input_1187_cast_fp16)[name = string("linear_205_cast_fp16")]; + tensor input_1191_cast_fp16 = add(x = input_1181_cast_fp16, y = linear_205_cast_fp16)[name = string("input_1191_cast_fp16")]; + tensor x_589_axes_0 = const()[name = string("x_589_axes_0"), val = tensor([-1])]; + tensor encoder_layers_22_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_22_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(491102336)))]; + tensor encoder_layers_22_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_22_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(491104448)))]; + tensor x_589_cast_fp16 = layer_norm(axes = x_589_axes_0, beta = encoder_layers_22_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_22_norm_conv_weight_to_fp16, x = input_1191_cast_fp16)[name = string("x_589_cast_fp16")]; + tensor input_1193_perm_0 = const()[name = string("input_1193_perm_0"), val = tensor([0, 2, 1])]; + string input_1195_pad_type_0 = const()[name = string("input_1195_pad_type_0"), val = string("valid")]; + tensor input_1195_strides_0 = const()[name = string("input_1195_strides_0"), val = tensor([1])]; + tensor input_1195_pad_0 = const()[name = string("input_1195_pad_0"), val = tensor([0, 0])]; + tensor input_1195_dilations_0 = const()[name = string("input_1195_dilations_0"), val = tensor([1])]; + int32 input_1195_groups_0 = const()[name = string("input_1195_groups_0"), val = int32(1)]; + tensor encoder_layers_22_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(491106560))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(493203776))))[name = string("encoder_layers_22_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_1193_cast_fp16 = transpose(perm = input_1193_perm_0, x = x_589_cast_fp16)[name = string("transpose_159")]; + tensor input_1195_cast_fp16 = conv(dilations = input_1195_dilations_0, groups = input_1195_groups_0, pad = input_1195_pad_0, pad_type = input_1195_pad_type_0, strides = input_1195_strides_0, weight = encoder_layers_22_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_1193_cast_fp16)[name = string("input_1195_cast_fp16")]; + int32 x_591_split_num_splits_0 = const()[name = string("x_591_split_num_splits_0"), val = int32(2)]; + int32 x_591_split_axis_0 = const()[name = string("x_591_split_axis_0"), val = int32(1)]; + tensor x_591_split_cast_fp16_0, tensor x_591_split_cast_fp16_1 = split(axis = x_591_split_axis_0, num_splits = x_591_split_num_splits_0, x = input_1195_cast_fp16)[name = string("x_591_split_cast_fp16")]; + tensor x_591_split_1_sigmoid_cast_fp16 = sigmoid(x = x_591_split_cast_fp16_1)[name = string("x_591_split_1_sigmoid_cast_fp16")]; + tensor x_591_cast_fp16 = mul(x = x_591_split_cast_fp16_0, y = x_591_split_1_sigmoid_cast_fp16)[name = string("x_591_cast_fp16")]; + tensor input_1197_cast_fp16 = select(a = var_44_to_fp16, b = x_591_cast_fp16, cond = var_575)[name = string("input_1197_cast_fp16")]; + bool new_x_91_interleave_0 = const()[name = string("new_x_91_interleave_0"), val = bool(false)]; + tensor new_x_91_cast_fp16 = concat(axis = var_59, interleave = new_x_91_interleave_0, values = (cache_91_cast_fp16, input_1197_cast_fp16))[name = string("new_x_91_cast_fp16")]; + tensor var_5274_begin_0 = const()[name = string("op_5274_begin_0"), val = tensor([0, 0, 28])]; + tensor var_5274_end_0 = const()[name = string("op_5274_end_0"), val = tensor([1, 1024, 36])]; + tensor var_5274_end_mask_0 = const()[name = string("op_5274_end_mask_0"), val = tensor([true, true, true])]; + tensor var_5274_cast_fp16 = slice_by_index(begin = var_5274_begin_0, end = var_5274_end_0, end_mask = var_5274_end_mask_0, x = new_x_91_cast_fp16)[name = string("op_5274_cast_fp16")]; + string x_593_pad_type_0 = const()[name = string("x_593_pad_type_0"), val = string("valid")]; + int32 x_593_groups_0 = const()[name = string("x_593_groups_0"), val = int32(1024)]; + tensor x_593_strides_0 = const()[name = string("x_593_strides_0"), val = tensor([1])]; + tensor x_593_pad_0 = const()[name = string("x_593_pad_0"), val = tensor([0, 0])]; + tensor x_593_dilations_0 = const()[name = string("x_593_dilations_0"), val = tensor([1])]; + tensor encoder_layers_22_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(493207936))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(493217216))))[name = string("encoder_layers_22_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_593_cast_fp16 = conv(dilations = x_593_dilations_0, groups = x_593_groups_0, pad = x_593_pad_0, pad_type = x_593_pad_type_0, strides = x_593_strides_0, weight = encoder_layers_22_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_91_cast_fp16)[name = string("x_593_cast_fp16")]; + tensor input_1199_perm_0 = const()[name = string("input_1199_perm_0"), val = tensor([0, 2, 1])]; + tensor x_595_axes_0 = const()[name = string("x_595_axes_0"), val = tensor([-1])]; + tensor encoder_layers_22_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_22_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(493219328)))]; + tensor encoder_layers_22_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_22_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(493221440)))]; + tensor input_1199_cast_fp16 = transpose(perm = input_1199_perm_0, x = x_593_cast_fp16)[name = string("transpose_158")]; + tensor x_595_cast_fp16 = layer_norm(axes = x_595_axes_0, beta = encoder_layers_22_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_22_conv_batch_norm_weight_to_fp16, x = input_1199_cast_fp16)[name = string("x_595_cast_fp16")]; + tensor input_1201_perm_0 = const()[name = string("input_1201_perm_0"), val = tensor([0, 2, 1])]; + tensor input_1201_cast_fp16 = transpose(perm = input_1201_perm_0, x = x_595_cast_fp16)[name = string("transpose_157")]; + tensor input_1203_cast_fp16 = silu(x = input_1201_cast_fp16)[name = string("input_1203_cast_fp16")]; + string x_597_pad_type_0 = const()[name = string("x_597_pad_type_0"), val = string("valid")]; + tensor x_597_strides_0 = const()[name = string("x_597_strides_0"), val = tensor([1])]; + tensor x_597_pad_0 = const()[name = string("x_597_pad_0"), val = tensor([0, 0])]; + tensor x_597_dilations_0 = const()[name = string("x_597_dilations_0"), val = tensor([1])]; + int32 x_597_groups_0 = const()[name = string("x_597_groups_0"), val = int32(1)]; + tensor encoder_layers_22_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(493223552))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(494272192))))[name = string("encoder_layers_22_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_597_cast_fp16 = conv(dilations = x_597_dilations_0, groups = x_597_groups_0, pad = x_597_pad_0, pad_type = x_597_pad_type_0, strides = x_597_strides_0, weight = encoder_layers_22_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_1203_cast_fp16)[name = string("x_597_cast_fp16")]; + tensor input_1205_perm_0 = const()[name = string("input_1205_perm_0"), val = tensor([0, 2, 1])]; + tensor input_1205_cast_fp16 = transpose(perm = input_1205_perm_0, x = x_597_cast_fp16)[name = string("transpose_156")]; + tensor input_1207_cast_fp16 = add(x = input_1191_cast_fp16, y = input_1205_cast_fp16)[name = string("input_1207_cast_fp16")]; + tensor input_1209_axes_0 = const()[name = string("input_1209_axes_0"), val = tensor([-1])]; + tensor encoder_layers_22_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_22_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(494274304)))]; + tensor encoder_layers_22_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_22_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(494276416)))]; + tensor input_1209_cast_fp16 = layer_norm(axes = input_1209_axes_0, beta = encoder_layers_22_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_22_norm_feed_forward2_weight_to_fp16, x = input_1207_cast_fp16)[name = string("input_1209_cast_fp16")]; + tensor encoder_layers_22_feed_forward2_linear1_weight_to_fp16 = const()[name = string("encoder_layers_22_feed_forward2_linear1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(494278528)))]; + tensor encoder_layers_22_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_22_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(502667200)))]; + tensor linear_206_cast_fp16 = linear(bias = encoder_layers_22_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_22_feed_forward2_linear1_weight_to_fp16, x = input_1209_cast_fp16)[name = string("linear_206_cast_fp16")]; + tensor input_1213_cast_fp16 = silu(x = linear_206_cast_fp16)[name = string("input_1213_cast_fp16")]; + tensor encoder_layers_22_feed_forward2_linear2_weight_to_fp16 = const()[name = string("encoder_layers_22_feed_forward2_linear2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(502675456)))]; + tensor encoder_layers_22_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_22_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(511064128)))]; + tensor linear_207_cast_fp16 = linear(bias = encoder_layers_22_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_22_feed_forward2_linear2_weight_to_fp16, x = input_1213_cast_fp16)[name = string("linear_207_cast_fp16")]; + fp16 var_5317_to_fp16 = const()[name = string("op_5317_to_fp16"), val = fp16(0x1p-1)]; + tensor var_5318_cast_fp16 = mul(x = linear_207_cast_fp16, y = var_5317_to_fp16)[name = string("op_5318_cast_fp16")]; + tensor input_1219_cast_fp16 = add(x = input_1207_cast_fp16, y = var_5318_cast_fp16)[name = string("input_1219_cast_fp16")]; + tensor input_1221_axes_0 = const()[name = string("input_1221_axes_0"), val = tensor([-1])]; + tensor encoder_layers_22_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_22_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(511066240)))]; + tensor encoder_layers_22_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_22_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(511068352)))]; + tensor input_1221_cast_fp16 = layer_norm(axes = input_1221_axes_0, beta = encoder_layers_22_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_22_norm_out_weight_to_fp16, x = input_1219_cast_fp16)[name = string("input_1221_cast_fp16")]; + tensor cache_93_begin_0 = const()[name = string("cache_93_begin_0"), val = tensor([23, 0, 0, 0])]; + tensor cache_93_end_0 = const()[name = string("cache_93_end_0"), val = tensor([24, 1, 42, 1024])]; + tensor cache_93_end_mask_0 = const()[name = string("cache_93_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_93_squeeze_mask_0 = const()[name = string("cache_93_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_93_cast_fp16 = slice_by_index(begin = cache_93_begin_0, end = cache_93_end_0, end_mask = cache_93_end_mask_0, squeeze_mask = cache_93_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_93_cast_fp16")]; + tensor cache_begin_0 = const()[name = string("cache_begin_0"), val = tensor([23, 0, 0, 0])]; + tensor cache_end_0 = const()[name = string("cache_end_0"), val = tensor([24, 1, 1024, 8])]; + tensor cache_end_mask_0 = const()[name = string("cache_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_squeeze_mask_0 = const()[name = string("cache_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_cast_fp16 = slice_by_index(begin = cache_begin_0, end = cache_end_0, end_mask = cache_end_mask_0, squeeze_mask = cache_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_cast_fp16")]; + tensor input_1223_axes_0 = const()[name = string("input_1223_axes_0"), val = tensor([-1])]; + tensor encoder_layers_23_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_23_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(511070464)))]; + tensor encoder_layers_23_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_23_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(511072576)))]; + tensor input_1223_cast_fp16 = layer_norm(axes = input_1223_axes_0, beta = encoder_layers_23_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_23_norm_feed_forward1_weight_to_fp16, x = input_1221_cast_fp16)[name = string("input_1223_cast_fp16")]; + tensor encoder_layers_23_feed_forward1_linear1_weight_to_fp16 = const()[name = string("encoder_layers_23_feed_forward1_linear1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(511074688)))]; + tensor encoder_layers_23_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_23_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(519463360)))]; + tensor linear_208_cast_fp16 = linear(bias = encoder_layers_23_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_23_feed_forward1_linear1_weight_to_fp16, x = input_1223_cast_fp16)[name = string("linear_208_cast_fp16")]; + tensor input_1227_cast_fp16 = silu(x = linear_208_cast_fp16)[name = string("input_1227_cast_fp16")]; + tensor encoder_layers_23_feed_forward1_linear2_weight_to_fp16 = const()[name = string("encoder_layers_23_feed_forward1_linear2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(519471616)))]; + tensor encoder_layers_23_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_23_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(527860288)))]; + tensor linear_209_cast_fp16 = linear(bias = encoder_layers_23_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_23_feed_forward1_linear2_weight_to_fp16, x = input_1227_cast_fp16)[name = string("linear_209_cast_fp16")]; + fp16 var_5354_to_fp16 = const()[name = string("op_5354_to_fp16"), val = fp16(0x1p-1)]; + tensor var_5355_cast_fp16 = mul(x = linear_209_cast_fp16, y = var_5354_to_fp16)[name = string("op_5355_cast_fp16")]; + tensor input_1233_cast_fp16 = add(x = input_1221_cast_fp16, y = var_5355_cast_fp16)[name = string("input_1233_cast_fp16")]; + tensor key_axes_0 = const()[name = string("key_axes_0"), val = tensor([-1])]; + tensor encoder_layers_23_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_23_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(527862400)))]; + tensor encoder_layers_23_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_23_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(527864512)))]; + tensor key_cast_fp16 = layer_norm(axes = key_axes_0, beta = encoder_layers_23_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_23_norm_self_att_weight_to_fp16, x = input_1233_cast_fp16)[name = string("key_cast_fp16")]; + bool input_1235_interleave_0 = const()[name = string("input_1235_interleave_0"), val = bool(false)]; + tensor input_1235_cast_fp16 = concat(axis = var_68, interleave = input_1235_interleave_0, values = (cache_93_cast_fp16, key_cast_fp16))[name = string("input_1235_cast_fp16")]; + tensor var_5377_begin_0 = const()[name = string("op_5377_begin_0"), val = tensor([0, 28, 0])]; + tensor var_5377_end_0 = const()[name = string("op_5377_end_0"), val = tensor([1, 42, 1024])]; + tensor var_5377_end_mask_0 = const()[name = string("op_5377_end_mask_0"), val = tensor([true, true, true])]; + tensor var_5377_cast_fp16 = slice_by_index(begin = var_5377_begin_0, end = var_5377_end_0, end_mask = var_5377_end_mask_0, x = cache_93_cast_fp16)[name = string("op_5377_cast_fp16")]; + bool cache_last_channel_cur_interleave_0 = const()[name = string("cache_last_channel_cur_interleave_0"), val = bool(false)]; + tensor cache_last_channel_cur_cast_fp16 = concat(axis = var_68, interleave = cache_last_channel_cur_interleave_0, values = (var_5377_cast_fp16, key_cast_fp16))[name = string("cache_last_channel_cur_cast_fp16")]; + tensor encoder_layers_23_self_attn_linear_q_weight_to_fp16 = const()[name = string("encoder_layers_23_self_attn_linear_q_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(527866624)))]; + tensor encoder_layers_23_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_23_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(529963840)))]; + tensor linear_210_cast_fp16 = linear(bias = encoder_layers_23_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_23_self_attn_linear_q_weight_to_fp16, x = key_cast_fp16)[name = string("linear_210_cast_fp16")]; + tensor var_5388 = const()[name = string("op_5388"), val = tensor([1, -1, 8, 128])]; + tensor q_139_cast_fp16 = reshape(shape = var_5388, x = linear_210_cast_fp16)[name = string("q_139_cast_fp16")]; + tensor encoder_layers_23_self_attn_linear_k_weight_to_fp16 = const()[name = string("encoder_layers_23_self_attn_linear_k_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(529965952)))]; + tensor encoder_layers_23_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_23_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(532063168)))]; + tensor linear_211_cast_fp16 = linear(bias = encoder_layers_23_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_23_self_attn_linear_k_weight_to_fp16, x = input_1235_cast_fp16)[name = string("linear_211_cast_fp16")]; + tensor var_5393 = const()[name = string("op_5393"), val = tensor([1, -1, 8, 128])]; + tensor k_93_cast_fp16 = reshape(shape = var_5393, x = linear_211_cast_fp16)[name = string("k_93_cast_fp16")]; + tensor encoder_layers_23_self_attn_linear_v_weight_to_fp16 = const()[name = string("encoder_layers_23_self_attn_linear_v_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(532065280)))]; + tensor encoder_layers_23_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_23_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(534162496)))]; + tensor linear_212_cast_fp16 = linear(bias = encoder_layers_23_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_23_self_attn_linear_v_weight_to_fp16, x = input_1235_cast_fp16)[name = string("linear_212_cast_fp16")]; + tensor var_5398 = const()[name = string("op_5398"), val = tensor([1, -1, 8, 128])]; + tensor v_cast_fp16 = reshape(shape = var_5398, x = linear_212_cast_fp16)[name = string("v_cast_fp16")]; + tensor value_perm_0 = const()[name = string("value_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_23_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_23_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(534164608)))]; + tensor var_5411_cast_fp16 = add(x = q_139_cast_fp16, y = encoder_layers_23_self_attn_pos_bias_u_to_fp16)[name = string("op_5411_cast_fp16")]; + tensor encoder_layers_23_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_23_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(534166720)))]; + tensor var_5413_cast_fp16 = add(x = q_139_cast_fp16, y = encoder_layers_23_self_attn_pos_bias_v_to_fp16)[name = string("op_5413_cast_fp16")]; + tensor q_with_bias_v_perm_0 = const()[name = string("q_with_bias_v_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_605_transpose_x_0 = const()[name = string("x_605_transpose_x_0"), val = bool(false)]; + bool x_605_transpose_y_0 = const()[name = string("x_605_transpose_y_0"), val = bool(false)]; + tensor op_5415_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(534168832))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(534311232))))[name = string("op_5415_to_fp16_quantized")]; + tensor q_with_bias_v_cast_fp16 = transpose(perm = q_with_bias_v_perm_0, x = var_5413_cast_fp16)[name = string("transpose_155")]; + tensor x_605_cast_fp16 = matmul(transpose_x = x_605_transpose_x_0, transpose_y = x_605_transpose_y_0, x = q_with_bias_v_cast_fp16, y = op_5415_to_fp16_quantized)[name = string("x_605_cast_fp16")]; + tensor x_607_pad_0 = const()[name = string("x_607_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_607_mode_0 = const()[name = string("x_607_mode_0"), val = string("constant")]; + fp16 const_378_to_fp16 = const()[name = string("const_378_to_fp16"), val = fp16(0x0p+0)]; + tensor x_607_cast_fp16 = pad(constant_val = const_378_to_fp16, mode = x_607_mode_0, pad = x_607_pad_0, x = x_605_cast_fp16)[name = string("x_607_cast_fp16")]; + tensor var_5423 = const()[name = string("op_5423"), val = tensor([1, 8, -1, 28])]; + tensor x_609_cast_fp16 = reshape(shape = var_5423, x = x_607_cast_fp16)[name = string("x_609_cast_fp16")]; + tensor var_5427_begin_0 = const()[name = string("op_5427_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_5427_end_0 = const()[name = string("op_5427_end_0"), val = tensor([1, 8, 140, 28])]; + tensor var_5427_end_mask_0 = const()[name = string("op_5427_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_5427_cast_fp16 = slice_by_index(begin = var_5427_begin_0, end = var_5427_end_0, end_mask = var_5427_end_mask_0, x = x_609_cast_fp16)[name = string("op_5427_cast_fp16")]; + tensor var_5428 = const()[name = string("op_5428"), val = tensor([1, 8, 28, 139])]; + tensor matrix_bd_93_cast_fp16 = reshape(shape = var_5428, x = var_5427_cast_fp16)[name = string("matrix_bd_93_cast_fp16")]; + bool matrix_ac_transpose_x_0 = const()[name = string("matrix_ac_transpose_x_0"), val = bool(false)]; + bool matrix_ac_transpose_y_0 = const()[name = string("matrix_ac_transpose_y_0"), val = bool(false)]; + tensor transpose_142_perm_0 = const()[name = string("transpose_142_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_143_perm_0 = const()[name = string("transpose_143_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_143 = transpose(perm = transpose_143_perm_0, x = k_93_cast_fp16)[name = string("transpose_153")]; + tensor transpose_142 = transpose(perm = transpose_142_perm_0, x = var_5411_cast_fp16)[name = string("transpose_154")]; + tensor matrix_ac_cast_fp16 = matmul(transpose_x = matrix_ac_transpose_x_0, transpose_y = matrix_ac_transpose_y_0, x = transpose_142, y = transpose_143)[name = string("matrix_ac_cast_fp16")]; + tensor matrix_bd_begin_0 = const()[name = string("matrix_bd_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_end_0 = const()[name = string("matrix_bd_end_0"), val = tensor([1, 8, 28, 70])]; + tensor matrix_bd_end_mask_0 = const()[name = string("matrix_bd_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_cast_fp16 = slice_by_index(begin = matrix_bd_begin_0, end = matrix_bd_end_0, end_mask = matrix_bd_end_mask_0, x = matrix_bd_93_cast_fp16)[name = string("matrix_bd_cast_fp16")]; + tensor var_5437_cast_fp16 = add(x = matrix_ac_cast_fp16, y = matrix_bd_cast_fp16)[name = string("op_5437_cast_fp16")]; + fp16 _inversed_scores_93_y_0_to_fp16 = const()[name = string("_inversed_scores_93_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_93_cast_fp16 = mul(x = var_5437_cast_fp16, y = _inversed_scores_93_y_0_to_fp16)[name = string("_inversed_scores_93_cast_fp16")]; + tensor scores_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_93_cast_fp16, cond = mask_11)[name = string("scores_cast_fp16")]; + tensor var_5443_cast_fp16 = softmax(axis = var_59, x = scores_cast_fp16)[name = string("op_5443_cast_fp16")]; + tensor input_1237_cast_fp16 = select(a = var_44_to_fp16, b = var_5443_cast_fp16, cond = mask_11)[name = string("input_1237_cast_fp16")]; + bool x_611_transpose_x_0 = const()[name = string("x_611_transpose_x_0"), val = bool(false)]; + bool x_611_transpose_y_0 = const()[name = string("x_611_transpose_y_0"), val = bool(false)]; + tensor value_cast_fp16 = transpose(perm = value_perm_0, x = v_cast_fp16)[name = string("transpose_152")]; + tensor x_611_cast_fp16 = matmul(transpose_x = x_611_transpose_x_0, transpose_y = x_611_transpose_y_0, x = input_1237_cast_fp16, y = value_cast_fp16)[name = string("x_611_cast_fp16")]; + tensor var_5447_perm_0 = const()[name = string("op_5447_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_5448 = const()[name = string("op_5448"), val = tensor([1, -1, 1024])]; + tensor var_5447_cast_fp16 = transpose(perm = var_5447_perm_0, x = x_611_cast_fp16)[name = string("transpose_151")]; + tensor input_1239_cast_fp16 = reshape(shape = var_5448, x = var_5447_cast_fp16)[name = string("input_1239_cast_fp16")]; + tensor encoder_layers_23_self_attn_linear_out_weight_to_fp16 = const()[name = string("encoder_layers_23_self_attn_linear_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(534311616)))]; + tensor encoder_layers_23_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_23_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(536408832)))]; + tensor linear_214_cast_fp16 = linear(bias = encoder_layers_23_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_23_self_attn_linear_out_weight_to_fp16, x = input_1239_cast_fp16)[name = string("linear_214_cast_fp16")]; + tensor input_1243_cast_fp16 = add(x = input_1233_cast_fp16, y = linear_214_cast_fp16)[name = string("input_1243_cast_fp16")]; + tensor x_615_axes_0 = const()[name = string("x_615_axes_0"), val = tensor([-1])]; + tensor encoder_layers_23_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_23_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(536410944)))]; + tensor encoder_layers_23_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_23_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(536413056)))]; + tensor x_615_cast_fp16 = layer_norm(axes = x_615_axes_0, beta = encoder_layers_23_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_23_norm_conv_weight_to_fp16, x = input_1243_cast_fp16)[name = string("x_615_cast_fp16")]; + tensor input_1245_perm_0 = const()[name = string("input_1245_perm_0"), val = tensor([0, 2, 1])]; + string input_1247_pad_type_0 = const()[name = string("input_1247_pad_type_0"), val = string("valid")]; + tensor input_1247_strides_0 = const()[name = string("input_1247_strides_0"), val = tensor([1])]; + tensor input_1247_pad_0 = const()[name = string("input_1247_pad_0"), val = tensor([0, 0])]; + tensor input_1247_dilations_0 = const()[name = string("input_1247_dilations_0"), val = tensor([1])]; + int32 input_1247_groups_0 = const()[name = string("input_1247_groups_0"), val = int32(1)]; + tensor encoder_layers_23_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(536415168))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(538512384))))[name = string("encoder_layers_23_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_1245_cast_fp16 = transpose(perm = input_1245_perm_0, x = x_615_cast_fp16)[name = string("transpose_150")]; + tensor input_1247_cast_fp16 = conv(dilations = input_1247_dilations_0, groups = input_1247_groups_0, pad = input_1247_pad_0, pad_type = input_1247_pad_type_0, strides = input_1247_strides_0, weight = encoder_layers_23_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_1245_cast_fp16)[name = string("input_1247_cast_fp16")]; + int32 x_617_split_num_splits_0 = const()[name = string("x_617_split_num_splits_0"), val = int32(2)]; + int32 x_617_split_axis_0 = const()[name = string("x_617_split_axis_0"), val = int32(1)]; + tensor x_617_split_cast_fp16_0, tensor x_617_split_cast_fp16_1 = split(axis = x_617_split_axis_0, num_splits = x_617_split_num_splits_0, x = input_1247_cast_fp16)[name = string("x_617_split_cast_fp16")]; + tensor x_617_split_1_sigmoid_cast_fp16 = sigmoid(x = x_617_split_cast_fp16_1)[name = string("x_617_split_1_sigmoid_cast_fp16")]; + tensor x_617_cast_fp16 = mul(x = x_617_split_cast_fp16_0, y = x_617_split_1_sigmoid_cast_fp16)[name = string("x_617_cast_fp16")]; + tensor input_1249_cast_fp16 = select(a = var_44_to_fp16, b = x_617_cast_fp16, cond = var_575)[name = string("input_1249_cast_fp16")]; + bool new_x_interleave_0 = const()[name = string("new_x_interleave_0"), val = bool(false)]; + tensor new_x_cast_fp16 = concat(axis = var_59, interleave = new_x_interleave_0, values = (cache_cast_fp16, input_1249_cast_fp16))[name = string("new_x_cast_fp16")]; + tensor cache_last_time_cur_begin_0 = const()[name = string("cache_last_time_cur_begin_0"), val = tensor([0, 0, 28])]; + tensor cache_last_time_cur_end_0 = const()[name = string("cache_last_time_cur_end_0"), val = tensor([1, 1024, 36])]; + tensor cache_last_time_cur_end_mask_0 = const()[name = string("cache_last_time_cur_end_mask_0"), val = tensor([true, true, true])]; + tensor cache_last_time_cur_cast_fp16 = slice_by_index(begin = cache_last_time_cur_begin_0, end = cache_last_time_cur_end_0, end_mask = cache_last_time_cur_end_mask_0, x = new_x_cast_fp16)[name = string("cache_last_time_cur_cast_fp16")]; + string x_619_pad_type_0 = const()[name = string("x_619_pad_type_0"), val = string("valid")]; + int32 x_619_groups_0 = const()[name = string("x_619_groups_0"), val = int32(1024)]; + tensor x_619_strides_0 = const()[name = string("x_619_strides_0"), val = tensor([1])]; + tensor x_619_pad_0 = const()[name = string("x_619_pad_0"), val = tensor([0, 0])]; + tensor x_619_dilations_0 = const()[name = string("x_619_dilations_0"), val = tensor([1])]; + tensor encoder_layers_23_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(538516544))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(538525824))))[name = string("encoder_layers_23_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_619_cast_fp16 = conv(dilations = x_619_dilations_0, groups = x_619_groups_0, pad = x_619_pad_0, pad_type = x_619_pad_type_0, strides = x_619_strides_0, weight = encoder_layers_23_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_cast_fp16)[name = string("x_619_cast_fp16")]; + tensor input_1251_perm_0 = const()[name = string("input_1251_perm_0"), val = tensor([0, 2, 1])]; + tensor x_621_axes_0 = const()[name = string("x_621_axes_0"), val = tensor([-1])]; + tensor encoder_layers_23_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_23_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(538527936)))]; + tensor encoder_layers_23_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_23_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(538530048)))]; + tensor input_1251_cast_fp16 = transpose(perm = input_1251_perm_0, x = x_619_cast_fp16)[name = string("transpose_149")]; + tensor x_621_cast_fp16 = layer_norm(axes = x_621_axes_0, beta = encoder_layers_23_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_23_conv_batch_norm_weight_to_fp16, x = input_1251_cast_fp16)[name = string("x_621_cast_fp16")]; + tensor input_1253_perm_0 = const()[name = string("input_1253_perm_0"), val = tensor([0, 2, 1])]; + tensor input_1253_cast_fp16 = transpose(perm = input_1253_perm_0, x = x_621_cast_fp16)[name = string("transpose_148")]; + tensor input_1255_cast_fp16 = silu(x = input_1253_cast_fp16)[name = string("input_1255_cast_fp16")]; + string x_623_pad_type_0 = const()[name = string("x_623_pad_type_0"), val = string("valid")]; + tensor x_623_strides_0 = const()[name = string("x_623_strides_0"), val = tensor([1])]; + tensor x_623_pad_0 = const()[name = string("x_623_pad_0"), val = tensor([0, 0])]; + tensor x_623_dilations_0 = const()[name = string("x_623_dilations_0"), val = tensor([1])]; + int32 x_623_groups_0 = const()[name = string("x_623_groups_0"), val = int32(1)]; + tensor encoder_layers_23_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(538532160))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(539580800))))[name = string("encoder_layers_23_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_623_cast_fp16 = conv(dilations = x_623_dilations_0, groups = x_623_groups_0, pad = x_623_pad_0, pad_type = x_623_pad_type_0, strides = x_623_strides_0, weight = encoder_layers_23_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_1255_cast_fp16)[name = string("x_623_cast_fp16")]; + tensor input_1257_perm_0 = const()[name = string("input_1257_perm_0"), val = tensor([0, 2, 1])]; + tensor input_1257_cast_fp16 = transpose(perm = input_1257_perm_0, x = x_623_cast_fp16)[name = string("transpose_147")]; + tensor input_1259_cast_fp16 = add(x = input_1243_cast_fp16, y = input_1257_cast_fp16)[name = string("input_1259_cast_fp16")]; + tensor input_1261_axes_0 = const()[name = string("input_1261_axes_0"), val = tensor([-1])]; + tensor encoder_layers_23_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_23_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(539582912)))]; + tensor encoder_layers_23_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_23_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(539585024)))]; + tensor input_1261_cast_fp16 = layer_norm(axes = input_1261_axes_0, beta = encoder_layers_23_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_23_norm_feed_forward2_weight_to_fp16, x = input_1259_cast_fp16)[name = string("input_1261_cast_fp16")]; + tensor encoder_layers_23_feed_forward2_linear1_weight_to_fp16 = const()[name = string("encoder_layers_23_feed_forward2_linear1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(539587136)))]; + tensor encoder_layers_23_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_23_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(547975808)))]; + tensor linear_215_cast_fp16 = linear(bias = encoder_layers_23_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_23_feed_forward2_linear1_weight_to_fp16, x = input_1261_cast_fp16)[name = string("linear_215_cast_fp16")]; + tensor input_1265_cast_fp16 = silu(x = linear_215_cast_fp16)[name = string("input_1265_cast_fp16")]; + tensor encoder_layers_23_feed_forward2_linear2_weight_to_fp16 = const()[name = string("encoder_layers_23_feed_forward2_linear2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(547984064)))]; + tensor encoder_layers_23_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_23_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(556372736)))]; + tensor linear_216_cast_fp16 = linear(bias = encoder_layers_23_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_23_feed_forward2_linear2_weight_to_fp16, x = input_1265_cast_fp16)[name = string("linear_216_cast_fp16")]; + fp16 var_5530_to_fp16 = const()[name = string("op_5530_to_fp16"), val = fp16(0x1p-1)]; + tensor var_5531_cast_fp16 = mul(x = linear_216_cast_fp16, y = var_5530_to_fp16)[name = string("op_5531_cast_fp16")]; + tensor input_1271_cast_fp16 = add(x = input_1259_cast_fp16, y = var_5531_cast_fp16)[name = string("input_1271_cast_fp16")]; + tensor audio_signal_axes_0 = const()[name = string("audio_signal_axes_0"), val = tensor([-1])]; + tensor encoder_layers_23_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_23_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(556374848)))]; + tensor encoder_layers_23_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_23_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(556376960)))]; + tensor audio_signal_cast_fp16 = layer_norm(axes = audio_signal_axes_0, beta = encoder_layers_23_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_23_norm_out_weight_to_fp16, x = input_1271_cast_fp16)[name = string("audio_signal_cast_fp16")]; + int32 obj_5_axis_0 = const()[name = string("obj_5_axis_0"), val = int32(0)]; + tensor obj_5_cast_fp16 = stack(axis = obj_5_axis_0, values = (var_484_cast_fp16, var_697_cast_fp16, var_910_cast_fp16, var_1123_cast_fp16, var_1336_cast_fp16, var_1549_cast_fp16, var_1762_cast_fp16, var_1975_cast_fp16, var_2188_cast_fp16, var_2401_cast_fp16, var_2614_cast_fp16, var_2827_cast_fp16, var_3040_cast_fp16, var_3253_cast_fp16, var_3466_cast_fp16, var_3679_cast_fp16, var_3892_cast_fp16, var_4105_cast_fp16, var_4318_cast_fp16, var_4531_cast_fp16, var_4744_cast_fp16, var_4957_cast_fp16, var_5170_cast_fp16, cache_last_channel_cur_cast_fp16))[name = string("obj_5_cast_fp16")]; + int32 obj_7_axis_0 = const()[name = string("obj_7_axis_0"), val = int32(0)]; + tensor obj_7_cast_fp16 = stack(axis = obj_7_axis_0, values = (var_588_cast_fp16, var_801_cast_fp16, var_1014_cast_fp16, var_1227_cast_fp16, var_1440_cast_fp16, var_1653_cast_fp16, var_1866_cast_fp16, var_2079_cast_fp16, var_2292_cast_fp16, var_2505_cast_fp16, var_2718_cast_fp16, var_2931_cast_fp16, var_3144_cast_fp16, var_3357_cast_fp16, var_3570_cast_fp16, var_3783_cast_fp16, var_3996_cast_fp16, var_4209_cast_fp16, var_4422_cast_fp16, var_4635_cast_fp16, var_4848_cast_fp16, var_5061_cast_fp16, var_5274_cast_fp16, cache_last_time_cur_cast_fp16))[name = string("obj_7_cast_fp16")]; + tensor var_5547 = add(x = cache_len, y = max_audio_length_1)[name = string("op_5547")]; + string var_5547_promoted_to_fp16_dtype_0 = const()[name = string("op_5547_promoted_to_fp16_dtype_0"), val = string("fp16")]; + fp16 const_384_to_fp16 = const()[name = string("const_384_to_fp16"), val = fp16(-inf)]; + fp16 var_49_promoted_to_fp16 = const()[name = string("op_49_promoted_to_fp16"), val = fp16(0x1.5p+5)]; + tensor var_5547_to_fp16 = cast(dtype = var_5547_promoted_to_fp16_dtype_0, x = var_5547)[name = string("cast_9")]; + tensor clip_1_cast_fp16 = clip(alpha = const_384_to_fp16, beta = var_49_promoted_to_fp16, x = var_5547_to_fp16)[name = string("clip_1_cast_fp16")]; + int32 one_hot_1_batch_dims_0 = const()[name = string("one_hot_1_batch_dims_0"), val = int32(0)]; + bool one_hot_1_validate_indices_0 = const()[name = string("one_hot_1_validate_indices_0"), val = bool(false)]; + tensor to_onehot_identity_table_to_fp16 = const()[name = string("to_onehot_identity_table_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(556379072)))]; + string prompt_id_to_int16_dtype_0 = const()[name = string("prompt_id_to_int16_dtype_0"), val = string("int16")]; + string cast_230_dtype_0 = const()[name = string("cast_230_dtype_0"), val = string("int32")]; + int32 greater_equal_0_y_0 = const()[name = string("greater_equal_0_y_0"), val = int32(0)]; + tensor prompt_id_to_int16 = cast(dtype = prompt_id_to_int16_dtype_0, x = prompt_id)[name = string("cast_8")]; + tensor cast_230 = cast(dtype = cast_230_dtype_0, x = prompt_id_to_int16)[name = string("cast_7")]; + tensor greater_equal_0 = greater_equal(x = cast_230, y = greater_equal_0_y_0)[name = string("greater_equal_0")]; + int32 slice_by_index_2 = const()[name = string("slice_by_index_2"), val = int32(128)]; + tensor add_0 = add(x = cast_230, y = slice_by_index_2)[name = string("add_0")]; + tensor select_0 = select(a = cast_230, b = add_0, cond = greater_equal_0)[name = string("select_0")]; + string select_0_to_int16_dtype_0 = const()[name = string("select_0_to_int16_dtype_0"), val = string("int16")]; + string cast_0_dtype_0 = const()[name = string("cast_0_dtype_0"), val = string("int32")]; + int32 greater_equal_0_y_0_1 = const()[name = string("greater_equal_0_y_0_1"), val = int32(0)]; + tensor select_0_to_int16 = cast(dtype = select_0_to_int16_dtype_0, x = select_0)[name = string("cast_6")]; + tensor cast_0 = cast(dtype = cast_0_dtype_0, x = select_0_to_int16)[name = string("cast_5")]; + tensor greater_equal_0_1 = greater_equal(x = cast_0, y = greater_equal_0_y_0_1)[name = string("greater_equal_0_1")]; + int32 slice_by_index_0 = const()[name = string("slice_by_index_0"), val = int32(128)]; + tensor add_0_1 = add(x = cast_0, y = slice_by_index_0)[name = string("add_0_1")]; + tensor select_0_1 = select(a = cast_0, b = add_0_1, cond = greater_equal_0_1)[name = string("select_0_1")]; + int32 greater_equal_0_y_0_2 = const()[name = string("greater_equal_0_y_0_2"), val = int32(0)]; + tensor greater_equal_0_2 = greater_equal(x = select_0_1, y = greater_equal_0_y_0_2)[name = string("greater_equal_0_2")]; + int32 slice_by_index_0_1 = const()[name = string("slice_by_index_0_1"), val = int32(128)]; + tensor add_0_2 = add(x = select_0_1, y = slice_by_index_0_1)[name = string("add_0_2")]; + tensor select_0_2 = select(a = select_0_1, b = add_0_2, cond = greater_equal_0_2)[name = string("select_0_2")]; + int32 one_hot_1_cast_fp16_cast_uint16_cast_uint16_axis_0 = const()[name = string("one_hot_1_cast_fp16_cast_uint16_cast_uint16_axis_0"), val = int32(0)]; + tensor one_hot_1_cast_fp16_cast_uint16_cast_uint16 = gather(axis = one_hot_1_cast_fp16_cast_uint16_cast_uint16_axis_0, batch_dims = one_hot_1_batch_dims_0, indices = select_0_2, validate_indices = one_hot_1_validate_indices_0, x = to_onehot_identity_table_to_fp16)[name = string("one_hot_1_cast_fp16_cast_uint16_cast_uint16")]; + tensor var_5593_axes_0 = const()[name = string("op_5593_axes_0"), val = tensor([1])]; + tensor var_5593_cast_fp16 = expand_dims(axes = var_5593_axes_0, x = one_hot_1_cast_fp16_cast_uint16_cast_uint16)[name = string("op_5593_cast_fp16")]; + tensor one_hot_reps_0 = const()[name = string("one_hot_reps_0"), val = tensor([1, 28, 1])]; + tensor one_hot_cast_fp16 = tile(reps = one_hot_reps_0, x = var_5593_cast_fp16)[name = string("one_hot_cast_fp16")]; + int32 var_5602 = const()[name = string("op_5602"), val = int32(-1)]; + bool input_1273_interleave_0 = const()[name = string("input_1273_interleave_0"), val = bool(false)]; + tensor input_1273_cast_fp16 = concat(axis = var_5602, interleave = input_1273_interleave_0, values = (audio_signal_cast_fp16, one_hot_cast_fp16))[name = string("input_1273_cast_fp16")]; + tensor prompt_kernel_0_weight_to_fp16 = const()[name = string("prompt_kernel_0_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(556411904)))]; + tensor prompt_kernel_0_bias_to_fp16 = const()[name = string("prompt_kernel_0_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(561130560)))]; + tensor linear_217_cast_fp16 = linear(bias = prompt_kernel_0_bias_to_fp16, weight = prompt_kernel_0_weight_to_fp16, x = input_1273_cast_fp16)[name = string("linear_217_cast_fp16")]; + tensor input_cast_fp16 = relu(x = linear_217_cast_fp16)[name = string("input_cast_fp16")]; + tensor prompt_kernel_2_weight_to_fp16 = const()[name = string("prompt_kernel_2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(561134720)))]; + tensor prompt_kernel_2_bias_to_fp16 = const()[name = string("prompt_kernel_2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(565329088)))]; + tensor linear_218_cast_fp16 = linear(bias = prompt_kernel_2_bias_to_fp16, weight = prompt_kernel_2_weight_to_fp16, x = input_cast_fp16)[name = string("linear_218_cast_fp16")]; + tensor var_5615_perm_0 = const()[name = string("op_5615_perm_0"), val = tensor([0, 2, 1])]; + string var_5615_cast_fp16_to_fp32_dtype_0 = const()[name = string("op_5615_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + string var_5620_dtype_0 = const()[name = string("op_5620_dtype_0"), val = string("int32")]; + tensor var_5623_perm_0 = const()[name = string("op_5623_perm_0"), val = tensor([1, 0, 2, 3])]; + string var_5623_cast_fp16_to_fp32_dtype_0 = const()[name = string("op_5623_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + tensor var_5626_perm_0 = const()[name = string("op_5626_perm_0"), val = tensor([1, 0, 2, 3])]; + string var_5626_cast_fp16_to_fp32_dtype_0 = const()[name = string("op_5626_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + string var_5631_dtype_0 = const()[name = string("op_5631_dtype_0"), val = string("int32")]; + tensor cache_len_out = cast(dtype = var_5631_dtype_0, x = clip_1_cast_fp16)[name = string("cast_0")]; + tensor var_5626_cast_fp16 = transpose(perm = var_5626_perm_0, x = obj_7_cast_fp16)[name = string("transpose_144")]; + tensor cache_time_out = cast(dtype = var_5626_cast_fp16_to_fp32_dtype_0, x = var_5626_cast_fp16)[name = string("cast_1")]; + tensor var_5623_cast_fp16 = transpose(perm = var_5623_perm_0, x = obj_5_cast_fp16)[name = string("transpose_145")]; + tensor cache_channel_out = cast(dtype = var_5623_cast_fp16_to_fp32_dtype_0, x = var_5623_cast_fp16)[name = string("cast_2")]; + tensor encoded_length = cast(dtype = var_5620_dtype_0, x = clip_0_cast_fp16)[name = string("cast_3")]; + tensor var_5615_cast_fp16 = transpose(perm = var_5615_perm_0, x = linear_218_cast_fp16)[name = string("transpose_146")]; + tensor encoded = cast(dtype = var_5615_cast_fp16_to_fp32_dtype_0, x = var_5615_cast_fp16)[name = string("cast_4")]; + } -> (encoded, encoded_length, cache_channel_out, cache_time_out, cache_len_out); +} \ No newline at end of file diff --git a/ja/2240ms/encoder.mlmodelc/weights/weight.bin b/ja/2240ms/encoder.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..7a12bc272fe95f2080e5cebe63d7189329748a69 --- /dev/null +++ b/ja/2240ms/encoder.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:728cf78b85d85c573a2a33a8287ea44bc7a846fc8462d9038d1ed1b24a2c9ac8 +size 565331200 diff --git a/ja/2240ms/encoder.mlpackage/Data/com.apple.CoreML/model.mlmodel b/ja/2240ms/encoder.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..808c1ca295fe918d2a2372f1e2fadb3bc18930d1 --- /dev/null +++ b/ja/2240ms/encoder.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a0653e7c223390a9fe23c5e6f6266b97488e294bbc759705aaa66c68bb15ff8f +size 805589 diff --git a/ja/2240ms/encoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/ja/2240ms/encoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..7a12bc272fe95f2080e5cebe63d7189329748a69 --- /dev/null +++ b/ja/2240ms/encoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:728cf78b85d85c573a2a33a8287ea44bc7a846fc8462d9038d1ed1b24a2c9ac8 +size 565331200 diff --git a/ja/2240ms/encoder.mlpackage/Manifest.json b/ja/2240ms/encoder.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..955535c4db53bb00e2d26147a70036cd887e0451 --- /dev/null +++ b/ja/2240ms/encoder.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "00BBBFAC-20B6-427A-B978-654939D77446": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + }, + "5707DA36-A222-47FA-8727-5AED43CBE002": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + } + }, + "rootModelIdentifier": "5707DA36-A222-47FA-8727-5AED43CBE002" +} diff --git a/ja/2240ms/joint.mlmodelc/analytics/coremldata.bin b/ja/2240ms/joint.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..a7ba5a90671da5c40e03362f44f23df528bc6d93 --- /dev/null +++ b/ja/2240ms/joint.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e342ce20383866520d2c6c860c2bf14d887b9e7fef53606661b41a23ad09472e +size 243 diff --git a/ja/2240ms/joint.mlmodelc/coremldata.bin b/ja/2240ms/joint.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..c211d8e905ed0db72831520a5614821702fb0f96 --- /dev/null +++ b/ja/2240ms/joint.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:78bc08520decca929c09521c2fea47b72335d3294a1642ffa2d797737e4f1b82 +size 401 diff --git a/ja/2240ms/joint.mlmodelc/model.mil b/ja/2240ms/joint.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..827f6cd71b5910ea07d4f6ba43462967d8b86410 --- /dev/null +++ b/ja/2240ms/joint.mlmodelc/model.mil @@ -0,0 +1,31 @@ +program(1.3) +[buildInfo = dict({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.10.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor decoder, tensor encoder) { + tensor input_1_perm_0 = const()[name = string("input_1_perm_0"), val = tensor([0, 2, 1])]; + string encoder_to_fp16_dtype_0 = const()[name = string("encoder_to_fp16_dtype_0"), val = string("fp16")]; + tensor input_3_perm_0 = const()[name = string("input_3_perm_0"), val = tensor([0, 2, 1])]; + string decoder_to_fp16_dtype_0 = const()[name = string("decoder_to_fp16_dtype_0"), val = string("fp16")]; + tensor module_enc_weight_to_fp16 = const()[name = string("module_enc_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))]; + tensor module_enc_bias_to_fp16 = const()[name = string("module_enc_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1310848)))]; + tensor encoder_to_fp16 = cast(dtype = encoder_to_fp16_dtype_0, x = encoder)[name = string("cast_2")]; + tensor input_1_cast_fp16 = transpose(perm = input_1_perm_0, x = encoder_to_fp16)[name = string("transpose_1")]; + tensor linear_0_cast_fp16 = linear(bias = module_enc_bias_to_fp16, weight = module_enc_weight_to_fp16, x = input_1_cast_fp16)[name = string("linear_0_cast_fp16")]; + tensor module_pred_weight_to_fp16 = const()[name = string("module_pred_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1312192)))]; + tensor module_pred_bias_to_fp16 = const()[name = string("module_pred_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2131456)))]; + tensor decoder_to_fp16 = cast(dtype = decoder_to_fp16_dtype_0, x = decoder)[name = string("cast_1")]; + tensor input_3_cast_fp16 = transpose(perm = input_3_perm_0, x = decoder_to_fp16)[name = string("transpose_0")]; + tensor linear_1_cast_fp16 = linear(bias = module_pred_bias_to_fp16, weight = module_pred_weight_to_fp16, x = input_3_cast_fp16)[name = string("linear_1_cast_fp16")]; + tensor var_23_axes_0 = const()[name = string("op_23_axes_0"), val = tensor([2])]; + tensor var_23_cast_fp16 = expand_dims(axes = var_23_axes_0, x = linear_0_cast_fp16)[name = string("op_23_cast_fp16")]; + tensor var_25_axes_0 = const()[name = string("op_25_axes_0"), val = tensor([1])]; + tensor var_25_cast_fp16 = expand_dims(axes = var_25_axes_0, x = linear_1_cast_fp16)[name = string("op_25_cast_fp16")]; + tensor input_5_cast_fp16 = add(x = var_23_cast_fp16, y = var_25_cast_fp16)[name = string("input_5_cast_fp16")]; + tensor input_7_cast_fp16 = relu(x = input_5_cast_fp16)[name = string("input_7_cast_fp16")]; + tensor module_joint_net_2_weight_to_fp16 = const()[name = string("module_joint_net_2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2132800)))]; + tensor module_joint_net_2_bias_to_fp16 = const()[name = string("module_joint_net_2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3929984)))]; + tensor linear_2_cast_fp16 = linear(bias = module_joint_net_2_bias_to_fp16, weight = module_joint_net_2_weight_to_fp16, x = input_7_cast_fp16)[name = string("linear_2_cast_fp16")]; + string linear_2_cast_fp16_to_fp32_dtype_0 = const()[name = string("linear_2_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + tensor logits = cast(dtype = linear_2_cast_fp16_to_fp32_dtype_0, x = linear_2_cast_fp16)[name = string("cast_0")]; + } -> (logits); +} \ No newline at end of file diff --git a/ja/2240ms/joint.mlmodelc/weights/weight.bin b/ja/2240ms/joint.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..31da4412e1214e0b52e77c023e0490150a12e242 --- /dev/null +++ b/ja/2240ms/joint.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:44c15c8be3c89e5d531258b840e560e467084d55ed400453a9d21325757fb111 +size 3932856 diff --git a/ja/2240ms/joint.mlpackage/Data/com.apple.CoreML/model.mlmodel b/ja/2240ms/joint.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..1bc98711fb995d49c835ce43242e33bc518a943d --- /dev/null +++ b/ja/2240ms/joint.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1df49e8dd61d86e19d95b935020748db2eda7ae1273f39988001a30535c5be45 +size 4545 diff --git a/ja/2240ms/joint.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/ja/2240ms/joint.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..31da4412e1214e0b52e77c023e0490150a12e242 --- /dev/null +++ b/ja/2240ms/joint.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:44c15c8be3c89e5d531258b840e560e467084d55ed400453a9d21325757fb111 +size 3932856 diff --git a/ja/2240ms/joint.mlpackage/Manifest.json b/ja/2240ms/joint.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..ef5aca0b538b9de8e541f27a29fff913203df8c7 --- /dev/null +++ b/ja/2240ms/joint.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "8348189F-7CB9-4961-A35F-4049C53D63B6": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + }, + "AA6A8B4F-747E-4EC1-87E1-2B387F1149D8": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + } + }, + "rootModelIdentifier": "8348189F-7CB9-4961-A35F-4049C53D63B6" +} diff --git a/ja/2240ms/metadata.json b/ja/2240ms/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..a6c602a6ea51251c05f93e8a0ae5c3c9de3b54c3 --- /dev/null +++ b/ja/2240ms/metadata.json @@ -0,0 +1,199 @@ +{ + "model": "nvidia/nemotron-asr-streaming-multilingual-0.6b", + "model_class": "nemo.collections.asr.models.rnnt_bpe_models_prompt.EncDecRNNTBPEModelWithPrompt", + "sample_rate": 16000, + "mel_features": 128, + "chunk_mel_frames": 224, + "pre_encode_cache": 9, + "total_mel_frames": 233, + "att_context_size": [ + 42, + 13 + ], + "vocab_size": 1403, + "blank_idx": 1403, + "vocab_pruned": true, + "vocab_pruned_original_size": 13087, + "cache_channel_shape": [ + 1, + 24, + 42, + 1024 + ], + "cache_time_shape": [ + 1, + 24, + 1024, + 8 + ], + "decoder_hidden": 640, + "decoder_layers": 2, + "encoder_dim": 1024, + "num_prompts": 128, + "prompt_dictionary": { + "af-ZA": 54, + "am-ET": 49, + "ar": 7, + "ar-AR": 7, + "auto": 101, + "ay-BO": 81, + "az-AZ": 66, + "bg": 30, + "bg-BG": 30, + "bn-IN": 36, + "cs": 22, + "cs-CZ": 22, + "da": 25, + "da-DK": 25, + "de": 9, + "de-DE": 9, + "el": 21, + "el-GR": 21, + "en": 0, + "en-GB": 1, + "en-US": 0, + "enGB": 1, + "es": 3, + "es-ES": 2, + "es-US": 3, + "esES": 2, + "et": 60, + "et-EE": 60, + "fa-IR": 38, + "fi": 26, + "fi-FI": 26, + "fr": 8, + "fr-CA": 100, + "fr-FR": 8, + "gn-PY": 82, + "gu-IN": 42, + "ha-NG": 50, + "haw-US": 97, + "he-IL": 64, + "hi": 6, + "hi-HI": 6, + "hi-IN": 6, + "hr": 29, + "hr-HR": 29, + "hu": 23, + "hu-HU": 23, + "hy-AM": 68, + "id-ID": 34, + "ig-NG": 53, + "it": 15, + "it-IT": 15, + "ja-JA": 10, + "ja-JP": 10, + "ka-GE": 67, + "km-KH": 47, + "kn-IN": 43, + "ko": 14, + "ko-KO": 14, + "ko-KR": 14, + "ku-TR": 65, + "ky-KG": 71, + "ln-CD": 58, + "lt": 31, + "lt-LT": 31, + "lv": 61, + "lv-LV": 61, + "mi-NZ": 96, + "ml-IN": 44, + "mr-IN": 41, + "ms-MY": 35, + "mt-MT": 102, + "nah-MX": 83, + "nb": 103, + "nb-NO": 103, + "ne-NP": 46, + "nl": 16, + "nl-NL": 16, + "nn": 104, + "nn-NO": 104, + "no": 27, + "no-NO": 27, + "ny-MW": 57, + "or-KE": 59, + "pl": 17, + "pl-PL": 17, + "pt": 13, + "pt-BR": 12, + "pt-PT": 13, + "qu-PE": 80, + "ro": 20, + "ro-RO": 20, + "ru": 11, + "ru-RU": 11, + "rw-RW": 55, + "si-LK": 45, + "sk": 28, + "sk-SK": 28, + "sl": 62, + "sl-SI": 62, + "sm-WS": 98, + "so-SO": 56, + "sv": 24, + "sv-SE": 24, + "sw-KE": 48, + "ta-IN": 39, + "te-IN": 40, + "tg-TJ": 70, + "th-TH": 32, + "to-TO": 99, + "tr": 18, + "tr-TR": 18, + "uk": 19, + "uk-UA": 19, + "ur-PK": 37, + "uz-UZ": 69, + "vi-VN": 33, + "yo-NG": 52, + "zh-CN": 4, + "zh-TW": 5, + "zh-ZH": 4, + "zu-ZA": 51 + }, + "default_prompt_id": 101, + "lang_tag_token_ids": [ + 1, + 52, + 62, + 66, + 69, + 70, + 75, + 76, + 77, + 79, + 81, + 83, + 86, + 88, + 89, + 90, + 92, + 94, + 95, + 96, + 97, + 99, + 100, + 103, + 107, + 109, + 111, + 112, + 114, + 115, + 117, + 1389, + 1390, + 1391, + 1392, + 1393, + 1394, + 1395, + 1402 + ], + "chunk_ms": 2240 +} \ No newline at end of file diff --git a/ja/2240ms/preprocessor.mlmodelc/analytics/coremldata.bin b/ja/2240ms/preprocessor.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..d4b1561ab413a9d87db506bc842f077779dcbded --- /dev/null +++ b/ja/2240ms/preprocessor.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3b8033bec5ee01649f325b8f4c5aeef1b31c99b469ce56d46039c1b73f09585d +size 243 diff --git a/ja/2240ms/preprocessor.mlmodelc/coremldata.bin b/ja/2240ms/preprocessor.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..aae12b21bc72a550074c203b5fbe68672b810d91 --- /dev/null +++ b/ja/2240ms/preprocessor.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:eb85daae5bdb39bd64daab8073f8bb8d8a8b2db735276f9e516142b85046e9c1 +size 431 diff --git a/ja/2240ms/preprocessor.mlmodelc/model.mil b/ja/2240ms/preprocessor.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..b1a0b2b9193c992de42e51245fc1ef433d345afc --- /dev/null +++ b/ja/2240ms/preprocessor.mlmodelc/model.mil @@ -0,0 +1,122 @@ +program(1.3) +[buildInfo = dict({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.10.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor audio, tensor audio_length) [FlexibleShapeInformation = tuple>>, tuple, ?>>>>((("DefaultShapes", {{"audio", [1, 1]}}), ("RangeDims", {{"audio", [[1, 1], [1, 1280000]]}})))] { + int32 var_9 = const()[name = string("op_9"), val = int32(1)]; + int32 var_10 = const()[name = string("op_10"), val = int32(160)]; + int32 var_12 = const()[name = string("op_12"), val = int32(0)]; + int32 var_33 = const()[name = string("op_33"), val = int32(512)]; + tensor var_34 = add(x = audio_length, y = var_33)[name = string("op_34")]; + int32 var_35 = const()[name = string("op_35"), val = int32(512)]; + tensor var_36 = sub(x = var_34, y = var_35)[name = string("op_36")]; + tensor floor_div_0 = floor_div(x = var_36, y = var_10)[name = string("floor_div_0")]; + tensor var_39 = equal(x = audio_length, y = var_12)[name = string("op_39")]; + tensor var_40 = const()[name = string("op_40"), val = tensor([0])]; + tensor mel_length = select(a = var_40, b = floor_div_0, cond = var_39)[name = string("seq_len")]; + string audio_to_fp16_dtype_0 = const()[name = string("audio_to_fp16_dtype_0"), val = string("fp16")]; + tensor audio_to_fp16 = cast(dtype = audio_to_fp16_dtype_0, x = audio)[name = string("cast_10")]; + tensor var_42_shape_cast_fp16 = shape(x = audio_to_fp16)[name = string("op_42_shape_cast_fp16")]; + int32 gather_0_axis_0 = const()[name = string("gather_0_axis_0"), val = int32(0)]; + int32 gather_0_batch_dims_0 = const()[name = string("gather_0_batch_dims_0"), val = int32(0)]; + bool gather_0_validate_indices_0 = const()[name = string("gather_0_validate_indices_0"), val = bool(false)]; + string var_42_shape_cast_fp16_to_int16_dtype_0 = const()[name = string("op_42_shape_cast_fp16_to_int16_dtype_0"), val = string("int16")]; + uint16 gather_0_indices_0_to_uint16 = const()[name = string("gather_0_indices_0_to_uint16"), val = uint16(1)]; + tensor var_42_shape_cast_fp16_to_int16 = cast(dtype = var_42_shape_cast_fp16_to_int16_dtype_0, x = var_42_shape_cast_fp16)[name = string("cast_9")]; + int16 gather_0_cast_uint16 = gather(axis = gather_0_axis_0, batch_dims = gather_0_batch_dims_0, indices = gather_0_indices_0_to_uint16, validate_indices = gather_0_validate_indices_0, x = var_42_shape_cast_fp16_to_int16)[name = string("gather_0_cast_uint16")]; + string gather_0_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_0_cast_uint16_to_int32_dtype_0"), val = string("int32")]; + int32 const_0 = const()[name = string("const_0"), val = int32(0)]; + int32 const_1 = const()[name = string("const_1"), val = int32(1)]; + int32 gather_0_cast_uint16_to_int32 = cast(dtype = gather_0_cast_uint16_to_int32_dtype_0, x = gather_0_cast_uint16)[name = string("cast_8")]; + tensor var_43 = range_1d(end = gather_0_cast_uint16_to_int32, start = const_0, step = const_1)[name = string("op_43")]; + tensor var_44_axes_0 = const()[name = string("op_44_axes_0"), val = tensor([0])]; + tensor var_44 = expand_dims(axes = var_44_axes_0, x = var_43)[name = string("op_44")]; + tensor var_45_axes_0 = const()[name = string("op_45_axes_0"), val = tensor([1])]; + tensor var_45 = expand_dims(axes = var_45_axes_0, x = audio_length)[name = string("op_45")]; + tensor timemask = less(x = var_44, y = var_45)[name = string("timemask")]; + tensor var_48_begin_0 = const()[name = string("op_48_begin_0"), val = tensor([0, 0])]; + tensor var_48_end_0 = const()[name = string("op_48_end_0"), val = tensor([1, 1])]; + tensor var_48_end_mask_0 = const()[name = string("op_48_end_mask_0"), val = tensor([true, false])]; + tensor var_48_squeeze_mask_0 = const()[name = string("op_48_squeeze_mask_0"), val = tensor([false, true])]; + tensor var_48_cast_fp16 = slice_by_index(begin = var_48_begin_0, end = var_48_end_0, end_mask = var_48_end_mask_0, squeeze_mask = var_48_squeeze_mask_0, x = audio_to_fp16)[name = string("op_48_cast_fp16")]; + tensor var_49_axes_0 = const()[name = string("op_49_axes_0"), val = tensor([1])]; + tensor var_49_cast_fp16 = expand_dims(axes = var_49_axes_0, x = var_48_cast_fp16)[name = string("op_49_cast_fp16")]; + tensor var_51_begin_0 = const()[name = string("op_51_begin_0"), val = tensor([0, 1])]; + tensor var_51_end_0 = const()[name = string("op_51_end_0"), val = tensor([1, 0])]; + tensor var_51_end_mask_0 = const()[name = string("op_51_end_mask_0"), val = tensor([true, true])]; + tensor var_51_cast_fp16 = slice_by_index(begin = var_51_begin_0, end = var_51_end_0, end_mask = var_51_end_mask_0, x = audio_to_fp16)[name = string("op_51_cast_fp16")]; + tensor var_53_begin_0 = const()[name = string("op_53_begin_0"), val = tensor([0, 0])]; + tensor var_53_end_0 = const()[name = string("op_53_end_0"), val = tensor([1, -1])]; + tensor var_53_end_mask_0 = const()[name = string("op_53_end_mask_0"), val = tensor([true, false])]; + tensor var_53_cast_fp16 = slice_by_index(begin = var_53_begin_0, end = var_53_end_0, end_mask = var_53_end_mask_0, x = audio_to_fp16)[name = string("op_53_cast_fp16")]; + fp16 var_54_to_fp16 = const()[name = string("op_54_to_fp16"), val = fp16(0x1.f0cp-1)]; + tensor var_55_cast_fp16 = mul(x = var_53_cast_fp16, y = var_54_to_fp16)[name = string("op_55_cast_fp16")]; + tensor var_56_cast_fp16 = sub(x = var_51_cast_fp16, y = var_55_cast_fp16)[name = string("op_56_cast_fp16")]; + bool x_3_interleave_0 = const()[name = string("x_3_interleave_0"), val = bool(false)]; + tensor x_3_cast_fp16 = concat(axis = var_9, interleave = x_3_interleave_0, values = (var_49_cast_fp16, var_56_cast_fp16))[name = string("x_3_cast_fp16")]; + tensor var_59 = logical_not(x = timemask)[name = string("op_59")]; + fp16 var_16_to_fp16 = const()[name = string("op_16_to_fp16"), val = fp16(0x0p+0)]; + tensor input_1_cast_fp16 = select(a = var_16_to_fp16, b = x_3_cast_fp16, cond = var_59)[name = string("input_1_cast_fp16")]; + tensor concat_1x = const()[name = string("concat_1x"), val = tensor([1, 1, -1])]; + tensor input_3_cast_fp16 = reshape(shape = concat_1x, x = input_1_cast_fp16)[name = string("input_3_cast_fp16")]; + tensor input_5_pad_0 = const()[name = string("input_5_pad_0"), val = tensor([0, 0, 0, 0, 256, 256])]; + string input_5_mode_0 = const()[name = string("input_5_mode_0"), val = string("constant")]; + fp16 const_3_to_fp16 = const()[name = string("const_3_to_fp16"), val = fp16(0x0p+0)]; + tensor input_5_cast_fp16 = pad(constant_val = const_3_to_fp16, mode = input_5_mode_0, pad = input_5_pad_0, x = input_3_cast_fp16)[name = string("input_5_cast_fp16")]; + tensor concat_2x = const()[name = string("concat_2x"), val = tensor([1, -1])]; + tensor input_cast_fp16 = reshape(shape = concat_2x, x = input_5_cast_fp16)[name = string("input_cast_fp16")]; + tensor expand_dims_3 = const()[name = string("expand_dims_3"), val = tensor([160])]; + tensor expand_dims_4_axes_0 = const()[name = string("expand_dims_4_axes_0"), val = tensor([1])]; + tensor expand_dims_4_cast_fp16 = expand_dims(axes = expand_dims_4_axes_0, x = input_cast_fp16)[name = string("expand_dims_4_cast_fp16")]; + string conv_0_pad_type_0 = const()[name = string("conv_0_pad_type_0"), val = string("valid")]; + tensor conv_0_pad_0 = const()[name = string("conv_0_pad_0"), val = tensor([0, 0])]; + tensor conv_0_dilations_0 = const()[name = string("conv_0_dilations_0"), val = tensor([1])]; + int32 conv_0_groups_0 = const()[name = string("conv_0_groups_0"), val = int32(1)]; + tensor expand_dims_1_to_fp16 = const()[name = string("expand_dims_1_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))]; + tensor conv_0_cast_fp16 = conv(dilations = conv_0_dilations_0, groups = conv_0_groups_0, pad = conv_0_pad_0, pad_type = conv_0_pad_type_0, strides = expand_dims_3, weight = expand_dims_1_to_fp16, x = expand_dims_4_cast_fp16)[name = string("conv_0_cast_fp16")]; + string conv_1_pad_type_0 = const()[name = string("conv_1_pad_type_0"), val = string("valid")]; + tensor conv_1_pad_0 = const()[name = string("conv_1_pad_0"), val = tensor([0, 0])]; + tensor conv_1_dilations_0 = const()[name = string("conv_1_dilations_0"), val = tensor([1])]; + int32 conv_1_groups_0 = const()[name = string("conv_1_groups_0"), val = int32(1)]; + tensor expand_dims_2_to_fp16 = const()[name = string("expand_dims_2_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(263296)))]; + tensor conv_1_cast_fp16 = conv(dilations = conv_1_dilations_0, groups = conv_1_groups_0, pad = conv_1_pad_0, pad_type = conv_1_pad_type_0, strides = expand_dims_3, weight = expand_dims_2_to_fp16, x = expand_dims_4_cast_fp16)[name = string("conv_1_cast_fp16")]; + int32 stack_0_axis_0 = const()[name = string("stack_0_axis_0"), val = int32(-1)]; + tensor stack_0_cast_fp16 = stack(axis = stack_0_axis_0, values = (conv_0_cast_fp16, conv_1_cast_fp16))[name = string("stack_0_cast_fp16")]; + fp16 var_19_promoted_to_fp16 = const()[name = string("op_19_promoted_to_fp16"), val = fp16(0x1p+1)]; + tensor var_74_cast_fp16 = pow(x = stack_0_cast_fp16, y = var_19_promoted_to_fp16)[name = string("op_74_cast_fp16")]; + tensor var_76_axes_0 = const()[name = string("op_76_axes_0"), val = tensor([-1])]; + bool var_76_keep_dims_0 = const()[name = string("op_76_keep_dims_0"), val = bool(false)]; + tensor var_76_cast_fp16 = reduce_sum(axes = var_76_axes_0, keep_dims = var_76_keep_dims_0, x = var_74_cast_fp16)[name = string("op_76_cast_fp16")]; + tensor x_11_cast_fp16 = identity(x = var_76_cast_fp16)[name = string("x_11_cast_fp16")]; + bool x_13_transpose_x_0 = const()[name = string("x_13_transpose_x_0"), val = bool(false)]; + bool x_13_transpose_y_0 = const()[name = string("x_13_transpose_y_0"), val = bool(false)]; + tensor const_4_to_fp16 = const()[name = string("const_4_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(526528)))]; + tensor x_13_cast_fp16 = matmul(transpose_x = x_13_transpose_x_0, transpose_y = x_13_transpose_y_0, x = const_4_to_fp16, y = x_11_cast_fp16)[name = string("x_13_cast_fp16")]; + fp16 var_83_to_fp16 = const()[name = string("op_83_to_fp16"), val = fp16(0x1p-24)]; + tensor var_84_cast_fp16 = add(x = x_13_cast_fp16, y = var_83_to_fp16)[name = string("op_84_cast_fp16")]; + fp32 x_epsilon_0 = const()[name = string("x_epsilon_0"), val = fp32(0x1p-149)]; + tensor x_cast_fp16 = log(epsilon = x_epsilon_0, x = var_84_cast_fp16)[name = string("x_cast_fp16")]; + tensor var_86_shape_cast_fp16 = shape(x = x_cast_fp16)[name = string("op_86_shape_cast_fp16")]; + int32 gather_5_batch_dims_0 = const()[name = string("gather_5_batch_dims_0"), val = int32(0)]; + bool gather_5_validate_indices_0 = const()[name = string("gather_5_validate_indices_0"), val = bool(false)]; + string var_86_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_86_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")]; + int32 gather_5_cast_uint16_axis_0 = const()[name = string("gather_5_cast_uint16_axis_0"), val = int32(0)]; + uint16 select_0_to_uint16 = const()[name = string("select_0_to_uint16"), val = uint16(2)]; + tensor var_86_shape_cast_fp16_to_uint16 = cast(dtype = var_86_shape_cast_fp16_to_uint16_dtype_0, x = var_86_shape_cast_fp16)[name = string("cast_7")]; + uint16 gather_5_cast_uint16_cast_uint16 = gather(axis = gather_5_cast_uint16_axis_0, batch_dims = gather_5_batch_dims_0, indices = select_0_to_uint16, validate_indices = gather_5_validate_indices_0, x = var_86_shape_cast_fp16_to_uint16)[name = string("gather_5_cast_uint16_cast_uint16")]; + string gather_5_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_5_cast_uint16_to_int32_dtype_0"), val = string("int32")]; + int32 const_5 = const()[name = string("const_5"), val = int32(0)]; + int32 const_6 = const()[name = string("const_6"), val = int32(1)]; + int32 gather_5_cast_uint16_to_int32 = cast(dtype = gather_5_cast_uint16_to_int32_dtype_0, x = gather_5_cast_uint16_cast_uint16)[name = string("cast_6")]; + tensor mask_1 = range_1d(end = gather_5_cast_uint16_to_int32, start = const_5, step = const_6)[name = string("mask_1")]; + tensor expand_dims_0_axes_0 = const()[name = string("expand_dims_0_axes_0"), val = tensor([0])]; + tensor expand_dims_0 = expand_dims(axes = expand_dims_0_axes_0, x = mask_1)[name = string("expand_dims_0")]; + tensor var_91_axes_0 = const()[name = string("op_91_axes_0"), val = tensor([1])]; + tensor var_91 = expand_dims(axes = var_91_axes_0, x = mel_length)[name = string("op_91")]; + tensor mask = greater_equal(x = expand_dims_0, y = var_91)[name = string("mask")]; + tensor var_93_axes_0 = const()[name = string("op_93_axes_0"), val = tensor([1])]; + tensor var_93 = expand_dims(axes = var_93_axes_0, x = mask)[name = string("op_93")]; + tensor processed_signal_cast_fp16 = select(a = var_16_to_fp16, b = x_cast_fp16, cond = var_93)[name = string("processed_signal_cast_fp16")]; + string processed_signal_cast_fp16_to_fp32_dtype_0 = const()[name = string("processed_signal_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + tensor mel = cast(dtype = processed_signal_cast_fp16_to_fp32_dtype_0, x = processed_signal_cast_fp16)[name = string("cast_5")]; + } -> (mel, mel_length); +} \ No newline at end of file diff --git a/ja/2240ms/preprocessor.mlmodelc/weights/weight.bin b/ja/2240ms/preprocessor.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..86dd375f6649d262d58c9cf8b89006ceab216411 --- /dev/null +++ b/ja/2240ms/preprocessor.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:297514e2b211d14b0e53cb97193d679bb89ead98d28e578f3f1d049ddbcc36b3 +size 592384 diff --git a/ja/2240ms/preprocessor.mlpackage/Data/com.apple.CoreML/model.mlmodel b/ja/2240ms/preprocessor.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..050fa97ca7a2aa4b7c4fa318f4fa2a51914287c4 --- /dev/null +++ b/ja/2240ms/preprocessor.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0feb90f2cf9a5bc749fa763b6eb78a23755b6184b41a262fe08faebe4e709b3e +size 16035 diff --git a/ja/2240ms/preprocessor.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/ja/2240ms/preprocessor.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..86dd375f6649d262d58c9cf8b89006ceab216411 --- /dev/null +++ b/ja/2240ms/preprocessor.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:297514e2b211d14b0e53cb97193d679bb89ead98d28e578f3f1d049ddbcc36b3 +size 592384 diff --git a/ja/2240ms/preprocessor.mlpackage/Manifest.json b/ja/2240ms/preprocessor.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..667b90978564849cb1c6d3c8b10f535ad524d4be --- /dev/null +++ b/ja/2240ms/preprocessor.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "02ED6D10-6658-4E52-83B8-E0D3DEA2B8AC": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + }, + "8D115BDF-817A-4C02-9F24-4AF6137B0210": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + } + }, + "rootModelIdentifier": "8D115BDF-817A-4C02-9F24-4AF6137B0210" +} diff --git a/ja/2240ms/tokenizer.json b/ja/2240ms/tokenizer.json new file mode 100644 index 0000000000000000000000000000000000000000..b05c35042625729d82c27359c8b8cf9ae4e2e4c9 --- /dev/null +++ b/ja/2240ms/tokenizer.json @@ -0,0 +1,1406 @@ +{ + "0": "", + "1": "", + "2": "▁", + "3": ".", + "4": ",", + "5": "e", + "6": "t", + "7": "a", + "8": "s", + "9": "o", + "10": "i", + "11": "r", + "12": "h", + "13": "m", + "14": "p", + "15": "n", + "16": "g", + "17": "en", + "18": "on", + "19": "y", + "20": "0", + "21": "b", + "22": "2", + "23": "1", + "24": "k", + "25": "3", + "26": "5", + "27": "I", + "28": "4", + "29": "z", + "30": "A", + "31": "6", + "32": "E", + "33": "7", + "34": "T", + "35": "8", + "36": "9", + "37": "S", + "38": "C", + "39": "M", + "40": "P", + "41": "D", + "42": "B", + "43": "U", + "44": "N", + "45": "O", + "46": "G", + "47": "F", + "48": "L", + "49": "R", + "50": "H", + "51": "K", + "52": "", + "53": "le", + "54": "li", + "55": "to", + "56": "na", + "57": "ci", + "58": "▁A", + "59": "vi", + "60": "Z", + "61": "X", + "62": "", + "63": "et", + "64": "ar", + "65": "el", + "66": "", + "67": "ie", + "68": "▁G", + "69": "", + "70": "", + "71": "ta", + "72": "ks", + "73": "gi", + "74": "di", + "75": "", + "76": "", + "77": "", + "78": "cs", + "79": "", + "80": "sti", + "81": "", + "82": "pe", + "83": "", + "84": "as", + "85": "▁Tai", + "86": "", + "87": "▁20", + "88": "", + "89": "", + "90": "", + "91": "-", + "92": "", + "93": "▁M", + "94": "", + "95": "", + "96": "", + "97": "", + "98": "pp", + "99": "", + "100": "", + "101": "▁N", + "102": "mm", + "103": "", + "104": "▁and", + "105": "th", + "106": "ex", + "107": "", + "108": "ng", + "109": "", + "110": "▁La", + "111": "", + "112": "", + "113": "avec", + "114": "", + "115": "", + "116": "!", + "117": "", + "118": "▁あ", + "119": "▁す", + "120": "▁し", + "121": "▁そ", + "122": "▁ま", + "123": "▁こ", + "124": "以", + "125": "仕", + "126": "▁文", + "127": "▁強", + "128": "▁月", + "129": "▁連", + "130": "▁ペ", + "131": "▁海", + "132": "▁科", + "133": "▁ゴ", + "134": "▁原", + "135": "▁ギ", + "136": "▁ヤ", + "137": "▁特", + "138": "▁参", + "139": "▁男", + "140": "普", + "141": "受", + "142": "▁都", + "143": "▁水", + "144": "▁終", + "145": "▁南", + "146": "▁例", + "147": "▁利", + "148": "▁現", + "149": "▁真", + "150": "▁古", + "151": "▁彼", + "152": "▁交", + "153": "専", + "154": "▁悪", + "155": "▁旅", + "156": "▁技", + "157": "▁再", + "158": "企", + "159": "登", + "160": "総", + "161": "改", + "162": "探", + "163": "異", + "164": "完", + "165": "頑", + "166": "深", + "167": "推", + "168": "昨", + "169": "危", + "170": "詳", + "171": "余", + "172": "描", + "173": "被", + "174": "固", + "175": "副", + "176": "混", + "177": "救", + "178": "互", + "179": "閉", + "180": "亜", + "181": "植", + "182": "責", + "183": "仮", + "184": "脱", + "185": "恐", + "186": "契", + "187": "隠", + "188": "麻", + "189": "哲", + "190": "宣", + "191": "浮", + "192": "諸", + "193": "巨", + "194": "午", + "195": "占", + "196": "距", + "197": "汚", + "198": "延", + "199": "訪", + "200": "沿", + "201": "唯", + "202": "批", + "203": "傾", + "204": "邦", + "205": "購", + "206": "維", + "207": "燃", + "208": "既", + "209": "緊", + "210": "獲", + "211": "炭", + "212": "搭", + "213": "努", + "214": "顧", + "215": "狭", + "216": "雰", + "217": "塗", + "218": "排", + "219": "奪", + "220": "埋", + "221": "履", + "222": "侵", + "223": "該", + "224": "頻", + "225": "勧", + "226": "魅", + "227": "抑", + "228": "驚", + "229": "衝", + "230": "耐", + "231": "披", + "232": "透", + "233": "促", + "234": "撤", + "235": "潜", + "236": "吐", + "237": "睡", + "238": "孤", + "239": "昆", + "240": "遭", + "241": "溶", + "242": "貿", + "243": "疾", + "244": "拒", + "245": "脅", + "246": "挟", + "247": "漏", + "248": "覆", + "249": "畜", + "250": "紛", + "251": "絞", + "252": "擬", + "253": "疎", + "254": "跳", + "255": "廷", + "256": "郊", + "257": "克", + "258": "措", + "259": "虐", + "260": "陥", + "261": "徐", + "262": "漠", + "263": "愚", + "264": "噴", + "265": "卿", + "266": "胡", + "267": "溢", + "268": "妨", + "269": "沸", + "270": "阻", + "271": "妃", + "272": "舶", + "273": "濡", + "274": "剥", + "275": "洪", + "276": "鋭", + "277": "娯", + "278": "腎", + "279": "訂", + "280": "肖", + "281": "枢", + "282": "髄", + "283": "傑", + "284": "砕", + "285": "幣", + "286": "勾", + "287": "臆", + "288": "歪", + "289": "喝", + "290": "彗", + "291": "礁", + "292": "蚊", + "293": "焉", + "294": "塵", + "295": "脊", + "296": "鱗", + "297": "瞑", + "298": "峻", + "299": "徊", + "300": "徘", + "301": "猥", + "302": "竣", + "303": "穣", + "304": "隻", + "305": "睨", + "306": "瀕", + "307": "褻", + "308": "踵", + "309": "奢", + "310": "董", + "311": "毅", + "312": "醒", + "313": "墜", + "314": "窟", + "315": "綿", + "316": "遮", + "317": "騰", + "318": "顎", + "319": "遍", + "320": "蓋", + "321": "澄", + "322": "棲", + "323": "虎", + "324": "粒", + "325": "胎", + "326": "爪", + "327": "殻", + "328": "鉛", + "329": "枕", + "330": "堤", + "331": "膝", + "332": "峠", + "333": "搬", + "334": "晋", + "335": "餌", + "336": "狼", + "337": "衰", + "338": "閲", + "339": "筒", + "340": "軸", + "341": "緯", + "342": "暦", + "343": "刃", + "344": "潔", + "345": "裂", + "346": "叫", + "347": "殖", + "348": "媒", + "349": "癌", + "350": "鎖", + "351": "磁", + "352": "洞", + "353": "遇", + "354": "枝", + "355": "紫", + "356": "縦", + "357": "胞", + "358": "釈", + "359": "威", + "360": "晶", + "361": "砲", + "362": "焦", + "363": "尿", + "364": "魂", + "365": "潮", + "366": "旬", + "367": "慎", + "368": "噂", + "369": "隔", + "370": "穴", + "371": "慮", + "372": "即", + "373": "滑", + "374": "雷", + "375": "摘", + "376": "鏡", + "377": "棒", + "378": "悟", + "379": "葬", + "380": "序", + "381": "貫", + "382": "氷", + "383": "針", + "384": "煮", + "385": "棄", + "386": "銃", + "387": "汁", + "388": "封", + "389": "湿", + "390": "靴", + "391": "豚", + "392": "締", + "393": "豪", + "394": "票", + "395": "皮", + "396": "縮", + "397": "徹", + "398": "較", + "399": "忍", + "400": "核", + "401": "儀", + "402": "到", + "403": "削", + "404": "駆", + "405": "繁", + "406": "陰", + "407": "浄", + "408": "脈", + "409": "滞", + "410": "至", + "411": "枚", + "412": "偉", + "413": "致", + "414": "貨", + "415": "漢", + "416": "己", + "417": "握", + "418": "欧", + "419": "薄", + "420": "献", + "421": "預", + "422": "龍", + "423": "快", + "424": "句", + "425": "縁", + "426": "微", + "427": "妙", + "428": "晩", + "429": "粉", + "430": "卓", + "431": "圏", + "432": "兼", + "433": "脳", + "434": "竜", + "435": "鳴", + "436": "騒", + "437": "請", + "438": "卵", + "439": "唱", + "440": "嵐", + "441": "臓", + "442": "箱", + "443": "祖", + "444": "浴", + "445": "壁", + "446": "析", + "447": "厚", + "448": "筆", + "449": "承", + "450": "均", + "451": "律", + "452": "否", + "453": "脚", + "454": "湖", + "455": "乳", + "456": "揮", + "457": "滅", + "458": "乾", + "459": "羽", + "460": "候", + "461": "拡", + "462": "貸", + "463": "砂", + "464": "敬", + "465": "庁", + "466": "煙", + "467": "底", + "468": "露", + "469": "骨", + "470": "倍", + "471": "殿", + "472": "易", + "473": "層", + "474": "幕", + "475": "毛", + "476": "爆", + "477": "暇", + "478": "械", + "479": "隣", + "480": "輸", + "481": "柄", + "482": "範", + "483": "掲", + "484": "嘘", + "485": "剤", + "486": "墓", + "487": "衣", + "488": "射", + "489": "菓", + "490": "募", + "491": "乱", + "492": "迎", + "493": "抱", + "494": "酸", + "495": "雄", + "496": "虫", + "497": "複", + "498": "為", + "499": "泳", + "500": "宝", + "501": "激", + "502": "暑", + "503": "疑", + "504": "誘", + "505": "暴", + "506": "聖", + "507": "捨", + "508": "破", + "509": "革", + "510": "希", + "511": "折", + "512": "惑", + "513": "測", + "514": "紀", + "515": "舎", + "516": "署", + "517": "患", + "518": "岸", + "519": "秀", + "520": "免", + "521": "禁", + "522": "躍", + "523": "聴", + "524": "抗", + "525": "税", + "526": "奏", + "527": "弾", + "528": "礼", + "529": "童", + "530": "裏", + "531": "吹", + "532": "眠", + "533": "歯", + "534": "拠", + "535": "慣", + "536": "触", + "537": "飼", + "538": "群", + "539": "宗", + "540": "傷", + "541": "額", + "542": "塩", + "543": "静", + "544": "留", + "545": "罪", + "546": "純", + "547": "壊", + "548": "闘", + "549": "弱", + "550": "刻", + "551": "航", + "552": "栄", + "553": "姿", + "554": "亡", + "555": "織", + "556": "敗", + "557": "章", + "558": "吸", + "559": "令", + "560": "捜", + "561": "模", + "562": "絵", + "563": "申", + "564": "盤", + "565": "積", + "566": "標", + "567": "階", + "568": "省", + "569": "項", + "570": "猫", + "571": "従", + "572": "非", + "573": "季", + "574": "捕", + "575": "党", + "576": "圧", + "577": "香", + "578": "操", + "579": "暗", + "580": "症", + "581": "散", + "582": "突", + "583": "適", + "584": "雑", + "585": "跡", + "586": "厳", + "587": "鳥", + "588": "逃", + "589": "講", + "590": "晴", + "591": "徴", + "592": "困", + "593": "短", + "594": "婦", + "595": "略", + "596": "齢", + "597": "震", + "598": "敵", + "599": "博", + "600": "血", + "601": "満", + "602": "舗", + "603": "宙", + "604": "寿", + "605": "遺", + "606": "極", + "607": "里", + "608": "因", + "609": "典", + "610": "染", + "611": "徒", + "612": "巻", + "613": "頂", + "614": "超", + "615": "河", + "616": "盛", + "617": "犬", + "618": "豊", + "619": "端", + "620": "紹", + "621": "首", + "622": "陽", + "623": "歳", + "624": "印", + "625": "紙", + "626": "払", + "627": "求", + "628": "障", + "629": "簡", + "630": "途", + "631": "創", + "632": "船", + "633": "菜", + "634": "ゥ", + "635": "勤", + "636": "痛", + "637": "並", + "638": "景", + "639": "雪", + "640": "節", + "641": "浜", + "642": "清", + "643": "抜", + "644": "勢", + "645": "暮", + "646": "銀", + "647": "盟", + "648": "魚", + "649": "率", + "650": "洋", + "651": "渡", + "652": "順", + "653": "況", + "654": "談", + "655": "舞", + "656": "案", + "657": "岩", + "658": "負", + "659": "旧", + "660": "財", + "661": "故", + "662": "冬", + "663": "横", + "664": "奥", + "665": "比", + "666": "囲", + "667": "停", + "668": "築", + "669": "波", + "670": "林", + "671": "暖", + "672": "索", + "673": "赤", + "674": "給", + "675": "末", + "676": "催", + "677": "遅", + "678": "述", + "679": "黒", + "680": "細", + "681": "与", + "682": "減", + "683": "級", + "684": "費", + "685": "越", + "686": "差", + "687": "領", + "688": "衛", + "689": "隊", + "690": "薬", + "691": "氏", + "692": "望", + "693": "似", + "694": "就", + "695": "条", + "696": "処", + "697": "谷", + "698": "策", + "699": "効", + "700": "熱", + "701": "復", + "702": "ヌ", + "703": "振", + "704": "規", + "705": "港", + "706": "注", + "707": "森", + "708": "防", + "709": "継", + "710": "退", + "711": "火", + "712": "陸", + "713": "去", + "714": "視", + "715": "整", + "716": "準", + "717": "庭", + "718": "ゾ", + "719": "独", + "720": "撃", + "721": "児", + "722": "橋", + "723": "換", + "724": "念", + "725": "識", + "726": "打", + "727": "津", + "728": "雨", + "729": "幸", + "730": "含", + "731": "響", + "732": "労", + "733": "官", + "734": "追", + "735": "遠", + "736": "未", + "737": "販", + "738": "街", + "739": "曜", + "740": "程", + "741": "提", + "742": "玉", + "743": "判", + "744": "移", + "745": "攻", + "746": "低", + "747": "装", + "748": "断", + "749": "及", + "750": "証", + "751": "象", + "752": "守", + "753": "戻", + "754": "詞", + "755": "投", + "756": "載", + "757": "具", + "758": "除", + "759": "環", + "760": "展", + "761": "争", + "762": "失", + "763": "春", + "764": "挙", + "765": "返", + "766": "馬", + "767": "欲", + "768": "材", + "769": "図", + "770": "養", + "771": "焼", + "772": "導", + "773": "夢", + "774": "米", + "775": "冷", + "776": "息", + "777": "兵", + "778": "済", + "779": "劇", + "780": "央", + "781": "険", + "782": "服", + "783": "態", + "784": "走", + "785": "評", + "786": "権", + "787": "論", + "788": "ゅ", + "789": "境", + "790": "察", + "791": "授", + "792": "頼", + "793": "派", + "794": "撮", + "795": "素", + "796": "修", + "797": "第", + "798": "質", + "799": "告", + "800": "興", + "801": "秒", + "802": "宇", + "803": "肉", + "804": "像", + "805": "称", + "806": "値", + "807": "頭", + "808": "週", + "809": "督", + "810": "消", + "811": "芸", + "812": "顔", + "813": "読", + "814": "仲", + "815": "遊", + "816": "試", + "817": "酒", + "818": "離", + "819": "増", + "820": "殺", + "821": "鉄", + "822": "害", + "823": "割", + "824": "石", + "825": "夏", + "826": "助", + "827": "英", + "828": "想", + "829": "管", + "830": "急", + "831": "頃", + "832": "づ", + "833": "造", + "834": "史", + "835": "量", + "836": "製", + "837": "府", + "838": "足", + "839": "王", + "840": "委", + "841": "両", + "842": "辺", + "843": "残", + "844": "逆", + "845": "備", + "846": "軍", + "847": "警", + "848": "査", + "849": "列", + "850": "編", + "851": "段", + "852": "反", + "853": "ゼ", + "854": "携", + "855": "歩", + "856": "座", + "857": "飛", + "858": "丈", + "859": "価", + "860": "監", + "861": "ヘ", + "862": "周", + "863": "毎", + "864": "統", + "865": "収", + "866": "落", + "867": "星", + "868": "降", + "869": "側", + "870": "療", + "871": "師", + "872": "写", + "873": "類", + "874": "命", + "875": "介", + "876": "護", + "877": "死", + "878": "果", + "879": "任", + "880": "更", + "881": "常", + "882": "検", + "883": "過", + "884": "資", + "885": "働", + "886": "認", + "887": "般", + "888": "示", + "889": "客", + "890": "習", + "891": "究", + "892": "半", + "893": "録", + "894": "字", + "895": "昔", + "896": "影", + "897": "覚", + "898": "型", + "899": "声", + "900": "件", + "901": "義", + "902": "施", + "903": "容", + "904": "路", + "905": "呼", + "906": "役", + "907": "単", + "908": "状", + "909": "建", + "910": "由", + "911": "属", + "912": "土", + "913": "葉", + "914": "起", + "915": "覧", + "916": "配", + "917": "張", + "918": "接", + "919": "込", + "920": "待", + "921": "室", + "922": "病", + "923": "帯", + "924": "婚", + "925": "光", + "926": "個", + "927": "職", + "928": "営", + "929": "ぼ", + "930": "研", + "931": "計", + "932": "直", + "933": "難", + "934": "絶", + "935": "ヨ", + "936": "照", + "937": "西", + "938": "約", + "939": "存", + "940": "験", + "941": "治", + "942": "解", + "943": "転", + "944": "商", + "945": "進", + "946": "係", + "947": "説", + "948": "観", + "949": "球", + "950": "乗", + "951": "支", + "952": "得", + "953": "議", + "954": "門", + "955": "止", + "956": "重", + "957": "温", + "958": "着", + "959": "飲", + "960": "母", + "961": "士", + "962": "ざ", + "963": "集", + "964": "万", + "965": "太", + "966": "続", + "967": "線", + "968": "種", + "969": "格", + "970": "位", + "971": "ユ", + "972": "歌", + "973": "夜", + "974": "共", + "975": "正", + "976": "必", + "977": "ヒ", + "978": "色", + "979": "問", + "980": "再", + "981": "域", + "982": "ゆ", + "983": "勝", + "984": "台", + "985": "技", + "986": "旅", + "987": "引", + "988": "系", + "989": "院", + "990": "悪", + "991": "基", + "992": "神", + "993": "産", + "994": "決", + "995": "民", + "996": "交", + "997": "政", + "998": "賞", + "999": "空", + "1000": "医", + "1001": "彼", + "1002": "夫", + "1003": "可", + "1004": "誰", + "1005": "古", + "1006": "帰", + "1007": "術", + "1008": "相", + "1009": "団", + "1010": "伝", + "1011": "住", + "1012": "題", + "1013": "平", + "1014": "予", + "1015": "音", + "1016": "朝", + "1017": "指", + "1018": "真", + "1019": "ヴ", + "1020": "務", + "1021": "点", + "1022": "各", + "1023": "館", + "1024": "応", + "1025": "現", + "1026": "利", + "1027": "天", + "1028": "等", + "1029": "木", + "1030": "白", + "1031": "形", + "1032": "供", + "1033": "経", + "1034": "族", + "1035": "早", + "1036": "例", + "1037": "不", + "1038": "切", + "1039": "南", + "1040": "加", + "1041": "際", + "1042": "終", + "1043": "様", + "1044": "放", + "1045": "和", + "1046": "州", + "1047": "水", + "1048": "協", + "1049": "在", + "1050": "組", + "1051": "向", + "1052": "広", + "1053": "身", + "1054": "界", + "1055": "工", + "1056": "選", + "1057": "始", + "1058": "元", + "1059": "々", + "1060": "親", + "1061": "美", + "1062": "信", + "1063": "都", + "1064": "置", + "1065": "局", + "1066": "運", + "1067": "送", + "1068": "風", + "1069": "口", + "1070": "演", + "1071": "調", + "1072": "ぎ", + "1073": "優", + "1074": "次", + "1075": "ォ", + "1076": "他", + "1077": "園", + "1078": "保", + "1079": "男", + "1080": "参", + "1081": "少", + "1082": "百", + "1083": "特", + "1084": "考", + "1085": "無", + "1086": "七", + "1087": "ヤ", + "1088": "ギ", + "1089": "良", + "1090": "ザ", + "1091": "制", + "1092": "売", + "1093": "能", + "1094": "原", + "1095": "ゲ", + "1096": "有", + "1097": "安", + "1098": "ゴ", + "1099": "育", + "1100": "科", + "1101": "要", + "1102": "料", + "1103": "書", + "1104": "語", + "1105": "設", + "1106": "海", + "1107": "期", + "1108": "流", + "1109": "確", + "1110": "ペ", + "1111": "区", + "1112": "む", + "1113": "連", + "1114": "買", + "1115": "ひ", + "1116": "ふ", + "1117": "付", + "1118": "町", + "1119": "活", + "1120": "情", + "1121": "月", + "1122": "表", + "1123": "曲", + "1124": "強", + "1125": "世", + "1126": "明", + "1127": "成", + "1128": "ノ", + "1129": "ァ", + "1130": "文", + "1131": "違", + "1132": "東", + "1133": "友", + "1134": "意", + "1135": "力", + "1136": "式", + "1137": "法", + "1138": "報", + "1139": "員", + "1140": "心", + "1141": "屋", + "1142": "品", + "1143": "北", + "1144": "先", + "1145": "島", + "1146": "味", + "1147": "川", + "1148": "開", + "1149": "千", + "1150": "関", + "1151": "電", + "1152": "然", + "1153": "度", + "1154": "達", + "1155": "面", + "1156": "九", + "1157": "数", + "1158": "取", + "1159": "楽", + "1160": "金", + "1161": "性", + "1162": "野", + "1163": "別", + "1164": "戦", + "1165": "公", + "1166": "機", + "1167": "道", + "1168": "目", + "1169": "記", + "1170": "び", + "1171": "発", + "1172": "対", + "1173": "立", + "1174": "初", + "1175": "化", + "1176": "ソ", + "1177": "ワ", + "1178": "田", + "1179": "持", + "1180": "ガ", + "1181": "車", + "1182": "番", + "1183": "ピ", + "1184": "聞", + "1185": "回", + "1186": "ぶ", + "1187": "ベ", + "1188": "げ", + "1189": "実", + "1190": "ボ", + "1191": "店", + "1192": "小", + "1193": "定", + "1194": "モ", + "1195": "長", + "1196": "新", + "1197": "ハ", + "1198": "ケ", + "1199": "外", + "1200": "ポ", + "1201": "近", + "1202": "所", + "1203": "へ", + "1204": "同", + "1205": "ネ", + "1206": "内", + "1207": "女", + "1208": "ホ", + "1209": "体", + "1210": "好", + "1211": "ツ", + "1212": "セ", + "1213": "知", + "1214": "山", + "1215": "来", + "1216": "ェ", + "1217": "使", + "1218": "ョ", + "1219": "ズ", + "1220": "主", + "1221": "動", + "1222": "理", + "1223": "物", + "1224": "映", + "1225": "者", + "1226": "ぐ", + "1227": "的", + "1228": "代", + "1229": "変", + "1230": "教", + "1231": "社", + "1232": "用", + "1233": "話", + "1234": "名", + "1235": "構", + "1236": "高", + "1237": "最", + "1238": "ず", + "1239": "ミ", + "1240": "校", + "1241": "ダ", + "1242": "食", + "1243": "後", + "1244": "手", + "1245": "三", + "1246": "通", + "1247": "感", + "1248": "合", + "1249": "多", + "1250": "業", + "1251": "入", + "1252": "エ", + "1253": "場", + "1254": "べ", + "1255": "上", + "1256": "家", + "1257": "私", + "1258": "年", + "1259": "間", + "1260": "画", + "1261": "前", + "1262": "下", + "1263": "ャ", + "1264": "地", + "1265": "二", + "1266": "ウ", + "1267": "ナ", + "1268": "ビ", + "1269": "自", + "1270": "全", + "1271": "パ", + "1272": "結", + "1273": "ブ", + "1274": "ュ", + "1275": "市", + "1276": "サ", + "1277": "気", + "1278": "方", + "1279": "デ", + "1280": "十", + "1281": "キ", + "1282": "当", + "1283": "国", + "1284": "作", + "1285": "ィ", + "1286": "部", + "1287": "オ", + "1288": "ニ", + "1289": "チ", + "1290": "ム", + "1291": "グ", + "1292": "メ", + "1293": "ご", + "1294": "子", + "1295": "ば", + "1296": "生", + "1297": "ほ", + "1298": "せ", + "1299": "何", + "1300": "出", + "1301": "言", + "1302": "今", + "1303": "バ", + "1304": "事", + "1305": "中", + "1306": "プ", + "1307": "時", + "1308": "コ", + "1309": "見", + "1310": "テ", + "1311": "会", + "1312": "マ", + "1313": "カ", + "1314": "思", + "1315": "ロ", + "1316": "ジ", + "1317": "フ", + "1318": "シ", + "1319": "め", + "1320": "レ", + "1321": "ド", + "1322": "分", + "1323": "ょ", + "1324": "ろ", + "1325": "学", + "1326": "行", + "1327": "タ", + "1328": "大", + "1329": "つ", + "1330": "本", + "1331": "日", + "1332": "わ", + "1333": "一", + "1334": "ク", + "1335": "み", + "1336": "リ", + "1337": "ア", + "1338": "ッ", + "1339": "人", + "1340": "ラ", + "1341": "お", + "1342": "じ", + "1343": "イ", + "1344": "ル", + "1345": "ト", + "1346": "ゃ", + "1347": "き", + "1348": "さ", + "1349": "ち", + "1350": "や", + "1351": "ス", + "1352": "ど", + "1353": "け", + "1354": "く", + "1355": "え", + "1356": "を", + "1357": "り", + "1358": "よ", + "1359": "こ", + "1360": "ン", + "1361": "だ", + "1362": "れ", + "1363": "ら", + "1364": "ね", + "1365": "が", + "1366": "ま", + "1367": "ー", + "1368": "も", + "1369": "そ", + "1370": "し", + "1371": "に", + "1372": "は", + "1373": "る", + "1374": "す", + "1375": "と", + "1376": "た", + "1377": "あ", + "1378": "て", + "1379": "っ", + "1380": "で", + "1381": "か", + "1382": "な", + "1383": "ん", + "1384": "う", + "1385": "の", + "1386": "、", + "1387": "。", + "1388": "い", + "1389": "", + "1390": "", + "1391": "", + "1392": "", + "1393": "", + "1394": "", + "1395": "", + "1396": "▁香", + "1397": "▁群", + "1398": "▁米", + "1399": "咆", + "1400": "哮", + "1401": "翅", + "1402": "", + "1403": "" +} \ No newline at end of file diff --git a/ja/4480ms/decoder.mlmodelc/analytics/coremldata.bin b/ja/4480ms/decoder.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..a23f14dd8e4d2bccc2844d3d81c6c9ca86ea3cba --- /dev/null +++ b/ja/4480ms/decoder.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2fcae710f3db79230f47be6daadc8af085539067285a96f89b2a4c0fd0cb3808 +size 243 diff --git a/ja/4480ms/decoder.mlmodelc/coremldata.bin b/ja/4480ms/decoder.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..b5a0365702bcdfd90328e91463eb94b1d4d5216f --- /dev/null +++ b/ja/4480ms/decoder.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0446e60685dd495222ef0266e71f310d3f9224bfb8bc34cec4daf296c2991f5f +size 493 diff --git a/ja/4480ms/decoder.mlmodelc/model.mil b/ja/4480ms/decoder.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..9f4b6b4cebc16f759164ca05a77b06fb57dedbce --- /dev/null +++ b/ja/4480ms/decoder.mlmodelc/model.mil @@ -0,0 +1,73 @@ +program(1.3) +[buildInfo = dict({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.10.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor c_in, tensor h_in, tensor token, tensor token_length) { + int32 y_batch_dims_0 = const()[name = string("y_batch_dims_0"), val = int32(0)]; + bool y_validate_indices_0 = const()[name = string("y_validate_indices_0"), val = bool(false)]; + tensor module_prediction_embed_weight_to_fp16 = const()[name = string("module_prediction_embed_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))]; + string token_to_int16_dtype_0 = const()[name = string("token_to_int16_dtype_0"), val = string("int16")]; + string cast_1_dtype_0 = const()[name = string("cast_1_dtype_0"), val = string("int32")]; + int32 greater_equal_0_y_0 = const()[name = string("greater_equal_0_y_0"), val = int32(0)]; + tensor token_to_int16 = cast(dtype = token_to_int16_dtype_0, x = token)[name = string("cast_9")]; + tensor cast_1 = cast(dtype = cast_1_dtype_0, x = token_to_int16)[name = string("cast_8")]; + tensor greater_equal_0 = greater_equal(x = cast_1, y = greater_equal_0_y_0)[name = string("greater_equal_0")]; + int32 slice_by_index_0 = const()[name = string("slice_by_index_0"), val = int32(1404)]; + tensor add_2 = add(x = cast_1, y = slice_by_index_0)[name = string("add_2")]; + tensor select_0 = select(a = cast_1, b = add_2, cond = greater_equal_0)[name = string("select_0")]; + int32 y_cast_fp16_cast_uint16_axis_0 = const()[name = string("y_cast_fp16_cast_uint16_axis_0"), val = int32(0)]; + string select_0_to_int16_dtype_0 = const()[name = string("select_0_to_int16_dtype_0"), val = string("int16")]; + tensor select_0_to_int16 = cast(dtype = select_0_to_int16_dtype_0, x = select_0)[name = string("cast_7")]; + tensor y_cast_fp16_cast_uint16_cast_uint16 = gather(axis = y_cast_fp16_cast_uint16_axis_0, batch_dims = y_batch_dims_0, indices = select_0_to_int16, validate_indices = y_validate_indices_0, x = module_prediction_embed_weight_to_fp16)[name = string("y_cast_fp16_cast_uint16_cast_uint16")]; + tensor input_3_perm_0 = const()[name = string("input_3_perm_0"), val = tensor([1, 0, 2])]; + int32 split_0_num_splits_0 = const()[name = string("split_0_num_splits_0"), val = int32(2)]; + int32 split_0_axis_0 = const()[name = string("split_0_axis_0"), val = int32(0)]; + string h_in_to_fp16_dtype_0 = const()[name = string("h_in_to_fp16_dtype_0"), val = string("fp16")]; + tensor h_in_to_fp16 = cast(dtype = h_in_to_fp16_dtype_0, x = h_in)[name = string("cast_6")]; + tensor split_0_cast_fp16_0, tensor split_0_cast_fp16_1 = split(axis = split_0_axis_0, num_splits = split_0_num_splits_0, x = h_in_to_fp16)[name = string("split_0_cast_fp16")]; + int32 split_1_num_splits_0 = const()[name = string("split_1_num_splits_0"), val = int32(2)]; + int32 split_1_axis_0 = const()[name = string("split_1_axis_0"), val = int32(0)]; + string c_in_to_fp16_dtype_0 = const()[name = string("c_in_to_fp16_dtype_0"), val = string("fp16")]; + tensor c_in_to_fp16 = cast(dtype = c_in_to_fp16_dtype_0, x = c_in)[name = string("cast_5")]; + tensor split_1_cast_fp16_0, tensor split_1_cast_fp16_1 = split(axis = split_1_axis_0, num_splits = split_1_num_splits_0, x = c_in_to_fp16)[name = string("split_1_cast_fp16")]; + tensor input_lstm_layer_0_lstm_h0_squeeze_axes_0 = const()[name = string("input_lstm_layer_0_lstm_h0_squeeze_axes_0"), val = tensor([0])]; + tensor input_lstm_layer_0_lstm_h0_squeeze_cast_fp16 = squeeze(axes = input_lstm_layer_0_lstm_h0_squeeze_axes_0, x = split_0_cast_fp16_0)[name = string("input_lstm_layer_0_lstm_h0_squeeze_cast_fp16")]; + tensor input_lstm_layer_0_lstm_c0_squeeze_axes_0 = const()[name = string("input_lstm_layer_0_lstm_c0_squeeze_axes_0"), val = tensor([0])]; + tensor input_lstm_layer_0_lstm_c0_squeeze_cast_fp16 = squeeze(axes = input_lstm_layer_0_lstm_c0_squeeze_axes_0, x = split_1_cast_fp16_0)[name = string("input_lstm_layer_0_lstm_c0_squeeze_cast_fp16")]; + string input_lstm_layer_0_direction_0 = const()[name = string("input_lstm_layer_0_direction_0"), val = string("forward")]; + bool input_lstm_layer_0_output_sequence_0 = const()[name = string("input_lstm_layer_0_output_sequence_0"), val = bool(true)]; + string input_lstm_layer_0_recurrent_activation_0 = const()[name = string("input_lstm_layer_0_recurrent_activation_0"), val = string("sigmoid")]; + string input_lstm_layer_0_cell_activation_0 = const()[name = string("input_lstm_layer_0_cell_activation_0"), val = string("tanh")]; + string input_lstm_layer_0_activation_0 = const()[name = string("input_lstm_layer_0_activation_0"), val = string("tanh")]; + tensor concat_1_to_fp16 = const()[name = string("concat_1_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1797248)))]; + tensor concat_2_to_fp16 = const()[name = string("concat_2_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5074112)))]; + tensor concat_0_to_fp16 = const()[name = string("concat_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8350976)))]; + tensor input_3_cast_fp16 = transpose(perm = input_3_perm_0, x = y_cast_fp16_cast_uint16_cast_uint16)[name = string("transpose_2")]; + tensor input_lstm_layer_0_cast_fp16_0, tensor input_lstm_layer_0_cast_fp16_1, tensor input_lstm_layer_0_cast_fp16_2 = lstm(activation = input_lstm_layer_0_activation_0, bias = concat_0_to_fp16, cell_activation = input_lstm_layer_0_cell_activation_0, direction = input_lstm_layer_0_direction_0, initial_c = input_lstm_layer_0_lstm_c0_squeeze_cast_fp16, initial_h = input_lstm_layer_0_lstm_h0_squeeze_cast_fp16, output_sequence = input_lstm_layer_0_output_sequence_0, recurrent_activation = input_lstm_layer_0_recurrent_activation_0, weight_hh = concat_2_to_fp16, weight_ih = concat_1_to_fp16, x = input_3_cast_fp16)[name = string("input_lstm_layer_0_cast_fp16")]; + tensor input_lstm_h0_squeeze_axes_0 = const()[name = string("input_lstm_h0_squeeze_axes_0"), val = tensor([0])]; + tensor input_lstm_h0_squeeze_cast_fp16 = squeeze(axes = input_lstm_h0_squeeze_axes_0, x = split_0_cast_fp16_1)[name = string("input_lstm_h0_squeeze_cast_fp16")]; + tensor input_lstm_c0_squeeze_axes_0 = const()[name = string("input_lstm_c0_squeeze_axes_0"), val = tensor([0])]; + tensor input_lstm_c0_squeeze_cast_fp16 = squeeze(axes = input_lstm_c0_squeeze_axes_0, x = split_1_cast_fp16_1)[name = string("input_lstm_c0_squeeze_cast_fp16")]; + string input_direction_0 = const()[name = string("input_direction_0"), val = string("forward")]; + bool input_output_sequence_0 = const()[name = string("input_output_sequence_0"), val = bool(true)]; + string input_recurrent_activation_0 = const()[name = string("input_recurrent_activation_0"), val = string("sigmoid")]; + string input_cell_activation_0 = const()[name = string("input_cell_activation_0"), val = string("tanh")]; + string input_activation_0 = const()[name = string("input_activation_0"), val = string("tanh")]; + tensor concat_4_to_fp16 = const()[name = string("concat_4_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8356160)))]; + tensor concat_5_to_fp16 = const()[name = string("concat_5_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11633024)))]; + tensor concat_3_to_fp16 = const()[name = string("concat_3_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14909888)))]; + tensor input_cast_fp16_0, tensor input_cast_fp16_1, tensor input_cast_fp16_2 = lstm(activation = input_activation_0, bias = concat_3_to_fp16, cell_activation = input_cell_activation_0, direction = input_direction_0, initial_c = input_lstm_c0_squeeze_cast_fp16, initial_h = input_lstm_h0_squeeze_cast_fp16, output_sequence = input_output_sequence_0, recurrent_activation = input_recurrent_activation_0, weight_hh = concat_5_to_fp16, weight_ih = concat_4_to_fp16, x = input_lstm_layer_0_cast_fp16_0)[name = string("input_cast_fp16")]; + int32 obj_3_axis_0 = const()[name = string("obj_3_axis_0"), val = int32(0)]; + tensor obj_3_cast_fp16 = stack(axis = obj_3_axis_0, values = (input_lstm_layer_0_cast_fp16_1, input_cast_fp16_1))[name = string("obj_3_cast_fp16")]; + string obj_3_cast_fp16_to_fp32_dtype_0 = const()[name = string("obj_3_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + int32 obj_axis_0 = const()[name = string("obj_axis_0"), val = int32(0)]; + tensor obj_cast_fp16 = stack(axis = obj_axis_0, values = (input_lstm_layer_0_cast_fp16_2, input_cast_fp16_2))[name = string("obj_cast_fp16")]; + string obj_cast_fp16_to_fp32_dtype_0 = const()[name = string("obj_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + tensor transpose_0_perm_0 = const()[name = string("transpose_0_perm_0"), val = tensor([1, 2, 0])]; + string transpose_0_cast_fp16_to_fp32_dtype_0 = const()[name = string("transpose_0_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + tensor transpose_0_cast_fp16 = transpose(perm = transpose_0_perm_0, x = input_cast_fp16_0)[name = string("transpose_1")]; + tensor decoder_out = cast(dtype = transpose_0_cast_fp16_to_fp32_dtype_0, x = transpose_0_cast_fp16)[name = string("cast_2")]; + tensor c_out = cast(dtype = obj_cast_fp16_to_fp32_dtype_0, x = obj_cast_fp16)[name = string("cast_3")]; + tensor h_out = cast(dtype = obj_3_cast_fp16_to_fp32_dtype_0, x = obj_3_cast_fp16)[name = string("cast_4")]; + tensor token_length_tmp = identity(x = token_length)[name = string("token_length_tmp")]; + } -> (decoder_out, h_out, c_out); +} \ No newline at end of file diff --git a/ja/4480ms/decoder.mlmodelc/weights/weight.bin b/ja/4480ms/decoder.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..9bcbce3e617e1135c0460941dfc723b71230d030 --- /dev/null +++ b/ja/4480ms/decoder.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d5eeb17b5de822e941f66ce3bd6739aae939dfa054f61b69ff6ce89d0ac9e778 +size 14915072 diff --git a/ja/4480ms/decoder.mlpackage/Data/com.apple.CoreML/model.mlmodel b/ja/4480ms/decoder.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..931ee2253d124627cf1b2689c6e01d5cf3746838 --- /dev/null +++ b/ja/4480ms/decoder.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:25c461bd45595f33022b4ce50bf3d493d5b70ae73c50bd0a98598336bd38864a +size 11598 diff --git a/ja/4480ms/decoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/ja/4480ms/decoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..9bcbce3e617e1135c0460941dfc723b71230d030 --- /dev/null +++ b/ja/4480ms/decoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d5eeb17b5de822e941f66ce3bd6739aae939dfa054f61b69ff6ce89d0ac9e778 +size 14915072 diff --git a/ja/4480ms/decoder.mlpackage/Manifest.json b/ja/4480ms/decoder.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..b4431e1d0753f3273eeff30f89de1232349486a7 --- /dev/null +++ b/ja/4480ms/decoder.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "8C20B369-4E12-4E4E-B3E8-A79B91D9CAFC": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + }, + "9356FC01-CF91-4D74-A142-118AF15703DD": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + } + }, + "rootModelIdentifier": "9356FC01-CF91-4D74-A142-118AF15703DD" +} diff --git a/ja/4480ms/decoder_joint.mlmodelc/analytics/coremldata.bin b/ja/4480ms/decoder_joint.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..30ee1bc4e73ed57bede1d9e6315c983146d06e8c --- /dev/null +++ b/ja/4480ms/decoder_joint.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:adba69d0e8e1547064d062072f64e8a9f1da383a6d09e2986a28268dd78cb23c +size 243 diff --git a/ja/4480ms/decoder_joint.mlmodelc/coremldata.bin b/ja/4480ms/decoder_joint.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..ff735d9d794bd3717f0344022771f29df72a633d --- /dev/null +++ b/ja/4480ms/decoder_joint.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:85c85ecec674a9bee9777b7cf93682fd4cb5ea9bed388a030224a6f90dd72cde +size 514 diff --git a/ja/4480ms/decoder_joint.mlmodelc/model.mil b/ja/4480ms/decoder_joint.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..2cb099751b91d7bd911aeac28092cc495bcaf315 --- /dev/null +++ b/ja/4480ms/decoder_joint.mlmodelc/model.mil @@ -0,0 +1,92 @@ +program(1.3) +[buildInfo = dict({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.10.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor c_in, tensor encoder, tensor h_in, tensor token, tensor token_length) { + int32 y_batch_dims_0 = const()[name = string("y_batch_dims_0"), val = int32(0)]; + bool y_validate_indices_0 = const()[name = string("y_validate_indices_0"), val = bool(false)]; + tensor decoder_module_prediction_embed_weight_to_fp16 = const()[name = string("decoder_module_prediction_embed_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))]; + string token_to_int16_dtype_0 = const()[name = string("token_to_int16_dtype_0"), val = string("int16")]; + string cast_1_dtype_0 = const()[name = string("cast_1_dtype_0"), val = string("int32")]; + int32 greater_equal_0_y_0 = const()[name = string("greater_equal_0_y_0"), val = int32(0)]; + tensor token_to_int16 = cast(dtype = token_to_int16_dtype_0, x = token)[name = string("cast_10")]; + tensor cast_1 = cast(dtype = cast_1_dtype_0, x = token_to_int16)[name = string("cast_9")]; + tensor greater_equal_0 = greater_equal(x = cast_1, y = greater_equal_0_y_0)[name = string("greater_equal_0")]; + int32 slice_by_index_0 = const()[name = string("slice_by_index_0"), val = int32(1404)]; + tensor add_2 = add(x = cast_1, y = slice_by_index_0)[name = string("add_2")]; + tensor select_0 = select(a = cast_1, b = add_2, cond = greater_equal_0)[name = string("select_0")]; + int32 y_cast_fp16_cast_uint16_axis_0 = const()[name = string("y_cast_fp16_cast_uint16_axis_0"), val = int32(0)]; + string select_0_to_int16_dtype_0 = const()[name = string("select_0_to_int16_dtype_0"), val = string("int16")]; + tensor select_0_to_int16 = cast(dtype = select_0_to_int16_dtype_0, x = select_0)[name = string("cast_8")]; + tensor y_cast_fp16_cast_uint16_cast_uint16 = gather(axis = y_cast_fp16_cast_uint16_axis_0, batch_dims = y_batch_dims_0, indices = select_0_to_int16, validate_indices = y_validate_indices_0, x = decoder_module_prediction_embed_weight_to_fp16)[name = string("y_cast_fp16_cast_uint16_cast_uint16")]; + tensor input_3_perm_0 = const()[name = string("input_3_perm_0"), val = tensor([1, 0, 2])]; + int32 split_0_num_splits_0 = const()[name = string("split_0_num_splits_0"), val = int32(2)]; + int32 split_0_axis_0 = const()[name = string("split_0_axis_0"), val = int32(0)]; + string h_in_to_fp16_dtype_0 = const()[name = string("h_in_to_fp16_dtype_0"), val = string("fp16")]; + tensor h_in_to_fp16 = cast(dtype = h_in_to_fp16_dtype_0, x = h_in)[name = string("cast_7")]; + tensor split_0_cast_fp16_0, tensor split_0_cast_fp16_1 = split(axis = split_0_axis_0, num_splits = split_0_num_splits_0, x = h_in_to_fp16)[name = string("split_0_cast_fp16")]; + int32 split_1_num_splits_0 = const()[name = string("split_1_num_splits_0"), val = int32(2)]; + int32 split_1_axis_0 = const()[name = string("split_1_axis_0"), val = int32(0)]; + string c_in_to_fp16_dtype_0 = const()[name = string("c_in_to_fp16_dtype_0"), val = string("fp16")]; + tensor c_in_to_fp16 = cast(dtype = c_in_to_fp16_dtype_0, x = c_in)[name = string("cast_6")]; + tensor split_1_cast_fp16_0, tensor split_1_cast_fp16_1 = split(axis = split_1_axis_0, num_splits = split_1_num_splits_0, x = c_in_to_fp16)[name = string("split_1_cast_fp16")]; + tensor input_5_lstm_layer_0_lstm_h0_squeeze_axes_0 = const()[name = string("input_5_lstm_layer_0_lstm_h0_squeeze_axes_0"), val = tensor([0])]; + tensor input_5_lstm_layer_0_lstm_h0_squeeze_cast_fp16 = squeeze(axes = input_5_lstm_layer_0_lstm_h0_squeeze_axes_0, x = split_0_cast_fp16_0)[name = string("input_5_lstm_layer_0_lstm_h0_squeeze_cast_fp16")]; + tensor input_5_lstm_layer_0_lstm_c0_squeeze_axes_0 = const()[name = string("input_5_lstm_layer_0_lstm_c0_squeeze_axes_0"), val = tensor([0])]; + tensor input_5_lstm_layer_0_lstm_c0_squeeze_cast_fp16 = squeeze(axes = input_5_lstm_layer_0_lstm_c0_squeeze_axes_0, x = split_1_cast_fp16_0)[name = string("input_5_lstm_layer_0_lstm_c0_squeeze_cast_fp16")]; + string input_5_lstm_layer_0_direction_0 = const()[name = string("input_5_lstm_layer_0_direction_0"), val = string("forward")]; + bool input_5_lstm_layer_0_output_sequence_0 = const()[name = string("input_5_lstm_layer_0_output_sequence_0"), val = bool(true)]; + string input_5_lstm_layer_0_recurrent_activation_0 = const()[name = string("input_5_lstm_layer_0_recurrent_activation_0"), val = string("sigmoid")]; + string input_5_lstm_layer_0_cell_activation_0 = const()[name = string("input_5_lstm_layer_0_cell_activation_0"), val = string("tanh")]; + string input_5_lstm_layer_0_activation_0 = const()[name = string("input_5_lstm_layer_0_activation_0"), val = string("tanh")]; + tensor concat_1_to_fp16 = const()[name = string("concat_1_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1797248)))]; + tensor concat_2_to_fp16 = const()[name = string("concat_2_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5074112)))]; + tensor concat_0_to_fp16 = const()[name = string("concat_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8350976)))]; + tensor input_3_cast_fp16 = transpose(perm = input_3_perm_0, x = y_cast_fp16_cast_uint16_cast_uint16)[name = string("transpose_4")]; + tensor input_5_lstm_layer_0_cast_fp16_0, tensor input_5_lstm_layer_0_cast_fp16_1, tensor input_5_lstm_layer_0_cast_fp16_2 = lstm(activation = input_5_lstm_layer_0_activation_0, bias = concat_0_to_fp16, cell_activation = input_5_lstm_layer_0_cell_activation_0, direction = input_5_lstm_layer_0_direction_0, initial_c = input_5_lstm_layer_0_lstm_c0_squeeze_cast_fp16, initial_h = input_5_lstm_layer_0_lstm_h0_squeeze_cast_fp16, output_sequence = input_5_lstm_layer_0_output_sequence_0, recurrent_activation = input_5_lstm_layer_0_recurrent_activation_0, weight_hh = concat_2_to_fp16, weight_ih = concat_1_to_fp16, x = input_3_cast_fp16)[name = string("input_5_lstm_layer_0_cast_fp16")]; + tensor input_5_lstm_h0_squeeze_axes_0 = const()[name = string("input_5_lstm_h0_squeeze_axes_0"), val = tensor([0])]; + tensor input_5_lstm_h0_squeeze_cast_fp16 = squeeze(axes = input_5_lstm_h0_squeeze_axes_0, x = split_0_cast_fp16_1)[name = string("input_5_lstm_h0_squeeze_cast_fp16")]; + tensor input_5_lstm_c0_squeeze_axes_0 = const()[name = string("input_5_lstm_c0_squeeze_axes_0"), val = tensor([0])]; + tensor input_5_lstm_c0_squeeze_cast_fp16 = squeeze(axes = input_5_lstm_c0_squeeze_axes_0, x = split_1_cast_fp16_1)[name = string("input_5_lstm_c0_squeeze_cast_fp16")]; + string input_5_direction_0 = const()[name = string("input_5_direction_0"), val = string("forward")]; + bool input_5_output_sequence_0 = const()[name = string("input_5_output_sequence_0"), val = bool(true)]; + string input_5_recurrent_activation_0 = const()[name = string("input_5_recurrent_activation_0"), val = string("sigmoid")]; + string input_5_cell_activation_0 = const()[name = string("input_5_cell_activation_0"), val = string("tanh")]; + string input_5_activation_0 = const()[name = string("input_5_activation_0"), val = string("tanh")]; + tensor concat_4_to_fp16 = const()[name = string("concat_4_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8356160)))]; + tensor concat_5_to_fp16 = const()[name = string("concat_5_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11633024)))]; + tensor concat_3_to_fp16 = const()[name = string("concat_3_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14909888)))]; + tensor input_5_cast_fp16_0, tensor input_5_cast_fp16_1, tensor input_5_cast_fp16_2 = lstm(activation = input_5_activation_0, bias = concat_3_to_fp16, cell_activation = input_5_cell_activation_0, direction = input_5_direction_0, initial_c = input_5_lstm_c0_squeeze_cast_fp16, initial_h = input_5_lstm_h0_squeeze_cast_fp16, output_sequence = input_5_output_sequence_0, recurrent_activation = input_5_recurrent_activation_0, weight_hh = concat_5_to_fp16, weight_ih = concat_4_to_fp16, x = input_5_lstm_layer_0_cast_fp16_0)[name = string("input_5_cast_fp16")]; + int32 obj_3_axis_0 = const()[name = string("obj_3_axis_0"), val = int32(0)]; + tensor obj_3_cast_fp16 = stack(axis = obj_3_axis_0, values = (input_5_lstm_layer_0_cast_fp16_1, input_5_cast_fp16_1))[name = string("obj_3_cast_fp16")]; + string obj_3_cast_fp16_to_fp32_dtype_0 = const()[name = string("obj_3_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + int32 obj_axis_0 = const()[name = string("obj_axis_0"), val = int32(0)]; + tensor obj_cast_fp16 = stack(axis = obj_axis_0, values = (input_5_lstm_layer_0_cast_fp16_2, input_5_cast_fp16_2))[name = string("obj_cast_fp16")]; + string obj_cast_fp16_to_fp32_dtype_0 = const()[name = string("obj_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + tensor transpose_1_perm_0 = const()[name = string("transpose_1_perm_0"), val = tensor([1, 0, 2])]; + tensor input_7_perm_0 = const()[name = string("input_7_perm_0"), val = tensor([0, 2, 1])]; + string encoder_to_fp16_dtype_0 = const()[name = string("encoder_to_fp16_dtype_0"), val = string("fp16")]; + tensor joint_module_enc_weight_to_fp16 = const()[name = string("joint_module_enc_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14915072)))]; + tensor joint_module_enc_bias_to_fp16 = const()[name = string("joint_module_enc_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16225856)))]; + tensor encoder_to_fp16 = cast(dtype = encoder_to_fp16_dtype_0, x = encoder)[name = string("cast_3")]; + tensor input_7_cast_fp16 = transpose(perm = input_7_perm_0, x = encoder_to_fp16)[name = string("transpose_2")]; + tensor linear_0_cast_fp16 = linear(bias = joint_module_enc_bias_to_fp16, weight = joint_module_enc_weight_to_fp16, x = input_7_cast_fp16)[name = string("linear_0_cast_fp16")]; + tensor joint_module_pred_weight_to_fp16 = const()[name = string("joint_module_pred_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16227200)))]; + tensor joint_module_pred_bias_to_fp16 = const()[name = string("joint_module_pred_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17046464)))]; + tensor transpose_1_cast_fp16 = transpose(perm = transpose_1_perm_0, x = input_5_cast_fp16_0)[name = string("transpose_3")]; + tensor linear_1_cast_fp16 = linear(bias = joint_module_pred_bias_to_fp16, weight = joint_module_pred_weight_to_fp16, x = transpose_1_cast_fp16)[name = string("linear_1_cast_fp16")]; + tensor var_79_axes_0 = const()[name = string("op_79_axes_0"), val = tensor([2])]; + tensor var_79_cast_fp16 = expand_dims(axes = var_79_axes_0, x = linear_0_cast_fp16)[name = string("op_79_cast_fp16")]; + tensor var_80_axes_0 = const()[name = string("op_80_axes_0"), val = tensor([1])]; + tensor var_80_cast_fp16 = expand_dims(axes = var_80_axes_0, x = linear_1_cast_fp16)[name = string("op_80_cast_fp16")]; + tensor input_11_cast_fp16 = add(x = var_79_cast_fp16, y = var_80_cast_fp16)[name = string("input_11_cast_fp16")]; + tensor input_13_cast_fp16 = relu(x = input_11_cast_fp16)[name = string("input_13_cast_fp16")]; + tensor joint_module_joint_net_2_weight_to_fp16 = const()[name = string("joint_module_joint_net_2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17047808)))]; + tensor joint_module_joint_net_2_bias_to_fp16 = const()[name = string("joint_module_joint_net_2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18844992)))]; + tensor linear_2_cast_fp16 = linear(bias = joint_module_joint_net_2_bias_to_fp16, weight = joint_module_joint_net_2_weight_to_fp16, x = input_13_cast_fp16)[name = string("linear_2_cast_fp16")]; + string linear_2_cast_fp16_to_fp32_dtype_0 = const()[name = string("linear_2_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + tensor logits = cast(dtype = linear_2_cast_fp16_to_fp32_dtype_0, x = linear_2_cast_fp16)[name = string("cast_2")]; + tensor c_out = cast(dtype = obj_cast_fp16_to_fp32_dtype_0, x = obj_cast_fp16)[name = string("cast_4")]; + tensor h_out = cast(dtype = obj_3_cast_fp16_to_fp32_dtype_0, x = obj_3_cast_fp16)[name = string("cast_5")]; + tensor token_length_tmp = identity(x = token_length)[name = string("token_length_tmp")]; + } -> (logits, h_out, c_out); +} \ No newline at end of file diff --git a/ja/4480ms/decoder_joint.mlmodelc/weights/weight.bin b/ja/4480ms/decoder_joint.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..c3bb2e494d21dfd602e504fdfe76da274071d914 --- /dev/null +++ b/ja/4480ms/decoder_joint.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bfc1c768dd0e0e61c0ab8806894ecc03902d2a5028e9c30f5d0a5e38d5139fd9 +size 18847864 diff --git a/ja/4480ms/decoder_joint.mlpackage/Data/com.apple.CoreML/model.mlmodel b/ja/4480ms/decoder_joint.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..4b6b70ea10ec445416345cff7ae90b536c38ce69 --- /dev/null +++ b/ja/4480ms/decoder_joint.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dad926552d028aa609be27dde32ede7e7e0687ce48d057ee7bae6d472f15229e +size 14986 diff --git a/ja/4480ms/decoder_joint.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/ja/4480ms/decoder_joint.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..c3bb2e494d21dfd602e504fdfe76da274071d914 --- /dev/null +++ b/ja/4480ms/decoder_joint.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bfc1c768dd0e0e61c0ab8806894ecc03902d2a5028e9c30f5d0a5e38d5139fd9 +size 18847864 diff --git a/ja/4480ms/decoder_joint.mlpackage/Manifest.json b/ja/4480ms/decoder_joint.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..8a2ce629ffcccdfce8e228506afd499877ee4507 --- /dev/null +++ b/ja/4480ms/decoder_joint.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "4DAC4947-89EA-4167-8E25-F5740051D450": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + }, + "D23EE53F-5E4C-405E-B75A-E4553E8229D1": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + } + }, + "rootModelIdentifier": "D23EE53F-5E4C-405E-B75A-E4553E8229D1" +} diff --git a/ja/4480ms/encoder.mlmodelc/analytics/coremldata.bin b/ja/4480ms/encoder.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..049f0a1486544209f9d00d2dce6dda2e7f54becd --- /dev/null +++ b/ja/4480ms/encoder.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:556c53b26f7ee3745e8d6bef611421aa4b891cfa3b23ebe6f3407ed97650a3f8 +size 243 diff --git a/ja/4480ms/encoder.mlmodelc/coremldata.bin b/ja/4480ms/encoder.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..6514818a831bc1cf19634c34cfa7b7e9ca69c7a0 --- /dev/null +++ b/ja/4480ms/encoder.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:300b2383c2b9cf6e2e8c7a7b18bccb39e715ba9d1d2d8cfc84843e03e9c5e306 +size 573 diff --git a/ja/4480ms/encoder.mlmodelc/model.mil b/ja/4480ms/encoder.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..ac7ff9f2bb92afa45e5a2fabeefa184a92678a46 --- /dev/null +++ b/ja/4480ms/encoder.mlmodelc/model.mil @@ -0,0 +1,4321 @@ +program(1.3) +[buildInfo = dict({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}})] +{ + func main(tensor cache_channel, tensor cache_len, tensor cache_time, tensor mel, tensor mel_length, tensor prompt_id) { + tensor value_3_perm_0 = const()[name = string("value_3_perm_0"), val = tensor([1, 0, 2, 3])]; + string cache_channel_to_fp16_dtype_0 = const()[name = string("cache_channel_to_fp16_dtype_0"), val = string("fp16")]; + tensor value_5_perm_0 = const()[name = string("value_5_perm_0"), val = tensor([1, 0, 2, 3])]; + string cache_time_to_fp16_dtype_0 = const()[name = string("cache_time_to_fp16_dtype_0"), val = string("fp16")]; + int32 var_58 = const()[name = string("op_58"), val = int32(-1)]; + int32 var_67 = const()[name = string("op_67"), val = int32(1)]; + tensor x_1_perm_0 = const()[name = string("x_1_perm_0"), val = tensor([0, 2, 1])]; + string mel_to_fp16_dtype_0 = const()[name = string("mel_to_fp16_dtype_0"), val = string("fp16")]; + tensor tensor_1_axes_0 = const()[name = string("tensor_1_axes_0"), val = tensor([1])]; + tensor mel_to_fp16 = cast(dtype = mel_to_fp16_dtype_0, x = mel)[name = string("cast_18")]; + tensor x_1_cast_fp16 = transpose(perm = x_1_perm_0, x = mel_to_fp16)[name = string("transpose_367")]; + tensor tensor_1_cast_fp16 = expand_dims(axes = tensor_1_axes_0, x = x_1_cast_fp16)[name = string("tensor_1_cast_fp16")]; + tensor expand_dims_0 = const()[name = string("expand_dims_0"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))]; + tensor var_136_axes_0 = const()[name = string("op_136_axes_0"), val = tensor([1])]; + tensor var_136 = expand_dims(axes = var_136_axes_0, x = mel_length)[name = string("op_136")]; + tensor time_mask_1 = less(x = expand_dims_0, y = var_136)[name = string("time_mask_1")]; + tensor var_138_axes_0 = const()[name = string("op_138_axes_0"), val = tensor([-1])]; + tensor var_138 = expand_dims(axes = var_138_axes_0, x = time_mask_1)[name = string("op_138")]; + tensor var_140_reps_0 = const()[name = string("op_140_reps_0"), val = tensor([1, 1, 128])]; + tensor var_140 = tile(reps = var_140_reps_0, x = var_138)[name = string("op_140")]; + tensor var_146_axes_0 = const()[name = string("op_146_axes_0"), val = tensor([1])]; + string cast_4_to_fp16_dtype_0 = const()[name = string("cast_4_to_fp16_dtype_0"), val = string("fp16")]; + tensor var_140_to_fp16 = cast(dtype = cast_4_to_fp16_dtype_0, x = var_140)[name = string("cast_17")]; + tensor var_146_cast_fp16 = expand_dims(axes = var_146_axes_0, x = var_140_to_fp16)[name = string("op_146_cast_fp16")]; + tensor input_1_cast_fp16 = mul(x = tensor_1_cast_fp16, y = var_146_cast_fp16)[name = string("input_1_cast_fp16")]; + tensor input_3_pad_0 = const()[name = string("input_3_pad_0"), val = tensor([0, 0, 0, 0, 2, 1, 2, 1])]; + string input_3_mode_0 = const()[name = string("input_3_mode_0"), val = string("constant")]; + fp16 const_9_to_fp16 = const()[name = string("const_9_to_fp16"), val = fp16(0x0p+0)]; + tensor input_3_cast_fp16 = pad(constant_val = const_9_to_fp16, mode = input_3_mode_0, pad = input_3_pad_0, x = input_1_cast_fp16)[name = string("input_3_cast_fp16")]; + string tensor_3_pad_type_0 = const()[name = string("tensor_3_pad_type_0"), val = string("valid")]; + tensor tensor_3_strides_0 = const()[name = string("tensor_3_strides_0"), val = tensor([2, 2])]; + tensor tensor_3_pad_0 = const()[name = string("tensor_3_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor tensor_3_dilations_0 = const()[name = string("tensor_3_dilations_0"), val = tensor([1, 1])]; + int32 tensor_3_groups_0 = const()[name = string("tensor_3_groups_0"), val = int32(1)]; + tensor encoder_pre_encode_conv_0_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1984))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4352))))[name = string("encoder_pre_encode_conv_0_weight_to_fp16_quantized")]; + tensor encoder_pre_encode_conv_0_bias_to_fp16 = const()[name = string("encoder_pre_encode_conv_0_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4928)))]; + tensor tensor_3_cast_fp16 = conv(bias = encoder_pre_encode_conv_0_bias_to_fp16, dilations = tensor_3_dilations_0, groups = tensor_3_groups_0, pad = tensor_3_pad_0, pad_type = tensor_3_pad_type_0, strides = tensor_3_strides_0, weight = encoder_pre_encode_conv_0_weight_to_fp16_quantized, x = input_3_cast_fp16)[name = string("tensor_3_cast_fp16")]; + string cast_2_to_fp16_dtype_0 = const()[name = string("cast_2_to_fp16_dtype_0"), val = string("fp16")]; + fp16 var_159_promoted_to_fp16 = const()[name = string("op_159_promoted_to_fp16"), val = fp16(0x1p+1)]; + tensor mel_length_to_fp16 = cast(dtype = cast_2_to_fp16_dtype_0, x = mel_length)[name = string("cast_16")]; + tensor var_160_cast_fp16 = add(x = mel_length_to_fp16, y = var_159_promoted_to_fp16)[name = string("op_160_cast_fp16")]; + fp16 var_161_promoted_to_fp16 = const()[name = string("op_161_promoted_to_fp16"), val = fp16(0x1p+0)]; + tensor var_162_cast_fp16 = add(x = var_160_cast_fp16, y = var_161_promoted_to_fp16)[name = string("op_162_cast_fp16")]; + fp16 var_163_promoted_to_fp16 = const()[name = string("op_163_promoted_to_fp16"), val = fp16(0x1.8p+1)]; + tensor var_164_cast_fp16 = sub(x = var_162_cast_fp16, y = var_163_promoted_to_fp16)[name = string("op_164_cast_fp16")]; + fp16 var_55_promoted_to_fp16 = const()[name = string("op_55_promoted_to_fp16"), val = fp16(0x1p+1)]; + tensor floor_div_0_cast_fp16 = floor_div(x = var_164_cast_fp16, y = var_55_promoted_to_fp16)[name = string("floor_div_0_cast_fp16")]; + fp16 var_166_promoted_to_fp16 = const()[name = string("op_166_promoted_to_fp16"), val = fp16(0x1p+0)]; + tensor current_lengths_3_cast_fp16 = add(x = floor_div_0_cast_fp16, y = var_166_promoted_to_fp16)[name = string("current_lengths_3_cast_fp16")]; + string cast_5_dtype_0 = const()[name = string("cast_5_dtype_0"), val = string("int32")]; + tensor expand_dims_1 = const()[name = string("expand_dims_1"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5504)))]; + tensor var_175_axes_0 = const()[name = string("op_175_axes_0"), val = tensor([1])]; + tensor current_lengths_3_cast_fp16_to_int32 = cast(dtype = cast_5_dtype_0, x = current_lengths_3_cast_fp16)[name = string("cast_15")]; + tensor var_175 = expand_dims(axes = var_175_axes_0, x = current_lengths_3_cast_fp16_to_int32)[name = string("op_175")]; + tensor time_mask_3 = less(x = expand_dims_1, y = var_175)[name = string("time_mask_3")]; + tensor var_177_axes_0 = const()[name = string("op_177_axes_0"), val = tensor([-1])]; + tensor var_177 = expand_dims(axes = var_177_axes_0, x = time_mask_3)[name = string("op_177")]; + tensor var_179_reps_0 = const()[name = string("op_179_reps_0"), val = tensor([1, 1, 65])]; + tensor var_179 = tile(reps = var_179_reps_0, x = var_177)[name = string("op_179")]; + tensor var_185_axes_0 = const()[name = string("op_185_axes_0"), val = tensor([1])]; + string cast_6_to_fp16_dtype_0 = const()[name = string("cast_6_to_fp16_dtype_0"), val = string("fp16")]; + tensor var_179_to_fp16 = cast(dtype = cast_6_to_fp16_dtype_0, x = var_179)[name = string("cast_14")]; + tensor var_185_cast_fp16 = expand_dims(axes = var_185_axes_0, x = var_179_to_fp16)[name = string("op_185_cast_fp16")]; + tensor expanded_mask_3_reps_0 = const()[name = string("expanded_mask_3_reps_0"), val = tensor([1, 256, 1, 1])]; + tensor expanded_mask_3_cast_fp16 = tile(reps = expanded_mask_3_reps_0, x = var_185_cast_fp16)[name = string("expanded_mask_3_cast_fp16")]; + tensor input_5_cast_fp16 = mul(x = tensor_3_cast_fp16, y = expanded_mask_3_cast_fp16)[name = string("input_5_cast_fp16")]; + tensor tensor_5_cast_fp16 = relu(x = input_5_cast_fp16)[name = string("tensor_5_cast_fp16")]; + tensor input_7_cast_fp16 = mul(x = tensor_5_cast_fp16, y = expanded_mask_3_cast_fp16)[name = string("input_7_cast_fp16")]; + tensor input_9_pad_0 = const()[name = string("input_9_pad_0"), val = tensor([0, 0, 0, 0, 2, 1, 2, 1])]; + string input_9_mode_0 = const()[name = string("input_9_mode_0"), val = string("constant")]; + fp16 const_23_to_fp16 = const()[name = string("const_23_to_fp16"), val = fp16(0x0p+0)]; + tensor input_9_cast_fp16 = pad(constant_val = const_23_to_fp16, mode = input_9_mode_0, pad = input_9_pad_0, x = input_7_cast_fp16)[name = string("input_9_cast_fp16")]; + string tensor_7_pad_type_0 = const()[name = string("tensor_7_pad_type_0"), val = string("valid")]; + tensor tensor_7_strides_0 = const()[name = string("tensor_7_strides_0"), val = tensor([2, 2])]; + int32 tensor_7_groups_0 = const()[name = string("tensor_7_groups_0"), val = int32(256)]; + tensor tensor_7_pad_0 = const()[name = string("tensor_7_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor tensor_7_dilations_0 = const()[name = string("tensor_7_dilations_0"), val = tensor([1, 1])]; + tensor encoder_pre_encode_conv_2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6528))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8896))))[name = string("encoder_pre_encode_conv_2_weight_to_fp16_quantized")]; + tensor encoder_pre_encode_conv_2_bias_to_fp16 = const()[name = string("encoder_pre_encode_conv_2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(9472)))]; + tensor tensor_7_cast_fp16 = conv(bias = encoder_pre_encode_conv_2_bias_to_fp16, dilations = tensor_7_dilations_0, groups = tensor_7_groups_0, pad = tensor_7_pad_0, pad_type = tensor_7_pad_type_0, strides = tensor_7_strides_0, weight = encoder_pre_encode_conv_2_weight_to_fp16_quantized, x = input_9_cast_fp16)[name = string("tensor_7_cast_fp16")]; + fp16 var_207_promoted_to_fp16 = const()[name = string("op_207_promoted_to_fp16"), val = fp16(0x1p+1)]; + tensor var_208_cast_fp16 = add(x = current_lengths_3_cast_fp16, y = var_207_promoted_to_fp16)[name = string("op_208_cast_fp16")]; + fp16 var_209_promoted_to_fp16 = const()[name = string("op_209_promoted_to_fp16"), val = fp16(0x1p+0)]; + tensor var_210_cast_fp16 = add(x = var_208_cast_fp16, y = var_209_promoted_to_fp16)[name = string("op_210_cast_fp16")]; + fp16 var_211_promoted_to_fp16 = const()[name = string("op_211_promoted_to_fp16"), val = fp16(0x1.8p+1)]; + tensor var_212_cast_fp16 = sub(x = var_210_cast_fp16, y = var_211_promoted_to_fp16)[name = string("op_212_cast_fp16")]; + fp16 var_55_promoted_1_to_fp16 = const()[name = string("op_55_promoted_1_to_fp16"), val = fp16(0x1p+1)]; + tensor floor_div_1_cast_fp16 = floor_div(x = var_212_cast_fp16, y = var_55_promoted_1_to_fp16)[name = string("floor_div_1_cast_fp16")]; + fp16 var_214_promoted_to_fp16 = const()[name = string("op_214_promoted_to_fp16"), val = fp16(0x1p+0)]; + tensor current_lengths_5_cast_fp16 = add(x = floor_div_1_cast_fp16, y = var_214_promoted_to_fp16)[name = string("current_lengths_5_cast_fp16")]; + string cast_7_dtype_0 = const()[name = string("cast_7_dtype_0"), val = string("int32")]; + tensor expand_dims_2 = const()[name = string("expand_dims_2"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10048)))]; + tensor var_223_axes_0 = const()[name = string("op_223_axes_0"), val = tensor([1])]; + tensor current_lengths_5_cast_fp16_to_int32 = cast(dtype = cast_7_dtype_0, x = current_lengths_5_cast_fp16)[name = string("cast_13")]; + tensor var_223 = expand_dims(axes = var_223_axes_0, x = current_lengths_5_cast_fp16_to_int32)[name = string("op_223")]; + tensor time_mask_5 = less(x = expand_dims_2, y = var_223)[name = string("time_mask_5")]; + tensor var_225_axes_0 = const()[name = string("op_225_axes_0"), val = tensor([-1])]; + tensor var_225 = expand_dims(axes = var_225_axes_0, x = time_mask_5)[name = string("op_225")]; + tensor var_227_reps_0 = const()[name = string("op_227_reps_0"), val = tensor([1, 1, 33])]; + tensor var_227 = tile(reps = var_227_reps_0, x = var_225)[name = string("op_227")]; + tensor var_233_axes_0 = const()[name = string("op_233_axes_0"), val = tensor([1])]; + string cast_8_to_fp16_dtype_0 = const()[name = string("cast_8_to_fp16_dtype_0"), val = string("fp16")]; + tensor var_227_to_fp16 = cast(dtype = cast_8_to_fp16_dtype_0, x = var_227)[name = string("cast_12")]; + tensor var_233_cast_fp16 = expand_dims(axes = var_233_axes_0, x = var_227_to_fp16)[name = string("op_233_cast_fp16")]; + tensor expanded_mask_7_reps_0 = const()[name = string("expanded_mask_7_reps_0"), val = tensor([1, 256, 1, 1])]; + tensor expanded_mask_7_cast_fp16 = tile(reps = expanded_mask_7_reps_0, x = var_233_cast_fp16)[name = string("expanded_mask_7_cast_fp16")]; + tensor input_11_cast_fp16 = mul(x = tensor_7_cast_fp16, y = expanded_mask_7_cast_fp16)[name = string("input_11_cast_fp16")]; + string tensor_9_pad_type_0 = const()[name = string("tensor_9_pad_type_0"), val = string("valid")]; + tensor tensor_9_strides_0 = const()[name = string("tensor_9_strides_0"), val = tensor([1, 1])]; + tensor tensor_9_pad_0 = const()[name = string("tensor_9_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor tensor_9_dilations_0 = const()[name = string("tensor_9_dilations_0"), val = tensor([1, 1])]; + int32 tensor_9_groups_0 = const()[name = string("tensor_9_groups_0"), val = int32(1)]; + tensor encoder_pre_encode_conv_3_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10624))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76224))))[name = string("encoder_pre_encode_conv_3_weight_to_fp16_quantized")]; + tensor encoder_pre_encode_conv_3_bias_to_fp16 = const()[name = string("encoder_pre_encode_conv_3_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76800)))]; + tensor tensor_9_cast_fp16 = conv(bias = encoder_pre_encode_conv_3_bias_to_fp16, dilations = tensor_9_dilations_0, groups = tensor_9_groups_0, pad = tensor_9_pad_0, pad_type = tensor_9_pad_type_0, strides = tensor_9_strides_0, weight = encoder_pre_encode_conv_3_weight_to_fp16_quantized, x = input_11_cast_fp16)[name = string("tensor_9_cast_fp16")]; + tensor input_13_cast_fp16 = mul(x = tensor_9_cast_fp16, y = expanded_mask_7_cast_fp16)[name = string("input_13_cast_fp16")]; + tensor tensor_11_cast_fp16 = relu(x = input_13_cast_fp16)[name = string("tensor_11_cast_fp16")]; + tensor input_15_cast_fp16 = mul(x = tensor_11_cast_fp16, y = expanded_mask_7_cast_fp16)[name = string("input_15_cast_fp16")]; + tensor input_17_pad_0 = const()[name = string("input_17_pad_0"), val = tensor([0, 0, 0, 0, 2, 1, 2, 1])]; + string input_17_mode_0 = const()[name = string("input_17_mode_0"), val = string("constant")]; + fp16 const_41_to_fp16 = const()[name = string("const_41_to_fp16"), val = fp16(0x0p+0)]; + tensor input_17_cast_fp16 = pad(constant_val = const_41_to_fp16, mode = input_17_mode_0, pad = input_17_pad_0, x = input_15_cast_fp16)[name = string("input_17_cast_fp16")]; + string tensor_13_pad_type_0 = const()[name = string("tensor_13_pad_type_0"), val = string("valid")]; + tensor tensor_13_strides_0 = const()[name = string("tensor_13_strides_0"), val = tensor([2, 2])]; + int32 tensor_13_groups_0 = const()[name = string("tensor_13_groups_0"), val = int32(256)]; + tensor tensor_13_pad_0 = const()[name = string("tensor_13_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor tensor_13_dilations_0 = const()[name = string("tensor_13_dilations_0"), val = tensor([1, 1])]; + tensor encoder_pre_encode_conv_5_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77376))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(79744))))[name = string("encoder_pre_encode_conv_5_weight_to_fp16_quantized")]; + tensor encoder_pre_encode_conv_5_bias_to_fp16 = const()[name = string("encoder_pre_encode_conv_5_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(80320)))]; + tensor tensor_13_cast_fp16 = conv(bias = encoder_pre_encode_conv_5_bias_to_fp16, dilations = tensor_13_dilations_0, groups = tensor_13_groups_0, pad = tensor_13_pad_0, pad_type = tensor_13_pad_type_0, strides = tensor_13_strides_0, weight = encoder_pre_encode_conv_5_weight_to_fp16_quantized, x = input_17_cast_fp16)[name = string("tensor_13_cast_fp16")]; + fp16 var_270_promoted_to_fp16 = const()[name = string("op_270_promoted_to_fp16"), val = fp16(0x1p+1)]; + tensor var_271_cast_fp16 = add(x = current_lengths_5_cast_fp16, y = var_270_promoted_to_fp16)[name = string("op_271_cast_fp16")]; + fp16 var_272_promoted_to_fp16 = const()[name = string("op_272_promoted_to_fp16"), val = fp16(0x1p+0)]; + tensor var_273_cast_fp16 = add(x = var_271_cast_fp16, y = var_272_promoted_to_fp16)[name = string("op_273_cast_fp16")]; + fp16 var_274_promoted_to_fp16 = const()[name = string("op_274_promoted_to_fp16"), val = fp16(0x1.8p+1)]; + tensor var_275_cast_fp16 = sub(x = var_273_cast_fp16, y = var_274_promoted_to_fp16)[name = string("op_275_cast_fp16")]; + fp16 var_55_promoted_2_to_fp16 = const()[name = string("op_55_promoted_2_to_fp16"), val = fp16(0x1p+1)]; + tensor floor_div_2_cast_fp16 = floor_div(x = var_275_cast_fp16, y = var_55_promoted_2_to_fp16)[name = string("floor_div_2_cast_fp16")]; + fp16 var_277_promoted_to_fp16 = const()[name = string("op_277_promoted_to_fp16"), val = fp16(0x1p+0)]; + tensor current_lengths_cast_fp16 = add(x = floor_div_2_cast_fp16, y = var_277_promoted_to_fp16)[name = string("current_lengths_cast_fp16")]; + string cast_9_dtype_0 = const()[name = string("cast_9_dtype_0"), val = string("int32")]; + tensor expand_dims_3 = const()[name = string("expand_dims_3"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(80896)))]; + tensor var_286_axes_0 = const()[name = string("op_286_axes_0"), val = tensor([1])]; + tensor current_lengths_cast_fp16_to_int32 = cast(dtype = cast_9_dtype_0, x = current_lengths_cast_fp16)[name = string("cast_11")]; + tensor var_286 = expand_dims(axes = var_286_axes_0, x = current_lengths_cast_fp16_to_int32)[name = string("op_286")]; + tensor time_mask = less(x = expand_dims_3, y = var_286)[name = string("time_mask")]; + tensor var_288_axes_0 = const()[name = string("op_288_axes_0"), val = tensor([-1])]; + tensor var_288 = expand_dims(axes = var_288_axes_0, x = time_mask)[name = string("op_288")]; + tensor var_290_reps_0 = const()[name = string("op_290_reps_0"), val = tensor([1, 1, 17])]; + tensor var_290 = tile(reps = var_290_reps_0, x = var_288)[name = string("op_290")]; + tensor var_296_axes_0 = const()[name = string("op_296_axes_0"), val = tensor([1])]; + string cast_10_to_fp16_dtype_0 = const()[name = string("cast_10_to_fp16_dtype_0"), val = string("fp16")]; + tensor var_290_to_fp16 = cast(dtype = cast_10_to_fp16_dtype_0, x = var_290)[name = string("cast_10")]; + tensor var_296_cast_fp16 = expand_dims(axes = var_296_axes_0, x = var_290_to_fp16)[name = string("op_296_cast_fp16")]; + tensor expanded_mask_13_reps_0 = const()[name = string("expanded_mask_13_reps_0"), val = tensor([1, 256, 1, 1])]; + tensor expanded_mask_13_cast_fp16 = tile(reps = expanded_mask_13_reps_0, x = var_296_cast_fp16)[name = string("expanded_mask_13_cast_fp16")]; + tensor input_19_cast_fp16 = mul(x = tensor_13_cast_fp16, y = expanded_mask_13_cast_fp16)[name = string("input_19_cast_fp16")]; + string tensor_15_pad_type_0 = const()[name = string("tensor_15_pad_type_0"), val = string("valid")]; + tensor tensor_15_strides_0 = const()[name = string("tensor_15_strides_0"), val = tensor([1, 1])]; + tensor tensor_15_pad_0 = const()[name = string("tensor_15_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor tensor_15_dilations_0 = const()[name = string("tensor_15_dilations_0"), val = tensor([1, 1])]; + int32 tensor_15_groups_0 = const()[name = string("tensor_15_groups_0"), val = int32(1)]; + tensor encoder_pre_encode_conv_6_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81216))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(146816))))[name = string("encoder_pre_encode_conv_6_weight_to_fp16_quantized")]; + tensor encoder_pre_encode_conv_6_bias_to_fp16 = const()[name = string("encoder_pre_encode_conv_6_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(147392)))]; + tensor tensor_15_cast_fp16 = conv(bias = encoder_pre_encode_conv_6_bias_to_fp16, dilations = tensor_15_dilations_0, groups = tensor_15_groups_0, pad = tensor_15_pad_0, pad_type = tensor_15_pad_type_0, strides = tensor_15_strides_0, weight = encoder_pre_encode_conv_6_weight_to_fp16_quantized, x = input_19_cast_fp16)[name = string("tensor_15_cast_fp16")]; + tensor input_21_cast_fp16 = mul(x = tensor_15_cast_fp16, y = expanded_mask_13_cast_fp16)[name = string("input_21_cast_fp16")]; + tensor tensor_cast_fp16 = relu(x = input_21_cast_fp16)[name = string("tensor_cast_fp16")]; + tensor x_3_cast_fp16 = mul(x = tensor_cast_fp16, y = expanded_mask_13_cast_fp16)[name = string("x_3_cast_fp16")]; + tensor var_330_perm_0 = const()[name = string("op_330_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_331 = const()[name = string("op_331"), val = tensor([1, 58, -1])]; + tensor var_330_cast_fp16 = transpose(perm = var_330_perm_0, x = x_3_cast_fp16)[name = string("transpose_366")]; + tensor input_23_cast_fp16 = reshape(shape = var_331, x = var_330_cast_fp16)[name = string("input_23_cast_fp16")]; + tensor encoder_pre_encode_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(147968))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4604480))))[name = string("encoder_pre_encode_out_weight_to_fp16_quantized")]; + tensor encoder_pre_encode_out_bias_to_fp16 = const()[name = string("encoder_pre_encode_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4606592)))]; + tensor linear_0_cast_fp16 = linear(bias = encoder_pre_encode_out_bias_to_fp16, weight = encoder_pre_encode_out_weight_to_fp16_quantized, x = input_23_cast_fp16)[name = string("linear_0_cast_fp16")]; + tensor var_341_begin_0 = const()[name = string("op_341_begin_0"), val = tensor([0, 2, 0])]; + tensor var_341_end_0 = const()[name = string("op_341_end_0"), val = tensor([1, 58, 1024])]; + tensor var_341_end_mask_0 = const()[name = string("op_341_end_mask_0"), val = tensor([true, true, true])]; + tensor var_341_cast_fp16 = slice_by_index(begin = var_341_begin_0, end = var_341_end_0, end_mask = var_341_end_mask_0, x = linear_0_cast_fp16)[name = string("op_341_cast_fp16")]; + int32 var_343 = const()[name = string("op_343"), val = int32(2)]; + tensor var_344 = sub(x = current_lengths_cast_fp16_to_int32, y = var_343)[name = string("op_344")]; + string var_344_promoted_to_fp16_dtype_0 = const()[name = string("op_344_promoted_to_fp16_dtype_0"), val = string("fp16")]; + fp16 var_61_promoted_to_fp16 = const()[name = string("op_61_promoted_to_fp16"), val = fp16(0x0p+0)]; + fp16 const_61_to_fp16 = const()[name = string("const_61_to_fp16"), val = fp16(inf)]; + tensor var_344_to_fp16 = cast(dtype = var_344_promoted_to_fp16_dtype_0, x = var_344)[name = string("cast_9")]; + tensor clip_0_cast_fp16 = clip(alpha = var_61_promoted_to_fp16, beta = const_61_to_fp16, x = var_344_to_fp16)[name = string("clip_0_cast_fp16")]; + tensor max_audio_length_1 = const()[name = string("max_audio_length_1"), val = tensor([56])]; + fp16 var_360_promoted_to_fp16 = const()[name = string("op_360_promoted_to_fp16"), val = fp16(0x1.5p+5)]; + tensor padding_length_cast_fp16 = add(x = clip_0_cast_fp16, y = var_360_promoted_to_fp16)[name = string("padding_length_cast_fp16")]; + int32 const_63 = const()[name = string("const_63"), val = int32(-1)]; + tensor var_362 = mul(x = cache_len, y = const_63)[name = string("op_362")]; + int32 var_363 = const()[name = string("op_363"), val = int32(42)]; + tensor offset = add(x = var_362, y = var_363)[name = string("offset")]; + tensor var_403_axes_0 = const()[name = string("op_403_axes_0"), val = tensor([-1])]; + tensor var_403_cast_fp16 = expand_dims(axes = var_403_axes_0, x = padding_length_cast_fp16)[name = string("op_403_cast_fp16")]; + tensor var_402_promoted_to_fp16 = const()[name = string("op_402_promoted_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4608704)))]; + tensor pad_mask_1_cast_fp16 = less(x = var_402_promoted_to_fp16, y = var_403_cast_fp16)[name = string("pad_mask_1_cast_fp16")]; + tensor expand_dims_5 = const()[name = string("expand_dims_5"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4609024)))]; + tensor var_409_axes_0 = const()[name = string("op_409_axes_0"), val = tensor([-1])]; + tensor var_409 = expand_dims(axes = var_409_axes_0, x = offset)[name = string("op_409")]; + tensor pad_mask_off = greater_equal(x = expand_dims_5, y = var_409)[name = string("pad_mask_off")]; + tensor pad_mask_3 = logical_and(x = pad_mask_off, y = pad_mask_1_cast_fp16)[name = string("pad_mask_3")]; + tensor var_412_axes_0 = const()[name = string("op_412_axes_0"), val = tensor([1])]; + tensor var_412 = expand_dims(axes = var_412_axes_0, x = pad_mask_3)[name = string("op_412")]; + tensor var_413 = const()[name = string("op_413"), val = tensor([1, 98, 1])]; + tensor pad_mask_for_att_mask_1 = tile(reps = var_413, x = var_412)[name = string("pad_mask_for_att_mask_1")]; + tensor var_415_perm_0 = const()[name = string("op_415_perm_0"), val = tensor([0, 2, 1])]; + tensor var_415 = transpose(perm = var_415_perm_0, x = pad_mask_for_att_mask_1)[name = string("transpose_365")]; + tensor pad_mask_for_att_mask = logical_and(x = pad_mask_for_att_mask_1, y = var_415)[name = string("pad_mask_for_att_mask")]; + tensor const_71 = const()[name = string("const_71"), val = tensor([[[true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true]]])]; + tensor att_mask_9 = logical_and(x = pad_mask_for_att_mask, y = const_71)[name = string("att_mask_9")]; + tensor att_mask = logical_not(x = att_mask_9)[name = string("att_mask")]; + tensor pad_mask_5 = logical_not(x = pad_mask_3)[name = string("pad_mask_5")]; + tensor pad_mask_begin_0 = const()[name = string("pad_mask_begin_0"), val = tensor([0, 42])]; + tensor pad_mask_end_0 = const()[name = string("pad_mask_end_0"), val = tensor([1, 98])]; + tensor pad_mask_end_mask_0 = const()[name = string("pad_mask_end_mask_0"), val = tensor([true, true])]; + tensor pad_mask = slice_by_index(begin = pad_mask_begin_0, end = pad_mask_end_0, end_mask = pad_mask_end_mask_0, x = pad_mask_5)[name = string("pad_mask")]; + tensor mask_9_begin_0 = const()[name = string("mask_9_begin_0"), val = tensor([0, 42, 0])]; + tensor mask_9_end_0 = const()[name = string("mask_9_end_0"), val = tensor([1, 98, 98])]; + tensor mask_9_end_mask_0 = const()[name = string("mask_9_end_mask_0"), val = tensor([true, true, true])]; + tensor mask_9 = slice_by_index(begin = mask_9_begin_0, end = mask_9_end_0, end_mask = mask_9_end_mask_0, x = att_mask)[name = string("mask_9")]; + tensor cache_1_begin_0 = const()[name = string("cache_1_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor cache_1_end_0 = const()[name = string("cache_1_end_0"), val = tensor([1, 1, 42, 1024])]; + tensor cache_1_end_mask_0 = const()[name = string("cache_1_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_1_squeeze_mask_0 = const()[name = string("cache_1_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_channel_to_fp16 = cast(dtype = cache_channel_to_fp16_dtype_0, x = cache_channel)[name = string("cast_8")]; + tensor value_3_cast_fp16 = transpose(perm = value_3_perm_0, x = cache_channel_to_fp16)[name = string("transpose_364")]; + tensor cache_1_cast_fp16 = slice_by_index(begin = cache_1_begin_0, end = cache_1_end_0, end_mask = cache_1_end_mask_0, squeeze_mask = cache_1_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_1_cast_fp16")]; + tensor cache_3_begin_0 = const()[name = string("cache_3_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor cache_3_end_0 = const()[name = string("cache_3_end_0"), val = tensor([1, 1, 1024, 8])]; + tensor cache_3_end_mask_0 = const()[name = string("cache_3_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_3_squeeze_mask_0 = const()[name = string("cache_3_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_time_to_fp16 = cast(dtype = cache_time_to_fp16_dtype_0, x = cache_time)[name = string("cast_7")]; + tensor value_5_cast_fp16 = transpose(perm = value_5_perm_0, x = cache_time_to_fp16)[name = string("transpose_363")]; + tensor cache_3_cast_fp16 = slice_by_index(begin = cache_3_begin_0, end = cache_3_end_0, end_mask = cache_3_end_mask_0, squeeze_mask = cache_3_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_3_cast_fp16")]; + tensor input_27_axes_0 = const()[name = string("input_27_axes_0"), val = tensor([-1])]; + tensor encoder_layers_0_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_0_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4609536)))]; + tensor encoder_layers_0_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_0_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4611648)))]; + fp16 var_41_to_fp16 = const()[name = string("op_41_to_fp16"), val = fp16(0x1.5p-17)]; + tensor input_27_cast_fp16 = layer_norm(axes = input_27_axes_0, beta = encoder_layers_0_norm_feed_forward1_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_0_norm_feed_forward1_weight_to_fp16, x = var_341_cast_fp16)[name = string("input_27_cast_fp16")]; + tensor encoder_layers_0_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4613760))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8808128))))[name = string("encoder_layers_0_feed_forward1_linear1_weight_to_fp16_quantized")]; + tensor encoder_layers_0_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_0_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8816384)))]; + tensor linear_1_cast_fp16 = linear(bias = encoder_layers_0_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_0_feed_forward1_linear1_weight_to_fp16_quantized, x = input_27_cast_fp16)[name = string("linear_1_cast_fp16")]; + tensor input_31_cast_fp16 = silu(x = linear_1_cast_fp16)[name = string("input_31_cast_fp16")]; + tensor encoder_layers_0_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8824640))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13019008))))[name = string("encoder_layers_0_feed_forward1_linear2_weight_to_fp16_quantized")]; + tensor encoder_layers_0_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_0_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13021120)))]; + tensor linear_2_cast_fp16 = linear(bias = encoder_layers_0_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_0_feed_forward1_linear2_weight_to_fp16_quantized, x = input_31_cast_fp16)[name = string("linear_2_cast_fp16")]; + fp16 var_454_to_fp16 = const()[name = string("op_454_to_fp16"), val = fp16(0x1p-1)]; + tensor var_455_cast_fp16 = mul(x = linear_2_cast_fp16, y = var_454_to_fp16)[name = string("op_455_cast_fp16")]; + tensor input_37_cast_fp16 = add(x = var_341_cast_fp16, y = var_455_cast_fp16)[name = string("input_37_cast_fp16")]; + tensor key_1_axes_0 = const()[name = string("key_1_axes_0"), val = tensor([-1])]; + tensor encoder_layers_0_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_0_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13023232)))]; + tensor encoder_layers_0_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_0_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13025344)))]; + tensor key_1_cast_fp16 = layer_norm(axes = key_1_axes_0, beta = encoder_layers_0_norm_self_att_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_0_norm_self_att_weight_to_fp16, x = input_37_cast_fp16)[name = string("key_1_cast_fp16")]; + bool input_39_interleave_0 = const()[name = string("input_39_interleave_0"), val = bool(false)]; + tensor input_39_cast_fp16 = concat(axis = var_67, interleave = input_39_interleave_0, values = (cache_1_cast_fp16, key_1_cast_fp16))[name = string("input_39_cast_fp16")]; + bool var_483_interleave_0 = const()[name = string("op_483_interleave_0"), val = bool(false)]; + tensor var_483_cast_fp16 = concat(axis = var_67, interleave = var_483_interleave_0, values = key_1_cast_fp16)[name = string("op_483_cast_fp16")]; + tensor encoder_layers_0_self_attn_linear_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13027456))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14076096))))[name = string("encoder_layers_0_self_attn_linear_q_weight_to_fp16_quantized")]; + tensor encoder_layers_0_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_0_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14078208)))]; + tensor linear_3_cast_fp16 = linear(bias = encoder_layers_0_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_0_self_attn_linear_q_weight_to_fp16_quantized, x = key_1_cast_fp16)[name = string("linear_3_cast_fp16")]; + tensor var_488 = const()[name = string("op_488"), val = tensor([1, -1, 8, 128])]; + tensor q_1_cast_fp16 = reshape(shape = var_488, x = linear_3_cast_fp16)[name = string("q_1_cast_fp16")]; + tensor encoder_layers_0_self_attn_linear_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14080320))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15128960))))[name = string("encoder_layers_0_self_attn_linear_k_weight_to_fp16_quantized")]; + tensor encoder_layers_0_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_0_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15131072)))]; + tensor linear_4_cast_fp16 = linear(bias = encoder_layers_0_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_0_self_attn_linear_k_weight_to_fp16_quantized, x = input_39_cast_fp16)[name = string("linear_4_cast_fp16")]; + tensor var_493 = const()[name = string("op_493"), val = tensor([1, -1, 8, 128])]; + tensor k_1_cast_fp16 = reshape(shape = var_493, x = linear_4_cast_fp16)[name = string("k_1_cast_fp16")]; + tensor encoder_layers_0_self_attn_linear_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15133184))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16181824))))[name = string("encoder_layers_0_self_attn_linear_v_weight_to_fp16_quantized")]; + tensor encoder_layers_0_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_0_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16183936)))]; + tensor linear_5_cast_fp16 = linear(bias = encoder_layers_0_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_0_self_attn_linear_v_weight_to_fp16_quantized, x = input_39_cast_fp16)[name = string("linear_5_cast_fp16")]; + tensor var_498 = const()[name = string("op_498"), val = tensor([1, -1, 8, 128])]; + tensor v_1_cast_fp16 = reshape(shape = var_498, x = linear_5_cast_fp16)[name = string("v_1_cast_fp16")]; + tensor value_9_perm_0 = const()[name = string("value_9_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_0_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_0_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16186048)))]; + tensor var_511_cast_fp16 = add(x = q_1_cast_fp16, y = encoder_layers_0_self_attn_pos_bias_u_to_fp16)[name = string("op_511_cast_fp16")]; + tensor encoder_layers_0_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_0_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16188160)))]; + tensor var_513_cast_fp16 = add(x = q_1_cast_fp16, y = encoder_layers_0_self_attn_pos_bias_v_to_fp16)[name = string("op_513_cast_fp16")]; + tensor q_with_bias_v_1_perm_0 = const()[name = string("q_with_bias_v_1_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_7_transpose_x_0 = const()[name = string("x_7_transpose_x_0"), val = bool(false)]; + bool x_7_transpose_y_0 = const()[name = string("x_7_transpose_y_0"), val = bool(false)]; + tensor op_515_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16190272))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16390016))))[name = string("op_515_to_fp16_quantized")]; + tensor q_with_bias_v_1_cast_fp16 = transpose(perm = q_with_bias_v_1_perm_0, x = var_513_cast_fp16)[name = string("transpose_362")]; + tensor x_7_cast_fp16 = matmul(transpose_x = x_7_transpose_x_0, transpose_y = x_7_transpose_y_0, x = q_with_bias_v_1_cast_fp16, y = op_515_to_fp16_quantized)[name = string("x_7_cast_fp16")]; + tensor x_9_pad_0 = const()[name = string("x_9_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_9_mode_0 = const()[name = string("x_9_mode_0"), val = string("constant")]; + fp16 const_79_to_fp16 = const()[name = string("const_79_to_fp16"), val = fp16(0x0p+0)]; + tensor x_9_cast_fp16 = pad(constant_val = const_79_to_fp16, mode = x_9_mode_0, pad = x_9_pad_0, x = x_7_cast_fp16)[name = string("x_9_cast_fp16")]; + tensor var_523 = const()[name = string("op_523"), val = tensor([1, 8, -1, 56])]; + tensor x_11_cast_fp16 = reshape(shape = var_523, x = x_9_cast_fp16)[name = string("x_11_cast_fp16")]; + tensor var_527_begin_0 = const()[name = string("op_527_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_527_end_0 = const()[name = string("op_527_end_0"), val = tensor([1, 8, 196, 56])]; + tensor var_527_end_mask_0 = const()[name = string("op_527_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_527_cast_fp16 = slice_by_index(begin = var_527_begin_0, end = var_527_end_0, end_mask = var_527_end_mask_0, x = x_11_cast_fp16)[name = string("op_527_cast_fp16")]; + tensor var_528 = const()[name = string("op_528"), val = tensor([1, 8, 56, 195])]; + tensor matrix_bd_1_cast_fp16 = reshape(shape = var_528, x = var_527_cast_fp16)[name = string("matrix_bd_1_cast_fp16")]; + bool matrix_ac_1_transpose_x_0 = const()[name = string("matrix_ac_1_transpose_x_0"), val = bool(false)]; + bool matrix_ac_1_transpose_y_0 = const()[name = string("matrix_ac_1_transpose_y_0"), val = bool(false)]; + tensor transpose_96_perm_0 = const()[name = string("transpose_96_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_97_perm_0 = const()[name = string("transpose_97_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_97 = transpose(perm = transpose_97_perm_0, x = k_1_cast_fp16)[name = string("transpose_360")]; + tensor transpose_96 = transpose(perm = transpose_96_perm_0, x = var_511_cast_fp16)[name = string("transpose_361")]; + tensor matrix_ac_1_cast_fp16 = matmul(transpose_x = matrix_ac_1_transpose_x_0, transpose_y = matrix_ac_1_transpose_y_0, x = transpose_96, y = transpose_97)[name = string("matrix_ac_1_cast_fp16")]; + tensor matrix_bd_3_begin_0 = const()[name = string("matrix_bd_3_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_3_end_0 = const()[name = string("matrix_bd_3_end_0"), val = tensor([1, 8, 56, 98])]; + tensor matrix_bd_3_end_mask_0 = const()[name = string("matrix_bd_3_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_3_cast_fp16 = slice_by_index(begin = matrix_bd_3_begin_0, end = matrix_bd_3_end_0, end_mask = matrix_bd_3_end_mask_0, x = matrix_bd_1_cast_fp16)[name = string("matrix_bd_3_cast_fp16")]; + tensor var_537_cast_fp16 = add(x = matrix_ac_1_cast_fp16, y = matrix_bd_3_cast_fp16)[name = string("op_537_cast_fp16")]; + fp16 _inversed_scores_1_y_0_to_fp16 = const()[name = string("_inversed_scores_1_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_1_cast_fp16 = mul(x = var_537_cast_fp16, y = _inversed_scores_1_y_0_to_fp16)[name = string("_inversed_scores_1_cast_fp16")]; + tensor mask_11_axes_0 = const()[name = string("mask_11_axes_0"), val = tensor([1])]; + tensor mask_11 = expand_dims(axes = mask_11_axes_0, x = mask_9)[name = string("mask_11")]; + fp16 var_44_to_fp16 = const()[name = string("op_44_to_fp16"), val = fp16(-0x1.388p+13)]; + tensor scores_3_cast_fp16 = select(a = var_44_to_fp16, b = _inversed_scores_1_cast_fp16, cond = mask_11)[name = string("scores_3_cast_fp16")]; + tensor var_543_cast_fp16 = softmax(axis = var_58, x = scores_3_cast_fp16)[name = string("op_543_cast_fp16")]; + fp16 var_43_to_fp16 = const()[name = string("op_43_to_fp16"), val = fp16(0x0p+0)]; + tensor input_41_cast_fp16 = select(a = var_43_to_fp16, b = var_543_cast_fp16, cond = mask_11)[name = string("input_41_cast_fp16")]; + bool x_13_transpose_x_0 = const()[name = string("x_13_transpose_x_0"), val = bool(false)]; + bool x_13_transpose_y_0 = const()[name = string("x_13_transpose_y_0"), val = bool(false)]; + tensor value_9_cast_fp16 = transpose(perm = value_9_perm_0, x = v_1_cast_fp16)[name = string("transpose_359")]; + tensor x_13_cast_fp16 = matmul(transpose_x = x_13_transpose_x_0, transpose_y = x_13_transpose_y_0, x = input_41_cast_fp16, y = value_9_cast_fp16)[name = string("x_13_cast_fp16")]; + tensor var_547_perm_0 = const()[name = string("op_547_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_548 = const()[name = string("op_548"), val = tensor([1, -1, 1024])]; + tensor var_547_cast_fp16 = transpose(perm = var_547_perm_0, x = x_13_cast_fp16)[name = string("transpose_358")]; + tensor input_43_cast_fp16 = reshape(shape = var_548, x = var_547_cast_fp16)[name = string("input_43_cast_fp16")]; + tensor encoder_layers_0_self_attn_linear_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16390528))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17439168))))[name = string("encoder_layers_0_self_attn_linear_out_weight_to_fp16_quantized")]; + tensor encoder_layers_0_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_0_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17441280)))]; + tensor linear_7_cast_fp16 = linear(bias = encoder_layers_0_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_0_self_attn_linear_out_weight_to_fp16_quantized, x = input_43_cast_fp16)[name = string("linear_7_cast_fp16")]; + tensor input_47_cast_fp16 = add(x = input_37_cast_fp16, y = linear_7_cast_fp16)[name = string("input_47_cast_fp16")]; + tensor x_17_axes_0 = const()[name = string("x_17_axes_0"), val = tensor([-1])]; + tensor encoder_layers_0_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_0_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17443392)))]; + tensor encoder_layers_0_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_0_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17445504)))]; + tensor x_17_cast_fp16 = layer_norm(axes = x_17_axes_0, beta = encoder_layers_0_norm_conv_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_0_norm_conv_weight_to_fp16, x = input_47_cast_fp16)[name = string("x_17_cast_fp16")]; + tensor input_49_perm_0 = const()[name = string("input_49_perm_0"), val = tensor([0, 2, 1])]; + string input_51_pad_type_0 = const()[name = string("input_51_pad_type_0"), val = string("valid")]; + tensor input_51_strides_0 = const()[name = string("input_51_strides_0"), val = tensor([1])]; + tensor input_51_pad_0 = const()[name = string("input_51_pad_0"), val = tensor([0, 0])]; + tensor input_51_dilations_0 = const()[name = string("input_51_dilations_0"), val = tensor([1])]; + int32 input_51_groups_0 = const()[name = string("input_51_groups_0"), val = int32(1)]; + tensor encoder_layers_0_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17447616))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19544832))))[name = string("encoder_layers_0_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_49_cast_fp16 = transpose(perm = input_49_perm_0, x = x_17_cast_fp16)[name = string("transpose_357")]; + tensor input_51_cast_fp16 = conv(dilations = input_51_dilations_0, groups = input_51_groups_0, pad = input_51_pad_0, pad_type = input_51_pad_type_0, strides = input_51_strides_0, weight = encoder_layers_0_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_49_cast_fp16)[name = string("input_51_cast_fp16")]; + int32 x_19_split_num_splits_0 = const()[name = string("x_19_split_num_splits_0"), val = int32(2)]; + int32 x_19_split_axis_0 = const()[name = string("x_19_split_axis_0"), val = int32(1)]; + tensor x_19_split_cast_fp16_0, tensor x_19_split_cast_fp16_1 = split(axis = x_19_split_axis_0, num_splits = x_19_split_num_splits_0, x = input_51_cast_fp16)[name = string("x_19_split_cast_fp16")]; + tensor x_19_split_1_sigmoid_cast_fp16 = sigmoid(x = x_19_split_cast_fp16_1)[name = string("x_19_split_1_sigmoid_cast_fp16")]; + tensor x_19_cast_fp16 = mul(x = x_19_split_cast_fp16_0, y = x_19_split_1_sigmoid_cast_fp16)[name = string("x_19_cast_fp16")]; + tensor var_574_axes_0 = const()[name = string("op_574_axes_0"), val = tensor([1])]; + tensor var_574 = expand_dims(axes = var_574_axes_0, x = pad_mask)[name = string("op_574")]; + tensor input_53_cast_fp16 = select(a = var_43_to_fp16, b = x_19_cast_fp16, cond = var_574)[name = string("input_53_cast_fp16")]; + bool new_x_3_interleave_0 = const()[name = string("new_x_3_interleave_0"), val = bool(false)]; + tensor new_x_3_cast_fp16 = concat(axis = var_58, interleave = new_x_3_interleave_0, values = (cache_3_cast_fp16, input_53_cast_fp16))[name = string("new_x_3_cast_fp16")]; + tensor var_587_begin_0 = const()[name = string("op_587_begin_0"), val = tensor([0, 0, 56])]; + tensor var_587_end_0 = const()[name = string("op_587_end_0"), val = tensor([1, 1024, 64])]; + tensor var_587_end_mask_0 = const()[name = string("op_587_end_mask_0"), val = tensor([true, true, true])]; + tensor var_587_cast_fp16 = slice_by_index(begin = var_587_begin_0, end = var_587_end_0, end_mask = var_587_end_mask_0, x = new_x_3_cast_fp16)[name = string("op_587_cast_fp16")]; + string x_21_pad_type_0 = const()[name = string("x_21_pad_type_0"), val = string("valid")]; + int32 x_21_groups_0 = const()[name = string("x_21_groups_0"), val = int32(1024)]; + tensor x_21_strides_0 = const()[name = string("x_21_strides_0"), val = tensor([1])]; + tensor x_21_pad_0 = const()[name = string("x_21_pad_0"), val = tensor([0, 0])]; + tensor x_21_dilations_0 = const()[name = string("x_21_dilations_0"), val = tensor([1])]; + tensor encoder_layers_0_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19548992))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19558272))))[name = string("encoder_layers_0_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_21_cast_fp16 = conv(dilations = x_21_dilations_0, groups = x_21_groups_0, pad = x_21_pad_0, pad_type = x_21_pad_type_0, strides = x_21_strides_0, weight = encoder_layers_0_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_3_cast_fp16)[name = string("x_21_cast_fp16")]; + tensor input_55_perm_0 = const()[name = string("input_55_perm_0"), val = tensor([0, 2, 1])]; + tensor x_23_axes_0 = const()[name = string("x_23_axes_0"), val = tensor([-1])]; + tensor encoder_layers_0_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_0_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19560384)))]; + tensor encoder_layers_0_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_0_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19562496)))]; + tensor input_55_cast_fp16 = transpose(perm = input_55_perm_0, x = x_21_cast_fp16)[name = string("transpose_356")]; + tensor x_23_cast_fp16 = layer_norm(axes = x_23_axes_0, beta = encoder_layers_0_conv_batch_norm_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_0_conv_batch_norm_weight_to_fp16, x = input_55_cast_fp16)[name = string("x_23_cast_fp16")]; + tensor input_57_perm_0 = const()[name = string("input_57_perm_0"), val = tensor([0, 2, 1])]; + tensor input_57_cast_fp16 = transpose(perm = input_57_perm_0, x = x_23_cast_fp16)[name = string("transpose_355")]; + tensor input_59_cast_fp16 = silu(x = input_57_cast_fp16)[name = string("input_59_cast_fp16")]; + string x_25_pad_type_0 = const()[name = string("x_25_pad_type_0"), val = string("valid")]; + tensor x_25_strides_0 = const()[name = string("x_25_strides_0"), val = tensor([1])]; + tensor x_25_pad_0 = const()[name = string("x_25_pad_0"), val = tensor([0, 0])]; + tensor x_25_dilations_0 = const()[name = string("x_25_dilations_0"), val = tensor([1])]; + int32 x_25_groups_0 = const()[name = string("x_25_groups_0"), val = int32(1)]; + tensor encoder_layers_0_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19564608))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20613248))))[name = string("encoder_layers_0_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_25_cast_fp16 = conv(dilations = x_25_dilations_0, groups = x_25_groups_0, pad = x_25_pad_0, pad_type = x_25_pad_type_0, strides = x_25_strides_0, weight = encoder_layers_0_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_59_cast_fp16)[name = string("x_25_cast_fp16")]; + tensor input_61_perm_0 = const()[name = string("input_61_perm_0"), val = tensor([0, 2, 1])]; + tensor input_61_cast_fp16 = transpose(perm = input_61_perm_0, x = x_25_cast_fp16)[name = string("transpose_354")]; + tensor input_63_cast_fp16 = add(x = input_47_cast_fp16, y = input_61_cast_fp16)[name = string("input_63_cast_fp16")]; + tensor input_65_axes_0 = const()[name = string("input_65_axes_0"), val = tensor([-1])]; + tensor encoder_layers_0_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_0_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20615360)))]; + tensor encoder_layers_0_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_0_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20617472)))]; + tensor input_65_cast_fp16 = layer_norm(axes = input_65_axes_0, beta = encoder_layers_0_norm_feed_forward2_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_0_norm_feed_forward2_weight_to_fp16, x = input_63_cast_fp16)[name = string("input_65_cast_fp16")]; + tensor encoder_layers_0_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20619584))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24813952))))[name = string("encoder_layers_0_feed_forward2_linear1_weight_to_fp16_quantized")]; + tensor encoder_layers_0_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_0_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24822208)))]; + tensor linear_8_cast_fp16 = linear(bias = encoder_layers_0_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_0_feed_forward2_linear1_weight_to_fp16_quantized, x = input_65_cast_fp16)[name = string("linear_8_cast_fp16")]; + tensor input_69_cast_fp16 = silu(x = linear_8_cast_fp16)[name = string("input_69_cast_fp16")]; + tensor encoder_layers_0_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24830464))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29024832))))[name = string("encoder_layers_0_feed_forward2_linear2_weight_to_fp16_quantized")]; + tensor encoder_layers_0_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_0_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29026944)))]; + tensor linear_9_cast_fp16 = linear(bias = encoder_layers_0_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_0_feed_forward2_linear2_weight_to_fp16_quantized, x = input_69_cast_fp16)[name = string("linear_9_cast_fp16")]; + fp16 var_630_to_fp16 = const()[name = string("op_630_to_fp16"), val = fp16(0x1p-1)]; + tensor var_631_cast_fp16 = mul(x = linear_9_cast_fp16, y = var_630_to_fp16)[name = string("op_631_cast_fp16")]; + tensor input_75_cast_fp16 = add(x = input_63_cast_fp16, y = var_631_cast_fp16)[name = string("input_75_cast_fp16")]; + tensor input_77_axes_0 = const()[name = string("input_77_axes_0"), val = tensor([-1])]; + tensor encoder_layers_0_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_0_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29029056)))]; + tensor encoder_layers_0_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_0_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29031168)))]; + tensor input_77_cast_fp16 = layer_norm(axes = input_77_axes_0, beta = encoder_layers_0_norm_out_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_0_norm_out_weight_to_fp16, x = input_75_cast_fp16)[name = string("input_77_cast_fp16")]; + tensor cache_5_begin_0 = const()[name = string("cache_5_begin_0"), val = tensor([1, 0, 0, 0])]; + tensor cache_5_end_0 = const()[name = string("cache_5_end_0"), val = tensor([2, 1, 42, 1024])]; + tensor cache_5_end_mask_0 = const()[name = string("cache_5_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_5_squeeze_mask_0 = const()[name = string("cache_5_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_5_cast_fp16 = slice_by_index(begin = cache_5_begin_0, end = cache_5_end_0, end_mask = cache_5_end_mask_0, squeeze_mask = cache_5_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_5_cast_fp16")]; + tensor cache_7_begin_0 = const()[name = string("cache_7_begin_0"), val = tensor([1, 0, 0, 0])]; + tensor cache_7_end_0 = const()[name = string("cache_7_end_0"), val = tensor([2, 1, 1024, 8])]; + tensor cache_7_end_mask_0 = const()[name = string("cache_7_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_7_squeeze_mask_0 = const()[name = string("cache_7_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_7_cast_fp16 = slice_by_index(begin = cache_7_begin_0, end = cache_7_end_0, end_mask = cache_7_end_mask_0, squeeze_mask = cache_7_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_7_cast_fp16")]; + tensor input_79_axes_0 = const()[name = string("input_79_axes_0"), val = tensor([-1])]; + tensor encoder_layers_1_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_1_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29033280)))]; + tensor encoder_layers_1_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_1_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29035392)))]; + tensor input_79_cast_fp16 = layer_norm(axes = input_79_axes_0, beta = encoder_layers_1_norm_feed_forward1_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_1_norm_feed_forward1_weight_to_fp16, x = input_77_cast_fp16)[name = string("input_79_cast_fp16")]; + tensor encoder_layers_1_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29037504))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33231872))))[name = string("encoder_layers_1_feed_forward1_linear1_weight_to_fp16_quantized")]; + tensor encoder_layers_1_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_1_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33240128)))]; + tensor linear_10_cast_fp16 = linear(bias = encoder_layers_1_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_1_feed_forward1_linear1_weight_to_fp16_quantized, x = input_79_cast_fp16)[name = string("linear_10_cast_fp16")]; + tensor input_83_cast_fp16 = silu(x = linear_10_cast_fp16)[name = string("input_83_cast_fp16")]; + tensor encoder_layers_1_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33248384))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37442752))))[name = string("encoder_layers_1_feed_forward1_linear2_weight_to_fp16_quantized")]; + tensor encoder_layers_1_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_1_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37444864)))]; + tensor linear_11_cast_fp16 = linear(bias = encoder_layers_1_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_1_feed_forward1_linear2_weight_to_fp16_quantized, x = input_83_cast_fp16)[name = string("linear_11_cast_fp16")]; + fp16 var_667_to_fp16 = const()[name = string("op_667_to_fp16"), val = fp16(0x1p-1)]; + tensor var_668_cast_fp16 = mul(x = linear_11_cast_fp16, y = var_667_to_fp16)[name = string("op_668_cast_fp16")]; + tensor input_89_cast_fp16 = add(x = input_77_cast_fp16, y = var_668_cast_fp16)[name = string("input_89_cast_fp16")]; + tensor key_3_axes_0 = const()[name = string("key_3_axes_0"), val = tensor([-1])]; + tensor encoder_layers_1_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_1_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37446976)))]; + tensor encoder_layers_1_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_1_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37449088)))]; + tensor key_3_cast_fp16 = layer_norm(axes = key_3_axes_0, beta = encoder_layers_1_norm_self_att_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_1_norm_self_att_weight_to_fp16, x = input_89_cast_fp16)[name = string("key_3_cast_fp16")]; + bool input_91_interleave_0 = const()[name = string("input_91_interleave_0"), val = bool(false)]; + tensor input_91_cast_fp16 = concat(axis = var_67, interleave = input_91_interleave_0, values = (cache_5_cast_fp16, key_3_cast_fp16))[name = string("input_91_cast_fp16")]; + bool var_696_interleave_0 = const()[name = string("op_696_interleave_0"), val = bool(false)]; + tensor var_696_cast_fp16 = concat(axis = var_67, interleave = var_696_interleave_0, values = key_3_cast_fp16)[name = string("op_696_cast_fp16")]; + tensor encoder_layers_1_self_attn_linear_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37451200))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38499840))))[name = string("encoder_layers_1_self_attn_linear_q_weight_to_fp16_quantized")]; + tensor encoder_layers_1_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_1_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38501952)))]; + tensor linear_12_cast_fp16 = linear(bias = encoder_layers_1_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_1_self_attn_linear_q_weight_to_fp16_quantized, x = key_3_cast_fp16)[name = string("linear_12_cast_fp16")]; + tensor var_701 = const()[name = string("op_701"), val = tensor([1, -1, 8, 128])]; + tensor q_7_cast_fp16 = reshape(shape = var_701, x = linear_12_cast_fp16)[name = string("q_7_cast_fp16")]; + tensor encoder_layers_1_self_attn_linear_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38504064))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39552704))))[name = string("encoder_layers_1_self_attn_linear_k_weight_to_fp16_quantized")]; + tensor encoder_layers_1_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_1_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39554816)))]; + tensor linear_13_cast_fp16 = linear(bias = encoder_layers_1_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_1_self_attn_linear_k_weight_to_fp16_quantized, x = input_91_cast_fp16)[name = string("linear_13_cast_fp16")]; + tensor var_706 = const()[name = string("op_706"), val = tensor([1, -1, 8, 128])]; + tensor k_5_cast_fp16 = reshape(shape = var_706, x = linear_13_cast_fp16)[name = string("k_5_cast_fp16")]; + tensor encoder_layers_1_self_attn_linear_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39556928))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40605568))))[name = string("encoder_layers_1_self_attn_linear_v_weight_to_fp16_quantized")]; + tensor encoder_layers_1_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_1_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40607680)))]; + tensor linear_14_cast_fp16 = linear(bias = encoder_layers_1_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_1_self_attn_linear_v_weight_to_fp16_quantized, x = input_91_cast_fp16)[name = string("linear_14_cast_fp16")]; + tensor var_711 = const()[name = string("op_711"), val = tensor([1, -1, 8, 128])]; + tensor v_3_cast_fp16 = reshape(shape = var_711, x = linear_14_cast_fp16)[name = string("v_3_cast_fp16")]; + tensor value_11_perm_0 = const()[name = string("value_11_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_1_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_1_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40609792)))]; + tensor var_724_cast_fp16 = add(x = q_7_cast_fp16, y = encoder_layers_1_self_attn_pos_bias_u_to_fp16)[name = string("op_724_cast_fp16")]; + tensor encoder_layers_1_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_1_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40611904)))]; + tensor var_726_cast_fp16 = add(x = q_7_cast_fp16, y = encoder_layers_1_self_attn_pos_bias_v_to_fp16)[name = string("op_726_cast_fp16")]; + tensor q_with_bias_v_3_perm_0 = const()[name = string("q_with_bias_v_3_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_33_transpose_x_0 = const()[name = string("x_33_transpose_x_0"), val = bool(false)]; + bool x_33_transpose_y_0 = const()[name = string("x_33_transpose_y_0"), val = bool(false)]; + tensor op_728_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40614016))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40813760))))[name = string("op_728_to_fp16_quantized")]; + tensor q_with_bias_v_3_cast_fp16 = transpose(perm = q_with_bias_v_3_perm_0, x = var_726_cast_fp16)[name = string("transpose_353")]; + tensor x_33_cast_fp16 = matmul(transpose_x = x_33_transpose_x_0, transpose_y = x_33_transpose_y_0, x = q_with_bias_v_3_cast_fp16, y = op_728_to_fp16_quantized)[name = string("x_33_cast_fp16")]; + tensor x_35_pad_0 = const()[name = string("x_35_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_35_mode_0 = const()[name = string("x_35_mode_0"), val = string("constant")]; + fp16 const_92_to_fp16 = const()[name = string("const_92_to_fp16"), val = fp16(0x0p+0)]; + tensor x_35_cast_fp16 = pad(constant_val = const_92_to_fp16, mode = x_35_mode_0, pad = x_35_pad_0, x = x_33_cast_fp16)[name = string("x_35_cast_fp16")]; + tensor var_736 = const()[name = string("op_736"), val = tensor([1, 8, -1, 56])]; + tensor x_37_cast_fp16 = reshape(shape = var_736, x = x_35_cast_fp16)[name = string("x_37_cast_fp16")]; + tensor var_740_begin_0 = const()[name = string("op_740_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_740_end_0 = const()[name = string("op_740_end_0"), val = tensor([1, 8, 196, 56])]; + tensor var_740_end_mask_0 = const()[name = string("op_740_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_740_cast_fp16 = slice_by_index(begin = var_740_begin_0, end = var_740_end_0, end_mask = var_740_end_mask_0, x = x_37_cast_fp16)[name = string("op_740_cast_fp16")]; + tensor var_741 = const()[name = string("op_741"), val = tensor([1, 8, 56, 195])]; + tensor matrix_bd_5_cast_fp16 = reshape(shape = var_741, x = var_740_cast_fp16)[name = string("matrix_bd_5_cast_fp16")]; + bool matrix_ac_3_transpose_x_0 = const()[name = string("matrix_ac_3_transpose_x_0"), val = bool(false)]; + bool matrix_ac_3_transpose_y_0 = const()[name = string("matrix_ac_3_transpose_y_0"), val = bool(false)]; + tensor transpose_98_perm_0 = const()[name = string("transpose_98_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_99_perm_0 = const()[name = string("transpose_99_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_99 = transpose(perm = transpose_99_perm_0, x = k_5_cast_fp16)[name = string("transpose_351")]; + tensor transpose_98 = transpose(perm = transpose_98_perm_0, x = var_724_cast_fp16)[name = string("transpose_352")]; + tensor matrix_ac_3_cast_fp16 = matmul(transpose_x = matrix_ac_3_transpose_x_0, transpose_y = matrix_ac_3_transpose_y_0, x = transpose_98, y = transpose_99)[name = string("matrix_ac_3_cast_fp16")]; + tensor matrix_bd_7_begin_0 = const()[name = string("matrix_bd_7_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_7_end_0 = const()[name = string("matrix_bd_7_end_0"), val = tensor([1, 8, 56, 98])]; + tensor matrix_bd_7_end_mask_0 = const()[name = string("matrix_bd_7_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_7_cast_fp16 = slice_by_index(begin = matrix_bd_7_begin_0, end = matrix_bd_7_end_0, end_mask = matrix_bd_7_end_mask_0, x = matrix_bd_5_cast_fp16)[name = string("matrix_bd_7_cast_fp16")]; + tensor var_750_cast_fp16 = add(x = matrix_ac_3_cast_fp16, y = matrix_bd_7_cast_fp16)[name = string("op_750_cast_fp16")]; + fp16 _inversed_scores_5_y_0_to_fp16 = const()[name = string("_inversed_scores_5_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_5_cast_fp16 = mul(x = var_750_cast_fp16, y = _inversed_scores_5_y_0_to_fp16)[name = string("_inversed_scores_5_cast_fp16")]; + tensor scores_7_cast_fp16 = select(a = var_44_to_fp16, b = _inversed_scores_5_cast_fp16, cond = mask_11)[name = string("scores_7_cast_fp16")]; + tensor var_756_cast_fp16 = softmax(axis = var_58, x = scores_7_cast_fp16)[name = string("op_756_cast_fp16")]; + tensor input_93_cast_fp16 = select(a = var_43_to_fp16, b = var_756_cast_fp16, cond = mask_11)[name = string("input_93_cast_fp16")]; + bool x_39_transpose_x_0 = const()[name = string("x_39_transpose_x_0"), val = bool(false)]; + bool x_39_transpose_y_0 = const()[name = string("x_39_transpose_y_0"), val = bool(false)]; + tensor value_11_cast_fp16 = transpose(perm = value_11_perm_0, x = v_3_cast_fp16)[name = string("transpose_350")]; + tensor x_39_cast_fp16 = matmul(transpose_x = x_39_transpose_x_0, transpose_y = x_39_transpose_y_0, x = input_93_cast_fp16, y = value_11_cast_fp16)[name = string("x_39_cast_fp16")]; + tensor var_760_perm_0 = const()[name = string("op_760_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_761 = const()[name = string("op_761"), val = tensor([1, -1, 1024])]; + tensor var_760_cast_fp16 = transpose(perm = var_760_perm_0, x = x_39_cast_fp16)[name = string("transpose_349")]; + tensor input_95_cast_fp16 = reshape(shape = var_761, x = var_760_cast_fp16)[name = string("input_95_cast_fp16")]; + tensor encoder_layers_1_self_attn_linear_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40814272))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41862912))))[name = string("encoder_layers_1_self_attn_linear_out_weight_to_fp16_quantized")]; + tensor encoder_layers_1_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_1_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41865024)))]; + tensor linear_16_cast_fp16 = linear(bias = encoder_layers_1_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_1_self_attn_linear_out_weight_to_fp16_quantized, x = input_95_cast_fp16)[name = string("linear_16_cast_fp16")]; + tensor input_99_cast_fp16 = add(x = input_89_cast_fp16, y = linear_16_cast_fp16)[name = string("input_99_cast_fp16")]; + tensor x_43_axes_0 = const()[name = string("x_43_axes_0"), val = tensor([-1])]; + tensor encoder_layers_1_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_1_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41867136)))]; + tensor encoder_layers_1_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_1_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41869248)))]; + tensor x_43_cast_fp16 = layer_norm(axes = x_43_axes_0, beta = encoder_layers_1_norm_conv_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_1_norm_conv_weight_to_fp16, x = input_99_cast_fp16)[name = string("x_43_cast_fp16")]; + tensor input_101_perm_0 = const()[name = string("input_101_perm_0"), val = tensor([0, 2, 1])]; + string input_103_pad_type_0 = const()[name = string("input_103_pad_type_0"), val = string("valid")]; + tensor input_103_strides_0 = const()[name = string("input_103_strides_0"), val = tensor([1])]; + tensor input_103_pad_0 = const()[name = string("input_103_pad_0"), val = tensor([0, 0])]; + tensor input_103_dilations_0 = const()[name = string("input_103_dilations_0"), val = tensor([1])]; + int32 input_103_groups_0 = const()[name = string("input_103_groups_0"), val = int32(1)]; + tensor encoder_layers_1_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41871360))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43968576))))[name = string("encoder_layers_1_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_101_cast_fp16 = transpose(perm = input_101_perm_0, x = x_43_cast_fp16)[name = string("transpose_348")]; + tensor input_103_cast_fp16 = conv(dilations = input_103_dilations_0, groups = input_103_groups_0, pad = input_103_pad_0, pad_type = input_103_pad_type_0, strides = input_103_strides_0, weight = encoder_layers_1_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_101_cast_fp16)[name = string("input_103_cast_fp16")]; + int32 x_45_split_num_splits_0 = const()[name = string("x_45_split_num_splits_0"), val = int32(2)]; + int32 x_45_split_axis_0 = const()[name = string("x_45_split_axis_0"), val = int32(1)]; + tensor x_45_split_cast_fp16_0, tensor x_45_split_cast_fp16_1 = split(axis = x_45_split_axis_0, num_splits = x_45_split_num_splits_0, x = input_103_cast_fp16)[name = string("x_45_split_cast_fp16")]; + tensor x_45_split_1_sigmoid_cast_fp16 = sigmoid(x = x_45_split_cast_fp16_1)[name = string("x_45_split_1_sigmoid_cast_fp16")]; + tensor x_45_cast_fp16 = mul(x = x_45_split_cast_fp16_0, y = x_45_split_1_sigmoid_cast_fp16)[name = string("x_45_cast_fp16")]; + tensor input_105_cast_fp16 = select(a = var_43_to_fp16, b = x_45_cast_fp16, cond = var_574)[name = string("input_105_cast_fp16")]; + bool new_x_7_interleave_0 = const()[name = string("new_x_7_interleave_0"), val = bool(false)]; + tensor new_x_7_cast_fp16 = concat(axis = var_58, interleave = new_x_7_interleave_0, values = (cache_7_cast_fp16, input_105_cast_fp16))[name = string("new_x_7_cast_fp16")]; + tensor var_800_begin_0 = const()[name = string("op_800_begin_0"), val = tensor([0, 0, 56])]; + tensor var_800_end_0 = const()[name = string("op_800_end_0"), val = tensor([1, 1024, 64])]; + tensor var_800_end_mask_0 = const()[name = string("op_800_end_mask_0"), val = tensor([true, true, true])]; + tensor var_800_cast_fp16 = slice_by_index(begin = var_800_begin_0, end = var_800_end_0, end_mask = var_800_end_mask_0, x = new_x_7_cast_fp16)[name = string("op_800_cast_fp16")]; + string x_47_pad_type_0 = const()[name = string("x_47_pad_type_0"), val = string("valid")]; + int32 x_47_groups_0 = const()[name = string("x_47_groups_0"), val = int32(1024)]; + tensor x_47_strides_0 = const()[name = string("x_47_strides_0"), val = tensor([1])]; + tensor x_47_pad_0 = const()[name = string("x_47_pad_0"), val = tensor([0, 0])]; + tensor x_47_dilations_0 = const()[name = string("x_47_dilations_0"), val = tensor([1])]; + tensor encoder_layers_1_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43972736))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43982016))))[name = string("encoder_layers_1_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_47_cast_fp16 = conv(dilations = x_47_dilations_0, groups = x_47_groups_0, pad = x_47_pad_0, pad_type = x_47_pad_type_0, strides = x_47_strides_0, weight = encoder_layers_1_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_7_cast_fp16)[name = string("x_47_cast_fp16")]; + tensor input_107_perm_0 = const()[name = string("input_107_perm_0"), val = tensor([0, 2, 1])]; + tensor x_49_axes_0 = const()[name = string("x_49_axes_0"), val = tensor([-1])]; + tensor encoder_layers_1_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_1_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43984128)))]; + tensor encoder_layers_1_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_1_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43986240)))]; + tensor input_107_cast_fp16 = transpose(perm = input_107_perm_0, x = x_47_cast_fp16)[name = string("transpose_347")]; + tensor x_49_cast_fp16 = layer_norm(axes = x_49_axes_0, beta = encoder_layers_1_conv_batch_norm_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_1_conv_batch_norm_weight_to_fp16, x = input_107_cast_fp16)[name = string("x_49_cast_fp16")]; + tensor input_109_perm_0 = const()[name = string("input_109_perm_0"), val = tensor([0, 2, 1])]; + tensor input_109_cast_fp16 = transpose(perm = input_109_perm_0, x = x_49_cast_fp16)[name = string("transpose_346")]; + tensor input_111_cast_fp16 = silu(x = input_109_cast_fp16)[name = string("input_111_cast_fp16")]; + string x_51_pad_type_0 = const()[name = string("x_51_pad_type_0"), val = string("valid")]; + tensor x_51_strides_0 = const()[name = string("x_51_strides_0"), val = tensor([1])]; + tensor x_51_pad_0 = const()[name = string("x_51_pad_0"), val = tensor([0, 0])]; + tensor x_51_dilations_0 = const()[name = string("x_51_dilations_0"), val = tensor([1])]; + int32 x_51_groups_0 = const()[name = string("x_51_groups_0"), val = int32(1)]; + tensor encoder_layers_1_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43988352))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45036992))))[name = string("encoder_layers_1_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_51_cast_fp16 = conv(dilations = x_51_dilations_0, groups = x_51_groups_0, pad = x_51_pad_0, pad_type = x_51_pad_type_0, strides = x_51_strides_0, weight = encoder_layers_1_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_111_cast_fp16)[name = string("x_51_cast_fp16")]; + tensor input_113_perm_0 = const()[name = string("input_113_perm_0"), val = tensor([0, 2, 1])]; + tensor input_113_cast_fp16 = transpose(perm = input_113_perm_0, x = x_51_cast_fp16)[name = string("transpose_345")]; + tensor input_115_cast_fp16 = add(x = input_99_cast_fp16, y = input_113_cast_fp16)[name = string("input_115_cast_fp16")]; + tensor input_117_axes_0 = const()[name = string("input_117_axes_0"), val = tensor([-1])]; + tensor encoder_layers_1_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_1_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45039104)))]; + tensor encoder_layers_1_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_1_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45041216)))]; + tensor input_117_cast_fp16 = layer_norm(axes = input_117_axes_0, beta = encoder_layers_1_norm_feed_forward2_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_1_norm_feed_forward2_weight_to_fp16, x = input_115_cast_fp16)[name = string("input_117_cast_fp16")]; + tensor encoder_layers_1_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45043328))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49237696))))[name = string("encoder_layers_1_feed_forward2_linear1_weight_to_fp16_quantized")]; + tensor encoder_layers_1_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_1_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49245952)))]; + tensor linear_17_cast_fp16 = linear(bias = encoder_layers_1_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_1_feed_forward2_linear1_weight_to_fp16_quantized, x = input_117_cast_fp16)[name = string("linear_17_cast_fp16")]; + tensor input_121_cast_fp16 = silu(x = linear_17_cast_fp16)[name = string("input_121_cast_fp16")]; + tensor encoder_layers_1_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49254208))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53448576))))[name = string("encoder_layers_1_feed_forward2_linear2_weight_to_fp16_quantized")]; + tensor encoder_layers_1_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_1_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53450688)))]; + tensor linear_18_cast_fp16 = linear(bias = encoder_layers_1_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_1_feed_forward2_linear2_weight_to_fp16_quantized, x = input_121_cast_fp16)[name = string("linear_18_cast_fp16")]; + fp16 var_843_to_fp16 = const()[name = string("op_843_to_fp16"), val = fp16(0x1p-1)]; + tensor var_844_cast_fp16 = mul(x = linear_18_cast_fp16, y = var_843_to_fp16)[name = string("op_844_cast_fp16")]; + tensor input_127_cast_fp16 = add(x = input_115_cast_fp16, y = var_844_cast_fp16)[name = string("input_127_cast_fp16")]; + tensor input_129_axes_0 = const()[name = string("input_129_axes_0"), val = tensor([-1])]; + tensor encoder_layers_1_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_1_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53452800)))]; + tensor encoder_layers_1_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_1_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53454912)))]; + tensor input_129_cast_fp16 = layer_norm(axes = input_129_axes_0, beta = encoder_layers_1_norm_out_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_1_norm_out_weight_to_fp16, x = input_127_cast_fp16)[name = string("input_129_cast_fp16")]; + tensor cache_9_begin_0 = const()[name = string("cache_9_begin_0"), val = tensor([2, 0, 0, 0])]; + tensor cache_9_end_0 = const()[name = string("cache_9_end_0"), val = tensor([3, 1, 42, 1024])]; + tensor cache_9_end_mask_0 = const()[name = string("cache_9_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_9_squeeze_mask_0 = const()[name = string("cache_9_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_9_cast_fp16 = slice_by_index(begin = cache_9_begin_0, end = cache_9_end_0, end_mask = cache_9_end_mask_0, squeeze_mask = cache_9_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_9_cast_fp16")]; + tensor cache_11_begin_0 = const()[name = string("cache_11_begin_0"), val = tensor([2, 0, 0, 0])]; + tensor cache_11_end_0 = const()[name = string("cache_11_end_0"), val = tensor([3, 1, 1024, 8])]; + tensor cache_11_end_mask_0 = const()[name = string("cache_11_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_11_squeeze_mask_0 = const()[name = string("cache_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_11_cast_fp16 = slice_by_index(begin = cache_11_begin_0, end = cache_11_end_0, end_mask = cache_11_end_mask_0, squeeze_mask = cache_11_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_11_cast_fp16")]; + tensor input_131_axes_0 = const()[name = string("input_131_axes_0"), val = tensor([-1])]; + tensor encoder_layers_2_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_2_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53457024)))]; + tensor encoder_layers_2_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_2_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53459136)))]; + tensor input_131_cast_fp16 = layer_norm(axes = input_131_axes_0, beta = encoder_layers_2_norm_feed_forward1_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_2_norm_feed_forward1_weight_to_fp16, x = input_129_cast_fp16)[name = string("input_131_cast_fp16")]; + tensor encoder_layers_2_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53461248))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57655616))))[name = string("encoder_layers_2_feed_forward1_linear1_weight_to_fp16_quantized")]; + tensor encoder_layers_2_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_2_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57663872)))]; + tensor linear_19_cast_fp16 = linear(bias = encoder_layers_2_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_2_feed_forward1_linear1_weight_to_fp16_quantized, x = input_131_cast_fp16)[name = string("linear_19_cast_fp16")]; + tensor input_135_cast_fp16 = silu(x = linear_19_cast_fp16)[name = string("input_135_cast_fp16")]; + tensor encoder_layers_2_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57672128))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61866496))))[name = string("encoder_layers_2_feed_forward1_linear2_weight_to_fp16_quantized")]; + tensor encoder_layers_2_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_2_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61868608)))]; + tensor linear_20_cast_fp16 = linear(bias = encoder_layers_2_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_2_feed_forward1_linear2_weight_to_fp16_quantized, x = input_135_cast_fp16)[name = string("linear_20_cast_fp16")]; + fp16 var_880_to_fp16 = const()[name = string("op_880_to_fp16"), val = fp16(0x1p-1)]; + tensor var_881_cast_fp16 = mul(x = linear_20_cast_fp16, y = var_880_to_fp16)[name = string("op_881_cast_fp16")]; + tensor input_141_cast_fp16 = add(x = input_129_cast_fp16, y = var_881_cast_fp16)[name = string("input_141_cast_fp16")]; + tensor key_5_axes_0 = const()[name = string("key_5_axes_0"), val = tensor([-1])]; + tensor encoder_layers_2_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_2_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61870720)))]; + tensor encoder_layers_2_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_2_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61872832)))]; + tensor key_5_cast_fp16 = layer_norm(axes = key_5_axes_0, beta = encoder_layers_2_norm_self_att_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_2_norm_self_att_weight_to_fp16, x = input_141_cast_fp16)[name = string("key_5_cast_fp16")]; + bool input_143_interleave_0 = const()[name = string("input_143_interleave_0"), val = bool(false)]; + tensor input_143_cast_fp16 = concat(axis = var_67, interleave = input_143_interleave_0, values = (cache_9_cast_fp16, key_5_cast_fp16))[name = string("input_143_cast_fp16")]; + bool var_909_interleave_0 = const()[name = string("op_909_interleave_0"), val = bool(false)]; + tensor var_909_cast_fp16 = concat(axis = var_67, interleave = var_909_interleave_0, values = key_5_cast_fp16)[name = string("op_909_cast_fp16")]; + tensor encoder_layers_2_self_attn_linear_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61874944))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62923584))))[name = string("encoder_layers_2_self_attn_linear_q_weight_to_fp16_quantized")]; + tensor encoder_layers_2_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_2_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62925696)))]; + tensor linear_21_cast_fp16 = linear(bias = encoder_layers_2_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_2_self_attn_linear_q_weight_to_fp16_quantized, x = key_5_cast_fp16)[name = string("linear_21_cast_fp16")]; + tensor var_914 = const()[name = string("op_914"), val = tensor([1, -1, 8, 128])]; + tensor q_13_cast_fp16 = reshape(shape = var_914, x = linear_21_cast_fp16)[name = string("q_13_cast_fp16")]; + tensor encoder_layers_2_self_attn_linear_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62927808))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(63976448))))[name = string("encoder_layers_2_self_attn_linear_k_weight_to_fp16_quantized")]; + tensor encoder_layers_2_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_2_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(63978560)))]; + tensor linear_22_cast_fp16 = linear(bias = encoder_layers_2_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_2_self_attn_linear_k_weight_to_fp16_quantized, x = input_143_cast_fp16)[name = string("linear_22_cast_fp16")]; + tensor var_919 = const()[name = string("op_919"), val = tensor([1, -1, 8, 128])]; + tensor k_9_cast_fp16 = reshape(shape = var_919, x = linear_22_cast_fp16)[name = string("k_9_cast_fp16")]; + tensor encoder_layers_2_self_attn_linear_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(63980672))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65029312))))[name = string("encoder_layers_2_self_attn_linear_v_weight_to_fp16_quantized")]; + tensor encoder_layers_2_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_2_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65031424)))]; + tensor linear_23_cast_fp16 = linear(bias = encoder_layers_2_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_2_self_attn_linear_v_weight_to_fp16_quantized, x = input_143_cast_fp16)[name = string("linear_23_cast_fp16")]; + tensor var_924 = const()[name = string("op_924"), val = tensor([1, -1, 8, 128])]; + tensor v_5_cast_fp16 = reshape(shape = var_924, x = linear_23_cast_fp16)[name = string("v_5_cast_fp16")]; + tensor value_13_perm_0 = const()[name = string("value_13_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_2_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_2_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65033536)))]; + tensor var_937_cast_fp16 = add(x = q_13_cast_fp16, y = encoder_layers_2_self_attn_pos_bias_u_to_fp16)[name = string("op_937_cast_fp16")]; + tensor encoder_layers_2_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_2_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65035648)))]; + tensor var_939_cast_fp16 = add(x = q_13_cast_fp16, y = encoder_layers_2_self_attn_pos_bias_v_to_fp16)[name = string("op_939_cast_fp16")]; + tensor q_with_bias_v_5_perm_0 = const()[name = string("q_with_bias_v_5_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_59_transpose_x_0 = const()[name = string("x_59_transpose_x_0"), val = bool(false)]; + bool x_59_transpose_y_0 = const()[name = string("x_59_transpose_y_0"), val = bool(false)]; + tensor op_941_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65037760))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65237504))))[name = string("op_941_to_fp16_quantized")]; + tensor q_with_bias_v_5_cast_fp16 = transpose(perm = q_with_bias_v_5_perm_0, x = var_939_cast_fp16)[name = string("transpose_344")]; + tensor x_59_cast_fp16 = matmul(transpose_x = x_59_transpose_x_0, transpose_y = x_59_transpose_y_0, x = q_with_bias_v_5_cast_fp16, y = op_941_to_fp16_quantized)[name = string("x_59_cast_fp16")]; + tensor x_61_pad_0 = const()[name = string("x_61_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_61_mode_0 = const()[name = string("x_61_mode_0"), val = string("constant")]; + fp16 const_105_to_fp16 = const()[name = string("const_105_to_fp16"), val = fp16(0x0p+0)]; + tensor x_61_cast_fp16 = pad(constant_val = const_105_to_fp16, mode = x_61_mode_0, pad = x_61_pad_0, x = x_59_cast_fp16)[name = string("x_61_cast_fp16")]; + tensor var_949 = const()[name = string("op_949"), val = tensor([1, 8, -1, 56])]; + tensor x_63_cast_fp16 = reshape(shape = var_949, x = x_61_cast_fp16)[name = string("x_63_cast_fp16")]; + tensor var_953_begin_0 = const()[name = string("op_953_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_953_end_0 = const()[name = string("op_953_end_0"), val = tensor([1, 8, 196, 56])]; + tensor var_953_end_mask_0 = const()[name = string("op_953_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_953_cast_fp16 = slice_by_index(begin = var_953_begin_0, end = var_953_end_0, end_mask = var_953_end_mask_0, x = x_63_cast_fp16)[name = string("op_953_cast_fp16")]; + tensor var_954 = const()[name = string("op_954"), val = tensor([1, 8, 56, 195])]; + tensor matrix_bd_9_cast_fp16 = reshape(shape = var_954, x = var_953_cast_fp16)[name = string("matrix_bd_9_cast_fp16")]; + bool matrix_ac_5_transpose_x_0 = const()[name = string("matrix_ac_5_transpose_x_0"), val = bool(false)]; + bool matrix_ac_5_transpose_y_0 = const()[name = string("matrix_ac_5_transpose_y_0"), val = bool(false)]; + tensor transpose_100_perm_0 = const()[name = string("transpose_100_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_101_perm_0 = const()[name = string("transpose_101_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_101 = transpose(perm = transpose_101_perm_0, x = k_9_cast_fp16)[name = string("transpose_342")]; + tensor transpose_100 = transpose(perm = transpose_100_perm_0, x = var_937_cast_fp16)[name = string("transpose_343")]; + tensor matrix_ac_5_cast_fp16 = matmul(transpose_x = matrix_ac_5_transpose_x_0, transpose_y = matrix_ac_5_transpose_y_0, x = transpose_100, y = transpose_101)[name = string("matrix_ac_5_cast_fp16")]; + tensor matrix_bd_11_begin_0 = const()[name = string("matrix_bd_11_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_11_end_0 = const()[name = string("matrix_bd_11_end_0"), val = tensor([1, 8, 56, 98])]; + tensor matrix_bd_11_end_mask_0 = const()[name = string("matrix_bd_11_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_11_cast_fp16 = slice_by_index(begin = matrix_bd_11_begin_0, end = matrix_bd_11_end_0, end_mask = matrix_bd_11_end_mask_0, x = matrix_bd_9_cast_fp16)[name = string("matrix_bd_11_cast_fp16")]; + tensor var_963_cast_fp16 = add(x = matrix_ac_5_cast_fp16, y = matrix_bd_11_cast_fp16)[name = string("op_963_cast_fp16")]; + fp16 _inversed_scores_9_y_0_to_fp16 = const()[name = string("_inversed_scores_9_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_9_cast_fp16 = mul(x = var_963_cast_fp16, y = _inversed_scores_9_y_0_to_fp16)[name = string("_inversed_scores_9_cast_fp16")]; + tensor scores_11_cast_fp16 = select(a = var_44_to_fp16, b = _inversed_scores_9_cast_fp16, cond = mask_11)[name = string("scores_11_cast_fp16")]; + tensor var_969_cast_fp16 = softmax(axis = var_58, x = scores_11_cast_fp16)[name = string("op_969_cast_fp16")]; + tensor input_145_cast_fp16 = select(a = var_43_to_fp16, b = var_969_cast_fp16, cond = mask_11)[name = string("input_145_cast_fp16")]; + bool x_65_transpose_x_0 = const()[name = string("x_65_transpose_x_0"), val = bool(false)]; + bool x_65_transpose_y_0 = const()[name = string("x_65_transpose_y_0"), val = bool(false)]; + tensor value_13_cast_fp16 = transpose(perm = value_13_perm_0, x = v_5_cast_fp16)[name = string("transpose_341")]; + tensor x_65_cast_fp16 = matmul(transpose_x = x_65_transpose_x_0, transpose_y = x_65_transpose_y_0, x = input_145_cast_fp16, y = value_13_cast_fp16)[name = string("x_65_cast_fp16")]; + tensor var_973_perm_0 = const()[name = string("op_973_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_974 = const()[name = string("op_974"), val = tensor([1, -1, 1024])]; + tensor var_973_cast_fp16 = transpose(perm = var_973_perm_0, x = x_65_cast_fp16)[name = string("transpose_340")]; + tensor input_147_cast_fp16 = reshape(shape = var_974, x = var_973_cast_fp16)[name = string("input_147_cast_fp16")]; + tensor encoder_layers_2_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65238016))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66024512))))[name = string("encoder_layers_2_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_2_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_2_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66024704)))]; + tensor linear_25_cast_fp16 = linear(bias = encoder_layers_2_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_2_self_attn_linear_out_weight_to_fp16_palettized, x = input_147_cast_fp16)[name = string("linear_25_cast_fp16")]; + tensor input_151_cast_fp16 = add(x = input_141_cast_fp16, y = linear_25_cast_fp16)[name = string("input_151_cast_fp16")]; + tensor x_69_axes_0 = const()[name = string("x_69_axes_0"), val = tensor([-1])]; + tensor encoder_layers_2_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_2_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66026816)))]; + tensor encoder_layers_2_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_2_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66028928)))]; + tensor x_69_cast_fp16 = layer_norm(axes = x_69_axes_0, beta = encoder_layers_2_norm_conv_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_2_norm_conv_weight_to_fp16, x = input_151_cast_fp16)[name = string("x_69_cast_fp16")]; + tensor input_153_perm_0 = const()[name = string("input_153_perm_0"), val = tensor([0, 2, 1])]; + string input_155_pad_type_0 = const()[name = string("input_155_pad_type_0"), val = string("valid")]; + tensor input_155_strides_0 = const()[name = string("input_155_strides_0"), val = tensor([1])]; + tensor input_155_pad_0 = const()[name = string("input_155_pad_0"), val = tensor([0, 0])]; + tensor input_155_dilations_0 = const()[name = string("input_155_dilations_0"), val = tensor([1])]; + int32 input_155_groups_0 = const()[name = string("input_155_groups_0"), val = int32(1)]; + tensor encoder_layers_2_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66031040))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68128256))))[name = string("encoder_layers_2_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_153_cast_fp16 = transpose(perm = input_153_perm_0, x = x_69_cast_fp16)[name = string("transpose_339")]; + tensor input_155_cast_fp16 = conv(dilations = input_155_dilations_0, groups = input_155_groups_0, pad = input_155_pad_0, pad_type = input_155_pad_type_0, strides = input_155_strides_0, weight = encoder_layers_2_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_153_cast_fp16)[name = string("input_155_cast_fp16")]; + int32 x_71_split_num_splits_0 = const()[name = string("x_71_split_num_splits_0"), val = int32(2)]; + int32 x_71_split_axis_0 = const()[name = string("x_71_split_axis_0"), val = int32(1)]; + tensor x_71_split_cast_fp16_0, tensor x_71_split_cast_fp16_1 = split(axis = x_71_split_axis_0, num_splits = x_71_split_num_splits_0, x = input_155_cast_fp16)[name = string("x_71_split_cast_fp16")]; + tensor x_71_split_1_sigmoid_cast_fp16 = sigmoid(x = x_71_split_cast_fp16_1)[name = string("x_71_split_1_sigmoid_cast_fp16")]; + tensor x_71_cast_fp16 = mul(x = x_71_split_cast_fp16_0, y = x_71_split_1_sigmoid_cast_fp16)[name = string("x_71_cast_fp16")]; + tensor input_157_cast_fp16 = select(a = var_43_to_fp16, b = x_71_cast_fp16, cond = var_574)[name = string("input_157_cast_fp16")]; + bool new_x_11_interleave_0 = const()[name = string("new_x_11_interleave_0"), val = bool(false)]; + tensor new_x_11_cast_fp16 = concat(axis = var_58, interleave = new_x_11_interleave_0, values = (cache_11_cast_fp16, input_157_cast_fp16))[name = string("new_x_11_cast_fp16")]; + tensor var_1013_begin_0 = const()[name = string("op_1013_begin_0"), val = tensor([0, 0, 56])]; + tensor var_1013_end_0 = const()[name = string("op_1013_end_0"), val = tensor([1, 1024, 64])]; + tensor var_1013_end_mask_0 = const()[name = string("op_1013_end_mask_0"), val = tensor([true, true, true])]; + tensor var_1013_cast_fp16 = slice_by_index(begin = var_1013_begin_0, end = var_1013_end_0, end_mask = var_1013_end_mask_0, x = new_x_11_cast_fp16)[name = string("op_1013_cast_fp16")]; + string x_73_pad_type_0 = const()[name = string("x_73_pad_type_0"), val = string("valid")]; + int32 x_73_groups_0 = const()[name = string("x_73_groups_0"), val = int32(1024)]; + tensor x_73_strides_0 = const()[name = string("x_73_strides_0"), val = tensor([1])]; + tensor x_73_pad_0 = const()[name = string("x_73_pad_0"), val = tensor([0, 0])]; + tensor x_73_dilations_0 = const()[name = string("x_73_dilations_0"), val = tensor([1])]; + tensor encoder_layers_2_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68132416))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68141696))))[name = string("encoder_layers_2_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_73_cast_fp16 = conv(dilations = x_73_dilations_0, groups = x_73_groups_0, pad = x_73_pad_0, pad_type = x_73_pad_type_0, strides = x_73_strides_0, weight = encoder_layers_2_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_11_cast_fp16)[name = string("x_73_cast_fp16")]; + tensor input_159_perm_0 = const()[name = string("input_159_perm_0"), val = tensor([0, 2, 1])]; + tensor x_75_axes_0 = const()[name = string("x_75_axes_0"), val = tensor([-1])]; + tensor encoder_layers_2_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_2_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68143808)))]; + tensor encoder_layers_2_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_2_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68145920)))]; + tensor input_159_cast_fp16 = transpose(perm = input_159_perm_0, x = x_73_cast_fp16)[name = string("transpose_338")]; + tensor x_75_cast_fp16 = layer_norm(axes = x_75_axes_0, beta = encoder_layers_2_conv_batch_norm_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_2_conv_batch_norm_weight_to_fp16, x = input_159_cast_fp16)[name = string("x_75_cast_fp16")]; + tensor input_161_perm_0 = const()[name = string("input_161_perm_0"), val = tensor([0, 2, 1])]; + tensor input_161_cast_fp16 = transpose(perm = input_161_perm_0, x = x_75_cast_fp16)[name = string("transpose_337")]; + tensor input_163_cast_fp16 = silu(x = input_161_cast_fp16)[name = string("input_163_cast_fp16")]; + string x_77_pad_type_0 = const()[name = string("x_77_pad_type_0"), val = string("valid")]; + tensor x_77_strides_0 = const()[name = string("x_77_strides_0"), val = tensor([1])]; + tensor x_77_pad_0 = const()[name = string("x_77_pad_0"), val = tensor([0, 0])]; + tensor x_77_dilations_0 = const()[name = string("x_77_dilations_0"), val = tensor([1])]; + int32 x_77_groups_0 = const()[name = string("x_77_groups_0"), val = int32(1)]; + tensor encoder_layers_2_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68148032))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69196672))))[name = string("encoder_layers_2_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_77_cast_fp16 = conv(dilations = x_77_dilations_0, groups = x_77_groups_0, pad = x_77_pad_0, pad_type = x_77_pad_type_0, strides = x_77_strides_0, weight = encoder_layers_2_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_163_cast_fp16)[name = string("x_77_cast_fp16")]; + tensor input_165_perm_0 = const()[name = string("input_165_perm_0"), val = tensor([0, 2, 1])]; + tensor input_165_cast_fp16 = transpose(perm = input_165_perm_0, x = x_77_cast_fp16)[name = string("transpose_336")]; + tensor input_167_cast_fp16 = add(x = input_151_cast_fp16, y = input_165_cast_fp16)[name = string("input_167_cast_fp16")]; + tensor input_169_axes_0 = const()[name = string("input_169_axes_0"), val = tensor([-1])]; + tensor encoder_layers_2_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_2_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69198784)))]; + tensor encoder_layers_2_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_2_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69200896)))]; + tensor input_169_cast_fp16 = layer_norm(axes = input_169_axes_0, beta = encoder_layers_2_norm_feed_forward2_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_2_norm_feed_forward2_weight_to_fp16, x = input_167_cast_fp16)[name = string("input_169_cast_fp16")]; + tensor encoder_layers_2_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69203008))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72348800))))[name = string("encoder_layers_2_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_2_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_2_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72348992)))]; + tensor linear_26_cast_fp16 = linear(bias = encoder_layers_2_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_2_feed_forward2_linear1_weight_to_fp16_palettized, x = input_169_cast_fp16)[name = string("linear_26_cast_fp16")]; + tensor input_173_cast_fp16 = silu(x = linear_26_cast_fp16)[name = string("input_173_cast_fp16")]; + tensor encoder_layers_2_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72357248))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75503040))))[name = string("encoder_layers_2_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_2_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_2_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75503232)))]; + tensor linear_27_cast_fp16 = linear(bias = encoder_layers_2_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_2_feed_forward2_linear2_weight_to_fp16_palettized, x = input_173_cast_fp16)[name = string("linear_27_cast_fp16")]; + fp16 var_1056_to_fp16 = const()[name = string("op_1056_to_fp16"), val = fp16(0x1p-1)]; + tensor var_1057_cast_fp16 = mul(x = linear_27_cast_fp16, y = var_1056_to_fp16)[name = string("op_1057_cast_fp16")]; + tensor input_179_cast_fp16 = add(x = input_167_cast_fp16, y = var_1057_cast_fp16)[name = string("input_179_cast_fp16")]; + tensor input_181_axes_0 = const()[name = string("input_181_axes_0"), val = tensor([-1])]; + tensor encoder_layers_2_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_2_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75505344)))]; + tensor encoder_layers_2_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_2_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75507456)))]; + tensor input_181_cast_fp16 = layer_norm(axes = input_181_axes_0, beta = encoder_layers_2_norm_out_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_2_norm_out_weight_to_fp16, x = input_179_cast_fp16)[name = string("input_181_cast_fp16")]; + tensor cache_13_begin_0 = const()[name = string("cache_13_begin_0"), val = tensor([3, 0, 0, 0])]; + tensor cache_13_end_0 = const()[name = string("cache_13_end_0"), val = tensor([4, 1, 42, 1024])]; + tensor cache_13_end_mask_0 = const()[name = string("cache_13_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_13_squeeze_mask_0 = const()[name = string("cache_13_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_13_cast_fp16 = slice_by_index(begin = cache_13_begin_0, end = cache_13_end_0, end_mask = cache_13_end_mask_0, squeeze_mask = cache_13_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_13_cast_fp16")]; + tensor cache_15_begin_0 = const()[name = string("cache_15_begin_0"), val = tensor([3, 0, 0, 0])]; + tensor cache_15_end_0 = const()[name = string("cache_15_end_0"), val = tensor([4, 1, 1024, 8])]; + tensor cache_15_end_mask_0 = const()[name = string("cache_15_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_15_squeeze_mask_0 = const()[name = string("cache_15_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_15_cast_fp16 = slice_by_index(begin = cache_15_begin_0, end = cache_15_end_0, end_mask = cache_15_end_mask_0, squeeze_mask = cache_15_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_15_cast_fp16")]; + tensor input_183_axes_0 = const()[name = string("input_183_axes_0"), val = tensor([-1])]; + tensor encoder_layers_3_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_3_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75509568)))]; + tensor encoder_layers_3_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_3_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75511680)))]; + tensor input_183_cast_fp16 = layer_norm(axes = input_183_axes_0, beta = encoder_layers_3_norm_feed_forward1_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_3_norm_feed_forward1_weight_to_fp16, x = input_181_cast_fp16)[name = string("input_183_cast_fp16")]; + tensor encoder_layers_3_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75513792))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78659584))))[name = string("encoder_layers_3_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_3_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_3_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78659776)))]; + tensor linear_28_cast_fp16 = linear(bias = encoder_layers_3_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_3_feed_forward1_linear1_weight_to_fp16_palettized, x = input_183_cast_fp16)[name = string("linear_28_cast_fp16")]; + tensor input_187_cast_fp16 = silu(x = linear_28_cast_fp16)[name = string("input_187_cast_fp16")]; + tensor encoder_layers_3_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78668032))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81813824))))[name = string("encoder_layers_3_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_3_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_3_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81814016)))]; + tensor linear_29_cast_fp16 = linear(bias = encoder_layers_3_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_3_feed_forward1_linear2_weight_to_fp16_palettized, x = input_187_cast_fp16)[name = string("linear_29_cast_fp16")]; + fp16 var_1093_to_fp16 = const()[name = string("op_1093_to_fp16"), val = fp16(0x1p-1)]; + tensor var_1094_cast_fp16 = mul(x = linear_29_cast_fp16, y = var_1093_to_fp16)[name = string("op_1094_cast_fp16")]; + tensor input_193_cast_fp16 = add(x = input_181_cast_fp16, y = var_1094_cast_fp16)[name = string("input_193_cast_fp16")]; + tensor key_7_axes_0 = const()[name = string("key_7_axes_0"), val = tensor([-1])]; + tensor encoder_layers_3_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_3_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81816128)))]; + tensor encoder_layers_3_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_3_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81818240)))]; + tensor key_7_cast_fp16 = layer_norm(axes = key_7_axes_0, beta = encoder_layers_3_norm_self_att_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_3_norm_self_att_weight_to_fp16, x = input_193_cast_fp16)[name = string("key_7_cast_fp16")]; + bool input_195_interleave_0 = const()[name = string("input_195_interleave_0"), val = bool(false)]; + tensor input_195_cast_fp16 = concat(axis = var_67, interleave = input_195_interleave_0, values = (cache_13_cast_fp16, key_7_cast_fp16))[name = string("input_195_cast_fp16")]; + bool var_1122_interleave_0 = const()[name = string("op_1122_interleave_0"), val = bool(false)]; + tensor var_1122_cast_fp16 = concat(axis = var_67, interleave = var_1122_interleave_0, values = key_7_cast_fp16)[name = string("op_1122_cast_fp16")]; + tensor encoder_layers_3_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81820352))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82606848))))[name = string("encoder_layers_3_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_3_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_3_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82607040)))]; + tensor linear_30_cast_fp16 = linear(bias = encoder_layers_3_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_3_self_attn_linear_q_weight_to_fp16_palettized, x = key_7_cast_fp16)[name = string("linear_30_cast_fp16")]; + tensor var_1127 = const()[name = string("op_1127"), val = tensor([1, -1, 8, 128])]; + tensor q_19_cast_fp16 = reshape(shape = var_1127, x = linear_30_cast_fp16)[name = string("q_19_cast_fp16")]; + tensor encoder_layers_3_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82609152))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83395648))))[name = string("encoder_layers_3_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_3_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_3_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83395840)))]; + tensor linear_31_cast_fp16 = linear(bias = encoder_layers_3_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_3_self_attn_linear_k_weight_to_fp16_palettized, x = input_195_cast_fp16)[name = string("linear_31_cast_fp16")]; + tensor var_1132 = const()[name = string("op_1132"), val = tensor([1, -1, 8, 128])]; + tensor k_13_cast_fp16 = reshape(shape = var_1132, x = linear_31_cast_fp16)[name = string("k_13_cast_fp16")]; + tensor encoder_layers_3_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83397952))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84184448))))[name = string("encoder_layers_3_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_3_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_3_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84184640)))]; + tensor linear_32_cast_fp16 = linear(bias = encoder_layers_3_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_3_self_attn_linear_v_weight_to_fp16_palettized, x = input_195_cast_fp16)[name = string("linear_32_cast_fp16")]; + tensor var_1137 = const()[name = string("op_1137"), val = tensor([1, -1, 8, 128])]; + tensor v_7_cast_fp16 = reshape(shape = var_1137, x = linear_32_cast_fp16)[name = string("v_7_cast_fp16")]; + tensor value_15_perm_0 = const()[name = string("value_15_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_3_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_3_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84186752)))]; + tensor var_1150_cast_fp16 = add(x = q_19_cast_fp16, y = encoder_layers_3_self_attn_pos_bias_u_to_fp16)[name = string("op_1150_cast_fp16")]; + tensor encoder_layers_3_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_3_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84188864)))]; + tensor var_1152_cast_fp16 = add(x = q_19_cast_fp16, y = encoder_layers_3_self_attn_pos_bias_v_to_fp16)[name = string("op_1152_cast_fp16")]; + tensor q_with_bias_v_7_perm_0 = const()[name = string("q_with_bias_v_7_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_85_transpose_x_0 = const()[name = string("x_85_transpose_x_0"), val = bool(false)]; + bool x_85_transpose_y_0 = const()[name = string("x_85_transpose_y_0"), val = bool(false)]; + tensor op_1154_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84190976))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84390720))))[name = string("op_1154_to_fp16_quantized")]; + tensor q_with_bias_v_7_cast_fp16 = transpose(perm = q_with_bias_v_7_perm_0, x = var_1152_cast_fp16)[name = string("transpose_335")]; + tensor x_85_cast_fp16 = matmul(transpose_x = x_85_transpose_x_0, transpose_y = x_85_transpose_y_0, x = q_with_bias_v_7_cast_fp16, y = op_1154_to_fp16_quantized)[name = string("x_85_cast_fp16")]; + tensor x_87_pad_0 = const()[name = string("x_87_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_87_mode_0 = const()[name = string("x_87_mode_0"), val = string("constant")]; + fp16 const_118_to_fp16 = const()[name = string("const_118_to_fp16"), val = fp16(0x0p+0)]; + tensor x_87_cast_fp16 = pad(constant_val = const_118_to_fp16, mode = x_87_mode_0, pad = x_87_pad_0, x = x_85_cast_fp16)[name = string("x_87_cast_fp16")]; + tensor var_1162 = const()[name = string("op_1162"), val = tensor([1, 8, -1, 56])]; + tensor x_89_cast_fp16 = reshape(shape = var_1162, x = x_87_cast_fp16)[name = string("x_89_cast_fp16")]; + tensor var_1166_begin_0 = const()[name = string("op_1166_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_1166_end_0 = const()[name = string("op_1166_end_0"), val = tensor([1, 8, 196, 56])]; + tensor var_1166_end_mask_0 = const()[name = string("op_1166_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_1166_cast_fp16 = slice_by_index(begin = var_1166_begin_0, end = var_1166_end_0, end_mask = var_1166_end_mask_0, x = x_89_cast_fp16)[name = string("op_1166_cast_fp16")]; + tensor var_1167 = const()[name = string("op_1167"), val = tensor([1, 8, 56, 195])]; + tensor matrix_bd_13_cast_fp16 = reshape(shape = var_1167, x = var_1166_cast_fp16)[name = string("matrix_bd_13_cast_fp16")]; + bool matrix_ac_7_transpose_x_0 = const()[name = string("matrix_ac_7_transpose_x_0"), val = bool(false)]; + bool matrix_ac_7_transpose_y_0 = const()[name = string("matrix_ac_7_transpose_y_0"), val = bool(false)]; + tensor transpose_102_perm_0 = const()[name = string("transpose_102_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_103_perm_0 = const()[name = string("transpose_103_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_103 = transpose(perm = transpose_103_perm_0, x = k_13_cast_fp16)[name = string("transpose_333")]; + tensor transpose_102 = transpose(perm = transpose_102_perm_0, x = var_1150_cast_fp16)[name = string("transpose_334")]; + tensor matrix_ac_7_cast_fp16 = matmul(transpose_x = matrix_ac_7_transpose_x_0, transpose_y = matrix_ac_7_transpose_y_0, x = transpose_102, y = transpose_103)[name = string("matrix_ac_7_cast_fp16")]; + tensor matrix_bd_15_begin_0 = const()[name = string("matrix_bd_15_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_15_end_0 = const()[name = string("matrix_bd_15_end_0"), val = tensor([1, 8, 56, 98])]; + tensor matrix_bd_15_end_mask_0 = const()[name = string("matrix_bd_15_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_15_cast_fp16 = slice_by_index(begin = matrix_bd_15_begin_0, end = matrix_bd_15_end_0, end_mask = matrix_bd_15_end_mask_0, x = matrix_bd_13_cast_fp16)[name = string("matrix_bd_15_cast_fp16")]; + tensor var_1176_cast_fp16 = add(x = matrix_ac_7_cast_fp16, y = matrix_bd_15_cast_fp16)[name = string("op_1176_cast_fp16")]; + fp16 _inversed_scores_13_y_0_to_fp16 = const()[name = string("_inversed_scores_13_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_13_cast_fp16 = mul(x = var_1176_cast_fp16, y = _inversed_scores_13_y_0_to_fp16)[name = string("_inversed_scores_13_cast_fp16")]; + tensor scores_15_cast_fp16 = select(a = var_44_to_fp16, b = _inversed_scores_13_cast_fp16, cond = mask_11)[name = string("scores_15_cast_fp16")]; + tensor var_1182_cast_fp16 = softmax(axis = var_58, x = scores_15_cast_fp16)[name = string("op_1182_cast_fp16")]; + tensor input_197_cast_fp16 = select(a = var_43_to_fp16, b = var_1182_cast_fp16, cond = mask_11)[name = string("input_197_cast_fp16")]; + bool x_91_transpose_x_0 = const()[name = string("x_91_transpose_x_0"), val = bool(false)]; + bool x_91_transpose_y_0 = const()[name = string("x_91_transpose_y_0"), val = bool(false)]; + tensor value_15_cast_fp16 = transpose(perm = value_15_perm_0, x = v_7_cast_fp16)[name = string("transpose_332")]; + tensor x_91_cast_fp16 = matmul(transpose_x = x_91_transpose_x_0, transpose_y = x_91_transpose_y_0, x = input_197_cast_fp16, y = value_15_cast_fp16)[name = string("x_91_cast_fp16")]; + tensor var_1186_perm_0 = const()[name = string("op_1186_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1187 = const()[name = string("op_1187"), val = tensor([1, -1, 1024])]; + tensor var_1186_cast_fp16 = transpose(perm = var_1186_perm_0, x = x_91_cast_fp16)[name = string("transpose_331")]; + tensor input_199_cast_fp16 = reshape(shape = var_1187, x = var_1186_cast_fp16)[name = string("input_199_cast_fp16")]; + tensor encoder_layers_3_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84391232))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85177728))))[name = string("encoder_layers_3_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_3_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_3_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85177920)))]; + tensor linear_34_cast_fp16 = linear(bias = encoder_layers_3_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_3_self_attn_linear_out_weight_to_fp16_palettized, x = input_199_cast_fp16)[name = string("linear_34_cast_fp16")]; + tensor input_203_cast_fp16 = add(x = input_193_cast_fp16, y = linear_34_cast_fp16)[name = string("input_203_cast_fp16")]; + tensor x_95_axes_0 = const()[name = string("x_95_axes_0"), val = tensor([-1])]; + tensor encoder_layers_3_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_3_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85180032)))]; + tensor encoder_layers_3_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_3_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85182144)))]; + tensor x_95_cast_fp16 = layer_norm(axes = x_95_axes_0, beta = encoder_layers_3_norm_conv_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_3_norm_conv_weight_to_fp16, x = input_203_cast_fp16)[name = string("x_95_cast_fp16")]; + tensor input_205_perm_0 = const()[name = string("input_205_perm_0"), val = tensor([0, 2, 1])]; + string input_207_pad_type_0 = const()[name = string("input_207_pad_type_0"), val = string("valid")]; + tensor input_207_strides_0 = const()[name = string("input_207_strides_0"), val = tensor([1])]; + tensor input_207_pad_0 = const()[name = string("input_207_pad_0"), val = tensor([0, 0])]; + tensor input_207_dilations_0 = const()[name = string("input_207_dilations_0"), val = tensor([1])]; + int32 input_207_groups_0 = const()[name = string("input_207_groups_0"), val = int32(1)]; + tensor encoder_layers_3_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85184256))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87281472))))[name = string("encoder_layers_3_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_205_cast_fp16 = transpose(perm = input_205_perm_0, x = x_95_cast_fp16)[name = string("transpose_330")]; + tensor input_207_cast_fp16 = conv(dilations = input_207_dilations_0, groups = input_207_groups_0, pad = input_207_pad_0, pad_type = input_207_pad_type_0, strides = input_207_strides_0, weight = encoder_layers_3_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_205_cast_fp16)[name = string("input_207_cast_fp16")]; + int32 x_97_split_num_splits_0 = const()[name = string("x_97_split_num_splits_0"), val = int32(2)]; + int32 x_97_split_axis_0 = const()[name = string("x_97_split_axis_0"), val = int32(1)]; + tensor x_97_split_cast_fp16_0, tensor x_97_split_cast_fp16_1 = split(axis = x_97_split_axis_0, num_splits = x_97_split_num_splits_0, x = input_207_cast_fp16)[name = string("x_97_split_cast_fp16")]; + tensor x_97_split_1_sigmoid_cast_fp16 = sigmoid(x = x_97_split_cast_fp16_1)[name = string("x_97_split_1_sigmoid_cast_fp16")]; + tensor x_97_cast_fp16 = mul(x = x_97_split_cast_fp16_0, y = x_97_split_1_sigmoid_cast_fp16)[name = string("x_97_cast_fp16")]; + tensor input_209_cast_fp16 = select(a = var_43_to_fp16, b = x_97_cast_fp16, cond = var_574)[name = string("input_209_cast_fp16")]; + bool new_x_15_interleave_0 = const()[name = string("new_x_15_interleave_0"), val = bool(false)]; + tensor new_x_15_cast_fp16 = concat(axis = var_58, interleave = new_x_15_interleave_0, values = (cache_15_cast_fp16, input_209_cast_fp16))[name = string("new_x_15_cast_fp16")]; + tensor var_1226_begin_0 = const()[name = string("op_1226_begin_0"), val = tensor([0, 0, 56])]; + tensor var_1226_end_0 = const()[name = string("op_1226_end_0"), val = tensor([1, 1024, 64])]; + tensor var_1226_end_mask_0 = const()[name = string("op_1226_end_mask_0"), val = tensor([true, true, true])]; + tensor var_1226_cast_fp16 = slice_by_index(begin = var_1226_begin_0, end = var_1226_end_0, end_mask = var_1226_end_mask_0, x = new_x_15_cast_fp16)[name = string("op_1226_cast_fp16")]; + string x_99_pad_type_0 = const()[name = string("x_99_pad_type_0"), val = string("valid")]; + int32 x_99_groups_0 = const()[name = string("x_99_groups_0"), val = int32(1024)]; + tensor x_99_strides_0 = const()[name = string("x_99_strides_0"), val = tensor([1])]; + tensor x_99_pad_0 = const()[name = string("x_99_pad_0"), val = tensor([0, 0])]; + tensor x_99_dilations_0 = const()[name = string("x_99_dilations_0"), val = tensor([1])]; + tensor encoder_layers_3_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87285632))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87294912))))[name = string("encoder_layers_3_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_99_cast_fp16 = conv(dilations = x_99_dilations_0, groups = x_99_groups_0, pad = x_99_pad_0, pad_type = x_99_pad_type_0, strides = x_99_strides_0, weight = encoder_layers_3_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_15_cast_fp16)[name = string("x_99_cast_fp16")]; + tensor input_211_perm_0 = const()[name = string("input_211_perm_0"), val = tensor([0, 2, 1])]; + tensor x_101_axes_0 = const()[name = string("x_101_axes_0"), val = tensor([-1])]; + tensor encoder_layers_3_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_3_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87297024)))]; + tensor encoder_layers_3_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_3_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87299136)))]; + tensor input_211_cast_fp16 = transpose(perm = input_211_perm_0, x = x_99_cast_fp16)[name = string("transpose_329")]; + tensor x_101_cast_fp16 = layer_norm(axes = x_101_axes_0, beta = encoder_layers_3_conv_batch_norm_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_3_conv_batch_norm_weight_to_fp16, x = input_211_cast_fp16)[name = string("x_101_cast_fp16")]; + tensor input_213_perm_0 = const()[name = string("input_213_perm_0"), val = tensor([0, 2, 1])]; + tensor input_213_cast_fp16 = transpose(perm = input_213_perm_0, x = x_101_cast_fp16)[name = string("transpose_328")]; + tensor input_215_cast_fp16 = silu(x = input_213_cast_fp16)[name = string("input_215_cast_fp16")]; + string x_103_pad_type_0 = const()[name = string("x_103_pad_type_0"), val = string("valid")]; + tensor x_103_strides_0 = const()[name = string("x_103_strides_0"), val = tensor([1])]; + tensor x_103_pad_0 = const()[name = string("x_103_pad_0"), val = tensor([0, 0])]; + tensor x_103_dilations_0 = const()[name = string("x_103_dilations_0"), val = tensor([1])]; + int32 x_103_groups_0 = const()[name = string("x_103_groups_0"), val = int32(1)]; + tensor encoder_layers_3_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87301248))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88349888))))[name = string("encoder_layers_3_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_103_cast_fp16 = conv(dilations = x_103_dilations_0, groups = x_103_groups_0, pad = x_103_pad_0, pad_type = x_103_pad_type_0, strides = x_103_strides_0, weight = encoder_layers_3_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_215_cast_fp16)[name = string("x_103_cast_fp16")]; + tensor input_217_perm_0 = const()[name = string("input_217_perm_0"), val = tensor([0, 2, 1])]; + tensor input_217_cast_fp16 = transpose(perm = input_217_perm_0, x = x_103_cast_fp16)[name = string("transpose_327")]; + tensor input_219_cast_fp16 = add(x = input_203_cast_fp16, y = input_217_cast_fp16)[name = string("input_219_cast_fp16")]; + tensor input_221_axes_0 = const()[name = string("input_221_axes_0"), val = tensor([-1])]; + tensor encoder_layers_3_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_3_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88352000)))]; + tensor encoder_layers_3_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_3_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88354112)))]; + tensor input_221_cast_fp16 = layer_norm(axes = input_221_axes_0, beta = encoder_layers_3_norm_feed_forward2_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_3_norm_feed_forward2_weight_to_fp16, x = input_219_cast_fp16)[name = string("input_221_cast_fp16")]; + tensor encoder_layers_3_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88356224))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(91502016))))[name = string("encoder_layers_3_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_3_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_3_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(91502208)))]; + tensor linear_35_cast_fp16 = linear(bias = encoder_layers_3_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_3_feed_forward2_linear1_weight_to_fp16_palettized, x = input_221_cast_fp16)[name = string("linear_35_cast_fp16")]; + tensor input_225_cast_fp16 = silu(x = linear_35_cast_fp16)[name = string("input_225_cast_fp16")]; + tensor encoder_layers_3_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(91510464))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94656256))))[name = string("encoder_layers_3_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_3_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_3_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94656448)))]; + tensor linear_36_cast_fp16 = linear(bias = encoder_layers_3_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_3_feed_forward2_linear2_weight_to_fp16_palettized, x = input_225_cast_fp16)[name = string("linear_36_cast_fp16")]; + fp16 var_1269_to_fp16 = const()[name = string("op_1269_to_fp16"), val = fp16(0x1p-1)]; + tensor var_1270_cast_fp16 = mul(x = linear_36_cast_fp16, y = var_1269_to_fp16)[name = string("op_1270_cast_fp16")]; + tensor input_231_cast_fp16 = add(x = input_219_cast_fp16, y = var_1270_cast_fp16)[name = string("input_231_cast_fp16")]; + tensor input_233_axes_0 = const()[name = string("input_233_axes_0"), val = tensor([-1])]; + tensor encoder_layers_3_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_3_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94658560)))]; + tensor encoder_layers_3_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_3_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94660672)))]; + tensor input_233_cast_fp16 = layer_norm(axes = input_233_axes_0, beta = encoder_layers_3_norm_out_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_3_norm_out_weight_to_fp16, x = input_231_cast_fp16)[name = string("input_233_cast_fp16")]; + tensor cache_17_begin_0 = const()[name = string("cache_17_begin_0"), val = tensor([4, 0, 0, 0])]; + tensor cache_17_end_0 = const()[name = string("cache_17_end_0"), val = tensor([5, 1, 42, 1024])]; + tensor cache_17_end_mask_0 = const()[name = string("cache_17_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_17_squeeze_mask_0 = const()[name = string("cache_17_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_17_cast_fp16 = slice_by_index(begin = cache_17_begin_0, end = cache_17_end_0, end_mask = cache_17_end_mask_0, squeeze_mask = cache_17_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_17_cast_fp16")]; + tensor cache_19_begin_0 = const()[name = string("cache_19_begin_0"), val = tensor([4, 0, 0, 0])]; + tensor cache_19_end_0 = const()[name = string("cache_19_end_0"), val = tensor([5, 1, 1024, 8])]; + tensor cache_19_end_mask_0 = const()[name = string("cache_19_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_19_squeeze_mask_0 = const()[name = string("cache_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_19_cast_fp16 = slice_by_index(begin = cache_19_begin_0, end = cache_19_end_0, end_mask = cache_19_end_mask_0, squeeze_mask = cache_19_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_19_cast_fp16")]; + tensor input_235_axes_0 = const()[name = string("input_235_axes_0"), val = tensor([-1])]; + tensor encoder_layers_4_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_4_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94662784)))]; + tensor encoder_layers_4_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_4_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94664896)))]; + tensor input_235_cast_fp16 = layer_norm(axes = input_235_axes_0, beta = encoder_layers_4_norm_feed_forward1_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_4_norm_feed_forward1_weight_to_fp16, x = input_233_cast_fp16)[name = string("input_235_cast_fp16")]; + tensor encoder_layers_4_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94667008))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97812800))))[name = string("encoder_layers_4_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_4_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_4_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97812992)))]; + tensor linear_37_cast_fp16 = linear(bias = encoder_layers_4_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_4_feed_forward1_linear1_weight_to_fp16_palettized, x = input_235_cast_fp16)[name = string("linear_37_cast_fp16")]; + tensor input_239_cast_fp16 = silu(x = linear_37_cast_fp16)[name = string("input_239_cast_fp16")]; + tensor encoder_layers_4_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97821248))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100967040))))[name = string("encoder_layers_4_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_4_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_4_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100967232)))]; + tensor linear_38_cast_fp16 = linear(bias = encoder_layers_4_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_4_feed_forward1_linear2_weight_to_fp16_palettized, x = input_239_cast_fp16)[name = string("linear_38_cast_fp16")]; + fp16 var_1306_to_fp16 = const()[name = string("op_1306_to_fp16"), val = fp16(0x1p-1)]; + tensor var_1307_cast_fp16 = mul(x = linear_38_cast_fp16, y = var_1306_to_fp16)[name = string("op_1307_cast_fp16")]; + tensor input_245_cast_fp16 = add(x = input_233_cast_fp16, y = var_1307_cast_fp16)[name = string("input_245_cast_fp16")]; + tensor key_9_axes_0 = const()[name = string("key_9_axes_0"), val = tensor([-1])]; + tensor encoder_layers_4_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_4_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100969344)))]; + tensor encoder_layers_4_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_4_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100971456)))]; + tensor key_9_cast_fp16 = layer_norm(axes = key_9_axes_0, beta = encoder_layers_4_norm_self_att_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_4_norm_self_att_weight_to_fp16, x = input_245_cast_fp16)[name = string("key_9_cast_fp16")]; + bool input_247_interleave_0 = const()[name = string("input_247_interleave_0"), val = bool(false)]; + tensor input_247_cast_fp16 = concat(axis = var_67, interleave = input_247_interleave_0, values = (cache_17_cast_fp16, key_9_cast_fp16))[name = string("input_247_cast_fp16")]; + bool var_1335_interleave_0 = const()[name = string("op_1335_interleave_0"), val = bool(false)]; + tensor var_1335_cast_fp16 = concat(axis = var_67, interleave = var_1335_interleave_0, values = key_9_cast_fp16)[name = string("op_1335_cast_fp16")]; + tensor encoder_layers_4_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100973568))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(101760064))))[name = string("encoder_layers_4_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_4_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_4_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(101760256)))]; + tensor linear_39_cast_fp16 = linear(bias = encoder_layers_4_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_4_self_attn_linear_q_weight_to_fp16_palettized, x = key_9_cast_fp16)[name = string("linear_39_cast_fp16")]; + tensor var_1340 = const()[name = string("op_1340"), val = tensor([1, -1, 8, 128])]; + tensor q_25_cast_fp16 = reshape(shape = var_1340, x = linear_39_cast_fp16)[name = string("q_25_cast_fp16")]; + tensor encoder_layers_4_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(101762368))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102548864))))[name = string("encoder_layers_4_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_4_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_4_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102549056)))]; + tensor linear_40_cast_fp16 = linear(bias = encoder_layers_4_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_4_self_attn_linear_k_weight_to_fp16_palettized, x = input_247_cast_fp16)[name = string("linear_40_cast_fp16")]; + tensor var_1345 = const()[name = string("op_1345"), val = tensor([1, -1, 8, 128])]; + tensor k_17_cast_fp16 = reshape(shape = var_1345, x = linear_40_cast_fp16)[name = string("k_17_cast_fp16")]; + tensor encoder_layers_4_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102551168))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103337664))))[name = string("encoder_layers_4_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_4_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_4_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103337856)))]; + tensor linear_41_cast_fp16 = linear(bias = encoder_layers_4_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_4_self_attn_linear_v_weight_to_fp16_palettized, x = input_247_cast_fp16)[name = string("linear_41_cast_fp16")]; + tensor var_1350 = const()[name = string("op_1350"), val = tensor([1, -1, 8, 128])]; + tensor v_9_cast_fp16 = reshape(shape = var_1350, x = linear_41_cast_fp16)[name = string("v_9_cast_fp16")]; + tensor value_17_perm_0 = const()[name = string("value_17_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_4_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_4_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103339968)))]; + tensor var_1363_cast_fp16 = add(x = q_25_cast_fp16, y = encoder_layers_4_self_attn_pos_bias_u_to_fp16)[name = string("op_1363_cast_fp16")]; + tensor encoder_layers_4_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_4_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103342080)))]; + tensor var_1365_cast_fp16 = add(x = q_25_cast_fp16, y = encoder_layers_4_self_attn_pos_bias_v_to_fp16)[name = string("op_1365_cast_fp16")]; + tensor q_with_bias_v_9_perm_0 = const()[name = string("q_with_bias_v_9_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_111_transpose_x_0 = const()[name = string("x_111_transpose_x_0"), val = bool(false)]; + bool x_111_transpose_y_0 = const()[name = string("x_111_transpose_y_0"), val = bool(false)]; + tensor op_1367_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103344192))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103543936))))[name = string("op_1367_to_fp16_quantized")]; + tensor q_with_bias_v_9_cast_fp16 = transpose(perm = q_with_bias_v_9_perm_0, x = var_1365_cast_fp16)[name = string("transpose_326")]; + tensor x_111_cast_fp16 = matmul(transpose_x = x_111_transpose_x_0, transpose_y = x_111_transpose_y_0, x = q_with_bias_v_9_cast_fp16, y = op_1367_to_fp16_quantized)[name = string("x_111_cast_fp16")]; + tensor x_113_pad_0 = const()[name = string("x_113_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_113_mode_0 = const()[name = string("x_113_mode_0"), val = string("constant")]; + fp16 const_131_to_fp16 = const()[name = string("const_131_to_fp16"), val = fp16(0x0p+0)]; + tensor x_113_cast_fp16 = pad(constant_val = const_131_to_fp16, mode = x_113_mode_0, pad = x_113_pad_0, x = x_111_cast_fp16)[name = string("x_113_cast_fp16")]; + tensor var_1375 = const()[name = string("op_1375"), val = tensor([1, 8, -1, 56])]; + tensor x_115_cast_fp16 = reshape(shape = var_1375, x = x_113_cast_fp16)[name = string("x_115_cast_fp16")]; + tensor var_1379_begin_0 = const()[name = string("op_1379_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_1379_end_0 = const()[name = string("op_1379_end_0"), val = tensor([1, 8, 196, 56])]; + tensor var_1379_end_mask_0 = const()[name = string("op_1379_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_1379_cast_fp16 = slice_by_index(begin = var_1379_begin_0, end = var_1379_end_0, end_mask = var_1379_end_mask_0, x = x_115_cast_fp16)[name = string("op_1379_cast_fp16")]; + tensor var_1380 = const()[name = string("op_1380"), val = tensor([1, 8, 56, 195])]; + tensor matrix_bd_17_cast_fp16 = reshape(shape = var_1380, x = var_1379_cast_fp16)[name = string("matrix_bd_17_cast_fp16")]; + bool matrix_ac_9_transpose_x_0 = const()[name = string("matrix_ac_9_transpose_x_0"), val = bool(false)]; + bool matrix_ac_9_transpose_y_0 = const()[name = string("matrix_ac_9_transpose_y_0"), val = bool(false)]; + tensor transpose_104_perm_0 = const()[name = string("transpose_104_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_105_perm_0 = const()[name = string("transpose_105_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_105 = transpose(perm = transpose_105_perm_0, x = k_17_cast_fp16)[name = string("transpose_324")]; + tensor transpose_104 = transpose(perm = transpose_104_perm_0, x = var_1363_cast_fp16)[name = string("transpose_325")]; + tensor matrix_ac_9_cast_fp16 = matmul(transpose_x = matrix_ac_9_transpose_x_0, transpose_y = matrix_ac_9_transpose_y_0, x = transpose_104, y = transpose_105)[name = string("matrix_ac_9_cast_fp16")]; + tensor matrix_bd_19_begin_0 = const()[name = string("matrix_bd_19_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_19_end_0 = const()[name = string("matrix_bd_19_end_0"), val = tensor([1, 8, 56, 98])]; + tensor matrix_bd_19_end_mask_0 = const()[name = string("matrix_bd_19_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_19_cast_fp16 = slice_by_index(begin = matrix_bd_19_begin_0, end = matrix_bd_19_end_0, end_mask = matrix_bd_19_end_mask_0, x = matrix_bd_17_cast_fp16)[name = string("matrix_bd_19_cast_fp16")]; + tensor var_1389_cast_fp16 = add(x = matrix_ac_9_cast_fp16, y = matrix_bd_19_cast_fp16)[name = string("op_1389_cast_fp16")]; + fp16 _inversed_scores_17_y_0_to_fp16 = const()[name = string("_inversed_scores_17_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_17_cast_fp16 = mul(x = var_1389_cast_fp16, y = _inversed_scores_17_y_0_to_fp16)[name = string("_inversed_scores_17_cast_fp16")]; + tensor scores_19_cast_fp16 = select(a = var_44_to_fp16, b = _inversed_scores_17_cast_fp16, cond = mask_11)[name = string("scores_19_cast_fp16")]; + tensor var_1395_cast_fp16 = softmax(axis = var_58, x = scores_19_cast_fp16)[name = string("op_1395_cast_fp16")]; + tensor input_249_cast_fp16 = select(a = var_43_to_fp16, b = var_1395_cast_fp16, cond = mask_11)[name = string("input_249_cast_fp16")]; + bool x_117_transpose_x_0 = const()[name = string("x_117_transpose_x_0"), val = bool(false)]; + bool x_117_transpose_y_0 = const()[name = string("x_117_transpose_y_0"), val = bool(false)]; + tensor value_17_cast_fp16 = transpose(perm = value_17_perm_0, x = v_9_cast_fp16)[name = string("transpose_323")]; + tensor x_117_cast_fp16 = matmul(transpose_x = x_117_transpose_x_0, transpose_y = x_117_transpose_y_0, x = input_249_cast_fp16, y = value_17_cast_fp16)[name = string("x_117_cast_fp16")]; + tensor var_1399_perm_0 = const()[name = string("op_1399_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1400 = const()[name = string("op_1400"), val = tensor([1, -1, 1024])]; + tensor var_1399_cast_fp16 = transpose(perm = var_1399_perm_0, x = x_117_cast_fp16)[name = string("transpose_322")]; + tensor input_251_cast_fp16 = reshape(shape = var_1400, x = var_1399_cast_fp16)[name = string("input_251_cast_fp16")]; + tensor encoder_layers_4_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103544448))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104330944))))[name = string("encoder_layers_4_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_4_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_4_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104331136)))]; + tensor linear_43_cast_fp16 = linear(bias = encoder_layers_4_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_4_self_attn_linear_out_weight_to_fp16_palettized, x = input_251_cast_fp16)[name = string("linear_43_cast_fp16")]; + tensor input_255_cast_fp16 = add(x = input_245_cast_fp16, y = linear_43_cast_fp16)[name = string("input_255_cast_fp16")]; + tensor x_121_axes_0 = const()[name = string("x_121_axes_0"), val = tensor([-1])]; + tensor encoder_layers_4_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_4_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104333248)))]; + tensor encoder_layers_4_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_4_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104335360)))]; + tensor x_121_cast_fp16 = layer_norm(axes = x_121_axes_0, beta = encoder_layers_4_norm_conv_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_4_norm_conv_weight_to_fp16, x = input_255_cast_fp16)[name = string("x_121_cast_fp16")]; + tensor input_257_perm_0 = const()[name = string("input_257_perm_0"), val = tensor([0, 2, 1])]; + string input_259_pad_type_0 = const()[name = string("input_259_pad_type_0"), val = string("valid")]; + tensor input_259_strides_0 = const()[name = string("input_259_strides_0"), val = tensor([1])]; + tensor input_259_pad_0 = const()[name = string("input_259_pad_0"), val = tensor([0, 0])]; + tensor input_259_dilations_0 = const()[name = string("input_259_dilations_0"), val = tensor([1])]; + int32 input_259_groups_0 = const()[name = string("input_259_groups_0"), val = int32(1)]; + tensor encoder_layers_4_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104337472))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106434688))))[name = string("encoder_layers_4_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_257_cast_fp16 = transpose(perm = input_257_perm_0, x = x_121_cast_fp16)[name = string("transpose_321")]; + tensor input_259_cast_fp16 = conv(dilations = input_259_dilations_0, groups = input_259_groups_0, pad = input_259_pad_0, pad_type = input_259_pad_type_0, strides = input_259_strides_0, weight = encoder_layers_4_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_257_cast_fp16)[name = string("input_259_cast_fp16")]; + int32 x_123_split_num_splits_0 = const()[name = string("x_123_split_num_splits_0"), val = int32(2)]; + int32 x_123_split_axis_0 = const()[name = string("x_123_split_axis_0"), val = int32(1)]; + tensor x_123_split_cast_fp16_0, tensor x_123_split_cast_fp16_1 = split(axis = x_123_split_axis_0, num_splits = x_123_split_num_splits_0, x = input_259_cast_fp16)[name = string("x_123_split_cast_fp16")]; + tensor x_123_split_1_sigmoid_cast_fp16 = sigmoid(x = x_123_split_cast_fp16_1)[name = string("x_123_split_1_sigmoid_cast_fp16")]; + tensor x_123_cast_fp16 = mul(x = x_123_split_cast_fp16_0, y = x_123_split_1_sigmoid_cast_fp16)[name = string("x_123_cast_fp16")]; + tensor input_261_cast_fp16 = select(a = var_43_to_fp16, b = x_123_cast_fp16, cond = var_574)[name = string("input_261_cast_fp16")]; + bool new_x_19_interleave_0 = const()[name = string("new_x_19_interleave_0"), val = bool(false)]; + tensor new_x_19_cast_fp16 = concat(axis = var_58, interleave = new_x_19_interleave_0, values = (cache_19_cast_fp16, input_261_cast_fp16))[name = string("new_x_19_cast_fp16")]; + tensor var_1439_begin_0 = const()[name = string("op_1439_begin_0"), val = tensor([0, 0, 56])]; + tensor var_1439_end_0 = const()[name = string("op_1439_end_0"), val = tensor([1, 1024, 64])]; + tensor var_1439_end_mask_0 = const()[name = string("op_1439_end_mask_0"), val = tensor([true, true, true])]; + tensor var_1439_cast_fp16 = slice_by_index(begin = var_1439_begin_0, end = var_1439_end_0, end_mask = var_1439_end_mask_0, x = new_x_19_cast_fp16)[name = string("op_1439_cast_fp16")]; + string x_125_pad_type_0 = const()[name = string("x_125_pad_type_0"), val = string("valid")]; + int32 x_125_groups_0 = const()[name = string("x_125_groups_0"), val = int32(1024)]; + tensor x_125_strides_0 = const()[name = string("x_125_strides_0"), val = tensor([1])]; + tensor x_125_pad_0 = const()[name = string("x_125_pad_0"), val = tensor([0, 0])]; + tensor x_125_dilations_0 = const()[name = string("x_125_dilations_0"), val = tensor([1])]; + tensor encoder_layers_4_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106438848))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106448128))))[name = string("encoder_layers_4_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_125_cast_fp16 = conv(dilations = x_125_dilations_0, groups = x_125_groups_0, pad = x_125_pad_0, pad_type = x_125_pad_type_0, strides = x_125_strides_0, weight = encoder_layers_4_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_19_cast_fp16)[name = string("x_125_cast_fp16")]; + tensor input_263_perm_0 = const()[name = string("input_263_perm_0"), val = tensor([0, 2, 1])]; + tensor x_127_axes_0 = const()[name = string("x_127_axes_0"), val = tensor([-1])]; + tensor encoder_layers_4_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_4_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106450240)))]; + tensor encoder_layers_4_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_4_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106452352)))]; + tensor input_263_cast_fp16 = transpose(perm = input_263_perm_0, x = x_125_cast_fp16)[name = string("transpose_320")]; + tensor x_127_cast_fp16 = layer_norm(axes = x_127_axes_0, beta = encoder_layers_4_conv_batch_norm_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_4_conv_batch_norm_weight_to_fp16, x = input_263_cast_fp16)[name = string("x_127_cast_fp16")]; + tensor input_265_perm_0 = const()[name = string("input_265_perm_0"), val = tensor([0, 2, 1])]; + tensor input_265_cast_fp16 = transpose(perm = input_265_perm_0, x = x_127_cast_fp16)[name = string("transpose_319")]; + tensor input_267_cast_fp16 = silu(x = input_265_cast_fp16)[name = string("input_267_cast_fp16")]; + string x_129_pad_type_0 = const()[name = string("x_129_pad_type_0"), val = string("valid")]; + tensor x_129_strides_0 = const()[name = string("x_129_strides_0"), val = tensor([1])]; + tensor x_129_pad_0 = const()[name = string("x_129_pad_0"), val = tensor([0, 0])]; + tensor x_129_dilations_0 = const()[name = string("x_129_dilations_0"), val = tensor([1])]; + int32 x_129_groups_0 = const()[name = string("x_129_groups_0"), val = int32(1)]; + tensor encoder_layers_4_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106454464))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107503104))))[name = string("encoder_layers_4_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_129_cast_fp16 = conv(dilations = x_129_dilations_0, groups = x_129_groups_0, pad = x_129_pad_0, pad_type = x_129_pad_type_0, strides = x_129_strides_0, weight = encoder_layers_4_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_267_cast_fp16)[name = string("x_129_cast_fp16")]; + tensor input_269_perm_0 = const()[name = string("input_269_perm_0"), val = tensor([0, 2, 1])]; + tensor input_269_cast_fp16 = transpose(perm = input_269_perm_0, x = x_129_cast_fp16)[name = string("transpose_318")]; + tensor input_271_cast_fp16 = add(x = input_255_cast_fp16, y = input_269_cast_fp16)[name = string("input_271_cast_fp16")]; + tensor input_273_axes_0 = const()[name = string("input_273_axes_0"), val = tensor([-1])]; + tensor encoder_layers_4_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_4_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107505216)))]; + tensor encoder_layers_4_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_4_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107507328)))]; + tensor input_273_cast_fp16 = layer_norm(axes = input_273_axes_0, beta = encoder_layers_4_norm_feed_forward2_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_4_norm_feed_forward2_weight_to_fp16, x = input_271_cast_fp16)[name = string("input_273_cast_fp16")]; + tensor encoder_layers_4_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107509440))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110655232))))[name = string("encoder_layers_4_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_4_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_4_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110655424)))]; + tensor linear_44_cast_fp16 = linear(bias = encoder_layers_4_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_4_feed_forward2_linear1_weight_to_fp16_palettized, x = input_273_cast_fp16)[name = string("linear_44_cast_fp16")]; + tensor input_277_cast_fp16 = silu(x = linear_44_cast_fp16)[name = string("input_277_cast_fp16")]; + tensor encoder_layers_4_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110663680))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113809472))))[name = string("encoder_layers_4_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_4_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_4_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113809664)))]; + tensor linear_45_cast_fp16 = linear(bias = encoder_layers_4_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_4_feed_forward2_linear2_weight_to_fp16_palettized, x = input_277_cast_fp16)[name = string("linear_45_cast_fp16")]; + fp16 var_1482_to_fp16 = const()[name = string("op_1482_to_fp16"), val = fp16(0x1p-1)]; + tensor var_1483_cast_fp16 = mul(x = linear_45_cast_fp16, y = var_1482_to_fp16)[name = string("op_1483_cast_fp16")]; + tensor input_283_cast_fp16 = add(x = input_271_cast_fp16, y = var_1483_cast_fp16)[name = string("input_283_cast_fp16")]; + tensor input_285_axes_0 = const()[name = string("input_285_axes_0"), val = tensor([-1])]; + tensor encoder_layers_4_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_4_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113811776)))]; + tensor encoder_layers_4_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_4_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113813888)))]; + tensor input_285_cast_fp16 = layer_norm(axes = input_285_axes_0, beta = encoder_layers_4_norm_out_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_4_norm_out_weight_to_fp16, x = input_283_cast_fp16)[name = string("input_285_cast_fp16")]; + tensor cache_21_begin_0 = const()[name = string("cache_21_begin_0"), val = tensor([5, 0, 0, 0])]; + tensor cache_21_end_0 = const()[name = string("cache_21_end_0"), val = tensor([6, 1, 42, 1024])]; + tensor cache_21_end_mask_0 = const()[name = string("cache_21_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_21_squeeze_mask_0 = const()[name = string("cache_21_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_21_cast_fp16 = slice_by_index(begin = cache_21_begin_0, end = cache_21_end_0, end_mask = cache_21_end_mask_0, squeeze_mask = cache_21_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_21_cast_fp16")]; + tensor cache_23_begin_0 = const()[name = string("cache_23_begin_0"), val = tensor([5, 0, 0, 0])]; + tensor cache_23_end_0 = const()[name = string("cache_23_end_0"), val = tensor([6, 1, 1024, 8])]; + tensor cache_23_end_mask_0 = const()[name = string("cache_23_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_23_squeeze_mask_0 = const()[name = string("cache_23_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_23_cast_fp16 = slice_by_index(begin = cache_23_begin_0, end = cache_23_end_0, end_mask = cache_23_end_mask_0, squeeze_mask = cache_23_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_23_cast_fp16")]; + tensor input_287_axes_0 = const()[name = string("input_287_axes_0"), val = tensor([-1])]; + tensor encoder_layers_5_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_5_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113816000)))]; + tensor encoder_layers_5_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_5_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113818112)))]; + tensor input_287_cast_fp16 = layer_norm(axes = input_287_axes_0, beta = encoder_layers_5_norm_feed_forward1_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_5_norm_feed_forward1_weight_to_fp16, x = input_285_cast_fp16)[name = string("input_287_cast_fp16")]; + tensor encoder_layers_5_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113820224))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116966016))))[name = string("encoder_layers_5_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_5_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_5_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116966208)))]; + tensor linear_46_cast_fp16 = linear(bias = encoder_layers_5_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_5_feed_forward1_linear1_weight_to_fp16_palettized, x = input_287_cast_fp16)[name = string("linear_46_cast_fp16")]; + tensor input_291_cast_fp16 = silu(x = linear_46_cast_fp16)[name = string("input_291_cast_fp16")]; + tensor encoder_layers_5_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116974464))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120120256))))[name = string("encoder_layers_5_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_5_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_5_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120120448)))]; + tensor linear_47_cast_fp16 = linear(bias = encoder_layers_5_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_5_feed_forward1_linear2_weight_to_fp16_palettized, x = input_291_cast_fp16)[name = string("linear_47_cast_fp16")]; + fp16 var_1519_to_fp16 = const()[name = string("op_1519_to_fp16"), val = fp16(0x1p-1)]; + tensor var_1520_cast_fp16 = mul(x = linear_47_cast_fp16, y = var_1519_to_fp16)[name = string("op_1520_cast_fp16")]; + tensor input_297_cast_fp16 = add(x = input_285_cast_fp16, y = var_1520_cast_fp16)[name = string("input_297_cast_fp16")]; + tensor key_11_axes_0 = const()[name = string("key_11_axes_0"), val = tensor([-1])]; + tensor encoder_layers_5_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_5_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120122560)))]; + tensor encoder_layers_5_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_5_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120124672)))]; + tensor key_11_cast_fp16 = layer_norm(axes = key_11_axes_0, beta = encoder_layers_5_norm_self_att_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_5_norm_self_att_weight_to_fp16, x = input_297_cast_fp16)[name = string("key_11_cast_fp16")]; + bool input_299_interleave_0 = const()[name = string("input_299_interleave_0"), val = bool(false)]; + tensor input_299_cast_fp16 = concat(axis = var_67, interleave = input_299_interleave_0, values = (cache_21_cast_fp16, key_11_cast_fp16))[name = string("input_299_cast_fp16")]; + bool var_1548_interleave_0 = const()[name = string("op_1548_interleave_0"), val = bool(false)]; + tensor var_1548_cast_fp16 = concat(axis = var_67, interleave = var_1548_interleave_0, values = key_11_cast_fp16)[name = string("op_1548_cast_fp16")]; + tensor encoder_layers_5_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120126784))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120913280))))[name = string("encoder_layers_5_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_5_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_5_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120913472)))]; + tensor linear_48_cast_fp16 = linear(bias = encoder_layers_5_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_5_self_attn_linear_q_weight_to_fp16_palettized, x = key_11_cast_fp16)[name = string("linear_48_cast_fp16")]; + tensor var_1553 = const()[name = string("op_1553"), val = tensor([1, -1, 8, 128])]; + tensor q_31_cast_fp16 = reshape(shape = var_1553, x = linear_48_cast_fp16)[name = string("q_31_cast_fp16")]; + tensor encoder_layers_5_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120915584))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121702080))))[name = string("encoder_layers_5_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_5_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_5_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121702272)))]; + tensor linear_49_cast_fp16 = linear(bias = encoder_layers_5_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_5_self_attn_linear_k_weight_to_fp16_palettized, x = input_299_cast_fp16)[name = string("linear_49_cast_fp16")]; + tensor var_1558 = const()[name = string("op_1558"), val = tensor([1, -1, 8, 128])]; + tensor k_21_cast_fp16 = reshape(shape = var_1558, x = linear_49_cast_fp16)[name = string("k_21_cast_fp16")]; + tensor encoder_layers_5_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121704384))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122490880))))[name = string("encoder_layers_5_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_5_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_5_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122491072)))]; + tensor linear_50_cast_fp16 = linear(bias = encoder_layers_5_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_5_self_attn_linear_v_weight_to_fp16_palettized, x = input_299_cast_fp16)[name = string("linear_50_cast_fp16")]; + tensor var_1563 = const()[name = string("op_1563"), val = tensor([1, -1, 8, 128])]; + tensor v_11_cast_fp16 = reshape(shape = var_1563, x = linear_50_cast_fp16)[name = string("v_11_cast_fp16")]; + tensor value_19_perm_0 = const()[name = string("value_19_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_5_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_5_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122493184)))]; + tensor var_1576_cast_fp16 = add(x = q_31_cast_fp16, y = encoder_layers_5_self_attn_pos_bias_u_to_fp16)[name = string("op_1576_cast_fp16")]; + tensor encoder_layers_5_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_5_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122495296)))]; + tensor var_1578_cast_fp16 = add(x = q_31_cast_fp16, y = encoder_layers_5_self_attn_pos_bias_v_to_fp16)[name = string("op_1578_cast_fp16")]; + tensor q_with_bias_v_11_perm_0 = const()[name = string("q_with_bias_v_11_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_137_transpose_x_0 = const()[name = string("x_137_transpose_x_0"), val = bool(false)]; + bool x_137_transpose_y_0 = const()[name = string("x_137_transpose_y_0"), val = bool(false)]; + tensor op_1580_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122497408))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122697152))))[name = string("op_1580_to_fp16_quantized")]; + tensor q_with_bias_v_11_cast_fp16 = transpose(perm = q_with_bias_v_11_perm_0, x = var_1578_cast_fp16)[name = string("transpose_317")]; + tensor x_137_cast_fp16 = matmul(transpose_x = x_137_transpose_x_0, transpose_y = x_137_transpose_y_0, x = q_with_bias_v_11_cast_fp16, y = op_1580_to_fp16_quantized)[name = string("x_137_cast_fp16")]; + tensor x_139_pad_0 = const()[name = string("x_139_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_139_mode_0 = const()[name = string("x_139_mode_0"), val = string("constant")]; + fp16 const_144_to_fp16 = const()[name = string("const_144_to_fp16"), val = fp16(0x0p+0)]; + tensor x_139_cast_fp16 = pad(constant_val = const_144_to_fp16, mode = x_139_mode_0, pad = x_139_pad_0, x = x_137_cast_fp16)[name = string("x_139_cast_fp16")]; + tensor var_1588 = const()[name = string("op_1588"), val = tensor([1, 8, -1, 56])]; + tensor x_141_cast_fp16 = reshape(shape = var_1588, x = x_139_cast_fp16)[name = string("x_141_cast_fp16")]; + tensor var_1592_begin_0 = const()[name = string("op_1592_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_1592_end_0 = const()[name = string("op_1592_end_0"), val = tensor([1, 8, 196, 56])]; + tensor var_1592_end_mask_0 = const()[name = string("op_1592_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_1592_cast_fp16 = slice_by_index(begin = var_1592_begin_0, end = var_1592_end_0, end_mask = var_1592_end_mask_0, x = x_141_cast_fp16)[name = string("op_1592_cast_fp16")]; + tensor var_1593 = const()[name = string("op_1593"), val = tensor([1, 8, 56, 195])]; + tensor matrix_bd_21_cast_fp16 = reshape(shape = var_1593, x = var_1592_cast_fp16)[name = string("matrix_bd_21_cast_fp16")]; + bool matrix_ac_11_transpose_x_0 = const()[name = string("matrix_ac_11_transpose_x_0"), val = bool(false)]; + bool matrix_ac_11_transpose_y_0 = const()[name = string("matrix_ac_11_transpose_y_0"), val = bool(false)]; + tensor transpose_106_perm_0 = const()[name = string("transpose_106_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_107_perm_0 = const()[name = string("transpose_107_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_107 = transpose(perm = transpose_107_perm_0, x = k_21_cast_fp16)[name = string("transpose_315")]; + tensor transpose_106 = transpose(perm = transpose_106_perm_0, x = var_1576_cast_fp16)[name = string("transpose_316")]; + tensor matrix_ac_11_cast_fp16 = matmul(transpose_x = matrix_ac_11_transpose_x_0, transpose_y = matrix_ac_11_transpose_y_0, x = transpose_106, y = transpose_107)[name = string("matrix_ac_11_cast_fp16")]; + tensor matrix_bd_23_begin_0 = const()[name = string("matrix_bd_23_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_23_end_0 = const()[name = string("matrix_bd_23_end_0"), val = tensor([1, 8, 56, 98])]; + tensor matrix_bd_23_end_mask_0 = const()[name = string("matrix_bd_23_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_23_cast_fp16 = slice_by_index(begin = matrix_bd_23_begin_0, end = matrix_bd_23_end_0, end_mask = matrix_bd_23_end_mask_0, x = matrix_bd_21_cast_fp16)[name = string("matrix_bd_23_cast_fp16")]; + tensor var_1602_cast_fp16 = add(x = matrix_ac_11_cast_fp16, y = matrix_bd_23_cast_fp16)[name = string("op_1602_cast_fp16")]; + fp16 _inversed_scores_21_y_0_to_fp16 = const()[name = string("_inversed_scores_21_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_21_cast_fp16 = mul(x = var_1602_cast_fp16, y = _inversed_scores_21_y_0_to_fp16)[name = string("_inversed_scores_21_cast_fp16")]; + tensor scores_23_cast_fp16 = select(a = var_44_to_fp16, b = _inversed_scores_21_cast_fp16, cond = mask_11)[name = string("scores_23_cast_fp16")]; + tensor var_1608_cast_fp16 = softmax(axis = var_58, x = scores_23_cast_fp16)[name = string("op_1608_cast_fp16")]; + tensor input_301_cast_fp16 = select(a = var_43_to_fp16, b = var_1608_cast_fp16, cond = mask_11)[name = string("input_301_cast_fp16")]; + bool x_143_transpose_x_0 = const()[name = string("x_143_transpose_x_0"), val = bool(false)]; + bool x_143_transpose_y_0 = const()[name = string("x_143_transpose_y_0"), val = bool(false)]; + tensor value_19_cast_fp16 = transpose(perm = value_19_perm_0, x = v_11_cast_fp16)[name = string("transpose_314")]; + tensor x_143_cast_fp16 = matmul(transpose_x = x_143_transpose_x_0, transpose_y = x_143_transpose_y_0, x = input_301_cast_fp16, y = value_19_cast_fp16)[name = string("x_143_cast_fp16")]; + tensor var_1612_perm_0 = const()[name = string("op_1612_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1613 = const()[name = string("op_1613"), val = tensor([1, -1, 1024])]; + tensor var_1612_cast_fp16 = transpose(perm = var_1612_perm_0, x = x_143_cast_fp16)[name = string("transpose_313")]; + tensor input_303_cast_fp16 = reshape(shape = var_1613, x = var_1612_cast_fp16)[name = string("input_303_cast_fp16")]; + tensor encoder_layers_5_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122697664))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123484160))))[name = string("encoder_layers_5_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_5_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_5_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123484352)))]; + tensor linear_52_cast_fp16 = linear(bias = encoder_layers_5_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_5_self_attn_linear_out_weight_to_fp16_palettized, x = input_303_cast_fp16)[name = string("linear_52_cast_fp16")]; + tensor input_307_cast_fp16 = add(x = input_297_cast_fp16, y = linear_52_cast_fp16)[name = string("input_307_cast_fp16")]; + tensor x_147_axes_0 = const()[name = string("x_147_axes_0"), val = tensor([-1])]; + tensor encoder_layers_5_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_5_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123486464)))]; + tensor encoder_layers_5_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_5_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123488576)))]; + tensor x_147_cast_fp16 = layer_norm(axes = x_147_axes_0, beta = encoder_layers_5_norm_conv_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_5_norm_conv_weight_to_fp16, x = input_307_cast_fp16)[name = string("x_147_cast_fp16")]; + tensor input_309_perm_0 = const()[name = string("input_309_perm_0"), val = tensor([0, 2, 1])]; + string input_311_pad_type_0 = const()[name = string("input_311_pad_type_0"), val = string("valid")]; + tensor input_311_strides_0 = const()[name = string("input_311_strides_0"), val = tensor([1])]; + tensor input_311_pad_0 = const()[name = string("input_311_pad_0"), val = tensor([0, 0])]; + tensor input_311_dilations_0 = const()[name = string("input_311_dilations_0"), val = tensor([1])]; + int32 input_311_groups_0 = const()[name = string("input_311_groups_0"), val = int32(1)]; + tensor encoder_layers_5_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123490688))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125587904))))[name = string("encoder_layers_5_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_309_cast_fp16 = transpose(perm = input_309_perm_0, x = x_147_cast_fp16)[name = string("transpose_312")]; + tensor input_311_cast_fp16 = conv(dilations = input_311_dilations_0, groups = input_311_groups_0, pad = input_311_pad_0, pad_type = input_311_pad_type_0, strides = input_311_strides_0, weight = encoder_layers_5_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_309_cast_fp16)[name = string("input_311_cast_fp16")]; + int32 x_149_split_num_splits_0 = const()[name = string("x_149_split_num_splits_0"), val = int32(2)]; + int32 x_149_split_axis_0 = const()[name = string("x_149_split_axis_0"), val = int32(1)]; + tensor x_149_split_cast_fp16_0, tensor x_149_split_cast_fp16_1 = split(axis = x_149_split_axis_0, num_splits = x_149_split_num_splits_0, x = input_311_cast_fp16)[name = string("x_149_split_cast_fp16")]; + tensor x_149_split_1_sigmoid_cast_fp16 = sigmoid(x = x_149_split_cast_fp16_1)[name = string("x_149_split_1_sigmoid_cast_fp16")]; + tensor x_149_cast_fp16 = mul(x = x_149_split_cast_fp16_0, y = x_149_split_1_sigmoid_cast_fp16)[name = string("x_149_cast_fp16")]; + tensor input_313_cast_fp16 = select(a = var_43_to_fp16, b = x_149_cast_fp16, cond = var_574)[name = string("input_313_cast_fp16")]; + bool new_x_23_interleave_0 = const()[name = string("new_x_23_interleave_0"), val = bool(false)]; + tensor new_x_23_cast_fp16 = concat(axis = var_58, interleave = new_x_23_interleave_0, values = (cache_23_cast_fp16, input_313_cast_fp16))[name = string("new_x_23_cast_fp16")]; + tensor var_1652_begin_0 = const()[name = string("op_1652_begin_0"), val = tensor([0, 0, 56])]; + tensor var_1652_end_0 = const()[name = string("op_1652_end_0"), val = tensor([1, 1024, 64])]; + tensor var_1652_end_mask_0 = const()[name = string("op_1652_end_mask_0"), val = tensor([true, true, true])]; + tensor var_1652_cast_fp16 = slice_by_index(begin = var_1652_begin_0, end = var_1652_end_0, end_mask = var_1652_end_mask_0, x = new_x_23_cast_fp16)[name = string("op_1652_cast_fp16")]; + string x_151_pad_type_0 = const()[name = string("x_151_pad_type_0"), val = string("valid")]; + int32 x_151_groups_0 = const()[name = string("x_151_groups_0"), val = int32(1024)]; + tensor x_151_strides_0 = const()[name = string("x_151_strides_0"), val = tensor([1])]; + tensor x_151_pad_0 = const()[name = string("x_151_pad_0"), val = tensor([0, 0])]; + tensor x_151_dilations_0 = const()[name = string("x_151_dilations_0"), val = tensor([1])]; + tensor encoder_layers_5_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125592064))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125601344))))[name = string("encoder_layers_5_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_151_cast_fp16 = conv(dilations = x_151_dilations_0, groups = x_151_groups_0, pad = x_151_pad_0, pad_type = x_151_pad_type_0, strides = x_151_strides_0, weight = encoder_layers_5_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_23_cast_fp16)[name = string("x_151_cast_fp16")]; + tensor input_315_perm_0 = const()[name = string("input_315_perm_0"), val = tensor([0, 2, 1])]; + tensor x_153_axes_0 = const()[name = string("x_153_axes_0"), val = tensor([-1])]; + tensor encoder_layers_5_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_5_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125603456)))]; + tensor encoder_layers_5_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_5_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125605568)))]; + tensor input_315_cast_fp16 = transpose(perm = input_315_perm_0, x = x_151_cast_fp16)[name = string("transpose_311")]; + tensor x_153_cast_fp16 = layer_norm(axes = x_153_axes_0, beta = encoder_layers_5_conv_batch_norm_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_5_conv_batch_norm_weight_to_fp16, x = input_315_cast_fp16)[name = string("x_153_cast_fp16")]; + tensor input_317_perm_0 = const()[name = string("input_317_perm_0"), val = tensor([0, 2, 1])]; + tensor input_317_cast_fp16 = transpose(perm = input_317_perm_0, x = x_153_cast_fp16)[name = string("transpose_310")]; + tensor input_319_cast_fp16 = silu(x = input_317_cast_fp16)[name = string("input_319_cast_fp16")]; + string x_155_pad_type_0 = const()[name = string("x_155_pad_type_0"), val = string("valid")]; + tensor x_155_strides_0 = const()[name = string("x_155_strides_0"), val = tensor([1])]; + tensor x_155_pad_0 = const()[name = string("x_155_pad_0"), val = tensor([0, 0])]; + tensor x_155_dilations_0 = const()[name = string("x_155_dilations_0"), val = tensor([1])]; + int32 x_155_groups_0 = const()[name = string("x_155_groups_0"), val = int32(1)]; + tensor encoder_layers_5_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125607680))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126656320))))[name = string("encoder_layers_5_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_155_cast_fp16 = conv(dilations = x_155_dilations_0, groups = x_155_groups_0, pad = x_155_pad_0, pad_type = x_155_pad_type_0, strides = x_155_strides_0, weight = encoder_layers_5_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_319_cast_fp16)[name = string("x_155_cast_fp16")]; + tensor input_321_perm_0 = const()[name = string("input_321_perm_0"), val = tensor([0, 2, 1])]; + tensor input_321_cast_fp16 = transpose(perm = input_321_perm_0, x = x_155_cast_fp16)[name = string("transpose_309")]; + tensor input_323_cast_fp16 = add(x = input_307_cast_fp16, y = input_321_cast_fp16)[name = string("input_323_cast_fp16")]; + tensor input_325_axes_0 = const()[name = string("input_325_axes_0"), val = tensor([-1])]; + tensor encoder_layers_5_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_5_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126658432)))]; + tensor encoder_layers_5_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_5_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126660544)))]; + tensor input_325_cast_fp16 = layer_norm(axes = input_325_axes_0, beta = encoder_layers_5_norm_feed_forward2_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_5_norm_feed_forward2_weight_to_fp16, x = input_323_cast_fp16)[name = string("input_325_cast_fp16")]; + tensor encoder_layers_5_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126662656))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(129808448))))[name = string("encoder_layers_5_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_5_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_5_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(129808640)))]; + tensor linear_53_cast_fp16 = linear(bias = encoder_layers_5_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_5_feed_forward2_linear1_weight_to_fp16_palettized, x = input_325_cast_fp16)[name = string("linear_53_cast_fp16")]; + tensor input_329_cast_fp16 = silu(x = linear_53_cast_fp16)[name = string("input_329_cast_fp16")]; + tensor encoder_layers_5_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(129816896))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132962688))))[name = string("encoder_layers_5_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_5_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_5_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132962880)))]; + tensor linear_54_cast_fp16 = linear(bias = encoder_layers_5_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_5_feed_forward2_linear2_weight_to_fp16_palettized, x = input_329_cast_fp16)[name = string("linear_54_cast_fp16")]; + fp16 var_1695_to_fp16 = const()[name = string("op_1695_to_fp16"), val = fp16(0x1p-1)]; + tensor var_1696_cast_fp16 = mul(x = linear_54_cast_fp16, y = var_1695_to_fp16)[name = string("op_1696_cast_fp16")]; + tensor input_335_cast_fp16 = add(x = input_323_cast_fp16, y = var_1696_cast_fp16)[name = string("input_335_cast_fp16")]; + tensor input_337_axes_0 = const()[name = string("input_337_axes_0"), val = tensor([-1])]; + tensor encoder_layers_5_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_5_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132964992)))]; + tensor encoder_layers_5_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_5_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132967104)))]; + tensor input_337_cast_fp16 = layer_norm(axes = input_337_axes_0, beta = encoder_layers_5_norm_out_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_5_norm_out_weight_to_fp16, x = input_335_cast_fp16)[name = string("input_337_cast_fp16")]; + tensor cache_25_begin_0 = const()[name = string("cache_25_begin_0"), val = tensor([6, 0, 0, 0])]; + tensor cache_25_end_0 = const()[name = string("cache_25_end_0"), val = tensor([7, 1, 42, 1024])]; + tensor cache_25_end_mask_0 = const()[name = string("cache_25_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_25_squeeze_mask_0 = const()[name = string("cache_25_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_25_cast_fp16 = slice_by_index(begin = cache_25_begin_0, end = cache_25_end_0, end_mask = cache_25_end_mask_0, squeeze_mask = cache_25_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_25_cast_fp16")]; + tensor cache_27_begin_0 = const()[name = string("cache_27_begin_0"), val = tensor([6, 0, 0, 0])]; + tensor cache_27_end_0 = const()[name = string("cache_27_end_0"), val = tensor([7, 1, 1024, 8])]; + tensor cache_27_end_mask_0 = const()[name = string("cache_27_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_27_squeeze_mask_0 = const()[name = string("cache_27_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_27_cast_fp16 = slice_by_index(begin = cache_27_begin_0, end = cache_27_end_0, end_mask = cache_27_end_mask_0, squeeze_mask = cache_27_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_27_cast_fp16")]; + tensor input_339_axes_0 = const()[name = string("input_339_axes_0"), val = tensor([-1])]; + tensor encoder_layers_6_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_6_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132969216)))]; + tensor encoder_layers_6_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_6_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132971328)))]; + tensor input_339_cast_fp16 = layer_norm(axes = input_339_axes_0, beta = encoder_layers_6_norm_feed_forward1_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_6_norm_feed_forward1_weight_to_fp16, x = input_337_cast_fp16)[name = string("input_339_cast_fp16")]; + tensor encoder_layers_6_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132973440))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136119232))))[name = string("encoder_layers_6_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_6_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_6_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136119424)))]; + tensor linear_55_cast_fp16 = linear(bias = encoder_layers_6_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_6_feed_forward1_linear1_weight_to_fp16_palettized, x = input_339_cast_fp16)[name = string("linear_55_cast_fp16")]; + tensor input_343_cast_fp16 = silu(x = linear_55_cast_fp16)[name = string("input_343_cast_fp16")]; + tensor encoder_layers_6_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136127680))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(139273472))))[name = string("encoder_layers_6_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_6_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_6_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(139273664)))]; + tensor linear_56_cast_fp16 = linear(bias = encoder_layers_6_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_6_feed_forward1_linear2_weight_to_fp16_palettized, x = input_343_cast_fp16)[name = string("linear_56_cast_fp16")]; + fp16 var_1732_to_fp16 = const()[name = string("op_1732_to_fp16"), val = fp16(0x1p-1)]; + tensor var_1733_cast_fp16 = mul(x = linear_56_cast_fp16, y = var_1732_to_fp16)[name = string("op_1733_cast_fp16")]; + tensor input_349_cast_fp16 = add(x = input_337_cast_fp16, y = var_1733_cast_fp16)[name = string("input_349_cast_fp16")]; + tensor key_13_axes_0 = const()[name = string("key_13_axes_0"), val = tensor([-1])]; + tensor encoder_layers_6_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_6_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(139275776)))]; + tensor encoder_layers_6_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_6_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(139277888)))]; + tensor key_13_cast_fp16 = layer_norm(axes = key_13_axes_0, beta = encoder_layers_6_norm_self_att_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_6_norm_self_att_weight_to_fp16, x = input_349_cast_fp16)[name = string("key_13_cast_fp16")]; + bool input_351_interleave_0 = const()[name = string("input_351_interleave_0"), val = bool(false)]; + tensor input_351_cast_fp16 = concat(axis = var_67, interleave = input_351_interleave_0, values = (cache_25_cast_fp16, key_13_cast_fp16))[name = string("input_351_cast_fp16")]; + bool var_1761_interleave_0 = const()[name = string("op_1761_interleave_0"), val = bool(false)]; + tensor var_1761_cast_fp16 = concat(axis = var_67, interleave = var_1761_interleave_0, values = key_13_cast_fp16)[name = string("op_1761_cast_fp16")]; + tensor encoder_layers_6_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(139280000))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140066496))))[name = string("encoder_layers_6_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_6_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_6_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140066688)))]; + tensor linear_57_cast_fp16 = linear(bias = encoder_layers_6_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_6_self_attn_linear_q_weight_to_fp16_palettized, x = key_13_cast_fp16)[name = string("linear_57_cast_fp16")]; + tensor var_1766 = const()[name = string("op_1766"), val = tensor([1, -1, 8, 128])]; + tensor q_37_cast_fp16 = reshape(shape = var_1766, x = linear_57_cast_fp16)[name = string("q_37_cast_fp16")]; + tensor encoder_layers_6_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140068800))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140855296))))[name = string("encoder_layers_6_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_6_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_6_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140855488)))]; + tensor linear_58_cast_fp16 = linear(bias = encoder_layers_6_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_6_self_attn_linear_k_weight_to_fp16_palettized, x = input_351_cast_fp16)[name = string("linear_58_cast_fp16")]; + tensor var_1771 = const()[name = string("op_1771"), val = tensor([1, -1, 8, 128])]; + tensor k_25_cast_fp16 = reshape(shape = var_1771, x = linear_58_cast_fp16)[name = string("k_25_cast_fp16")]; + tensor encoder_layers_6_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140857600))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141644096))))[name = string("encoder_layers_6_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_6_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_6_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141644288)))]; + tensor linear_59_cast_fp16 = linear(bias = encoder_layers_6_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_6_self_attn_linear_v_weight_to_fp16_palettized, x = input_351_cast_fp16)[name = string("linear_59_cast_fp16")]; + tensor var_1776 = const()[name = string("op_1776"), val = tensor([1, -1, 8, 128])]; + tensor v_13_cast_fp16 = reshape(shape = var_1776, x = linear_59_cast_fp16)[name = string("v_13_cast_fp16")]; + tensor value_21_perm_0 = const()[name = string("value_21_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_6_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_6_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141646400)))]; + tensor var_1789_cast_fp16 = add(x = q_37_cast_fp16, y = encoder_layers_6_self_attn_pos_bias_u_to_fp16)[name = string("op_1789_cast_fp16")]; + tensor encoder_layers_6_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_6_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141648512)))]; + tensor var_1791_cast_fp16 = add(x = q_37_cast_fp16, y = encoder_layers_6_self_attn_pos_bias_v_to_fp16)[name = string("op_1791_cast_fp16")]; + tensor q_with_bias_v_13_perm_0 = const()[name = string("q_with_bias_v_13_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_163_transpose_x_0 = const()[name = string("x_163_transpose_x_0"), val = bool(false)]; + bool x_163_transpose_y_0 = const()[name = string("x_163_transpose_y_0"), val = bool(false)]; + tensor op_1793_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141650624))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141850368))))[name = string("op_1793_to_fp16_quantized")]; + tensor q_with_bias_v_13_cast_fp16 = transpose(perm = q_with_bias_v_13_perm_0, x = var_1791_cast_fp16)[name = string("transpose_308")]; + tensor x_163_cast_fp16 = matmul(transpose_x = x_163_transpose_x_0, transpose_y = x_163_transpose_y_0, x = q_with_bias_v_13_cast_fp16, y = op_1793_to_fp16_quantized)[name = string("x_163_cast_fp16")]; + tensor x_165_pad_0 = const()[name = string("x_165_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_165_mode_0 = const()[name = string("x_165_mode_0"), val = string("constant")]; + fp16 const_157_to_fp16 = const()[name = string("const_157_to_fp16"), val = fp16(0x0p+0)]; + tensor x_165_cast_fp16 = pad(constant_val = const_157_to_fp16, mode = x_165_mode_0, pad = x_165_pad_0, x = x_163_cast_fp16)[name = string("x_165_cast_fp16")]; + tensor var_1801 = const()[name = string("op_1801"), val = tensor([1, 8, -1, 56])]; + tensor x_167_cast_fp16 = reshape(shape = var_1801, x = x_165_cast_fp16)[name = string("x_167_cast_fp16")]; + tensor var_1805_begin_0 = const()[name = string("op_1805_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_1805_end_0 = const()[name = string("op_1805_end_0"), val = tensor([1, 8, 196, 56])]; + tensor var_1805_end_mask_0 = const()[name = string("op_1805_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_1805_cast_fp16 = slice_by_index(begin = var_1805_begin_0, end = var_1805_end_0, end_mask = var_1805_end_mask_0, x = x_167_cast_fp16)[name = string("op_1805_cast_fp16")]; + tensor var_1806 = const()[name = string("op_1806"), val = tensor([1, 8, 56, 195])]; + tensor matrix_bd_25_cast_fp16 = reshape(shape = var_1806, x = var_1805_cast_fp16)[name = string("matrix_bd_25_cast_fp16")]; + bool matrix_ac_13_transpose_x_0 = const()[name = string("matrix_ac_13_transpose_x_0"), val = bool(false)]; + bool matrix_ac_13_transpose_y_0 = const()[name = string("matrix_ac_13_transpose_y_0"), val = bool(false)]; + tensor transpose_108_perm_0 = const()[name = string("transpose_108_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_109_perm_0 = const()[name = string("transpose_109_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_109 = transpose(perm = transpose_109_perm_0, x = k_25_cast_fp16)[name = string("transpose_306")]; + tensor transpose_108 = transpose(perm = transpose_108_perm_0, x = var_1789_cast_fp16)[name = string("transpose_307")]; + tensor matrix_ac_13_cast_fp16 = matmul(transpose_x = matrix_ac_13_transpose_x_0, transpose_y = matrix_ac_13_transpose_y_0, x = transpose_108, y = transpose_109)[name = string("matrix_ac_13_cast_fp16")]; + tensor matrix_bd_27_begin_0 = const()[name = string("matrix_bd_27_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_27_end_0 = const()[name = string("matrix_bd_27_end_0"), val = tensor([1, 8, 56, 98])]; + tensor matrix_bd_27_end_mask_0 = const()[name = string("matrix_bd_27_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_27_cast_fp16 = slice_by_index(begin = matrix_bd_27_begin_0, end = matrix_bd_27_end_0, end_mask = matrix_bd_27_end_mask_0, x = matrix_bd_25_cast_fp16)[name = string("matrix_bd_27_cast_fp16")]; + tensor var_1815_cast_fp16 = add(x = matrix_ac_13_cast_fp16, y = matrix_bd_27_cast_fp16)[name = string("op_1815_cast_fp16")]; + fp16 _inversed_scores_25_y_0_to_fp16 = const()[name = string("_inversed_scores_25_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_25_cast_fp16 = mul(x = var_1815_cast_fp16, y = _inversed_scores_25_y_0_to_fp16)[name = string("_inversed_scores_25_cast_fp16")]; + tensor scores_27_cast_fp16 = select(a = var_44_to_fp16, b = _inversed_scores_25_cast_fp16, cond = mask_11)[name = string("scores_27_cast_fp16")]; + tensor var_1821_cast_fp16 = softmax(axis = var_58, x = scores_27_cast_fp16)[name = string("op_1821_cast_fp16")]; + tensor input_353_cast_fp16 = select(a = var_43_to_fp16, b = var_1821_cast_fp16, cond = mask_11)[name = string("input_353_cast_fp16")]; + bool x_169_transpose_x_0 = const()[name = string("x_169_transpose_x_0"), val = bool(false)]; + bool x_169_transpose_y_0 = const()[name = string("x_169_transpose_y_0"), val = bool(false)]; + tensor value_21_cast_fp16 = transpose(perm = value_21_perm_0, x = v_13_cast_fp16)[name = string("transpose_305")]; + tensor x_169_cast_fp16 = matmul(transpose_x = x_169_transpose_x_0, transpose_y = x_169_transpose_y_0, x = input_353_cast_fp16, y = value_21_cast_fp16)[name = string("x_169_cast_fp16")]; + tensor var_1825_perm_0 = const()[name = string("op_1825_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1826 = const()[name = string("op_1826"), val = tensor([1, -1, 1024])]; + tensor var_1825_cast_fp16 = transpose(perm = var_1825_perm_0, x = x_169_cast_fp16)[name = string("transpose_304")]; + tensor input_355_cast_fp16 = reshape(shape = var_1826, x = var_1825_cast_fp16)[name = string("input_355_cast_fp16")]; + tensor encoder_layers_6_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141850880))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142637376))))[name = string("encoder_layers_6_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_6_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_6_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142637568)))]; + tensor linear_61_cast_fp16 = linear(bias = encoder_layers_6_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_6_self_attn_linear_out_weight_to_fp16_palettized, x = input_355_cast_fp16)[name = string("linear_61_cast_fp16")]; + tensor input_359_cast_fp16 = add(x = input_349_cast_fp16, y = linear_61_cast_fp16)[name = string("input_359_cast_fp16")]; + tensor x_173_axes_0 = const()[name = string("x_173_axes_0"), val = tensor([-1])]; + tensor encoder_layers_6_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_6_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142639680)))]; + tensor encoder_layers_6_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_6_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142641792)))]; + tensor x_173_cast_fp16 = layer_norm(axes = x_173_axes_0, beta = encoder_layers_6_norm_conv_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_6_norm_conv_weight_to_fp16, x = input_359_cast_fp16)[name = string("x_173_cast_fp16")]; + tensor input_361_perm_0 = const()[name = string("input_361_perm_0"), val = tensor([0, 2, 1])]; + string input_363_pad_type_0 = const()[name = string("input_363_pad_type_0"), val = string("valid")]; + tensor input_363_strides_0 = const()[name = string("input_363_strides_0"), val = tensor([1])]; + tensor input_363_pad_0 = const()[name = string("input_363_pad_0"), val = tensor([0, 0])]; + tensor input_363_dilations_0 = const()[name = string("input_363_dilations_0"), val = tensor([1])]; + int32 input_363_groups_0 = const()[name = string("input_363_groups_0"), val = int32(1)]; + tensor encoder_layers_6_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142643904))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144741120))))[name = string("encoder_layers_6_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_361_cast_fp16 = transpose(perm = input_361_perm_0, x = x_173_cast_fp16)[name = string("transpose_303")]; + tensor input_363_cast_fp16 = conv(dilations = input_363_dilations_0, groups = input_363_groups_0, pad = input_363_pad_0, pad_type = input_363_pad_type_0, strides = input_363_strides_0, weight = encoder_layers_6_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_361_cast_fp16)[name = string("input_363_cast_fp16")]; + int32 x_175_split_num_splits_0 = const()[name = string("x_175_split_num_splits_0"), val = int32(2)]; + int32 x_175_split_axis_0 = const()[name = string("x_175_split_axis_0"), val = int32(1)]; + tensor x_175_split_cast_fp16_0, tensor x_175_split_cast_fp16_1 = split(axis = x_175_split_axis_0, num_splits = x_175_split_num_splits_0, x = input_363_cast_fp16)[name = string("x_175_split_cast_fp16")]; + tensor x_175_split_1_sigmoid_cast_fp16 = sigmoid(x = x_175_split_cast_fp16_1)[name = string("x_175_split_1_sigmoid_cast_fp16")]; + tensor x_175_cast_fp16 = mul(x = x_175_split_cast_fp16_0, y = x_175_split_1_sigmoid_cast_fp16)[name = string("x_175_cast_fp16")]; + tensor input_365_cast_fp16 = select(a = var_43_to_fp16, b = x_175_cast_fp16, cond = var_574)[name = string("input_365_cast_fp16")]; + bool new_x_27_interleave_0 = const()[name = string("new_x_27_interleave_0"), val = bool(false)]; + tensor new_x_27_cast_fp16 = concat(axis = var_58, interleave = new_x_27_interleave_0, values = (cache_27_cast_fp16, input_365_cast_fp16))[name = string("new_x_27_cast_fp16")]; + tensor var_1865_begin_0 = const()[name = string("op_1865_begin_0"), val = tensor([0, 0, 56])]; + tensor var_1865_end_0 = const()[name = string("op_1865_end_0"), val = tensor([1, 1024, 64])]; + tensor var_1865_end_mask_0 = const()[name = string("op_1865_end_mask_0"), val = tensor([true, true, true])]; + tensor var_1865_cast_fp16 = slice_by_index(begin = var_1865_begin_0, end = var_1865_end_0, end_mask = var_1865_end_mask_0, x = new_x_27_cast_fp16)[name = string("op_1865_cast_fp16")]; + string x_177_pad_type_0 = const()[name = string("x_177_pad_type_0"), val = string("valid")]; + int32 x_177_groups_0 = const()[name = string("x_177_groups_0"), val = int32(1024)]; + tensor x_177_strides_0 = const()[name = string("x_177_strides_0"), val = tensor([1])]; + tensor x_177_pad_0 = const()[name = string("x_177_pad_0"), val = tensor([0, 0])]; + tensor x_177_dilations_0 = const()[name = string("x_177_dilations_0"), val = tensor([1])]; + tensor encoder_layers_6_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144745280))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144754560))))[name = string("encoder_layers_6_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_177_cast_fp16 = conv(dilations = x_177_dilations_0, groups = x_177_groups_0, pad = x_177_pad_0, pad_type = x_177_pad_type_0, strides = x_177_strides_0, weight = encoder_layers_6_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_27_cast_fp16)[name = string("x_177_cast_fp16")]; + tensor input_367_perm_0 = const()[name = string("input_367_perm_0"), val = tensor([0, 2, 1])]; + tensor x_179_axes_0 = const()[name = string("x_179_axes_0"), val = tensor([-1])]; + tensor encoder_layers_6_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_6_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144756672)))]; + tensor encoder_layers_6_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_6_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144758784)))]; + tensor input_367_cast_fp16 = transpose(perm = input_367_perm_0, x = x_177_cast_fp16)[name = string("transpose_302")]; + tensor x_179_cast_fp16 = layer_norm(axes = x_179_axes_0, beta = encoder_layers_6_conv_batch_norm_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_6_conv_batch_norm_weight_to_fp16, x = input_367_cast_fp16)[name = string("x_179_cast_fp16")]; + tensor input_369_perm_0 = const()[name = string("input_369_perm_0"), val = tensor([0, 2, 1])]; + tensor input_369_cast_fp16 = transpose(perm = input_369_perm_0, x = x_179_cast_fp16)[name = string("transpose_301")]; + tensor input_371_cast_fp16 = silu(x = input_369_cast_fp16)[name = string("input_371_cast_fp16")]; + string x_181_pad_type_0 = const()[name = string("x_181_pad_type_0"), val = string("valid")]; + tensor x_181_strides_0 = const()[name = string("x_181_strides_0"), val = tensor([1])]; + tensor x_181_pad_0 = const()[name = string("x_181_pad_0"), val = tensor([0, 0])]; + tensor x_181_dilations_0 = const()[name = string("x_181_dilations_0"), val = tensor([1])]; + int32 x_181_groups_0 = const()[name = string("x_181_groups_0"), val = int32(1)]; + tensor encoder_layers_6_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144760896))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(145809536))))[name = string("encoder_layers_6_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_181_cast_fp16 = conv(dilations = x_181_dilations_0, groups = x_181_groups_0, pad = x_181_pad_0, pad_type = x_181_pad_type_0, strides = x_181_strides_0, weight = encoder_layers_6_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_371_cast_fp16)[name = string("x_181_cast_fp16")]; + tensor input_373_perm_0 = const()[name = string("input_373_perm_0"), val = tensor([0, 2, 1])]; + tensor input_373_cast_fp16 = transpose(perm = input_373_perm_0, x = x_181_cast_fp16)[name = string("transpose_300")]; + tensor input_375_cast_fp16 = add(x = input_359_cast_fp16, y = input_373_cast_fp16)[name = string("input_375_cast_fp16")]; + tensor input_377_axes_0 = const()[name = string("input_377_axes_0"), val = tensor([-1])]; + tensor encoder_layers_6_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_6_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(145811648)))]; + tensor encoder_layers_6_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_6_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(145813760)))]; + tensor input_377_cast_fp16 = layer_norm(axes = input_377_axes_0, beta = encoder_layers_6_norm_feed_forward2_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_6_norm_feed_forward2_weight_to_fp16, x = input_375_cast_fp16)[name = string("input_377_cast_fp16")]; + tensor encoder_layers_6_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(145815872))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(148961664))))[name = string("encoder_layers_6_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_6_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_6_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(148961856)))]; + tensor linear_62_cast_fp16 = linear(bias = encoder_layers_6_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_6_feed_forward2_linear1_weight_to_fp16_palettized, x = input_377_cast_fp16)[name = string("linear_62_cast_fp16")]; + tensor input_381_cast_fp16 = silu(x = linear_62_cast_fp16)[name = string("input_381_cast_fp16")]; + tensor encoder_layers_6_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(148970112))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152115904))))[name = string("encoder_layers_6_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_6_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_6_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152116096)))]; + tensor linear_63_cast_fp16 = linear(bias = encoder_layers_6_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_6_feed_forward2_linear2_weight_to_fp16_palettized, x = input_381_cast_fp16)[name = string("linear_63_cast_fp16")]; + fp16 var_1908_to_fp16 = const()[name = string("op_1908_to_fp16"), val = fp16(0x1p-1)]; + tensor var_1909_cast_fp16 = mul(x = linear_63_cast_fp16, y = var_1908_to_fp16)[name = string("op_1909_cast_fp16")]; + tensor input_387_cast_fp16 = add(x = input_375_cast_fp16, y = var_1909_cast_fp16)[name = string("input_387_cast_fp16")]; + tensor input_389_axes_0 = const()[name = string("input_389_axes_0"), val = tensor([-1])]; + tensor encoder_layers_6_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_6_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152118208)))]; + tensor encoder_layers_6_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_6_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152120320)))]; + tensor input_389_cast_fp16 = layer_norm(axes = input_389_axes_0, beta = encoder_layers_6_norm_out_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_6_norm_out_weight_to_fp16, x = input_387_cast_fp16)[name = string("input_389_cast_fp16")]; + tensor cache_29_begin_0 = const()[name = string("cache_29_begin_0"), val = tensor([7, 0, 0, 0])]; + tensor cache_29_end_0 = const()[name = string("cache_29_end_0"), val = tensor([8, 1, 42, 1024])]; + tensor cache_29_end_mask_0 = const()[name = string("cache_29_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_29_squeeze_mask_0 = const()[name = string("cache_29_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_29_cast_fp16 = slice_by_index(begin = cache_29_begin_0, end = cache_29_end_0, end_mask = cache_29_end_mask_0, squeeze_mask = cache_29_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_29_cast_fp16")]; + tensor cache_31_begin_0 = const()[name = string("cache_31_begin_0"), val = tensor([7, 0, 0, 0])]; + tensor cache_31_end_0 = const()[name = string("cache_31_end_0"), val = tensor([8, 1, 1024, 8])]; + tensor cache_31_end_mask_0 = const()[name = string("cache_31_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_31_squeeze_mask_0 = const()[name = string("cache_31_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_31_cast_fp16 = slice_by_index(begin = cache_31_begin_0, end = cache_31_end_0, end_mask = cache_31_end_mask_0, squeeze_mask = cache_31_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_31_cast_fp16")]; + tensor input_391_axes_0 = const()[name = string("input_391_axes_0"), val = tensor([-1])]; + tensor encoder_layers_7_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_7_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152122432)))]; + tensor encoder_layers_7_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_7_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152124544)))]; + tensor input_391_cast_fp16 = layer_norm(axes = input_391_axes_0, beta = encoder_layers_7_norm_feed_forward1_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_7_norm_feed_forward1_weight_to_fp16, x = input_389_cast_fp16)[name = string("input_391_cast_fp16")]; + tensor encoder_layers_7_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152126656))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(155272448))))[name = string("encoder_layers_7_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_7_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_7_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(155272640)))]; + tensor linear_64_cast_fp16 = linear(bias = encoder_layers_7_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_7_feed_forward1_linear1_weight_to_fp16_palettized, x = input_391_cast_fp16)[name = string("linear_64_cast_fp16")]; + tensor input_395_cast_fp16 = silu(x = linear_64_cast_fp16)[name = string("input_395_cast_fp16")]; + tensor encoder_layers_7_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(155280896))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158426688))))[name = string("encoder_layers_7_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_7_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_7_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158426880)))]; + tensor linear_65_cast_fp16 = linear(bias = encoder_layers_7_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_7_feed_forward1_linear2_weight_to_fp16_palettized, x = input_395_cast_fp16)[name = string("linear_65_cast_fp16")]; + fp16 var_1945_to_fp16 = const()[name = string("op_1945_to_fp16"), val = fp16(0x1p-1)]; + tensor var_1946_cast_fp16 = mul(x = linear_65_cast_fp16, y = var_1945_to_fp16)[name = string("op_1946_cast_fp16")]; + tensor input_401_cast_fp16 = add(x = input_389_cast_fp16, y = var_1946_cast_fp16)[name = string("input_401_cast_fp16")]; + tensor key_15_axes_0 = const()[name = string("key_15_axes_0"), val = tensor([-1])]; + tensor encoder_layers_7_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_7_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158428992)))]; + tensor encoder_layers_7_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_7_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158431104)))]; + tensor key_15_cast_fp16 = layer_norm(axes = key_15_axes_0, beta = encoder_layers_7_norm_self_att_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_7_norm_self_att_weight_to_fp16, x = input_401_cast_fp16)[name = string("key_15_cast_fp16")]; + bool input_403_interleave_0 = const()[name = string("input_403_interleave_0"), val = bool(false)]; + tensor input_403_cast_fp16 = concat(axis = var_67, interleave = input_403_interleave_0, values = (cache_29_cast_fp16, key_15_cast_fp16))[name = string("input_403_cast_fp16")]; + bool var_1974_interleave_0 = const()[name = string("op_1974_interleave_0"), val = bool(false)]; + tensor var_1974_cast_fp16 = concat(axis = var_67, interleave = var_1974_interleave_0, values = key_15_cast_fp16)[name = string("op_1974_cast_fp16")]; + tensor encoder_layers_7_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158433216))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(159219712))))[name = string("encoder_layers_7_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_7_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_7_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(159219904)))]; + tensor linear_66_cast_fp16 = linear(bias = encoder_layers_7_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_7_self_attn_linear_q_weight_to_fp16_palettized, x = key_15_cast_fp16)[name = string("linear_66_cast_fp16")]; + tensor var_1979 = const()[name = string("op_1979"), val = tensor([1, -1, 8, 128])]; + tensor q_43_cast_fp16 = reshape(shape = var_1979, x = linear_66_cast_fp16)[name = string("q_43_cast_fp16")]; + tensor encoder_layers_7_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(159222016))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160008512))))[name = string("encoder_layers_7_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_7_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_7_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160008704)))]; + tensor linear_67_cast_fp16 = linear(bias = encoder_layers_7_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_7_self_attn_linear_k_weight_to_fp16_palettized, x = input_403_cast_fp16)[name = string("linear_67_cast_fp16")]; + tensor var_1984 = const()[name = string("op_1984"), val = tensor([1, -1, 8, 128])]; + tensor k_29_cast_fp16 = reshape(shape = var_1984, x = linear_67_cast_fp16)[name = string("k_29_cast_fp16")]; + tensor encoder_layers_7_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160010816))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160797312))))[name = string("encoder_layers_7_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_7_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_7_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160797504)))]; + tensor linear_68_cast_fp16 = linear(bias = encoder_layers_7_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_7_self_attn_linear_v_weight_to_fp16_palettized, x = input_403_cast_fp16)[name = string("linear_68_cast_fp16")]; + tensor var_1989 = const()[name = string("op_1989"), val = tensor([1, -1, 8, 128])]; + tensor v_15_cast_fp16 = reshape(shape = var_1989, x = linear_68_cast_fp16)[name = string("v_15_cast_fp16")]; + tensor value_23_perm_0 = const()[name = string("value_23_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_7_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_7_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160799616)))]; + tensor var_2002_cast_fp16 = add(x = q_43_cast_fp16, y = encoder_layers_7_self_attn_pos_bias_u_to_fp16)[name = string("op_2002_cast_fp16")]; + tensor encoder_layers_7_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_7_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160801728)))]; + tensor var_2004_cast_fp16 = add(x = q_43_cast_fp16, y = encoder_layers_7_self_attn_pos_bias_v_to_fp16)[name = string("op_2004_cast_fp16")]; + tensor q_with_bias_v_15_perm_0 = const()[name = string("q_with_bias_v_15_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_189_transpose_x_0 = const()[name = string("x_189_transpose_x_0"), val = bool(false)]; + bool x_189_transpose_y_0 = const()[name = string("x_189_transpose_y_0"), val = bool(false)]; + tensor op_2006_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160803840))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161003584))))[name = string("op_2006_to_fp16_quantized")]; + tensor q_with_bias_v_15_cast_fp16 = transpose(perm = q_with_bias_v_15_perm_0, x = var_2004_cast_fp16)[name = string("transpose_299")]; + tensor x_189_cast_fp16 = matmul(transpose_x = x_189_transpose_x_0, transpose_y = x_189_transpose_y_0, x = q_with_bias_v_15_cast_fp16, y = op_2006_to_fp16_quantized)[name = string("x_189_cast_fp16")]; + tensor x_191_pad_0 = const()[name = string("x_191_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_191_mode_0 = const()[name = string("x_191_mode_0"), val = string("constant")]; + fp16 const_170_to_fp16 = const()[name = string("const_170_to_fp16"), val = fp16(0x0p+0)]; + tensor x_191_cast_fp16 = pad(constant_val = const_170_to_fp16, mode = x_191_mode_0, pad = x_191_pad_0, x = x_189_cast_fp16)[name = string("x_191_cast_fp16")]; + tensor var_2014 = const()[name = string("op_2014"), val = tensor([1, 8, -1, 56])]; + tensor x_193_cast_fp16 = reshape(shape = var_2014, x = x_191_cast_fp16)[name = string("x_193_cast_fp16")]; + tensor var_2018_begin_0 = const()[name = string("op_2018_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_2018_end_0 = const()[name = string("op_2018_end_0"), val = tensor([1, 8, 196, 56])]; + tensor var_2018_end_mask_0 = const()[name = string("op_2018_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_2018_cast_fp16 = slice_by_index(begin = var_2018_begin_0, end = var_2018_end_0, end_mask = var_2018_end_mask_0, x = x_193_cast_fp16)[name = string("op_2018_cast_fp16")]; + tensor var_2019 = const()[name = string("op_2019"), val = tensor([1, 8, 56, 195])]; + tensor matrix_bd_29_cast_fp16 = reshape(shape = var_2019, x = var_2018_cast_fp16)[name = string("matrix_bd_29_cast_fp16")]; + bool matrix_ac_15_transpose_x_0 = const()[name = string("matrix_ac_15_transpose_x_0"), val = bool(false)]; + bool matrix_ac_15_transpose_y_0 = const()[name = string("matrix_ac_15_transpose_y_0"), val = bool(false)]; + tensor transpose_110_perm_0 = const()[name = string("transpose_110_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_111_perm_0 = const()[name = string("transpose_111_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_111 = transpose(perm = transpose_111_perm_0, x = k_29_cast_fp16)[name = string("transpose_297")]; + tensor transpose_110 = transpose(perm = transpose_110_perm_0, x = var_2002_cast_fp16)[name = string("transpose_298")]; + tensor matrix_ac_15_cast_fp16 = matmul(transpose_x = matrix_ac_15_transpose_x_0, transpose_y = matrix_ac_15_transpose_y_0, x = transpose_110, y = transpose_111)[name = string("matrix_ac_15_cast_fp16")]; + tensor matrix_bd_31_begin_0 = const()[name = string("matrix_bd_31_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_31_end_0 = const()[name = string("matrix_bd_31_end_0"), val = tensor([1, 8, 56, 98])]; + tensor matrix_bd_31_end_mask_0 = const()[name = string("matrix_bd_31_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_31_cast_fp16 = slice_by_index(begin = matrix_bd_31_begin_0, end = matrix_bd_31_end_0, end_mask = matrix_bd_31_end_mask_0, x = matrix_bd_29_cast_fp16)[name = string("matrix_bd_31_cast_fp16")]; + tensor var_2028_cast_fp16 = add(x = matrix_ac_15_cast_fp16, y = matrix_bd_31_cast_fp16)[name = string("op_2028_cast_fp16")]; + fp16 _inversed_scores_29_y_0_to_fp16 = const()[name = string("_inversed_scores_29_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_29_cast_fp16 = mul(x = var_2028_cast_fp16, y = _inversed_scores_29_y_0_to_fp16)[name = string("_inversed_scores_29_cast_fp16")]; + tensor scores_31_cast_fp16 = select(a = var_44_to_fp16, b = _inversed_scores_29_cast_fp16, cond = mask_11)[name = string("scores_31_cast_fp16")]; + tensor var_2034_cast_fp16 = softmax(axis = var_58, x = scores_31_cast_fp16)[name = string("op_2034_cast_fp16")]; + tensor input_405_cast_fp16 = select(a = var_43_to_fp16, b = var_2034_cast_fp16, cond = mask_11)[name = string("input_405_cast_fp16")]; + bool x_195_transpose_x_0 = const()[name = string("x_195_transpose_x_0"), val = bool(false)]; + bool x_195_transpose_y_0 = const()[name = string("x_195_transpose_y_0"), val = bool(false)]; + tensor value_23_cast_fp16 = transpose(perm = value_23_perm_0, x = v_15_cast_fp16)[name = string("transpose_296")]; + tensor x_195_cast_fp16 = matmul(transpose_x = x_195_transpose_x_0, transpose_y = x_195_transpose_y_0, x = input_405_cast_fp16, y = value_23_cast_fp16)[name = string("x_195_cast_fp16")]; + tensor var_2038_perm_0 = const()[name = string("op_2038_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_2039 = const()[name = string("op_2039"), val = tensor([1, -1, 1024])]; + tensor var_2038_cast_fp16 = transpose(perm = var_2038_perm_0, x = x_195_cast_fp16)[name = string("transpose_295")]; + tensor input_407_cast_fp16 = reshape(shape = var_2039, x = var_2038_cast_fp16)[name = string("input_407_cast_fp16")]; + tensor encoder_layers_7_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161004096))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161790592))))[name = string("encoder_layers_7_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_7_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_7_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161790784)))]; + tensor linear_70_cast_fp16 = linear(bias = encoder_layers_7_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_7_self_attn_linear_out_weight_to_fp16_palettized, x = input_407_cast_fp16)[name = string("linear_70_cast_fp16")]; + tensor input_411_cast_fp16 = add(x = input_401_cast_fp16, y = linear_70_cast_fp16)[name = string("input_411_cast_fp16")]; + tensor x_199_axes_0 = const()[name = string("x_199_axes_0"), val = tensor([-1])]; + tensor encoder_layers_7_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_7_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161792896)))]; + tensor encoder_layers_7_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_7_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161795008)))]; + tensor x_199_cast_fp16 = layer_norm(axes = x_199_axes_0, beta = encoder_layers_7_norm_conv_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_7_norm_conv_weight_to_fp16, x = input_411_cast_fp16)[name = string("x_199_cast_fp16")]; + tensor input_413_perm_0 = const()[name = string("input_413_perm_0"), val = tensor([0, 2, 1])]; + string input_415_pad_type_0 = const()[name = string("input_415_pad_type_0"), val = string("valid")]; + tensor input_415_strides_0 = const()[name = string("input_415_strides_0"), val = tensor([1])]; + tensor input_415_pad_0 = const()[name = string("input_415_pad_0"), val = tensor([0, 0])]; + tensor input_415_dilations_0 = const()[name = string("input_415_dilations_0"), val = tensor([1])]; + int32 input_415_groups_0 = const()[name = string("input_415_groups_0"), val = int32(1)]; + tensor encoder_layers_7_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161797120))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163894336))))[name = string("encoder_layers_7_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_413_cast_fp16 = transpose(perm = input_413_perm_0, x = x_199_cast_fp16)[name = string("transpose_294")]; + tensor input_415_cast_fp16 = conv(dilations = input_415_dilations_0, groups = input_415_groups_0, pad = input_415_pad_0, pad_type = input_415_pad_type_0, strides = input_415_strides_0, weight = encoder_layers_7_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_413_cast_fp16)[name = string("input_415_cast_fp16")]; + int32 x_201_split_num_splits_0 = const()[name = string("x_201_split_num_splits_0"), val = int32(2)]; + int32 x_201_split_axis_0 = const()[name = string("x_201_split_axis_0"), val = int32(1)]; + tensor x_201_split_cast_fp16_0, tensor x_201_split_cast_fp16_1 = split(axis = x_201_split_axis_0, num_splits = x_201_split_num_splits_0, x = input_415_cast_fp16)[name = string("x_201_split_cast_fp16")]; + tensor x_201_split_1_sigmoid_cast_fp16 = sigmoid(x = x_201_split_cast_fp16_1)[name = string("x_201_split_1_sigmoid_cast_fp16")]; + tensor x_201_cast_fp16 = mul(x = x_201_split_cast_fp16_0, y = x_201_split_1_sigmoid_cast_fp16)[name = string("x_201_cast_fp16")]; + tensor input_417_cast_fp16 = select(a = var_43_to_fp16, b = x_201_cast_fp16, cond = var_574)[name = string("input_417_cast_fp16")]; + bool new_x_31_interleave_0 = const()[name = string("new_x_31_interleave_0"), val = bool(false)]; + tensor new_x_31_cast_fp16 = concat(axis = var_58, interleave = new_x_31_interleave_0, values = (cache_31_cast_fp16, input_417_cast_fp16))[name = string("new_x_31_cast_fp16")]; + tensor var_2078_begin_0 = const()[name = string("op_2078_begin_0"), val = tensor([0, 0, 56])]; + tensor var_2078_end_0 = const()[name = string("op_2078_end_0"), val = tensor([1, 1024, 64])]; + tensor var_2078_end_mask_0 = const()[name = string("op_2078_end_mask_0"), val = tensor([true, true, true])]; + tensor var_2078_cast_fp16 = slice_by_index(begin = var_2078_begin_0, end = var_2078_end_0, end_mask = var_2078_end_mask_0, x = new_x_31_cast_fp16)[name = string("op_2078_cast_fp16")]; + string x_203_pad_type_0 = const()[name = string("x_203_pad_type_0"), val = string("valid")]; + int32 x_203_groups_0 = const()[name = string("x_203_groups_0"), val = int32(1024)]; + tensor x_203_strides_0 = const()[name = string("x_203_strides_0"), val = tensor([1])]; + tensor x_203_pad_0 = const()[name = string("x_203_pad_0"), val = tensor([0, 0])]; + tensor x_203_dilations_0 = const()[name = string("x_203_dilations_0"), val = tensor([1])]; + tensor encoder_layers_7_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163898496))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163907776))))[name = string("encoder_layers_7_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_203_cast_fp16 = conv(dilations = x_203_dilations_0, groups = x_203_groups_0, pad = x_203_pad_0, pad_type = x_203_pad_type_0, strides = x_203_strides_0, weight = encoder_layers_7_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_31_cast_fp16)[name = string("x_203_cast_fp16")]; + tensor input_419_perm_0 = const()[name = string("input_419_perm_0"), val = tensor([0, 2, 1])]; + tensor x_205_axes_0 = const()[name = string("x_205_axes_0"), val = tensor([-1])]; + tensor encoder_layers_7_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_7_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163909888)))]; + tensor encoder_layers_7_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_7_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163912000)))]; + tensor input_419_cast_fp16 = transpose(perm = input_419_perm_0, x = x_203_cast_fp16)[name = string("transpose_293")]; + tensor x_205_cast_fp16 = layer_norm(axes = x_205_axes_0, beta = encoder_layers_7_conv_batch_norm_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_7_conv_batch_norm_weight_to_fp16, x = input_419_cast_fp16)[name = string("x_205_cast_fp16")]; + tensor input_421_perm_0 = const()[name = string("input_421_perm_0"), val = tensor([0, 2, 1])]; + tensor input_421_cast_fp16 = transpose(perm = input_421_perm_0, x = x_205_cast_fp16)[name = string("transpose_292")]; + tensor input_423_cast_fp16 = silu(x = input_421_cast_fp16)[name = string("input_423_cast_fp16")]; + string x_207_pad_type_0 = const()[name = string("x_207_pad_type_0"), val = string("valid")]; + tensor x_207_strides_0 = const()[name = string("x_207_strides_0"), val = tensor([1])]; + tensor x_207_pad_0 = const()[name = string("x_207_pad_0"), val = tensor([0, 0])]; + tensor x_207_dilations_0 = const()[name = string("x_207_dilations_0"), val = tensor([1])]; + int32 x_207_groups_0 = const()[name = string("x_207_groups_0"), val = int32(1)]; + tensor encoder_layers_7_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163914112))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(164962752))))[name = string("encoder_layers_7_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_207_cast_fp16 = conv(dilations = x_207_dilations_0, groups = x_207_groups_0, pad = x_207_pad_0, pad_type = x_207_pad_type_0, strides = x_207_strides_0, weight = encoder_layers_7_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_423_cast_fp16)[name = string("x_207_cast_fp16")]; + tensor input_425_perm_0 = const()[name = string("input_425_perm_0"), val = tensor([0, 2, 1])]; + tensor input_425_cast_fp16 = transpose(perm = input_425_perm_0, x = x_207_cast_fp16)[name = string("transpose_291")]; + tensor input_427_cast_fp16 = add(x = input_411_cast_fp16, y = input_425_cast_fp16)[name = string("input_427_cast_fp16")]; + tensor input_429_axes_0 = const()[name = string("input_429_axes_0"), val = tensor([-1])]; + tensor encoder_layers_7_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_7_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(164964864)))]; + tensor encoder_layers_7_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_7_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(164966976)))]; + tensor input_429_cast_fp16 = layer_norm(axes = input_429_axes_0, beta = encoder_layers_7_norm_feed_forward2_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_7_norm_feed_forward2_weight_to_fp16, x = input_427_cast_fp16)[name = string("input_429_cast_fp16")]; + tensor encoder_layers_7_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(164969088))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(168114880))))[name = string("encoder_layers_7_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_7_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_7_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(168115072)))]; + tensor linear_71_cast_fp16 = linear(bias = encoder_layers_7_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_7_feed_forward2_linear1_weight_to_fp16_palettized, x = input_429_cast_fp16)[name = string("linear_71_cast_fp16")]; + tensor input_433_cast_fp16 = silu(x = linear_71_cast_fp16)[name = string("input_433_cast_fp16")]; + tensor encoder_layers_7_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(168123328))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(171269120))))[name = string("encoder_layers_7_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_7_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_7_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(171269312)))]; + tensor linear_72_cast_fp16 = linear(bias = encoder_layers_7_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_7_feed_forward2_linear2_weight_to_fp16_palettized, x = input_433_cast_fp16)[name = string("linear_72_cast_fp16")]; + fp16 var_2121_to_fp16 = const()[name = string("op_2121_to_fp16"), val = fp16(0x1p-1)]; + tensor var_2122_cast_fp16 = mul(x = linear_72_cast_fp16, y = var_2121_to_fp16)[name = string("op_2122_cast_fp16")]; + tensor input_439_cast_fp16 = add(x = input_427_cast_fp16, y = var_2122_cast_fp16)[name = string("input_439_cast_fp16")]; + tensor input_441_axes_0 = const()[name = string("input_441_axes_0"), val = tensor([-1])]; + tensor encoder_layers_7_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_7_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(171271424)))]; + tensor encoder_layers_7_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_7_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(171273536)))]; + tensor input_441_cast_fp16 = layer_norm(axes = input_441_axes_0, beta = encoder_layers_7_norm_out_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_7_norm_out_weight_to_fp16, x = input_439_cast_fp16)[name = string("input_441_cast_fp16")]; + tensor cache_33_begin_0 = const()[name = string("cache_33_begin_0"), val = tensor([8, 0, 0, 0])]; + tensor cache_33_end_0 = const()[name = string("cache_33_end_0"), val = tensor([9, 1, 42, 1024])]; + tensor cache_33_end_mask_0 = const()[name = string("cache_33_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_33_squeeze_mask_0 = const()[name = string("cache_33_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_33_cast_fp16 = slice_by_index(begin = cache_33_begin_0, end = cache_33_end_0, end_mask = cache_33_end_mask_0, squeeze_mask = cache_33_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_33_cast_fp16")]; + tensor cache_35_begin_0 = const()[name = string("cache_35_begin_0"), val = tensor([8, 0, 0, 0])]; + tensor cache_35_end_0 = const()[name = string("cache_35_end_0"), val = tensor([9, 1, 1024, 8])]; + tensor cache_35_end_mask_0 = const()[name = string("cache_35_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_35_squeeze_mask_0 = const()[name = string("cache_35_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_35_cast_fp16 = slice_by_index(begin = cache_35_begin_0, end = cache_35_end_0, end_mask = cache_35_end_mask_0, squeeze_mask = cache_35_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_35_cast_fp16")]; + tensor input_443_axes_0 = const()[name = string("input_443_axes_0"), val = tensor([-1])]; + tensor encoder_layers_8_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_8_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(171275648)))]; + tensor encoder_layers_8_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_8_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(171277760)))]; + tensor input_443_cast_fp16 = layer_norm(axes = input_443_axes_0, beta = encoder_layers_8_norm_feed_forward1_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_8_norm_feed_forward1_weight_to_fp16, x = input_441_cast_fp16)[name = string("input_443_cast_fp16")]; + tensor encoder_layers_8_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(171279872))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(174425664))))[name = string("encoder_layers_8_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_8_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_8_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(174425856)))]; + tensor linear_73_cast_fp16 = linear(bias = encoder_layers_8_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_8_feed_forward1_linear1_weight_to_fp16_palettized, x = input_443_cast_fp16)[name = string("linear_73_cast_fp16")]; + tensor input_447_cast_fp16 = silu(x = linear_73_cast_fp16)[name = string("input_447_cast_fp16")]; + tensor encoder_layers_8_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(174434112))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(177579904))))[name = string("encoder_layers_8_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_8_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_8_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(177580096)))]; + tensor linear_74_cast_fp16 = linear(bias = encoder_layers_8_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_8_feed_forward1_linear2_weight_to_fp16_palettized, x = input_447_cast_fp16)[name = string("linear_74_cast_fp16")]; + fp16 var_2158_to_fp16 = const()[name = string("op_2158_to_fp16"), val = fp16(0x1p-1)]; + tensor var_2159_cast_fp16 = mul(x = linear_74_cast_fp16, y = var_2158_to_fp16)[name = string("op_2159_cast_fp16")]; + tensor input_453_cast_fp16 = add(x = input_441_cast_fp16, y = var_2159_cast_fp16)[name = string("input_453_cast_fp16")]; + tensor key_17_axes_0 = const()[name = string("key_17_axes_0"), val = tensor([-1])]; + tensor encoder_layers_8_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_8_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(177582208)))]; + tensor encoder_layers_8_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_8_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(177584320)))]; + tensor key_17_cast_fp16 = layer_norm(axes = key_17_axes_0, beta = encoder_layers_8_norm_self_att_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_8_norm_self_att_weight_to_fp16, x = input_453_cast_fp16)[name = string("key_17_cast_fp16")]; + bool input_455_interleave_0 = const()[name = string("input_455_interleave_0"), val = bool(false)]; + tensor input_455_cast_fp16 = concat(axis = var_67, interleave = input_455_interleave_0, values = (cache_33_cast_fp16, key_17_cast_fp16))[name = string("input_455_cast_fp16")]; + bool var_2187_interleave_0 = const()[name = string("op_2187_interleave_0"), val = bool(false)]; + tensor var_2187_cast_fp16 = concat(axis = var_67, interleave = var_2187_interleave_0, values = key_17_cast_fp16)[name = string("op_2187_cast_fp16")]; + tensor encoder_layers_8_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(177586432))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(178372928))))[name = string("encoder_layers_8_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_8_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_8_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(178373120)))]; + tensor linear_75_cast_fp16 = linear(bias = encoder_layers_8_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_8_self_attn_linear_q_weight_to_fp16_palettized, x = key_17_cast_fp16)[name = string("linear_75_cast_fp16")]; + tensor var_2192 = const()[name = string("op_2192"), val = tensor([1, -1, 8, 128])]; + tensor q_49_cast_fp16 = reshape(shape = var_2192, x = linear_75_cast_fp16)[name = string("q_49_cast_fp16")]; + tensor encoder_layers_8_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(178375232))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179161728))))[name = string("encoder_layers_8_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_8_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_8_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179161920)))]; + tensor linear_76_cast_fp16 = linear(bias = encoder_layers_8_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_8_self_attn_linear_k_weight_to_fp16_palettized, x = input_455_cast_fp16)[name = string("linear_76_cast_fp16")]; + tensor var_2197 = const()[name = string("op_2197"), val = tensor([1, -1, 8, 128])]; + tensor k_33_cast_fp16 = reshape(shape = var_2197, x = linear_76_cast_fp16)[name = string("k_33_cast_fp16")]; + tensor encoder_layers_8_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179164032))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179950528))))[name = string("encoder_layers_8_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_8_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_8_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179950720)))]; + tensor linear_77_cast_fp16 = linear(bias = encoder_layers_8_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_8_self_attn_linear_v_weight_to_fp16_palettized, x = input_455_cast_fp16)[name = string("linear_77_cast_fp16")]; + tensor var_2202 = const()[name = string("op_2202"), val = tensor([1, -1, 8, 128])]; + tensor v_17_cast_fp16 = reshape(shape = var_2202, x = linear_77_cast_fp16)[name = string("v_17_cast_fp16")]; + tensor value_25_perm_0 = const()[name = string("value_25_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_8_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_8_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179952832)))]; + tensor var_2215_cast_fp16 = add(x = q_49_cast_fp16, y = encoder_layers_8_self_attn_pos_bias_u_to_fp16)[name = string("op_2215_cast_fp16")]; + tensor encoder_layers_8_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_8_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179954944)))]; + tensor var_2217_cast_fp16 = add(x = q_49_cast_fp16, y = encoder_layers_8_self_attn_pos_bias_v_to_fp16)[name = string("op_2217_cast_fp16")]; + tensor q_with_bias_v_17_perm_0 = const()[name = string("q_with_bias_v_17_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_215_transpose_x_0 = const()[name = string("x_215_transpose_x_0"), val = bool(false)]; + bool x_215_transpose_y_0 = const()[name = string("x_215_transpose_y_0"), val = bool(false)]; + tensor op_2219_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179957056))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(180156800))))[name = string("op_2219_to_fp16_quantized")]; + tensor q_with_bias_v_17_cast_fp16 = transpose(perm = q_with_bias_v_17_perm_0, x = var_2217_cast_fp16)[name = string("transpose_290")]; + tensor x_215_cast_fp16 = matmul(transpose_x = x_215_transpose_x_0, transpose_y = x_215_transpose_y_0, x = q_with_bias_v_17_cast_fp16, y = op_2219_to_fp16_quantized)[name = string("x_215_cast_fp16")]; + tensor x_217_pad_0 = const()[name = string("x_217_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_217_mode_0 = const()[name = string("x_217_mode_0"), val = string("constant")]; + fp16 const_183_to_fp16 = const()[name = string("const_183_to_fp16"), val = fp16(0x0p+0)]; + tensor x_217_cast_fp16 = pad(constant_val = const_183_to_fp16, mode = x_217_mode_0, pad = x_217_pad_0, x = x_215_cast_fp16)[name = string("x_217_cast_fp16")]; + tensor var_2227 = const()[name = string("op_2227"), val = tensor([1, 8, -1, 56])]; + tensor x_219_cast_fp16 = reshape(shape = var_2227, x = x_217_cast_fp16)[name = string("x_219_cast_fp16")]; + tensor var_2231_begin_0 = const()[name = string("op_2231_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_2231_end_0 = const()[name = string("op_2231_end_0"), val = tensor([1, 8, 196, 56])]; + tensor var_2231_end_mask_0 = const()[name = string("op_2231_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_2231_cast_fp16 = slice_by_index(begin = var_2231_begin_0, end = var_2231_end_0, end_mask = var_2231_end_mask_0, x = x_219_cast_fp16)[name = string("op_2231_cast_fp16")]; + tensor var_2232 = const()[name = string("op_2232"), val = tensor([1, 8, 56, 195])]; + tensor matrix_bd_33_cast_fp16 = reshape(shape = var_2232, x = var_2231_cast_fp16)[name = string("matrix_bd_33_cast_fp16")]; + bool matrix_ac_17_transpose_x_0 = const()[name = string("matrix_ac_17_transpose_x_0"), val = bool(false)]; + bool matrix_ac_17_transpose_y_0 = const()[name = string("matrix_ac_17_transpose_y_0"), val = bool(false)]; + tensor transpose_112_perm_0 = const()[name = string("transpose_112_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_113_perm_0 = const()[name = string("transpose_113_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_113 = transpose(perm = transpose_113_perm_0, x = k_33_cast_fp16)[name = string("transpose_288")]; + tensor transpose_112 = transpose(perm = transpose_112_perm_0, x = var_2215_cast_fp16)[name = string("transpose_289")]; + tensor matrix_ac_17_cast_fp16 = matmul(transpose_x = matrix_ac_17_transpose_x_0, transpose_y = matrix_ac_17_transpose_y_0, x = transpose_112, y = transpose_113)[name = string("matrix_ac_17_cast_fp16")]; + tensor matrix_bd_35_begin_0 = const()[name = string("matrix_bd_35_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_35_end_0 = const()[name = string("matrix_bd_35_end_0"), val = tensor([1, 8, 56, 98])]; + tensor matrix_bd_35_end_mask_0 = const()[name = string("matrix_bd_35_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_35_cast_fp16 = slice_by_index(begin = matrix_bd_35_begin_0, end = matrix_bd_35_end_0, end_mask = matrix_bd_35_end_mask_0, x = matrix_bd_33_cast_fp16)[name = string("matrix_bd_35_cast_fp16")]; + tensor var_2241_cast_fp16 = add(x = matrix_ac_17_cast_fp16, y = matrix_bd_35_cast_fp16)[name = string("op_2241_cast_fp16")]; + fp16 _inversed_scores_33_y_0_to_fp16 = const()[name = string("_inversed_scores_33_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_33_cast_fp16 = mul(x = var_2241_cast_fp16, y = _inversed_scores_33_y_0_to_fp16)[name = string("_inversed_scores_33_cast_fp16")]; + tensor scores_35_cast_fp16 = select(a = var_44_to_fp16, b = _inversed_scores_33_cast_fp16, cond = mask_11)[name = string("scores_35_cast_fp16")]; + tensor var_2247_cast_fp16 = softmax(axis = var_58, x = scores_35_cast_fp16)[name = string("op_2247_cast_fp16")]; + tensor input_457_cast_fp16 = select(a = var_43_to_fp16, b = var_2247_cast_fp16, cond = mask_11)[name = string("input_457_cast_fp16")]; + bool x_221_transpose_x_0 = const()[name = string("x_221_transpose_x_0"), val = bool(false)]; + bool x_221_transpose_y_0 = const()[name = string("x_221_transpose_y_0"), val = bool(false)]; + tensor value_25_cast_fp16 = transpose(perm = value_25_perm_0, x = v_17_cast_fp16)[name = string("transpose_287")]; + tensor x_221_cast_fp16 = matmul(transpose_x = x_221_transpose_x_0, transpose_y = x_221_transpose_y_0, x = input_457_cast_fp16, y = value_25_cast_fp16)[name = string("x_221_cast_fp16")]; + tensor var_2251_perm_0 = const()[name = string("op_2251_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_2252 = const()[name = string("op_2252"), val = tensor([1, -1, 1024])]; + tensor var_2251_cast_fp16 = transpose(perm = var_2251_perm_0, x = x_221_cast_fp16)[name = string("transpose_286")]; + tensor input_459_cast_fp16 = reshape(shape = var_2252, x = var_2251_cast_fp16)[name = string("input_459_cast_fp16")]; + tensor encoder_layers_8_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(180157312))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(180943808))))[name = string("encoder_layers_8_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_8_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_8_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(180944000)))]; + tensor linear_79_cast_fp16 = linear(bias = encoder_layers_8_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_8_self_attn_linear_out_weight_to_fp16_palettized, x = input_459_cast_fp16)[name = string("linear_79_cast_fp16")]; + tensor input_463_cast_fp16 = add(x = input_453_cast_fp16, y = linear_79_cast_fp16)[name = string("input_463_cast_fp16")]; + tensor x_225_axes_0 = const()[name = string("x_225_axes_0"), val = tensor([-1])]; + tensor encoder_layers_8_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_8_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(180946112)))]; + tensor encoder_layers_8_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_8_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(180948224)))]; + tensor x_225_cast_fp16 = layer_norm(axes = x_225_axes_0, beta = encoder_layers_8_norm_conv_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_8_norm_conv_weight_to_fp16, x = input_463_cast_fp16)[name = string("x_225_cast_fp16")]; + tensor input_465_perm_0 = const()[name = string("input_465_perm_0"), val = tensor([0, 2, 1])]; + string input_467_pad_type_0 = const()[name = string("input_467_pad_type_0"), val = string("valid")]; + tensor input_467_strides_0 = const()[name = string("input_467_strides_0"), val = tensor([1])]; + tensor input_467_pad_0 = const()[name = string("input_467_pad_0"), val = tensor([0, 0])]; + tensor input_467_dilations_0 = const()[name = string("input_467_dilations_0"), val = tensor([1])]; + int32 input_467_groups_0 = const()[name = string("input_467_groups_0"), val = int32(1)]; + tensor encoder_layers_8_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(180950336))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183047552))))[name = string("encoder_layers_8_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_465_cast_fp16 = transpose(perm = input_465_perm_0, x = x_225_cast_fp16)[name = string("transpose_285")]; + tensor input_467_cast_fp16 = conv(dilations = input_467_dilations_0, groups = input_467_groups_0, pad = input_467_pad_0, pad_type = input_467_pad_type_0, strides = input_467_strides_0, weight = encoder_layers_8_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_465_cast_fp16)[name = string("input_467_cast_fp16")]; + int32 x_227_split_num_splits_0 = const()[name = string("x_227_split_num_splits_0"), val = int32(2)]; + int32 x_227_split_axis_0 = const()[name = string("x_227_split_axis_0"), val = int32(1)]; + tensor x_227_split_cast_fp16_0, tensor x_227_split_cast_fp16_1 = split(axis = x_227_split_axis_0, num_splits = x_227_split_num_splits_0, x = input_467_cast_fp16)[name = string("x_227_split_cast_fp16")]; + tensor x_227_split_1_sigmoid_cast_fp16 = sigmoid(x = x_227_split_cast_fp16_1)[name = string("x_227_split_1_sigmoid_cast_fp16")]; + tensor x_227_cast_fp16 = mul(x = x_227_split_cast_fp16_0, y = x_227_split_1_sigmoid_cast_fp16)[name = string("x_227_cast_fp16")]; + tensor input_469_cast_fp16 = select(a = var_43_to_fp16, b = x_227_cast_fp16, cond = var_574)[name = string("input_469_cast_fp16")]; + bool new_x_35_interleave_0 = const()[name = string("new_x_35_interleave_0"), val = bool(false)]; + tensor new_x_35_cast_fp16 = concat(axis = var_58, interleave = new_x_35_interleave_0, values = (cache_35_cast_fp16, input_469_cast_fp16))[name = string("new_x_35_cast_fp16")]; + tensor var_2291_begin_0 = const()[name = string("op_2291_begin_0"), val = tensor([0, 0, 56])]; + tensor var_2291_end_0 = const()[name = string("op_2291_end_0"), val = tensor([1, 1024, 64])]; + tensor var_2291_end_mask_0 = const()[name = string("op_2291_end_mask_0"), val = tensor([true, true, true])]; + tensor var_2291_cast_fp16 = slice_by_index(begin = var_2291_begin_0, end = var_2291_end_0, end_mask = var_2291_end_mask_0, x = new_x_35_cast_fp16)[name = string("op_2291_cast_fp16")]; + string x_229_pad_type_0 = const()[name = string("x_229_pad_type_0"), val = string("valid")]; + int32 x_229_groups_0 = const()[name = string("x_229_groups_0"), val = int32(1024)]; + tensor x_229_strides_0 = const()[name = string("x_229_strides_0"), val = tensor([1])]; + tensor x_229_pad_0 = const()[name = string("x_229_pad_0"), val = tensor([0, 0])]; + tensor x_229_dilations_0 = const()[name = string("x_229_dilations_0"), val = tensor([1])]; + tensor encoder_layers_8_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183051712))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183060992))))[name = string("encoder_layers_8_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_229_cast_fp16 = conv(dilations = x_229_dilations_0, groups = x_229_groups_0, pad = x_229_pad_0, pad_type = x_229_pad_type_0, strides = x_229_strides_0, weight = encoder_layers_8_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_35_cast_fp16)[name = string("x_229_cast_fp16")]; + tensor input_471_perm_0 = const()[name = string("input_471_perm_0"), val = tensor([0, 2, 1])]; + tensor x_231_axes_0 = const()[name = string("x_231_axes_0"), val = tensor([-1])]; + tensor encoder_layers_8_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_8_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183063104)))]; + tensor encoder_layers_8_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_8_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183065216)))]; + tensor input_471_cast_fp16 = transpose(perm = input_471_perm_0, x = x_229_cast_fp16)[name = string("transpose_284")]; + tensor x_231_cast_fp16 = layer_norm(axes = x_231_axes_0, beta = encoder_layers_8_conv_batch_norm_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_8_conv_batch_norm_weight_to_fp16, x = input_471_cast_fp16)[name = string("x_231_cast_fp16")]; + tensor input_473_perm_0 = const()[name = string("input_473_perm_0"), val = tensor([0, 2, 1])]; + tensor input_473_cast_fp16 = transpose(perm = input_473_perm_0, x = x_231_cast_fp16)[name = string("transpose_283")]; + tensor input_475_cast_fp16 = silu(x = input_473_cast_fp16)[name = string("input_475_cast_fp16")]; + string x_233_pad_type_0 = const()[name = string("x_233_pad_type_0"), val = string("valid")]; + tensor x_233_strides_0 = const()[name = string("x_233_strides_0"), val = tensor([1])]; + tensor x_233_pad_0 = const()[name = string("x_233_pad_0"), val = tensor([0, 0])]; + tensor x_233_dilations_0 = const()[name = string("x_233_dilations_0"), val = tensor([1])]; + int32 x_233_groups_0 = const()[name = string("x_233_groups_0"), val = int32(1)]; + tensor encoder_layers_8_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183067328))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(184115968))))[name = string("encoder_layers_8_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_233_cast_fp16 = conv(dilations = x_233_dilations_0, groups = x_233_groups_0, pad = x_233_pad_0, pad_type = x_233_pad_type_0, strides = x_233_strides_0, weight = encoder_layers_8_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_475_cast_fp16)[name = string("x_233_cast_fp16")]; + tensor input_477_perm_0 = const()[name = string("input_477_perm_0"), val = tensor([0, 2, 1])]; + tensor input_477_cast_fp16 = transpose(perm = input_477_perm_0, x = x_233_cast_fp16)[name = string("transpose_282")]; + tensor input_479_cast_fp16 = add(x = input_463_cast_fp16, y = input_477_cast_fp16)[name = string("input_479_cast_fp16")]; + tensor input_481_axes_0 = const()[name = string("input_481_axes_0"), val = tensor([-1])]; + tensor encoder_layers_8_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_8_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(184118080)))]; + tensor encoder_layers_8_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_8_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(184120192)))]; + tensor input_481_cast_fp16 = layer_norm(axes = input_481_axes_0, beta = encoder_layers_8_norm_feed_forward2_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_8_norm_feed_forward2_weight_to_fp16, x = input_479_cast_fp16)[name = string("input_481_cast_fp16")]; + tensor encoder_layers_8_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(184122304))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(187268096))))[name = string("encoder_layers_8_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_8_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_8_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(187268288)))]; + tensor linear_80_cast_fp16 = linear(bias = encoder_layers_8_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_8_feed_forward2_linear1_weight_to_fp16_palettized, x = input_481_cast_fp16)[name = string("linear_80_cast_fp16")]; + tensor input_485_cast_fp16 = silu(x = linear_80_cast_fp16)[name = string("input_485_cast_fp16")]; + tensor encoder_layers_8_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(187276544))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(190422336))))[name = string("encoder_layers_8_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_8_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_8_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(190422528)))]; + tensor linear_81_cast_fp16 = linear(bias = encoder_layers_8_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_8_feed_forward2_linear2_weight_to_fp16_palettized, x = input_485_cast_fp16)[name = string("linear_81_cast_fp16")]; + fp16 var_2334_to_fp16 = const()[name = string("op_2334_to_fp16"), val = fp16(0x1p-1)]; + tensor var_2335_cast_fp16 = mul(x = linear_81_cast_fp16, y = var_2334_to_fp16)[name = string("op_2335_cast_fp16")]; + tensor input_491_cast_fp16 = add(x = input_479_cast_fp16, y = var_2335_cast_fp16)[name = string("input_491_cast_fp16")]; + tensor input_493_axes_0 = const()[name = string("input_493_axes_0"), val = tensor([-1])]; + tensor encoder_layers_8_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_8_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(190424640)))]; + tensor encoder_layers_8_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_8_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(190426752)))]; + tensor input_493_cast_fp16 = layer_norm(axes = input_493_axes_0, beta = encoder_layers_8_norm_out_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_8_norm_out_weight_to_fp16, x = input_491_cast_fp16)[name = string("input_493_cast_fp16")]; + tensor cache_37_begin_0 = const()[name = string("cache_37_begin_0"), val = tensor([9, 0, 0, 0])]; + tensor cache_37_end_0 = const()[name = string("cache_37_end_0"), val = tensor([10, 1, 42, 1024])]; + tensor cache_37_end_mask_0 = const()[name = string("cache_37_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_37_squeeze_mask_0 = const()[name = string("cache_37_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_37_cast_fp16 = slice_by_index(begin = cache_37_begin_0, end = cache_37_end_0, end_mask = cache_37_end_mask_0, squeeze_mask = cache_37_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_37_cast_fp16")]; + tensor cache_39_begin_0 = const()[name = string("cache_39_begin_0"), val = tensor([9, 0, 0, 0])]; + tensor cache_39_end_0 = const()[name = string("cache_39_end_0"), val = tensor([10, 1, 1024, 8])]; + tensor cache_39_end_mask_0 = const()[name = string("cache_39_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_39_squeeze_mask_0 = const()[name = string("cache_39_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_39_cast_fp16 = slice_by_index(begin = cache_39_begin_0, end = cache_39_end_0, end_mask = cache_39_end_mask_0, squeeze_mask = cache_39_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_39_cast_fp16")]; + tensor input_495_axes_0 = const()[name = string("input_495_axes_0"), val = tensor([-1])]; + tensor encoder_layers_9_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_9_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(190428864)))]; + tensor encoder_layers_9_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_9_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(190430976)))]; + tensor input_495_cast_fp16 = layer_norm(axes = input_495_axes_0, beta = encoder_layers_9_norm_feed_forward1_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_9_norm_feed_forward1_weight_to_fp16, x = input_493_cast_fp16)[name = string("input_495_cast_fp16")]; + tensor encoder_layers_9_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(190433088))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(193578880))))[name = string("encoder_layers_9_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_9_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_9_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(193579072)))]; + tensor linear_82_cast_fp16 = linear(bias = encoder_layers_9_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_9_feed_forward1_linear1_weight_to_fp16_palettized, x = input_495_cast_fp16)[name = string("linear_82_cast_fp16")]; + tensor input_499_cast_fp16 = silu(x = linear_82_cast_fp16)[name = string("input_499_cast_fp16")]; + tensor encoder_layers_9_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(193587328))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196733120))))[name = string("encoder_layers_9_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_9_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_9_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196733312)))]; + tensor linear_83_cast_fp16 = linear(bias = encoder_layers_9_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_9_feed_forward1_linear2_weight_to_fp16_palettized, x = input_499_cast_fp16)[name = string("linear_83_cast_fp16")]; + fp16 var_2371_to_fp16 = const()[name = string("op_2371_to_fp16"), val = fp16(0x1p-1)]; + tensor var_2372_cast_fp16 = mul(x = linear_83_cast_fp16, y = var_2371_to_fp16)[name = string("op_2372_cast_fp16")]; + tensor input_505_cast_fp16 = add(x = input_493_cast_fp16, y = var_2372_cast_fp16)[name = string("input_505_cast_fp16")]; + tensor key_19_axes_0 = const()[name = string("key_19_axes_0"), val = tensor([-1])]; + tensor encoder_layers_9_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_9_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196735424)))]; + tensor encoder_layers_9_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_9_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196737536)))]; + tensor key_19_cast_fp16 = layer_norm(axes = key_19_axes_0, beta = encoder_layers_9_norm_self_att_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_9_norm_self_att_weight_to_fp16, x = input_505_cast_fp16)[name = string("key_19_cast_fp16")]; + bool input_507_interleave_0 = const()[name = string("input_507_interleave_0"), val = bool(false)]; + tensor input_507_cast_fp16 = concat(axis = var_67, interleave = input_507_interleave_0, values = (cache_37_cast_fp16, key_19_cast_fp16))[name = string("input_507_cast_fp16")]; + bool var_2400_interleave_0 = const()[name = string("op_2400_interleave_0"), val = bool(false)]; + tensor var_2400_cast_fp16 = concat(axis = var_67, interleave = var_2400_interleave_0, values = key_19_cast_fp16)[name = string("op_2400_cast_fp16")]; + tensor encoder_layers_9_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196739648))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197526144))))[name = string("encoder_layers_9_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_9_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_9_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197526336)))]; + tensor linear_84_cast_fp16 = linear(bias = encoder_layers_9_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_9_self_attn_linear_q_weight_to_fp16_palettized, x = key_19_cast_fp16)[name = string("linear_84_cast_fp16")]; + tensor var_2405 = const()[name = string("op_2405"), val = tensor([1, -1, 8, 128])]; + tensor q_55_cast_fp16 = reshape(shape = var_2405, x = linear_84_cast_fp16)[name = string("q_55_cast_fp16")]; + tensor encoder_layers_9_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197528448))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198314944))))[name = string("encoder_layers_9_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_9_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_9_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198315136)))]; + tensor linear_85_cast_fp16 = linear(bias = encoder_layers_9_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_9_self_attn_linear_k_weight_to_fp16_palettized, x = input_507_cast_fp16)[name = string("linear_85_cast_fp16")]; + tensor var_2410 = const()[name = string("op_2410"), val = tensor([1, -1, 8, 128])]; + tensor k_37_cast_fp16 = reshape(shape = var_2410, x = linear_85_cast_fp16)[name = string("k_37_cast_fp16")]; + tensor encoder_layers_9_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198317248))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199103744))))[name = string("encoder_layers_9_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_9_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_9_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199103936)))]; + tensor linear_86_cast_fp16 = linear(bias = encoder_layers_9_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_9_self_attn_linear_v_weight_to_fp16_palettized, x = input_507_cast_fp16)[name = string("linear_86_cast_fp16")]; + tensor var_2415 = const()[name = string("op_2415"), val = tensor([1, -1, 8, 128])]; + tensor v_19_cast_fp16 = reshape(shape = var_2415, x = linear_86_cast_fp16)[name = string("v_19_cast_fp16")]; + tensor value_27_perm_0 = const()[name = string("value_27_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_9_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_9_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199106048)))]; + tensor var_2428_cast_fp16 = add(x = q_55_cast_fp16, y = encoder_layers_9_self_attn_pos_bias_u_to_fp16)[name = string("op_2428_cast_fp16")]; + tensor encoder_layers_9_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_9_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199108160)))]; + tensor var_2430_cast_fp16 = add(x = q_55_cast_fp16, y = encoder_layers_9_self_attn_pos_bias_v_to_fp16)[name = string("op_2430_cast_fp16")]; + tensor q_with_bias_v_19_perm_0 = const()[name = string("q_with_bias_v_19_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_241_transpose_x_0 = const()[name = string("x_241_transpose_x_0"), val = bool(false)]; + bool x_241_transpose_y_0 = const()[name = string("x_241_transpose_y_0"), val = bool(false)]; + tensor op_2432_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199110272))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199310016))))[name = string("op_2432_to_fp16_quantized")]; + tensor q_with_bias_v_19_cast_fp16 = transpose(perm = q_with_bias_v_19_perm_0, x = var_2430_cast_fp16)[name = string("transpose_281")]; + tensor x_241_cast_fp16 = matmul(transpose_x = x_241_transpose_x_0, transpose_y = x_241_transpose_y_0, x = q_with_bias_v_19_cast_fp16, y = op_2432_to_fp16_quantized)[name = string("x_241_cast_fp16")]; + tensor x_243_pad_0 = const()[name = string("x_243_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_243_mode_0 = const()[name = string("x_243_mode_0"), val = string("constant")]; + fp16 const_196_to_fp16 = const()[name = string("const_196_to_fp16"), val = fp16(0x0p+0)]; + tensor x_243_cast_fp16 = pad(constant_val = const_196_to_fp16, mode = x_243_mode_0, pad = x_243_pad_0, x = x_241_cast_fp16)[name = string("x_243_cast_fp16")]; + tensor var_2440 = const()[name = string("op_2440"), val = tensor([1, 8, -1, 56])]; + tensor x_245_cast_fp16 = reshape(shape = var_2440, x = x_243_cast_fp16)[name = string("x_245_cast_fp16")]; + tensor var_2444_begin_0 = const()[name = string("op_2444_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_2444_end_0 = const()[name = string("op_2444_end_0"), val = tensor([1, 8, 196, 56])]; + tensor var_2444_end_mask_0 = const()[name = string("op_2444_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_2444_cast_fp16 = slice_by_index(begin = var_2444_begin_0, end = var_2444_end_0, end_mask = var_2444_end_mask_0, x = x_245_cast_fp16)[name = string("op_2444_cast_fp16")]; + tensor var_2445 = const()[name = string("op_2445"), val = tensor([1, 8, 56, 195])]; + tensor matrix_bd_37_cast_fp16 = reshape(shape = var_2445, x = var_2444_cast_fp16)[name = string("matrix_bd_37_cast_fp16")]; + bool matrix_ac_19_transpose_x_0 = const()[name = string("matrix_ac_19_transpose_x_0"), val = bool(false)]; + bool matrix_ac_19_transpose_y_0 = const()[name = string("matrix_ac_19_transpose_y_0"), val = bool(false)]; + tensor transpose_114_perm_0 = const()[name = string("transpose_114_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_115_perm_0 = const()[name = string("transpose_115_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_115 = transpose(perm = transpose_115_perm_0, x = k_37_cast_fp16)[name = string("transpose_279")]; + tensor transpose_114 = transpose(perm = transpose_114_perm_0, x = var_2428_cast_fp16)[name = string("transpose_280")]; + tensor matrix_ac_19_cast_fp16 = matmul(transpose_x = matrix_ac_19_transpose_x_0, transpose_y = matrix_ac_19_transpose_y_0, x = transpose_114, y = transpose_115)[name = string("matrix_ac_19_cast_fp16")]; + tensor matrix_bd_39_begin_0 = const()[name = string("matrix_bd_39_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_39_end_0 = const()[name = string("matrix_bd_39_end_0"), val = tensor([1, 8, 56, 98])]; + tensor matrix_bd_39_end_mask_0 = const()[name = string("matrix_bd_39_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_39_cast_fp16 = slice_by_index(begin = matrix_bd_39_begin_0, end = matrix_bd_39_end_0, end_mask = matrix_bd_39_end_mask_0, x = matrix_bd_37_cast_fp16)[name = string("matrix_bd_39_cast_fp16")]; + tensor var_2454_cast_fp16 = add(x = matrix_ac_19_cast_fp16, y = matrix_bd_39_cast_fp16)[name = string("op_2454_cast_fp16")]; + fp16 _inversed_scores_37_y_0_to_fp16 = const()[name = string("_inversed_scores_37_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_37_cast_fp16 = mul(x = var_2454_cast_fp16, y = _inversed_scores_37_y_0_to_fp16)[name = string("_inversed_scores_37_cast_fp16")]; + tensor scores_39_cast_fp16 = select(a = var_44_to_fp16, b = _inversed_scores_37_cast_fp16, cond = mask_11)[name = string("scores_39_cast_fp16")]; + tensor var_2460_cast_fp16 = softmax(axis = var_58, x = scores_39_cast_fp16)[name = string("op_2460_cast_fp16")]; + tensor input_509_cast_fp16 = select(a = var_43_to_fp16, b = var_2460_cast_fp16, cond = mask_11)[name = string("input_509_cast_fp16")]; + bool x_247_transpose_x_0 = const()[name = string("x_247_transpose_x_0"), val = bool(false)]; + bool x_247_transpose_y_0 = const()[name = string("x_247_transpose_y_0"), val = bool(false)]; + tensor value_27_cast_fp16 = transpose(perm = value_27_perm_0, x = v_19_cast_fp16)[name = string("transpose_278")]; + tensor x_247_cast_fp16 = matmul(transpose_x = x_247_transpose_x_0, transpose_y = x_247_transpose_y_0, x = input_509_cast_fp16, y = value_27_cast_fp16)[name = string("x_247_cast_fp16")]; + tensor var_2464_perm_0 = const()[name = string("op_2464_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_2465 = const()[name = string("op_2465"), val = tensor([1, -1, 1024])]; + tensor var_2464_cast_fp16 = transpose(perm = var_2464_perm_0, x = x_247_cast_fp16)[name = string("transpose_277")]; + tensor input_511_cast_fp16 = reshape(shape = var_2465, x = var_2464_cast_fp16)[name = string("input_511_cast_fp16")]; + tensor encoder_layers_9_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199310528))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(200097024))))[name = string("encoder_layers_9_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_9_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_9_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(200097216)))]; + tensor linear_88_cast_fp16 = linear(bias = encoder_layers_9_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_9_self_attn_linear_out_weight_to_fp16_palettized, x = input_511_cast_fp16)[name = string("linear_88_cast_fp16")]; + tensor input_515_cast_fp16 = add(x = input_505_cast_fp16, y = linear_88_cast_fp16)[name = string("input_515_cast_fp16")]; + tensor x_251_axes_0 = const()[name = string("x_251_axes_0"), val = tensor([-1])]; + tensor encoder_layers_9_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_9_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(200099328)))]; + tensor encoder_layers_9_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_9_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(200101440)))]; + tensor x_251_cast_fp16 = layer_norm(axes = x_251_axes_0, beta = encoder_layers_9_norm_conv_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_9_norm_conv_weight_to_fp16, x = input_515_cast_fp16)[name = string("x_251_cast_fp16")]; + tensor input_517_perm_0 = const()[name = string("input_517_perm_0"), val = tensor([0, 2, 1])]; + string input_519_pad_type_0 = const()[name = string("input_519_pad_type_0"), val = string("valid")]; + tensor input_519_strides_0 = const()[name = string("input_519_strides_0"), val = tensor([1])]; + tensor input_519_pad_0 = const()[name = string("input_519_pad_0"), val = tensor([0, 0])]; + tensor input_519_dilations_0 = const()[name = string("input_519_dilations_0"), val = tensor([1])]; + int32 input_519_groups_0 = const()[name = string("input_519_groups_0"), val = int32(1)]; + tensor encoder_layers_9_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(200103552))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202200768))))[name = string("encoder_layers_9_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_517_cast_fp16 = transpose(perm = input_517_perm_0, x = x_251_cast_fp16)[name = string("transpose_276")]; + tensor input_519_cast_fp16 = conv(dilations = input_519_dilations_0, groups = input_519_groups_0, pad = input_519_pad_0, pad_type = input_519_pad_type_0, strides = input_519_strides_0, weight = encoder_layers_9_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_517_cast_fp16)[name = string("input_519_cast_fp16")]; + int32 x_253_split_num_splits_0 = const()[name = string("x_253_split_num_splits_0"), val = int32(2)]; + int32 x_253_split_axis_0 = const()[name = string("x_253_split_axis_0"), val = int32(1)]; + tensor x_253_split_cast_fp16_0, tensor x_253_split_cast_fp16_1 = split(axis = x_253_split_axis_0, num_splits = x_253_split_num_splits_0, x = input_519_cast_fp16)[name = string("x_253_split_cast_fp16")]; + tensor x_253_split_1_sigmoid_cast_fp16 = sigmoid(x = x_253_split_cast_fp16_1)[name = string("x_253_split_1_sigmoid_cast_fp16")]; + tensor x_253_cast_fp16 = mul(x = x_253_split_cast_fp16_0, y = x_253_split_1_sigmoid_cast_fp16)[name = string("x_253_cast_fp16")]; + tensor input_521_cast_fp16 = select(a = var_43_to_fp16, b = x_253_cast_fp16, cond = var_574)[name = string("input_521_cast_fp16")]; + bool new_x_39_interleave_0 = const()[name = string("new_x_39_interleave_0"), val = bool(false)]; + tensor new_x_39_cast_fp16 = concat(axis = var_58, interleave = new_x_39_interleave_0, values = (cache_39_cast_fp16, input_521_cast_fp16))[name = string("new_x_39_cast_fp16")]; + tensor var_2504_begin_0 = const()[name = string("op_2504_begin_0"), val = tensor([0, 0, 56])]; + tensor var_2504_end_0 = const()[name = string("op_2504_end_0"), val = tensor([1, 1024, 64])]; + tensor var_2504_end_mask_0 = const()[name = string("op_2504_end_mask_0"), val = tensor([true, true, true])]; + tensor var_2504_cast_fp16 = slice_by_index(begin = var_2504_begin_0, end = var_2504_end_0, end_mask = var_2504_end_mask_0, x = new_x_39_cast_fp16)[name = string("op_2504_cast_fp16")]; + string x_255_pad_type_0 = const()[name = string("x_255_pad_type_0"), val = string("valid")]; + int32 x_255_groups_0 = const()[name = string("x_255_groups_0"), val = int32(1024)]; + tensor x_255_strides_0 = const()[name = string("x_255_strides_0"), val = tensor([1])]; + tensor x_255_pad_0 = const()[name = string("x_255_pad_0"), val = tensor([0, 0])]; + tensor x_255_dilations_0 = const()[name = string("x_255_dilations_0"), val = tensor([1])]; + tensor encoder_layers_9_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202204928))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202214208))))[name = string("encoder_layers_9_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_255_cast_fp16 = conv(dilations = x_255_dilations_0, groups = x_255_groups_0, pad = x_255_pad_0, pad_type = x_255_pad_type_0, strides = x_255_strides_0, weight = encoder_layers_9_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_39_cast_fp16)[name = string("x_255_cast_fp16")]; + tensor input_523_perm_0 = const()[name = string("input_523_perm_0"), val = tensor([0, 2, 1])]; + tensor x_257_axes_0 = const()[name = string("x_257_axes_0"), val = tensor([-1])]; + tensor encoder_layers_9_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_9_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202216320)))]; + tensor encoder_layers_9_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_9_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202218432)))]; + tensor input_523_cast_fp16 = transpose(perm = input_523_perm_0, x = x_255_cast_fp16)[name = string("transpose_275")]; + tensor x_257_cast_fp16 = layer_norm(axes = x_257_axes_0, beta = encoder_layers_9_conv_batch_norm_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_9_conv_batch_norm_weight_to_fp16, x = input_523_cast_fp16)[name = string("x_257_cast_fp16")]; + tensor input_525_perm_0 = const()[name = string("input_525_perm_0"), val = tensor([0, 2, 1])]; + tensor input_525_cast_fp16 = transpose(perm = input_525_perm_0, x = x_257_cast_fp16)[name = string("transpose_274")]; + tensor input_527_cast_fp16 = silu(x = input_525_cast_fp16)[name = string("input_527_cast_fp16")]; + string x_259_pad_type_0 = const()[name = string("x_259_pad_type_0"), val = string("valid")]; + tensor x_259_strides_0 = const()[name = string("x_259_strides_0"), val = tensor([1])]; + tensor x_259_pad_0 = const()[name = string("x_259_pad_0"), val = tensor([0, 0])]; + tensor x_259_dilations_0 = const()[name = string("x_259_dilations_0"), val = tensor([1])]; + int32 x_259_groups_0 = const()[name = string("x_259_groups_0"), val = int32(1)]; + tensor encoder_layers_9_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202220544))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(203269184))))[name = string("encoder_layers_9_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_259_cast_fp16 = conv(dilations = x_259_dilations_0, groups = x_259_groups_0, pad = x_259_pad_0, pad_type = x_259_pad_type_0, strides = x_259_strides_0, weight = encoder_layers_9_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_527_cast_fp16)[name = string("x_259_cast_fp16")]; + tensor input_529_perm_0 = const()[name = string("input_529_perm_0"), val = tensor([0, 2, 1])]; + tensor input_529_cast_fp16 = transpose(perm = input_529_perm_0, x = x_259_cast_fp16)[name = string("transpose_273")]; + tensor input_531_cast_fp16 = add(x = input_515_cast_fp16, y = input_529_cast_fp16)[name = string("input_531_cast_fp16")]; + tensor input_533_axes_0 = const()[name = string("input_533_axes_0"), val = tensor([-1])]; + tensor encoder_layers_9_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_9_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(203271296)))]; + tensor encoder_layers_9_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_9_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(203273408)))]; + tensor input_533_cast_fp16 = layer_norm(axes = input_533_axes_0, beta = encoder_layers_9_norm_feed_forward2_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_9_norm_feed_forward2_weight_to_fp16, x = input_531_cast_fp16)[name = string("input_533_cast_fp16")]; + tensor encoder_layers_9_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(203275520))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(206421312))))[name = string("encoder_layers_9_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_9_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_9_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(206421504)))]; + tensor linear_89_cast_fp16 = linear(bias = encoder_layers_9_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_9_feed_forward2_linear1_weight_to_fp16_palettized, x = input_533_cast_fp16)[name = string("linear_89_cast_fp16")]; + tensor input_537_cast_fp16 = silu(x = linear_89_cast_fp16)[name = string("input_537_cast_fp16")]; + tensor encoder_layers_9_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(206429760))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(209575552))))[name = string("encoder_layers_9_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_9_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_9_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(209575744)))]; + tensor linear_90_cast_fp16 = linear(bias = encoder_layers_9_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_9_feed_forward2_linear2_weight_to_fp16_palettized, x = input_537_cast_fp16)[name = string("linear_90_cast_fp16")]; + fp16 var_2547_to_fp16 = const()[name = string("op_2547_to_fp16"), val = fp16(0x1p-1)]; + tensor var_2548_cast_fp16 = mul(x = linear_90_cast_fp16, y = var_2547_to_fp16)[name = string("op_2548_cast_fp16")]; + tensor input_543_cast_fp16 = add(x = input_531_cast_fp16, y = var_2548_cast_fp16)[name = string("input_543_cast_fp16")]; + tensor input_545_axes_0 = const()[name = string("input_545_axes_0"), val = tensor([-1])]; + tensor encoder_layers_9_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_9_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(209577856)))]; + tensor encoder_layers_9_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_9_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(209579968)))]; + tensor input_545_cast_fp16 = layer_norm(axes = input_545_axes_0, beta = encoder_layers_9_norm_out_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_9_norm_out_weight_to_fp16, x = input_543_cast_fp16)[name = string("input_545_cast_fp16")]; + tensor cache_41_begin_0 = const()[name = string("cache_41_begin_0"), val = tensor([10, 0, 0, 0])]; + tensor cache_41_end_0 = const()[name = string("cache_41_end_0"), val = tensor([11, 1, 42, 1024])]; + tensor cache_41_end_mask_0 = const()[name = string("cache_41_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_41_squeeze_mask_0 = const()[name = string("cache_41_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_41_cast_fp16 = slice_by_index(begin = cache_41_begin_0, end = cache_41_end_0, end_mask = cache_41_end_mask_0, squeeze_mask = cache_41_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_41_cast_fp16")]; + tensor cache_43_begin_0 = const()[name = string("cache_43_begin_0"), val = tensor([10, 0, 0, 0])]; + tensor cache_43_end_0 = const()[name = string("cache_43_end_0"), val = tensor([11, 1, 1024, 8])]; + tensor cache_43_end_mask_0 = const()[name = string("cache_43_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_43_squeeze_mask_0 = const()[name = string("cache_43_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_43_cast_fp16 = slice_by_index(begin = cache_43_begin_0, end = cache_43_end_0, end_mask = cache_43_end_mask_0, squeeze_mask = cache_43_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_43_cast_fp16")]; + tensor input_547_axes_0 = const()[name = string("input_547_axes_0"), val = tensor([-1])]; + tensor encoder_layers_10_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_10_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(209582080)))]; + tensor encoder_layers_10_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_10_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(209584192)))]; + tensor input_547_cast_fp16 = layer_norm(axes = input_547_axes_0, beta = encoder_layers_10_norm_feed_forward1_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_10_norm_feed_forward1_weight_to_fp16, x = input_545_cast_fp16)[name = string("input_547_cast_fp16")]; + tensor encoder_layers_10_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(209586304))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212732096))))[name = string("encoder_layers_10_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_10_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_10_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212732288)))]; + tensor linear_91_cast_fp16 = linear(bias = encoder_layers_10_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_10_feed_forward1_linear1_weight_to_fp16_palettized, x = input_547_cast_fp16)[name = string("linear_91_cast_fp16")]; + tensor input_551_cast_fp16 = silu(x = linear_91_cast_fp16)[name = string("input_551_cast_fp16")]; + tensor encoder_layers_10_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212740544))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215886336))))[name = string("encoder_layers_10_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_10_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_10_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215886528)))]; + tensor linear_92_cast_fp16 = linear(bias = encoder_layers_10_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_10_feed_forward1_linear2_weight_to_fp16_palettized, x = input_551_cast_fp16)[name = string("linear_92_cast_fp16")]; + fp16 var_2584_to_fp16 = const()[name = string("op_2584_to_fp16"), val = fp16(0x1p-1)]; + tensor var_2585_cast_fp16 = mul(x = linear_92_cast_fp16, y = var_2584_to_fp16)[name = string("op_2585_cast_fp16")]; + tensor input_557_cast_fp16 = add(x = input_545_cast_fp16, y = var_2585_cast_fp16)[name = string("input_557_cast_fp16")]; + tensor key_21_axes_0 = const()[name = string("key_21_axes_0"), val = tensor([-1])]; + tensor encoder_layers_10_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_10_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215888640)))]; + tensor encoder_layers_10_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_10_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215890752)))]; + tensor key_21_cast_fp16 = layer_norm(axes = key_21_axes_0, beta = encoder_layers_10_norm_self_att_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_10_norm_self_att_weight_to_fp16, x = input_557_cast_fp16)[name = string("key_21_cast_fp16")]; + bool input_559_interleave_0 = const()[name = string("input_559_interleave_0"), val = bool(false)]; + tensor input_559_cast_fp16 = concat(axis = var_67, interleave = input_559_interleave_0, values = (cache_41_cast_fp16, key_21_cast_fp16))[name = string("input_559_cast_fp16")]; + bool var_2613_interleave_0 = const()[name = string("op_2613_interleave_0"), val = bool(false)]; + tensor var_2613_cast_fp16 = concat(axis = var_67, interleave = var_2613_interleave_0, values = key_21_cast_fp16)[name = string("op_2613_cast_fp16")]; + tensor encoder_layers_10_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215892864))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216679360))))[name = string("encoder_layers_10_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_10_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_10_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216679552)))]; + tensor linear_93_cast_fp16 = linear(bias = encoder_layers_10_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_10_self_attn_linear_q_weight_to_fp16_palettized, x = key_21_cast_fp16)[name = string("linear_93_cast_fp16")]; + tensor var_2618 = const()[name = string("op_2618"), val = tensor([1, -1, 8, 128])]; + tensor q_61_cast_fp16 = reshape(shape = var_2618, x = linear_93_cast_fp16)[name = string("q_61_cast_fp16")]; + tensor encoder_layers_10_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216681664))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217468160))))[name = string("encoder_layers_10_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_10_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_10_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217468352)))]; + tensor linear_94_cast_fp16 = linear(bias = encoder_layers_10_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_10_self_attn_linear_k_weight_to_fp16_palettized, x = input_559_cast_fp16)[name = string("linear_94_cast_fp16")]; + tensor var_2623 = const()[name = string("op_2623"), val = tensor([1, -1, 8, 128])]; + tensor k_41_cast_fp16 = reshape(shape = var_2623, x = linear_94_cast_fp16)[name = string("k_41_cast_fp16")]; + tensor encoder_layers_10_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217470464))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218256960))))[name = string("encoder_layers_10_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_10_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_10_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218257152)))]; + tensor linear_95_cast_fp16 = linear(bias = encoder_layers_10_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_10_self_attn_linear_v_weight_to_fp16_palettized, x = input_559_cast_fp16)[name = string("linear_95_cast_fp16")]; + tensor var_2628 = const()[name = string("op_2628"), val = tensor([1, -1, 8, 128])]; + tensor v_21_cast_fp16 = reshape(shape = var_2628, x = linear_95_cast_fp16)[name = string("v_21_cast_fp16")]; + tensor value_29_perm_0 = const()[name = string("value_29_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_10_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_10_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218259264)))]; + tensor var_2641_cast_fp16 = add(x = q_61_cast_fp16, y = encoder_layers_10_self_attn_pos_bias_u_to_fp16)[name = string("op_2641_cast_fp16")]; + tensor encoder_layers_10_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_10_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218261376)))]; + tensor var_2643_cast_fp16 = add(x = q_61_cast_fp16, y = encoder_layers_10_self_attn_pos_bias_v_to_fp16)[name = string("op_2643_cast_fp16")]; + tensor q_with_bias_v_21_perm_0 = const()[name = string("q_with_bias_v_21_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_267_transpose_x_0 = const()[name = string("x_267_transpose_x_0"), val = bool(false)]; + bool x_267_transpose_y_0 = const()[name = string("x_267_transpose_y_0"), val = bool(false)]; + tensor op_2645_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218263488))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218463232))))[name = string("op_2645_to_fp16_quantized")]; + tensor q_with_bias_v_21_cast_fp16 = transpose(perm = q_with_bias_v_21_perm_0, x = var_2643_cast_fp16)[name = string("transpose_272")]; + tensor x_267_cast_fp16 = matmul(transpose_x = x_267_transpose_x_0, transpose_y = x_267_transpose_y_0, x = q_with_bias_v_21_cast_fp16, y = op_2645_to_fp16_quantized)[name = string("x_267_cast_fp16")]; + tensor x_269_pad_0 = const()[name = string("x_269_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_269_mode_0 = const()[name = string("x_269_mode_0"), val = string("constant")]; + fp16 const_209_to_fp16 = const()[name = string("const_209_to_fp16"), val = fp16(0x0p+0)]; + tensor x_269_cast_fp16 = pad(constant_val = const_209_to_fp16, mode = x_269_mode_0, pad = x_269_pad_0, x = x_267_cast_fp16)[name = string("x_269_cast_fp16")]; + tensor var_2653 = const()[name = string("op_2653"), val = tensor([1, 8, -1, 56])]; + tensor x_271_cast_fp16 = reshape(shape = var_2653, x = x_269_cast_fp16)[name = string("x_271_cast_fp16")]; + tensor var_2657_begin_0 = const()[name = string("op_2657_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_2657_end_0 = const()[name = string("op_2657_end_0"), val = tensor([1, 8, 196, 56])]; + tensor var_2657_end_mask_0 = const()[name = string("op_2657_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_2657_cast_fp16 = slice_by_index(begin = var_2657_begin_0, end = var_2657_end_0, end_mask = var_2657_end_mask_0, x = x_271_cast_fp16)[name = string("op_2657_cast_fp16")]; + tensor var_2658 = const()[name = string("op_2658"), val = tensor([1, 8, 56, 195])]; + tensor matrix_bd_41_cast_fp16 = reshape(shape = var_2658, x = var_2657_cast_fp16)[name = string("matrix_bd_41_cast_fp16")]; + bool matrix_ac_21_transpose_x_0 = const()[name = string("matrix_ac_21_transpose_x_0"), val = bool(false)]; + bool matrix_ac_21_transpose_y_0 = const()[name = string("matrix_ac_21_transpose_y_0"), val = bool(false)]; + tensor transpose_116_perm_0 = const()[name = string("transpose_116_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_117_perm_0 = const()[name = string("transpose_117_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_117 = transpose(perm = transpose_117_perm_0, x = k_41_cast_fp16)[name = string("transpose_270")]; + tensor transpose_116 = transpose(perm = transpose_116_perm_0, x = var_2641_cast_fp16)[name = string("transpose_271")]; + tensor matrix_ac_21_cast_fp16 = matmul(transpose_x = matrix_ac_21_transpose_x_0, transpose_y = matrix_ac_21_transpose_y_0, x = transpose_116, y = transpose_117)[name = string("matrix_ac_21_cast_fp16")]; + tensor matrix_bd_43_begin_0 = const()[name = string("matrix_bd_43_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_43_end_0 = const()[name = string("matrix_bd_43_end_0"), val = tensor([1, 8, 56, 98])]; + tensor matrix_bd_43_end_mask_0 = const()[name = string("matrix_bd_43_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_43_cast_fp16 = slice_by_index(begin = matrix_bd_43_begin_0, end = matrix_bd_43_end_0, end_mask = matrix_bd_43_end_mask_0, x = matrix_bd_41_cast_fp16)[name = string("matrix_bd_43_cast_fp16")]; + tensor var_2667_cast_fp16 = add(x = matrix_ac_21_cast_fp16, y = matrix_bd_43_cast_fp16)[name = string("op_2667_cast_fp16")]; + fp16 _inversed_scores_41_y_0_to_fp16 = const()[name = string("_inversed_scores_41_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_41_cast_fp16 = mul(x = var_2667_cast_fp16, y = _inversed_scores_41_y_0_to_fp16)[name = string("_inversed_scores_41_cast_fp16")]; + tensor scores_43_cast_fp16 = select(a = var_44_to_fp16, b = _inversed_scores_41_cast_fp16, cond = mask_11)[name = string("scores_43_cast_fp16")]; + tensor var_2673_cast_fp16 = softmax(axis = var_58, x = scores_43_cast_fp16)[name = string("op_2673_cast_fp16")]; + tensor input_561_cast_fp16 = select(a = var_43_to_fp16, b = var_2673_cast_fp16, cond = mask_11)[name = string("input_561_cast_fp16")]; + bool x_273_transpose_x_0 = const()[name = string("x_273_transpose_x_0"), val = bool(false)]; + bool x_273_transpose_y_0 = const()[name = string("x_273_transpose_y_0"), val = bool(false)]; + tensor value_29_cast_fp16 = transpose(perm = value_29_perm_0, x = v_21_cast_fp16)[name = string("transpose_269")]; + tensor x_273_cast_fp16 = matmul(transpose_x = x_273_transpose_x_0, transpose_y = x_273_transpose_y_0, x = input_561_cast_fp16, y = value_29_cast_fp16)[name = string("x_273_cast_fp16")]; + tensor var_2677_perm_0 = const()[name = string("op_2677_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_2678 = const()[name = string("op_2678"), val = tensor([1, -1, 1024])]; + tensor var_2677_cast_fp16 = transpose(perm = var_2677_perm_0, x = x_273_cast_fp16)[name = string("transpose_268")]; + tensor input_563_cast_fp16 = reshape(shape = var_2678, x = var_2677_cast_fp16)[name = string("input_563_cast_fp16")]; + tensor encoder_layers_10_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218463744))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(219250240))))[name = string("encoder_layers_10_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_10_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_10_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(219250432)))]; + tensor linear_97_cast_fp16 = linear(bias = encoder_layers_10_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_10_self_attn_linear_out_weight_to_fp16_palettized, x = input_563_cast_fp16)[name = string("linear_97_cast_fp16")]; + tensor input_567_cast_fp16 = add(x = input_557_cast_fp16, y = linear_97_cast_fp16)[name = string("input_567_cast_fp16")]; + tensor x_277_axes_0 = const()[name = string("x_277_axes_0"), val = tensor([-1])]; + tensor encoder_layers_10_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_10_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(219252544)))]; + tensor encoder_layers_10_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_10_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(219254656)))]; + tensor x_277_cast_fp16 = layer_norm(axes = x_277_axes_0, beta = encoder_layers_10_norm_conv_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_10_norm_conv_weight_to_fp16, x = input_567_cast_fp16)[name = string("x_277_cast_fp16")]; + tensor input_569_perm_0 = const()[name = string("input_569_perm_0"), val = tensor([0, 2, 1])]; + string input_571_pad_type_0 = const()[name = string("input_571_pad_type_0"), val = string("valid")]; + tensor input_571_strides_0 = const()[name = string("input_571_strides_0"), val = tensor([1])]; + tensor input_571_pad_0 = const()[name = string("input_571_pad_0"), val = tensor([0, 0])]; + tensor input_571_dilations_0 = const()[name = string("input_571_dilations_0"), val = tensor([1])]; + int32 input_571_groups_0 = const()[name = string("input_571_groups_0"), val = int32(1)]; + tensor encoder_layers_10_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(219256768))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221353984))))[name = string("encoder_layers_10_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_569_cast_fp16 = transpose(perm = input_569_perm_0, x = x_277_cast_fp16)[name = string("transpose_267")]; + tensor input_571_cast_fp16 = conv(dilations = input_571_dilations_0, groups = input_571_groups_0, pad = input_571_pad_0, pad_type = input_571_pad_type_0, strides = input_571_strides_0, weight = encoder_layers_10_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_569_cast_fp16)[name = string("input_571_cast_fp16")]; + int32 x_279_split_num_splits_0 = const()[name = string("x_279_split_num_splits_0"), val = int32(2)]; + int32 x_279_split_axis_0 = const()[name = string("x_279_split_axis_0"), val = int32(1)]; + tensor x_279_split_cast_fp16_0, tensor x_279_split_cast_fp16_1 = split(axis = x_279_split_axis_0, num_splits = x_279_split_num_splits_0, x = input_571_cast_fp16)[name = string("x_279_split_cast_fp16")]; + tensor x_279_split_1_sigmoid_cast_fp16 = sigmoid(x = x_279_split_cast_fp16_1)[name = string("x_279_split_1_sigmoid_cast_fp16")]; + tensor x_279_cast_fp16 = mul(x = x_279_split_cast_fp16_0, y = x_279_split_1_sigmoid_cast_fp16)[name = string("x_279_cast_fp16")]; + tensor input_573_cast_fp16 = select(a = var_43_to_fp16, b = x_279_cast_fp16, cond = var_574)[name = string("input_573_cast_fp16")]; + bool new_x_43_interleave_0 = const()[name = string("new_x_43_interleave_0"), val = bool(false)]; + tensor new_x_43_cast_fp16 = concat(axis = var_58, interleave = new_x_43_interleave_0, values = (cache_43_cast_fp16, input_573_cast_fp16))[name = string("new_x_43_cast_fp16")]; + tensor var_2717_begin_0 = const()[name = string("op_2717_begin_0"), val = tensor([0, 0, 56])]; + tensor var_2717_end_0 = const()[name = string("op_2717_end_0"), val = tensor([1, 1024, 64])]; + tensor var_2717_end_mask_0 = const()[name = string("op_2717_end_mask_0"), val = tensor([true, true, true])]; + tensor var_2717_cast_fp16 = slice_by_index(begin = var_2717_begin_0, end = var_2717_end_0, end_mask = var_2717_end_mask_0, x = new_x_43_cast_fp16)[name = string("op_2717_cast_fp16")]; + string x_281_pad_type_0 = const()[name = string("x_281_pad_type_0"), val = string("valid")]; + int32 x_281_groups_0 = const()[name = string("x_281_groups_0"), val = int32(1024)]; + tensor x_281_strides_0 = const()[name = string("x_281_strides_0"), val = tensor([1])]; + tensor x_281_pad_0 = const()[name = string("x_281_pad_0"), val = tensor([0, 0])]; + tensor x_281_dilations_0 = const()[name = string("x_281_dilations_0"), val = tensor([1])]; + tensor encoder_layers_10_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221358144))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221367424))))[name = string("encoder_layers_10_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_281_cast_fp16 = conv(dilations = x_281_dilations_0, groups = x_281_groups_0, pad = x_281_pad_0, pad_type = x_281_pad_type_0, strides = x_281_strides_0, weight = encoder_layers_10_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_43_cast_fp16)[name = string("x_281_cast_fp16")]; + tensor input_575_perm_0 = const()[name = string("input_575_perm_0"), val = tensor([0, 2, 1])]; + tensor x_283_axes_0 = const()[name = string("x_283_axes_0"), val = tensor([-1])]; + tensor encoder_layers_10_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_10_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221369536)))]; + tensor encoder_layers_10_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_10_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221371648)))]; + tensor input_575_cast_fp16 = transpose(perm = input_575_perm_0, x = x_281_cast_fp16)[name = string("transpose_266")]; + tensor x_283_cast_fp16 = layer_norm(axes = x_283_axes_0, beta = encoder_layers_10_conv_batch_norm_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_10_conv_batch_norm_weight_to_fp16, x = input_575_cast_fp16)[name = string("x_283_cast_fp16")]; + tensor input_577_perm_0 = const()[name = string("input_577_perm_0"), val = tensor([0, 2, 1])]; + tensor input_577_cast_fp16 = transpose(perm = input_577_perm_0, x = x_283_cast_fp16)[name = string("transpose_265")]; + tensor input_579_cast_fp16 = silu(x = input_577_cast_fp16)[name = string("input_579_cast_fp16")]; + string x_285_pad_type_0 = const()[name = string("x_285_pad_type_0"), val = string("valid")]; + tensor x_285_strides_0 = const()[name = string("x_285_strides_0"), val = tensor([1])]; + tensor x_285_pad_0 = const()[name = string("x_285_pad_0"), val = tensor([0, 0])]; + tensor x_285_dilations_0 = const()[name = string("x_285_dilations_0"), val = tensor([1])]; + int32 x_285_groups_0 = const()[name = string("x_285_groups_0"), val = int32(1)]; + tensor encoder_layers_10_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221373760))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(222422400))))[name = string("encoder_layers_10_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_285_cast_fp16 = conv(dilations = x_285_dilations_0, groups = x_285_groups_0, pad = x_285_pad_0, pad_type = x_285_pad_type_0, strides = x_285_strides_0, weight = encoder_layers_10_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_579_cast_fp16)[name = string("x_285_cast_fp16")]; + tensor input_581_perm_0 = const()[name = string("input_581_perm_0"), val = tensor([0, 2, 1])]; + tensor input_581_cast_fp16 = transpose(perm = input_581_perm_0, x = x_285_cast_fp16)[name = string("transpose_264")]; + tensor input_583_cast_fp16 = add(x = input_567_cast_fp16, y = input_581_cast_fp16)[name = string("input_583_cast_fp16")]; + tensor input_585_axes_0 = const()[name = string("input_585_axes_0"), val = tensor([-1])]; + tensor encoder_layers_10_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_10_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(222424512)))]; + tensor encoder_layers_10_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_10_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(222426624)))]; + tensor input_585_cast_fp16 = layer_norm(axes = input_585_axes_0, beta = encoder_layers_10_norm_feed_forward2_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_10_norm_feed_forward2_weight_to_fp16, x = input_583_cast_fp16)[name = string("input_585_cast_fp16")]; + tensor encoder_layers_10_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(222428736))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(225574528))))[name = string("encoder_layers_10_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_10_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_10_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(225574720)))]; + tensor linear_98_cast_fp16 = linear(bias = encoder_layers_10_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_10_feed_forward2_linear1_weight_to_fp16_palettized, x = input_585_cast_fp16)[name = string("linear_98_cast_fp16")]; + tensor input_589_cast_fp16 = silu(x = linear_98_cast_fp16)[name = string("input_589_cast_fp16")]; + tensor encoder_layers_10_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(225582976))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(228728768))))[name = string("encoder_layers_10_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_10_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_10_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(228728960)))]; + tensor linear_99_cast_fp16 = linear(bias = encoder_layers_10_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_10_feed_forward2_linear2_weight_to_fp16_palettized, x = input_589_cast_fp16)[name = string("linear_99_cast_fp16")]; + fp16 var_2760_to_fp16 = const()[name = string("op_2760_to_fp16"), val = fp16(0x1p-1)]; + tensor var_2761_cast_fp16 = mul(x = linear_99_cast_fp16, y = var_2760_to_fp16)[name = string("op_2761_cast_fp16")]; + tensor input_595_cast_fp16 = add(x = input_583_cast_fp16, y = var_2761_cast_fp16)[name = string("input_595_cast_fp16")]; + tensor input_597_axes_0 = const()[name = string("input_597_axes_0"), val = tensor([-1])]; + tensor encoder_layers_10_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_10_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(228731072)))]; + tensor encoder_layers_10_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_10_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(228733184)))]; + tensor input_597_cast_fp16 = layer_norm(axes = input_597_axes_0, beta = encoder_layers_10_norm_out_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_10_norm_out_weight_to_fp16, x = input_595_cast_fp16)[name = string("input_597_cast_fp16")]; + tensor cache_45_begin_0 = const()[name = string("cache_45_begin_0"), val = tensor([11, 0, 0, 0])]; + tensor cache_45_end_0 = const()[name = string("cache_45_end_0"), val = tensor([12, 1, 42, 1024])]; + tensor cache_45_end_mask_0 = const()[name = string("cache_45_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_45_squeeze_mask_0 = const()[name = string("cache_45_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_45_cast_fp16 = slice_by_index(begin = cache_45_begin_0, end = cache_45_end_0, end_mask = cache_45_end_mask_0, squeeze_mask = cache_45_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_45_cast_fp16")]; + tensor cache_47_begin_0 = const()[name = string("cache_47_begin_0"), val = tensor([11, 0, 0, 0])]; + tensor cache_47_end_0 = const()[name = string("cache_47_end_0"), val = tensor([12, 1, 1024, 8])]; + tensor cache_47_end_mask_0 = const()[name = string("cache_47_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_47_squeeze_mask_0 = const()[name = string("cache_47_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_47_cast_fp16 = slice_by_index(begin = cache_47_begin_0, end = cache_47_end_0, end_mask = cache_47_end_mask_0, squeeze_mask = cache_47_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_47_cast_fp16")]; + tensor input_599_axes_0 = const()[name = string("input_599_axes_0"), val = tensor([-1])]; + tensor encoder_layers_11_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_11_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(228735296)))]; + tensor encoder_layers_11_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_11_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(228737408)))]; + tensor input_599_cast_fp16 = layer_norm(axes = input_599_axes_0, beta = encoder_layers_11_norm_feed_forward1_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_11_norm_feed_forward1_weight_to_fp16, x = input_597_cast_fp16)[name = string("input_599_cast_fp16")]; + tensor encoder_layers_11_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(228739520))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(231885312))))[name = string("encoder_layers_11_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_11_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_11_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(231885504)))]; + tensor linear_100_cast_fp16 = linear(bias = encoder_layers_11_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_11_feed_forward1_linear1_weight_to_fp16_palettized, x = input_599_cast_fp16)[name = string("linear_100_cast_fp16")]; + tensor input_603_cast_fp16 = silu(x = linear_100_cast_fp16)[name = string("input_603_cast_fp16")]; + tensor encoder_layers_11_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(231893760))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(235039552))))[name = string("encoder_layers_11_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_11_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_11_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(235039744)))]; + tensor linear_101_cast_fp16 = linear(bias = encoder_layers_11_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_11_feed_forward1_linear2_weight_to_fp16_palettized, x = input_603_cast_fp16)[name = string("linear_101_cast_fp16")]; + fp16 var_2797_to_fp16 = const()[name = string("op_2797_to_fp16"), val = fp16(0x1p-1)]; + tensor var_2798_cast_fp16 = mul(x = linear_101_cast_fp16, y = var_2797_to_fp16)[name = string("op_2798_cast_fp16")]; + tensor input_609_cast_fp16 = add(x = input_597_cast_fp16, y = var_2798_cast_fp16)[name = string("input_609_cast_fp16")]; + tensor key_23_axes_0 = const()[name = string("key_23_axes_0"), val = tensor([-1])]; + tensor encoder_layers_11_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_11_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(235041856)))]; + tensor encoder_layers_11_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_11_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(235043968)))]; + tensor key_23_cast_fp16 = layer_norm(axes = key_23_axes_0, beta = encoder_layers_11_norm_self_att_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_11_norm_self_att_weight_to_fp16, x = input_609_cast_fp16)[name = string("key_23_cast_fp16")]; + bool input_611_interleave_0 = const()[name = string("input_611_interleave_0"), val = bool(false)]; + tensor input_611_cast_fp16 = concat(axis = var_67, interleave = input_611_interleave_0, values = (cache_45_cast_fp16, key_23_cast_fp16))[name = string("input_611_cast_fp16")]; + bool var_2826_interleave_0 = const()[name = string("op_2826_interleave_0"), val = bool(false)]; + tensor var_2826_cast_fp16 = concat(axis = var_67, interleave = var_2826_interleave_0, values = key_23_cast_fp16)[name = string("op_2826_cast_fp16")]; + tensor encoder_layers_11_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(235046080))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(235832576))))[name = string("encoder_layers_11_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_11_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_11_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(235832768)))]; + tensor linear_102_cast_fp16 = linear(bias = encoder_layers_11_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_11_self_attn_linear_q_weight_to_fp16_palettized, x = key_23_cast_fp16)[name = string("linear_102_cast_fp16")]; + tensor var_2831 = const()[name = string("op_2831"), val = tensor([1, -1, 8, 128])]; + tensor q_67_cast_fp16 = reshape(shape = var_2831, x = linear_102_cast_fp16)[name = string("q_67_cast_fp16")]; + tensor encoder_layers_11_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(235834880))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236621376))))[name = string("encoder_layers_11_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_11_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_11_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236621568)))]; + tensor linear_103_cast_fp16 = linear(bias = encoder_layers_11_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_11_self_attn_linear_k_weight_to_fp16_palettized, x = input_611_cast_fp16)[name = string("linear_103_cast_fp16")]; + tensor var_2836 = const()[name = string("op_2836"), val = tensor([1, -1, 8, 128])]; + tensor k_45_cast_fp16 = reshape(shape = var_2836, x = linear_103_cast_fp16)[name = string("k_45_cast_fp16")]; + tensor encoder_layers_11_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236623680))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(237410176))))[name = string("encoder_layers_11_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_11_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_11_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(237410368)))]; + tensor linear_104_cast_fp16 = linear(bias = encoder_layers_11_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_11_self_attn_linear_v_weight_to_fp16_palettized, x = input_611_cast_fp16)[name = string("linear_104_cast_fp16")]; + tensor var_2841 = const()[name = string("op_2841"), val = tensor([1, -1, 8, 128])]; + tensor v_23_cast_fp16 = reshape(shape = var_2841, x = linear_104_cast_fp16)[name = string("v_23_cast_fp16")]; + tensor value_31_perm_0 = const()[name = string("value_31_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_11_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_11_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(237412480)))]; + tensor var_2854_cast_fp16 = add(x = q_67_cast_fp16, y = encoder_layers_11_self_attn_pos_bias_u_to_fp16)[name = string("op_2854_cast_fp16")]; + tensor encoder_layers_11_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_11_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(237414592)))]; + tensor var_2856_cast_fp16 = add(x = q_67_cast_fp16, y = encoder_layers_11_self_attn_pos_bias_v_to_fp16)[name = string("op_2856_cast_fp16")]; + tensor q_with_bias_v_23_perm_0 = const()[name = string("q_with_bias_v_23_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_293_transpose_x_0 = const()[name = string("x_293_transpose_x_0"), val = bool(false)]; + bool x_293_transpose_y_0 = const()[name = string("x_293_transpose_y_0"), val = bool(false)]; + tensor op_2858_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(237416704))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(237616448))))[name = string("op_2858_to_fp16_quantized")]; + tensor q_with_bias_v_23_cast_fp16 = transpose(perm = q_with_bias_v_23_perm_0, x = var_2856_cast_fp16)[name = string("transpose_263")]; + tensor x_293_cast_fp16 = matmul(transpose_x = x_293_transpose_x_0, transpose_y = x_293_transpose_y_0, x = q_with_bias_v_23_cast_fp16, y = op_2858_to_fp16_quantized)[name = string("x_293_cast_fp16")]; + tensor x_295_pad_0 = const()[name = string("x_295_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_295_mode_0 = const()[name = string("x_295_mode_0"), val = string("constant")]; + fp16 const_222_to_fp16 = const()[name = string("const_222_to_fp16"), val = fp16(0x0p+0)]; + tensor x_295_cast_fp16 = pad(constant_val = const_222_to_fp16, mode = x_295_mode_0, pad = x_295_pad_0, x = x_293_cast_fp16)[name = string("x_295_cast_fp16")]; + tensor var_2866 = const()[name = string("op_2866"), val = tensor([1, 8, -1, 56])]; + tensor x_297_cast_fp16 = reshape(shape = var_2866, x = x_295_cast_fp16)[name = string("x_297_cast_fp16")]; + tensor var_2870_begin_0 = const()[name = string("op_2870_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_2870_end_0 = const()[name = string("op_2870_end_0"), val = tensor([1, 8, 196, 56])]; + tensor var_2870_end_mask_0 = const()[name = string("op_2870_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_2870_cast_fp16 = slice_by_index(begin = var_2870_begin_0, end = var_2870_end_0, end_mask = var_2870_end_mask_0, x = x_297_cast_fp16)[name = string("op_2870_cast_fp16")]; + tensor var_2871 = const()[name = string("op_2871"), val = tensor([1, 8, 56, 195])]; + tensor matrix_bd_45_cast_fp16 = reshape(shape = var_2871, x = var_2870_cast_fp16)[name = string("matrix_bd_45_cast_fp16")]; + bool matrix_ac_23_transpose_x_0 = const()[name = string("matrix_ac_23_transpose_x_0"), val = bool(false)]; + bool matrix_ac_23_transpose_y_0 = const()[name = string("matrix_ac_23_transpose_y_0"), val = bool(false)]; + tensor transpose_118_perm_0 = const()[name = string("transpose_118_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_119_perm_0 = const()[name = string("transpose_119_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_119 = transpose(perm = transpose_119_perm_0, x = k_45_cast_fp16)[name = string("transpose_261")]; + tensor transpose_118 = transpose(perm = transpose_118_perm_0, x = var_2854_cast_fp16)[name = string("transpose_262")]; + tensor matrix_ac_23_cast_fp16 = matmul(transpose_x = matrix_ac_23_transpose_x_0, transpose_y = matrix_ac_23_transpose_y_0, x = transpose_118, y = transpose_119)[name = string("matrix_ac_23_cast_fp16")]; + tensor matrix_bd_47_begin_0 = const()[name = string("matrix_bd_47_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_47_end_0 = const()[name = string("matrix_bd_47_end_0"), val = tensor([1, 8, 56, 98])]; + tensor matrix_bd_47_end_mask_0 = const()[name = string("matrix_bd_47_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_47_cast_fp16 = slice_by_index(begin = matrix_bd_47_begin_0, end = matrix_bd_47_end_0, end_mask = matrix_bd_47_end_mask_0, x = matrix_bd_45_cast_fp16)[name = string("matrix_bd_47_cast_fp16")]; + tensor var_2880_cast_fp16 = add(x = matrix_ac_23_cast_fp16, y = matrix_bd_47_cast_fp16)[name = string("op_2880_cast_fp16")]; + fp16 _inversed_scores_45_y_0_to_fp16 = const()[name = string("_inversed_scores_45_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_45_cast_fp16 = mul(x = var_2880_cast_fp16, y = _inversed_scores_45_y_0_to_fp16)[name = string("_inversed_scores_45_cast_fp16")]; + tensor scores_47_cast_fp16 = select(a = var_44_to_fp16, b = _inversed_scores_45_cast_fp16, cond = mask_11)[name = string("scores_47_cast_fp16")]; + tensor var_2886_cast_fp16 = softmax(axis = var_58, x = scores_47_cast_fp16)[name = string("op_2886_cast_fp16")]; + tensor input_613_cast_fp16 = select(a = var_43_to_fp16, b = var_2886_cast_fp16, cond = mask_11)[name = string("input_613_cast_fp16")]; + bool x_299_transpose_x_0 = const()[name = string("x_299_transpose_x_0"), val = bool(false)]; + bool x_299_transpose_y_0 = const()[name = string("x_299_transpose_y_0"), val = bool(false)]; + tensor value_31_cast_fp16 = transpose(perm = value_31_perm_0, x = v_23_cast_fp16)[name = string("transpose_260")]; + tensor x_299_cast_fp16 = matmul(transpose_x = x_299_transpose_x_0, transpose_y = x_299_transpose_y_0, x = input_613_cast_fp16, y = value_31_cast_fp16)[name = string("x_299_cast_fp16")]; + tensor var_2890_perm_0 = const()[name = string("op_2890_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_2891 = const()[name = string("op_2891"), val = tensor([1, -1, 1024])]; + tensor var_2890_cast_fp16 = transpose(perm = var_2890_perm_0, x = x_299_cast_fp16)[name = string("transpose_259")]; + tensor input_615_cast_fp16 = reshape(shape = var_2891, x = var_2890_cast_fp16)[name = string("input_615_cast_fp16")]; + tensor encoder_layers_11_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(237616960))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(238403456))))[name = string("encoder_layers_11_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_11_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_11_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(238403648)))]; + tensor linear_106_cast_fp16 = linear(bias = encoder_layers_11_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_11_self_attn_linear_out_weight_to_fp16_palettized, x = input_615_cast_fp16)[name = string("linear_106_cast_fp16")]; + tensor input_619_cast_fp16 = add(x = input_609_cast_fp16, y = linear_106_cast_fp16)[name = string("input_619_cast_fp16")]; + tensor x_303_axes_0 = const()[name = string("x_303_axes_0"), val = tensor([-1])]; + tensor encoder_layers_11_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_11_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(238405760)))]; + tensor encoder_layers_11_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_11_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(238407872)))]; + tensor x_303_cast_fp16 = layer_norm(axes = x_303_axes_0, beta = encoder_layers_11_norm_conv_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_11_norm_conv_weight_to_fp16, x = input_619_cast_fp16)[name = string("x_303_cast_fp16")]; + tensor input_621_perm_0 = const()[name = string("input_621_perm_0"), val = tensor([0, 2, 1])]; + string input_623_pad_type_0 = const()[name = string("input_623_pad_type_0"), val = string("valid")]; + tensor input_623_strides_0 = const()[name = string("input_623_strides_0"), val = tensor([1])]; + tensor input_623_pad_0 = const()[name = string("input_623_pad_0"), val = tensor([0, 0])]; + tensor input_623_dilations_0 = const()[name = string("input_623_dilations_0"), val = tensor([1])]; + int32 input_623_groups_0 = const()[name = string("input_623_groups_0"), val = int32(1)]; + tensor encoder_layers_11_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(238409984))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240507200))))[name = string("encoder_layers_11_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_621_cast_fp16 = transpose(perm = input_621_perm_0, x = x_303_cast_fp16)[name = string("transpose_258")]; + tensor input_623_cast_fp16 = conv(dilations = input_623_dilations_0, groups = input_623_groups_0, pad = input_623_pad_0, pad_type = input_623_pad_type_0, strides = input_623_strides_0, weight = encoder_layers_11_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_621_cast_fp16)[name = string("input_623_cast_fp16")]; + int32 x_305_split_num_splits_0 = const()[name = string("x_305_split_num_splits_0"), val = int32(2)]; + int32 x_305_split_axis_0 = const()[name = string("x_305_split_axis_0"), val = int32(1)]; + tensor x_305_split_cast_fp16_0, tensor x_305_split_cast_fp16_1 = split(axis = x_305_split_axis_0, num_splits = x_305_split_num_splits_0, x = input_623_cast_fp16)[name = string("x_305_split_cast_fp16")]; + tensor x_305_split_1_sigmoid_cast_fp16 = sigmoid(x = x_305_split_cast_fp16_1)[name = string("x_305_split_1_sigmoid_cast_fp16")]; + tensor x_305_cast_fp16 = mul(x = x_305_split_cast_fp16_0, y = x_305_split_1_sigmoid_cast_fp16)[name = string("x_305_cast_fp16")]; + tensor input_625_cast_fp16 = select(a = var_43_to_fp16, b = x_305_cast_fp16, cond = var_574)[name = string("input_625_cast_fp16")]; + bool new_x_47_interleave_0 = const()[name = string("new_x_47_interleave_0"), val = bool(false)]; + tensor new_x_47_cast_fp16 = concat(axis = var_58, interleave = new_x_47_interleave_0, values = (cache_47_cast_fp16, input_625_cast_fp16))[name = string("new_x_47_cast_fp16")]; + tensor var_2930_begin_0 = const()[name = string("op_2930_begin_0"), val = tensor([0, 0, 56])]; + tensor var_2930_end_0 = const()[name = string("op_2930_end_0"), val = tensor([1, 1024, 64])]; + tensor var_2930_end_mask_0 = const()[name = string("op_2930_end_mask_0"), val = tensor([true, true, true])]; + tensor var_2930_cast_fp16 = slice_by_index(begin = var_2930_begin_0, end = var_2930_end_0, end_mask = var_2930_end_mask_0, x = new_x_47_cast_fp16)[name = string("op_2930_cast_fp16")]; + string x_307_pad_type_0 = const()[name = string("x_307_pad_type_0"), val = string("valid")]; + int32 x_307_groups_0 = const()[name = string("x_307_groups_0"), val = int32(1024)]; + tensor x_307_strides_0 = const()[name = string("x_307_strides_0"), val = tensor([1])]; + tensor x_307_pad_0 = const()[name = string("x_307_pad_0"), val = tensor([0, 0])]; + tensor x_307_dilations_0 = const()[name = string("x_307_dilations_0"), val = tensor([1])]; + tensor encoder_layers_11_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240511360))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240520640))))[name = string("encoder_layers_11_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_307_cast_fp16 = conv(dilations = x_307_dilations_0, groups = x_307_groups_0, pad = x_307_pad_0, pad_type = x_307_pad_type_0, strides = x_307_strides_0, weight = encoder_layers_11_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_47_cast_fp16)[name = string("x_307_cast_fp16")]; + tensor input_627_perm_0 = const()[name = string("input_627_perm_0"), val = tensor([0, 2, 1])]; + tensor x_309_axes_0 = const()[name = string("x_309_axes_0"), val = tensor([-1])]; + tensor encoder_layers_11_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_11_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240522752)))]; + tensor encoder_layers_11_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_11_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240524864)))]; + tensor input_627_cast_fp16 = transpose(perm = input_627_perm_0, x = x_307_cast_fp16)[name = string("transpose_257")]; + tensor x_309_cast_fp16 = layer_norm(axes = x_309_axes_0, beta = encoder_layers_11_conv_batch_norm_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_11_conv_batch_norm_weight_to_fp16, x = input_627_cast_fp16)[name = string("x_309_cast_fp16")]; + tensor input_629_perm_0 = const()[name = string("input_629_perm_0"), val = tensor([0, 2, 1])]; + tensor input_629_cast_fp16 = transpose(perm = input_629_perm_0, x = x_309_cast_fp16)[name = string("transpose_256")]; + tensor input_631_cast_fp16 = silu(x = input_629_cast_fp16)[name = string("input_631_cast_fp16")]; + string x_311_pad_type_0 = const()[name = string("x_311_pad_type_0"), val = string("valid")]; + tensor x_311_strides_0 = const()[name = string("x_311_strides_0"), val = tensor([1])]; + tensor x_311_pad_0 = const()[name = string("x_311_pad_0"), val = tensor([0, 0])]; + tensor x_311_dilations_0 = const()[name = string("x_311_dilations_0"), val = tensor([1])]; + int32 x_311_groups_0 = const()[name = string("x_311_groups_0"), val = int32(1)]; + tensor encoder_layers_11_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240526976))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241575616))))[name = string("encoder_layers_11_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_311_cast_fp16 = conv(dilations = x_311_dilations_0, groups = x_311_groups_0, pad = x_311_pad_0, pad_type = x_311_pad_type_0, strides = x_311_strides_0, weight = encoder_layers_11_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_631_cast_fp16)[name = string("x_311_cast_fp16")]; + tensor input_633_perm_0 = const()[name = string("input_633_perm_0"), val = tensor([0, 2, 1])]; + tensor input_633_cast_fp16 = transpose(perm = input_633_perm_0, x = x_311_cast_fp16)[name = string("transpose_255")]; + tensor input_635_cast_fp16 = add(x = input_619_cast_fp16, y = input_633_cast_fp16)[name = string("input_635_cast_fp16")]; + tensor input_637_axes_0 = const()[name = string("input_637_axes_0"), val = tensor([-1])]; + tensor encoder_layers_11_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_11_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241577728)))]; + tensor encoder_layers_11_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_11_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241579840)))]; + tensor input_637_cast_fp16 = layer_norm(axes = input_637_axes_0, beta = encoder_layers_11_norm_feed_forward2_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_11_norm_feed_forward2_weight_to_fp16, x = input_635_cast_fp16)[name = string("input_637_cast_fp16")]; + tensor encoder_layers_11_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241581952))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(244727744))))[name = string("encoder_layers_11_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_11_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_11_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(244727936)))]; + tensor linear_107_cast_fp16 = linear(bias = encoder_layers_11_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_11_feed_forward2_linear1_weight_to_fp16_palettized, x = input_637_cast_fp16)[name = string("linear_107_cast_fp16")]; + tensor input_641_cast_fp16 = silu(x = linear_107_cast_fp16)[name = string("input_641_cast_fp16")]; + tensor encoder_layers_11_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(244736192))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(247881984))))[name = string("encoder_layers_11_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_11_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_11_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(247882176)))]; + tensor linear_108_cast_fp16 = linear(bias = encoder_layers_11_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_11_feed_forward2_linear2_weight_to_fp16_palettized, x = input_641_cast_fp16)[name = string("linear_108_cast_fp16")]; + fp16 var_2973_to_fp16 = const()[name = string("op_2973_to_fp16"), val = fp16(0x1p-1)]; + tensor var_2974_cast_fp16 = mul(x = linear_108_cast_fp16, y = var_2973_to_fp16)[name = string("op_2974_cast_fp16")]; + tensor input_647_cast_fp16 = add(x = input_635_cast_fp16, y = var_2974_cast_fp16)[name = string("input_647_cast_fp16")]; + tensor input_649_axes_0 = const()[name = string("input_649_axes_0"), val = tensor([-1])]; + tensor encoder_layers_11_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_11_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(247884288)))]; + tensor encoder_layers_11_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_11_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(247886400)))]; + tensor input_649_cast_fp16 = layer_norm(axes = input_649_axes_0, beta = encoder_layers_11_norm_out_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_11_norm_out_weight_to_fp16, x = input_647_cast_fp16)[name = string("input_649_cast_fp16")]; + tensor cache_49_begin_0 = const()[name = string("cache_49_begin_0"), val = tensor([12, 0, 0, 0])]; + tensor cache_49_end_0 = const()[name = string("cache_49_end_0"), val = tensor([13, 1, 42, 1024])]; + tensor cache_49_end_mask_0 = const()[name = string("cache_49_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_49_squeeze_mask_0 = const()[name = string("cache_49_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_49_cast_fp16 = slice_by_index(begin = cache_49_begin_0, end = cache_49_end_0, end_mask = cache_49_end_mask_0, squeeze_mask = cache_49_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_49_cast_fp16")]; + tensor cache_51_begin_0 = const()[name = string("cache_51_begin_0"), val = tensor([12, 0, 0, 0])]; + tensor cache_51_end_0 = const()[name = string("cache_51_end_0"), val = tensor([13, 1, 1024, 8])]; + tensor cache_51_end_mask_0 = const()[name = string("cache_51_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_51_squeeze_mask_0 = const()[name = string("cache_51_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_51_cast_fp16 = slice_by_index(begin = cache_51_begin_0, end = cache_51_end_0, end_mask = cache_51_end_mask_0, squeeze_mask = cache_51_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_51_cast_fp16")]; + tensor input_651_axes_0 = const()[name = string("input_651_axes_0"), val = tensor([-1])]; + tensor encoder_layers_12_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_12_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(247888512)))]; + tensor encoder_layers_12_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_12_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(247890624)))]; + tensor input_651_cast_fp16 = layer_norm(axes = input_651_axes_0, beta = encoder_layers_12_norm_feed_forward1_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_12_norm_feed_forward1_weight_to_fp16, x = input_649_cast_fp16)[name = string("input_651_cast_fp16")]; + tensor encoder_layers_12_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(247892736))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(251038528))))[name = string("encoder_layers_12_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_12_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_12_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(251038720)))]; + tensor linear_109_cast_fp16 = linear(bias = encoder_layers_12_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_12_feed_forward1_linear1_weight_to_fp16_palettized, x = input_651_cast_fp16)[name = string("linear_109_cast_fp16")]; + tensor input_655_cast_fp16 = silu(x = linear_109_cast_fp16)[name = string("input_655_cast_fp16")]; + tensor encoder_layers_12_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(251046976))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(254192768))))[name = string("encoder_layers_12_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_12_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_12_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(254192960)))]; + tensor linear_110_cast_fp16 = linear(bias = encoder_layers_12_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_12_feed_forward1_linear2_weight_to_fp16_palettized, x = input_655_cast_fp16)[name = string("linear_110_cast_fp16")]; + fp16 var_3010_to_fp16 = const()[name = string("op_3010_to_fp16"), val = fp16(0x1p-1)]; + tensor var_3011_cast_fp16 = mul(x = linear_110_cast_fp16, y = var_3010_to_fp16)[name = string("op_3011_cast_fp16")]; + tensor input_661_cast_fp16 = add(x = input_649_cast_fp16, y = var_3011_cast_fp16)[name = string("input_661_cast_fp16")]; + tensor key_25_axes_0 = const()[name = string("key_25_axes_0"), val = tensor([-1])]; + tensor encoder_layers_12_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_12_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(254195072)))]; + tensor encoder_layers_12_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_12_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(254197184)))]; + tensor key_25_cast_fp16 = layer_norm(axes = key_25_axes_0, beta = encoder_layers_12_norm_self_att_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_12_norm_self_att_weight_to_fp16, x = input_661_cast_fp16)[name = string("key_25_cast_fp16")]; + bool input_663_interleave_0 = const()[name = string("input_663_interleave_0"), val = bool(false)]; + tensor input_663_cast_fp16 = concat(axis = var_67, interleave = input_663_interleave_0, values = (cache_49_cast_fp16, key_25_cast_fp16))[name = string("input_663_cast_fp16")]; + bool var_3039_interleave_0 = const()[name = string("op_3039_interleave_0"), val = bool(false)]; + tensor var_3039_cast_fp16 = concat(axis = var_67, interleave = var_3039_interleave_0, values = key_25_cast_fp16)[name = string("op_3039_cast_fp16")]; + tensor encoder_layers_12_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(254199296))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(254985792))))[name = string("encoder_layers_12_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_12_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_12_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(254985984)))]; + tensor linear_111_cast_fp16 = linear(bias = encoder_layers_12_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_12_self_attn_linear_q_weight_to_fp16_palettized, x = key_25_cast_fp16)[name = string("linear_111_cast_fp16")]; + tensor var_3044 = const()[name = string("op_3044"), val = tensor([1, -1, 8, 128])]; + tensor q_73_cast_fp16 = reshape(shape = var_3044, x = linear_111_cast_fp16)[name = string("q_73_cast_fp16")]; + tensor encoder_layers_12_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(254988096))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255774592))))[name = string("encoder_layers_12_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_12_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_12_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255774784)))]; + tensor linear_112_cast_fp16 = linear(bias = encoder_layers_12_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_12_self_attn_linear_k_weight_to_fp16_palettized, x = input_663_cast_fp16)[name = string("linear_112_cast_fp16")]; + tensor var_3049 = const()[name = string("op_3049"), val = tensor([1, -1, 8, 128])]; + tensor k_49_cast_fp16 = reshape(shape = var_3049, x = linear_112_cast_fp16)[name = string("k_49_cast_fp16")]; + tensor encoder_layers_12_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255776896))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256563392))))[name = string("encoder_layers_12_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_12_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_12_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256563584)))]; + tensor linear_113_cast_fp16 = linear(bias = encoder_layers_12_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_12_self_attn_linear_v_weight_to_fp16_palettized, x = input_663_cast_fp16)[name = string("linear_113_cast_fp16")]; + tensor var_3054 = const()[name = string("op_3054"), val = tensor([1, -1, 8, 128])]; + tensor v_25_cast_fp16 = reshape(shape = var_3054, x = linear_113_cast_fp16)[name = string("v_25_cast_fp16")]; + tensor value_33_perm_0 = const()[name = string("value_33_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_12_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_12_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256565696)))]; + tensor var_3067_cast_fp16 = add(x = q_73_cast_fp16, y = encoder_layers_12_self_attn_pos_bias_u_to_fp16)[name = string("op_3067_cast_fp16")]; + tensor encoder_layers_12_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_12_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256567808)))]; + tensor var_3069_cast_fp16 = add(x = q_73_cast_fp16, y = encoder_layers_12_self_attn_pos_bias_v_to_fp16)[name = string("op_3069_cast_fp16")]; + tensor q_with_bias_v_25_perm_0 = const()[name = string("q_with_bias_v_25_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_319_transpose_x_0 = const()[name = string("x_319_transpose_x_0"), val = bool(false)]; + bool x_319_transpose_y_0 = const()[name = string("x_319_transpose_y_0"), val = bool(false)]; + tensor op_3071_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256569920))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256769664))))[name = string("op_3071_to_fp16_quantized")]; + tensor q_with_bias_v_25_cast_fp16 = transpose(perm = q_with_bias_v_25_perm_0, x = var_3069_cast_fp16)[name = string("transpose_254")]; + tensor x_319_cast_fp16 = matmul(transpose_x = x_319_transpose_x_0, transpose_y = x_319_transpose_y_0, x = q_with_bias_v_25_cast_fp16, y = op_3071_to_fp16_quantized)[name = string("x_319_cast_fp16")]; + tensor x_321_pad_0 = const()[name = string("x_321_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_321_mode_0 = const()[name = string("x_321_mode_0"), val = string("constant")]; + fp16 const_235_to_fp16 = const()[name = string("const_235_to_fp16"), val = fp16(0x0p+0)]; + tensor x_321_cast_fp16 = pad(constant_val = const_235_to_fp16, mode = x_321_mode_0, pad = x_321_pad_0, x = x_319_cast_fp16)[name = string("x_321_cast_fp16")]; + tensor var_3079 = const()[name = string("op_3079"), val = tensor([1, 8, -1, 56])]; + tensor x_323_cast_fp16 = reshape(shape = var_3079, x = x_321_cast_fp16)[name = string("x_323_cast_fp16")]; + tensor var_3083_begin_0 = const()[name = string("op_3083_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_3083_end_0 = const()[name = string("op_3083_end_0"), val = tensor([1, 8, 196, 56])]; + tensor var_3083_end_mask_0 = const()[name = string("op_3083_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_3083_cast_fp16 = slice_by_index(begin = var_3083_begin_0, end = var_3083_end_0, end_mask = var_3083_end_mask_0, x = x_323_cast_fp16)[name = string("op_3083_cast_fp16")]; + tensor var_3084 = const()[name = string("op_3084"), val = tensor([1, 8, 56, 195])]; + tensor matrix_bd_49_cast_fp16 = reshape(shape = var_3084, x = var_3083_cast_fp16)[name = string("matrix_bd_49_cast_fp16")]; + bool matrix_ac_25_transpose_x_0 = const()[name = string("matrix_ac_25_transpose_x_0"), val = bool(false)]; + bool matrix_ac_25_transpose_y_0 = const()[name = string("matrix_ac_25_transpose_y_0"), val = bool(false)]; + tensor transpose_120_perm_0 = const()[name = string("transpose_120_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_121_perm_0 = const()[name = string("transpose_121_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_121 = transpose(perm = transpose_121_perm_0, x = k_49_cast_fp16)[name = string("transpose_252")]; + tensor transpose_120 = transpose(perm = transpose_120_perm_0, x = var_3067_cast_fp16)[name = string("transpose_253")]; + tensor matrix_ac_25_cast_fp16 = matmul(transpose_x = matrix_ac_25_transpose_x_0, transpose_y = matrix_ac_25_transpose_y_0, x = transpose_120, y = transpose_121)[name = string("matrix_ac_25_cast_fp16")]; + tensor matrix_bd_51_begin_0 = const()[name = string("matrix_bd_51_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_51_end_0 = const()[name = string("matrix_bd_51_end_0"), val = tensor([1, 8, 56, 98])]; + tensor matrix_bd_51_end_mask_0 = const()[name = string("matrix_bd_51_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_51_cast_fp16 = slice_by_index(begin = matrix_bd_51_begin_0, end = matrix_bd_51_end_0, end_mask = matrix_bd_51_end_mask_0, x = matrix_bd_49_cast_fp16)[name = string("matrix_bd_51_cast_fp16")]; + tensor var_3093_cast_fp16 = add(x = matrix_ac_25_cast_fp16, y = matrix_bd_51_cast_fp16)[name = string("op_3093_cast_fp16")]; + fp16 _inversed_scores_49_y_0_to_fp16 = const()[name = string("_inversed_scores_49_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_49_cast_fp16 = mul(x = var_3093_cast_fp16, y = _inversed_scores_49_y_0_to_fp16)[name = string("_inversed_scores_49_cast_fp16")]; + tensor scores_51_cast_fp16 = select(a = var_44_to_fp16, b = _inversed_scores_49_cast_fp16, cond = mask_11)[name = string("scores_51_cast_fp16")]; + tensor var_3099_cast_fp16 = softmax(axis = var_58, x = scores_51_cast_fp16)[name = string("op_3099_cast_fp16")]; + tensor input_665_cast_fp16 = select(a = var_43_to_fp16, b = var_3099_cast_fp16, cond = mask_11)[name = string("input_665_cast_fp16")]; + bool x_325_transpose_x_0 = const()[name = string("x_325_transpose_x_0"), val = bool(false)]; + bool x_325_transpose_y_0 = const()[name = string("x_325_transpose_y_0"), val = bool(false)]; + tensor value_33_cast_fp16 = transpose(perm = value_33_perm_0, x = v_25_cast_fp16)[name = string("transpose_251")]; + tensor x_325_cast_fp16 = matmul(transpose_x = x_325_transpose_x_0, transpose_y = x_325_transpose_y_0, x = input_665_cast_fp16, y = value_33_cast_fp16)[name = string("x_325_cast_fp16")]; + tensor var_3103_perm_0 = const()[name = string("op_3103_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_3104 = const()[name = string("op_3104"), val = tensor([1, -1, 1024])]; + tensor var_3103_cast_fp16 = transpose(perm = var_3103_perm_0, x = x_325_cast_fp16)[name = string("transpose_250")]; + tensor input_667_cast_fp16 = reshape(shape = var_3104, x = var_3103_cast_fp16)[name = string("input_667_cast_fp16")]; + tensor encoder_layers_12_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256770176))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257556672))))[name = string("encoder_layers_12_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_12_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_12_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257556864)))]; + tensor linear_115_cast_fp16 = linear(bias = encoder_layers_12_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_12_self_attn_linear_out_weight_to_fp16_palettized, x = input_667_cast_fp16)[name = string("linear_115_cast_fp16")]; + tensor input_671_cast_fp16 = add(x = input_661_cast_fp16, y = linear_115_cast_fp16)[name = string("input_671_cast_fp16")]; + tensor x_329_axes_0 = const()[name = string("x_329_axes_0"), val = tensor([-1])]; + tensor encoder_layers_12_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_12_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257558976)))]; + tensor encoder_layers_12_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_12_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257561088)))]; + tensor x_329_cast_fp16 = layer_norm(axes = x_329_axes_0, beta = encoder_layers_12_norm_conv_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_12_norm_conv_weight_to_fp16, x = input_671_cast_fp16)[name = string("x_329_cast_fp16")]; + tensor input_673_perm_0 = const()[name = string("input_673_perm_0"), val = tensor([0, 2, 1])]; + string input_675_pad_type_0 = const()[name = string("input_675_pad_type_0"), val = string("valid")]; + tensor input_675_strides_0 = const()[name = string("input_675_strides_0"), val = tensor([1])]; + tensor input_675_pad_0 = const()[name = string("input_675_pad_0"), val = tensor([0, 0])]; + tensor input_675_dilations_0 = const()[name = string("input_675_dilations_0"), val = tensor([1])]; + int32 input_675_groups_0 = const()[name = string("input_675_groups_0"), val = int32(1)]; + tensor encoder_layers_12_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257563200))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259660416))))[name = string("encoder_layers_12_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_673_cast_fp16 = transpose(perm = input_673_perm_0, x = x_329_cast_fp16)[name = string("transpose_249")]; + tensor input_675_cast_fp16 = conv(dilations = input_675_dilations_0, groups = input_675_groups_0, pad = input_675_pad_0, pad_type = input_675_pad_type_0, strides = input_675_strides_0, weight = encoder_layers_12_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_673_cast_fp16)[name = string("input_675_cast_fp16")]; + int32 x_331_split_num_splits_0 = const()[name = string("x_331_split_num_splits_0"), val = int32(2)]; + int32 x_331_split_axis_0 = const()[name = string("x_331_split_axis_0"), val = int32(1)]; + tensor x_331_split_cast_fp16_0, tensor x_331_split_cast_fp16_1 = split(axis = x_331_split_axis_0, num_splits = x_331_split_num_splits_0, x = input_675_cast_fp16)[name = string("x_331_split_cast_fp16")]; + tensor x_331_split_1_sigmoid_cast_fp16 = sigmoid(x = x_331_split_cast_fp16_1)[name = string("x_331_split_1_sigmoid_cast_fp16")]; + tensor x_331_cast_fp16 = mul(x = x_331_split_cast_fp16_0, y = x_331_split_1_sigmoid_cast_fp16)[name = string("x_331_cast_fp16")]; + tensor input_677_cast_fp16 = select(a = var_43_to_fp16, b = x_331_cast_fp16, cond = var_574)[name = string("input_677_cast_fp16")]; + bool new_x_51_interleave_0 = const()[name = string("new_x_51_interleave_0"), val = bool(false)]; + tensor new_x_51_cast_fp16 = concat(axis = var_58, interleave = new_x_51_interleave_0, values = (cache_51_cast_fp16, input_677_cast_fp16))[name = string("new_x_51_cast_fp16")]; + tensor var_3143_begin_0 = const()[name = string("op_3143_begin_0"), val = tensor([0, 0, 56])]; + tensor var_3143_end_0 = const()[name = string("op_3143_end_0"), val = tensor([1, 1024, 64])]; + tensor var_3143_end_mask_0 = const()[name = string("op_3143_end_mask_0"), val = tensor([true, true, true])]; + tensor var_3143_cast_fp16 = slice_by_index(begin = var_3143_begin_0, end = var_3143_end_0, end_mask = var_3143_end_mask_0, x = new_x_51_cast_fp16)[name = string("op_3143_cast_fp16")]; + string x_333_pad_type_0 = const()[name = string("x_333_pad_type_0"), val = string("valid")]; + int32 x_333_groups_0 = const()[name = string("x_333_groups_0"), val = int32(1024)]; + tensor x_333_strides_0 = const()[name = string("x_333_strides_0"), val = tensor([1])]; + tensor x_333_pad_0 = const()[name = string("x_333_pad_0"), val = tensor([0, 0])]; + tensor x_333_dilations_0 = const()[name = string("x_333_dilations_0"), val = tensor([1])]; + tensor encoder_layers_12_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259664576))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259673856))))[name = string("encoder_layers_12_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_333_cast_fp16 = conv(dilations = x_333_dilations_0, groups = x_333_groups_0, pad = x_333_pad_0, pad_type = x_333_pad_type_0, strides = x_333_strides_0, weight = encoder_layers_12_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_51_cast_fp16)[name = string("x_333_cast_fp16")]; + tensor input_679_perm_0 = const()[name = string("input_679_perm_0"), val = tensor([0, 2, 1])]; + tensor x_335_axes_0 = const()[name = string("x_335_axes_0"), val = tensor([-1])]; + tensor encoder_layers_12_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_12_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259675968)))]; + tensor encoder_layers_12_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_12_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259678080)))]; + tensor input_679_cast_fp16 = transpose(perm = input_679_perm_0, x = x_333_cast_fp16)[name = string("transpose_248")]; + tensor x_335_cast_fp16 = layer_norm(axes = x_335_axes_0, beta = encoder_layers_12_conv_batch_norm_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_12_conv_batch_norm_weight_to_fp16, x = input_679_cast_fp16)[name = string("x_335_cast_fp16")]; + tensor input_681_perm_0 = const()[name = string("input_681_perm_0"), val = tensor([0, 2, 1])]; + tensor input_681_cast_fp16 = transpose(perm = input_681_perm_0, x = x_335_cast_fp16)[name = string("transpose_247")]; + tensor input_683_cast_fp16 = silu(x = input_681_cast_fp16)[name = string("input_683_cast_fp16")]; + string x_337_pad_type_0 = const()[name = string("x_337_pad_type_0"), val = string("valid")]; + tensor x_337_strides_0 = const()[name = string("x_337_strides_0"), val = tensor([1])]; + tensor x_337_pad_0 = const()[name = string("x_337_pad_0"), val = tensor([0, 0])]; + tensor x_337_dilations_0 = const()[name = string("x_337_dilations_0"), val = tensor([1])]; + int32 x_337_groups_0 = const()[name = string("x_337_groups_0"), val = int32(1)]; + tensor encoder_layers_12_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259680192))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(260728832))))[name = string("encoder_layers_12_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_337_cast_fp16 = conv(dilations = x_337_dilations_0, groups = x_337_groups_0, pad = x_337_pad_0, pad_type = x_337_pad_type_0, strides = x_337_strides_0, weight = encoder_layers_12_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_683_cast_fp16)[name = string("x_337_cast_fp16")]; + tensor input_685_perm_0 = const()[name = string("input_685_perm_0"), val = tensor([0, 2, 1])]; + tensor input_685_cast_fp16 = transpose(perm = input_685_perm_0, x = x_337_cast_fp16)[name = string("transpose_246")]; + tensor input_687_cast_fp16 = add(x = input_671_cast_fp16, y = input_685_cast_fp16)[name = string("input_687_cast_fp16")]; + tensor input_689_axes_0 = const()[name = string("input_689_axes_0"), val = tensor([-1])]; + tensor encoder_layers_12_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_12_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(260730944)))]; + tensor encoder_layers_12_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_12_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(260733056)))]; + tensor input_689_cast_fp16 = layer_norm(axes = input_689_axes_0, beta = encoder_layers_12_norm_feed_forward2_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_12_norm_feed_forward2_weight_to_fp16, x = input_687_cast_fp16)[name = string("input_689_cast_fp16")]; + tensor encoder_layers_12_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(260735168))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(263880960))))[name = string("encoder_layers_12_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_12_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_12_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(263881152)))]; + tensor linear_116_cast_fp16 = linear(bias = encoder_layers_12_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_12_feed_forward2_linear1_weight_to_fp16_palettized, x = input_689_cast_fp16)[name = string("linear_116_cast_fp16")]; + tensor input_693_cast_fp16 = silu(x = linear_116_cast_fp16)[name = string("input_693_cast_fp16")]; + tensor encoder_layers_12_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(263889408))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(267035200))))[name = string("encoder_layers_12_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_12_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_12_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(267035392)))]; + tensor linear_117_cast_fp16 = linear(bias = encoder_layers_12_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_12_feed_forward2_linear2_weight_to_fp16_palettized, x = input_693_cast_fp16)[name = string("linear_117_cast_fp16")]; + fp16 var_3186_to_fp16 = const()[name = string("op_3186_to_fp16"), val = fp16(0x1p-1)]; + tensor var_3187_cast_fp16 = mul(x = linear_117_cast_fp16, y = var_3186_to_fp16)[name = string("op_3187_cast_fp16")]; + tensor input_699_cast_fp16 = add(x = input_687_cast_fp16, y = var_3187_cast_fp16)[name = string("input_699_cast_fp16")]; + tensor input_701_axes_0 = const()[name = string("input_701_axes_0"), val = tensor([-1])]; + tensor encoder_layers_12_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_12_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(267037504)))]; + tensor encoder_layers_12_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_12_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(267039616)))]; + tensor input_701_cast_fp16 = layer_norm(axes = input_701_axes_0, beta = encoder_layers_12_norm_out_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_12_norm_out_weight_to_fp16, x = input_699_cast_fp16)[name = string("input_701_cast_fp16")]; + tensor cache_53_begin_0 = const()[name = string("cache_53_begin_0"), val = tensor([13, 0, 0, 0])]; + tensor cache_53_end_0 = const()[name = string("cache_53_end_0"), val = tensor([14, 1, 42, 1024])]; + tensor cache_53_end_mask_0 = const()[name = string("cache_53_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_53_squeeze_mask_0 = const()[name = string("cache_53_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_53_cast_fp16 = slice_by_index(begin = cache_53_begin_0, end = cache_53_end_0, end_mask = cache_53_end_mask_0, squeeze_mask = cache_53_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_53_cast_fp16")]; + tensor cache_55_begin_0 = const()[name = string("cache_55_begin_0"), val = tensor([13, 0, 0, 0])]; + tensor cache_55_end_0 = const()[name = string("cache_55_end_0"), val = tensor([14, 1, 1024, 8])]; + tensor cache_55_end_mask_0 = const()[name = string("cache_55_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_55_squeeze_mask_0 = const()[name = string("cache_55_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_55_cast_fp16 = slice_by_index(begin = cache_55_begin_0, end = cache_55_end_0, end_mask = cache_55_end_mask_0, squeeze_mask = cache_55_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_55_cast_fp16")]; + tensor input_703_axes_0 = const()[name = string("input_703_axes_0"), val = tensor([-1])]; + tensor encoder_layers_13_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_13_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(267041728)))]; + tensor encoder_layers_13_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_13_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(267043840)))]; + tensor input_703_cast_fp16 = layer_norm(axes = input_703_axes_0, beta = encoder_layers_13_norm_feed_forward1_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_13_norm_feed_forward1_weight_to_fp16, x = input_701_cast_fp16)[name = string("input_703_cast_fp16")]; + tensor encoder_layers_13_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(267045952))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(270191744))))[name = string("encoder_layers_13_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_13_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_13_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(270191936)))]; + tensor linear_118_cast_fp16 = linear(bias = encoder_layers_13_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_13_feed_forward1_linear1_weight_to_fp16_palettized, x = input_703_cast_fp16)[name = string("linear_118_cast_fp16")]; + tensor input_707_cast_fp16 = silu(x = linear_118_cast_fp16)[name = string("input_707_cast_fp16")]; + tensor encoder_layers_13_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(270200192))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(273345984))))[name = string("encoder_layers_13_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_13_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_13_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(273346176)))]; + tensor linear_119_cast_fp16 = linear(bias = encoder_layers_13_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_13_feed_forward1_linear2_weight_to_fp16_palettized, x = input_707_cast_fp16)[name = string("linear_119_cast_fp16")]; + fp16 var_3223_to_fp16 = const()[name = string("op_3223_to_fp16"), val = fp16(0x1p-1)]; + tensor var_3224_cast_fp16 = mul(x = linear_119_cast_fp16, y = var_3223_to_fp16)[name = string("op_3224_cast_fp16")]; + tensor input_713_cast_fp16 = add(x = input_701_cast_fp16, y = var_3224_cast_fp16)[name = string("input_713_cast_fp16")]; + tensor key_27_axes_0 = const()[name = string("key_27_axes_0"), val = tensor([-1])]; + tensor encoder_layers_13_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_13_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(273348288)))]; + tensor encoder_layers_13_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_13_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(273350400)))]; + tensor key_27_cast_fp16 = layer_norm(axes = key_27_axes_0, beta = encoder_layers_13_norm_self_att_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_13_norm_self_att_weight_to_fp16, x = input_713_cast_fp16)[name = string("key_27_cast_fp16")]; + bool input_715_interleave_0 = const()[name = string("input_715_interleave_0"), val = bool(false)]; + tensor input_715_cast_fp16 = concat(axis = var_67, interleave = input_715_interleave_0, values = (cache_53_cast_fp16, key_27_cast_fp16))[name = string("input_715_cast_fp16")]; + bool var_3252_interleave_0 = const()[name = string("op_3252_interleave_0"), val = bool(false)]; + tensor var_3252_cast_fp16 = concat(axis = var_67, interleave = var_3252_interleave_0, values = key_27_cast_fp16)[name = string("op_3252_cast_fp16")]; + tensor encoder_layers_13_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(273352512))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274139008))))[name = string("encoder_layers_13_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_13_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_13_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274139200)))]; + tensor linear_120_cast_fp16 = linear(bias = encoder_layers_13_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_13_self_attn_linear_q_weight_to_fp16_palettized, x = key_27_cast_fp16)[name = string("linear_120_cast_fp16")]; + tensor var_3257 = const()[name = string("op_3257"), val = tensor([1, -1, 8, 128])]; + tensor q_79_cast_fp16 = reshape(shape = var_3257, x = linear_120_cast_fp16)[name = string("q_79_cast_fp16")]; + tensor encoder_layers_13_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274141312))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274927808))))[name = string("encoder_layers_13_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_13_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_13_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274928000)))]; + tensor linear_121_cast_fp16 = linear(bias = encoder_layers_13_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_13_self_attn_linear_k_weight_to_fp16_palettized, x = input_715_cast_fp16)[name = string("linear_121_cast_fp16")]; + tensor var_3262 = const()[name = string("op_3262"), val = tensor([1, -1, 8, 128])]; + tensor k_53_cast_fp16 = reshape(shape = var_3262, x = linear_121_cast_fp16)[name = string("k_53_cast_fp16")]; + tensor encoder_layers_13_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274930112))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275716608))))[name = string("encoder_layers_13_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_13_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_13_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275716800)))]; + tensor linear_122_cast_fp16 = linear(bias = encoder_layers_13_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_13_self_attn_linear_v_weight_to_fp16_palettized, x = input_715_cast_fp16)[name = string("linear_122_cast_fp16")]; + tensor var_3267 = const()[name = string("op_3267"), val = tensor([1, -1, 8, 128])]; + tensor v_27_cast_fp16 = reshape(shape = var_3267, x = linear_122_cast_fp16)[name = string("v_27_cast_fp16")]; + tensor value_35_perm_0 = const()[name = string("value_35_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_13_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_13_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275718912)))]; + tensor var_3280_cast_fp16 = add(x = q_79_cast_fp16, y = encoder_layers_13_self_attn_pos_bias_u_to_fp16)[name = string("op_3280_cast_fp16")]; + tensor encoder_layers_13_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_13_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275721024)))]; + tensor var_3282_cast_fp16 = add(x = q_79_cast_fp16, y = encoder_layers_13_self_attn_pos_bias_v_to_fp16)[name = string("op_3282_cast_fp16")]; + tensor q_with_bias_v_27_perm_0 = const()[name = string("q_with_bias_v_27_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_345_transpose_x_0 = const()[name = string("x_345_transpose_x_0"), val = bool(false)]; + bool x_345_transpose_y_0 = const()[name = string("x_345_transpose_y_0"), val = bool(false)]; + tensor op_3284_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275723136))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275922880))))[name = string("op_3284_to_fp16_quantized")]; + tensor q_with_bias_v_27_cast_fp16 = transpose(perm = q_with_bias_v_27_perm_0, x = var_3282_cast_fp16)[name = string("transpose_245")]; + tensor x_345_cast_fp16 = matmul(transpose_x = x_345_transpose_x_0, transpose_y = x_345_transpose_y_0, x = q_with_bias_v_27_cast_fp16, y = op_3284_to_fp16_quantized)[name = string("x_345_cast_fp16")]; + tensor x_347_pad_0 = const()[name = string("x_347_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_347_mode_0 = const()[name = string("x_347_mode_0"), val = string("constant")]; + fp16 const_248_to_fp16 = const()[name = string("const_248_to_fp16"), val = fp16(0x0p+0)]; + tensor x_347_cast_fp16 = pad(constant_val = const_248_to_fp16, mode = x_347_mode_0, pad = x_347_pad_0, x = x_345_cast_fp16)[name = string("x_347_cast_fp16")]; + tensor var_3292 = const()[name = string("op_3292"), val = tensor([1, 8, -1, 56])]; + tensor x_349_cast_fp16 = reshape(shape = var_3292, x = x_347_cast_fp16)[name = string("x_349_cast_fp16")]; + tensor var_3296_begin_0 = const()[name = string("op_3296_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_3296_end_0 = const()[name = string("op_3296_end_0"), val = tensor([1, 8, 196, 56])]; + tensor var_3296_end_mask_0 = const()[name = string("op_3296_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_3296_cast_fp16 = slice_by_index(begin = var_3296_begin_0, end = var_3296_end_0, end_mask = var_3296_end_mask_0, x = x_349_cast_fp16)[name = string("op_3296_cast_fp16")]; + tensor var_3297 = const()[name = string("op_3297"), val = tensor([1, 8, 56, 195])]; + tensor matrix_bd_53_cast_fp16 = reshape(shape = var_3297, x = var_3296_cast_fp16)[name = string("matrix_bd_53_cast_fp16")]; + bool matrix_ac_27_transpose_x_0 = const()[name = string("matrix_ac_27_transpose_x_0"), val = bool(false)]; + bool matrix_ac_27_transpose_y_0 = const()[name = string("matrix_ac_27_transpose_y_0"), val = bool(false)]; + tensor transpose_122_perm_0 = const()[name = string("transpose_122_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_123_perm_0 = const()[name = string("transpose_123_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_123 = transpose(perm = transpose_123_perm_0, x = k_53_cast_fp16)[name = string("transpose_243")]; + tensor transpose_122 = transpose(perm = transpose_122_perm_0, x = var_3280_cast_fp16)[name = string("transpose_244")]; + tensor matrix_ac_27_cast_fp16 = matmul(transpose_x = matrix_ac_27_transpose_x_0, transpose_y = matrix_ac_27_transpose_y_0, x = transpose_122, y = transpose_123)[name = string("matrix_ac_27_cast_fp16")]; + tensor matrix_bd_55_begin_0 = const()[name = string("matrix_bd_55_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_55_end_0 = const()[name = string("matrix_bd_55_end_0"), val = tensor([1, 8, 56, 98])]; + tensor matrix_bd_55_end_mask_0 = const()[name = string("matrix_bd_55_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_55_cast_fp16 = slice_by_index(begin = matrix_bd_55_begin_0, end = matrix_bd_55_end_0, end_mask = matrix_bd_55_end_mask_0, x = matrix_bd_53_cast_fp16)[name = string("matrix_bd_55_cast_fp16")]; + tensor var_3306_cast_fp16 = add(x = matrix_ac_27_cast_fp16, y = matrix_bd_55_cast_fp16)[name = string("op_3306_cast_fp16")]; + fp16 _inversed_scores_53_y_0_to_fp16 = const()[name = string("_inversed_scores_53_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_53_cast_fp16 = mul(x = var_3306_cast_fp16, y = _inversed_scores_53_y_0_to_fp16)[name = string("_inversed_scores_53_cast_fp16")]; + tensor scores_55_cast_fp16 = select(a = var_44_to_fp16, b = _inversed_scores_53_cast_fp16, cond = mask_11)[name = string("scores_55_cast_fp16")]; + tensor var_3312_cast_fp16 = softmax(axis = var_58, x = scores_55_cast_fp16)[name = string("op_3312_cast_fp16")]; + tensor input_717_cast_fp16 = select(a = var_43_to_fp16, b = var_3312_cast_fp16, cond = mask_11)[name = string("input_717_cast_fp16")]; + bool x_351_transpose_x_0 = const()[name = string("x_351_transpose_x_0"), val = bool(false)]; + bool x_351_transpose_y_0 = const()[name = string("x_351_transpose_y_0"), val = bool(false)]; + tensor value_35_cast_fp16 = transpose(perm = value_35_perm_0, x = v_27_cast_fp16)[name = string("transpose_242")]; + tensor x_351_cast_fp16 = matmul(transpose_x = x_351_transpose_x_0, transpose_y = x_351_transpose_y_0, x = input_717_cast_fp16, y = value_35_cast_fp16)[name = string("x_351_cast_fp16")]; + tensor var_3316_perm_0 = const()[name = string("op_3316_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_3317 = const()[name = string("op_3317"), val = tensor([1, -1, 1024])]; + tensor var_3316_cast_fp16 = transpose(perm = var_3316_perm_0, x = x_351_cast_fp16)[name = string("transpose_241")]; + tensor input_719_cast_fp16 = reshape(shape = var_3317, x = var_3316_cast_fp16)[name = string("input_719_cast_fp16")]; + tensor encoder_layers_13_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275923392))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(276709888))))[name = string("encoder_layers_13_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_13_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_13_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(276710080)))]; + tensor linear_124_cast_fp16 = linear(bias = encoder_layers_13_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_13_self_attn_linear_out_weight_to_fp16_palettized, x = input_719_cast_fp16)[name = string("linear_124_cast_fp16")]; + tensor input_723_cast_fp16 = add(x = input_713_cast_fp16, y = linear_124_cast_fp16)[name = string("input_723_cast_fp16")]; + tensor x_355_axes_0 = const()[name = string("x_355_axes_0"), val = tensor([-1])]; + tensor encoder_layers_13_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_13_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(276712192)))]; + tensor encoder_layers_13_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_13_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(276714304)))]; + tensor x_355_cast_fp16 = layer_norm(axes = x_355_axes_0, beta = encoder_layers_13_norm_conv_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_13_norm_conv_weight_to_fp16, x = input_723_cast_fp16)[name = string("x_355_cast_fp16")]; + tensor input_725_perm_0 = const()[name = string("input_725_perm_0"), val = tensor([0, 2, 1])]; + string input_727_pad_type_0 = const()[name = string("input_727_pad_type_0"), val = string("valid")]; + tensor input_727_strides_0 = const()[name = string("input_727_strides_0"), val = tensor([1])]; + tensor input_727_pad_0 = const()[name = string("input_727_pad_0"), val = tensor([0, 0])]; + tensor input_727_dilations_0 = const()[name = string("input_727_dilations_0"), val = tensor([1])]; + int32 input_727_groups_0 = const()[name = string("input_727_groups_0"), val = int32(1)]; + tensor encoder_layers_13_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(276716416))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278813632))))[name = string("encoder_layers_13_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_725_cast_fp16 = transpose(perm = input_725_perm_0, x = x_355_cast_fp16)[name = string("transpose_240")]; + tensor input_727_cast_fp16 = conv(dilations = input_727_dilations_0, groups = input_727_groups_0, pad = input_727_pad_0, pad_type = input_727_pad_type_0, strides = input_727_strides_0, weight = encoder_layers_13_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_725_cast_fp16)[name = string("input_727_cast_fp16")]; + int32 x_357_split_num_splits_0 = const()[name = string("x_357_split_num_splits_0"), val = int32(2)]; + int32 x_357_split_axis_0 = const()[name = string("x_357_split_axis_0"), val = int32(1)]; + tensor x_357_split_cast_fp16_0, tensor x_357_split_cast_fp16_1 = split(axis = x_357_split_axis_0, num_splits = x_357_split_num_splits_0, x = input_727_cast_fp16)[name = string("x_357_split_cast_fp16")]; + tensor x_357_split_1_sigmoid_cast_fp16 = sigmoid(x = x_357_split_cast_fp16_1)[name = string("x_357_split_1_sigmoid_cast_fp16")]; + tensor x_357_cast_fp16 = mul(x = x_357_split_cast_fp16_0, y = x_357_split_1_sigmoid_cast_fp16)[name = string("x_357_cast_fp16")]; + tensor input_729_cast_fp16 = select(a = var_43_to_fp16, b = x_357_cast_fp16, cond = var_574)[name = string("input_729_cast_fp16")]; + bool new_x_55_interleave_0 = const()[name = string("new_x_55_interleave_0"), val = bool(false)]; + tensor new_x_55_cast_fp16 = concat(axis = var_58, interleave = new_x_55_interleave_0, values = (cache_55_cast_fp16, input_729_cast_fp16))[name = string("new_x_55_cast_fp16")]; + tensor var_3356_begin_0 = const()[name = string("op_3356_begin_0"), val = tensor([0, 0, 56])]; + tensor var_3356_end_0 = const()[name = string("op_3356_end_0"), val = tensor([1, 1024, 64])]; + tensor var_3356_end_mask_0 = const()[name = string("op_3356_end_mask_0"), val = tensor([true, true, true])]; + tensor var_3356_cast_fp16 = slice_by_index(begin = var_3356_begin_0, end = var_3356_end_0, end_mask = var_3356_end_mask_0, x = new_x_55_cast_fp16)[name = string("op_3356_cast_fp16")]; + string x_359_pad_type_0 = const()[name = string("x_359_pad_type_0"), val = string("valid")]; + int32 x_359_groups_0 = const()[name = string("x_359_groups_0"), val = int32(1024)]; + tensor x_359_strides_0 = const()[name = string("x_359_strides_0"), val = tensor([1])]; + tensor x_359_pad_0 = const()[name = string("x_359_pad_0"), val = tensor([0, 0])]; + tensor x_359_dilations_0 = const()[name = string("x_359_dilations_0"), val = tensor([1])]; + tensor encoder_layers_13_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278817792))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278827072))))[name = string("encoder_layers_13_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_359_cast_fp16 = conv(dilations = x_359_dilations_0, groups = x_359_groups_0, pad = x_359_pad_0, pad_type = x_359_pad_type_0, strides = x_359_strides_0, weight = encoder_layers_13_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_55_cast_fp16)[name = string("x_359_cast_fp16")]; + tensor input_731_perm_0 = const()[name = string("input_731_perm_0"), val = tensor([0, 2, 1])]; + tensor x_361_axes_0 = const()[name = string("x_361_axes_0"), val = tensor([-1])]; + tensor encoder_layers_13_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_13_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278829184)))]; + tensor encoder_layers_13_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_13_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278831296)))]; + tensor input_731_cast_fp16 = transpose(perm = input_731_perm_0, x = x_359_cast_fp16)[name = string("transpose_239")]; + tensor x_361_cast_fp16 = layer_norm(axes = x_361_axes_0, beta = encoder_layers_13_conv_batch_norm_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_13_conv_batch_norm_weight_to_fp16, x = input_731_cast_fp16)[name = string("x_361_cast_fp16")]; + tensor input_733_perm_0 = const()[name = string("input_733_perm_0"), val = tensor([0, 2, 1])]; + tensor input_733_cast_fp16 = transpose(perm = input_733_perm_0, x = x_361_cast_fp16)[name = string("transpose_238")]; + tensor input_735_cast_fp16 = silu(x = input_733_cast_fp16)[name = string("input_735_cast_fp16")]; + string x_363_pad_type_0 = const()[name = string("x_363_pad_type_0"), val = string("valid")]; + tensor x_363_strides_0 = const()[name = string("x_363_strides_0"), val = tensor([1])]; + tensor x_363_pad_0 = const()[name = string("x_363_pad_0"), val = tensor([0, 0])]; + tensor x_363_dilations_0 = const()[name = string("x_363_dilations_0"), val = tensor([1])]; + int32 x_363_groups_0 = const()[name = string("x_363_groups_0"), val = int32(1)]; + tensor encoder_layers_13_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278833408))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(279882048))))[name = string("encoder_layers_13_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_363_cast_fp16 = conv(dilations = x_363_dilations_0, groups = x_363_groups_0, pad = x_363_pad_0, pad_type = x_363_pad_type_0, strides = x_363_strides_0, weight = encoder_layers_13_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_735_cast_fp16)[name = string("x_363_cast_fp16")]; + tensor input_737_perm_0 = const()[name = string("input_737_perm_0"), val = tensor([0, 2, 1])]; + tensor input_737_cast_fp16 = transpose(perm = input_737_perm_0, x = x_363_cast_fp16)[name = string("transpose_237")]; + tensor input_739_cast_fp16 = add(x = input_723_cast_fp16, y = input_737_cast_fp16)[name = string("input_739_cast_fp16")]; + tensor input_741_axes_0 = const()[name = string("input_741_axes_0"), val = tensor([-1])]; + tensor encoder_layers_13_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_13_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(279884160)))]; + tensor encoder_layers_13_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_13_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(279886272)))]; + tensor input_741_cast_fp16 = layer_norm(axes = input_741_axes_0, beta = encoder_layers_13_norm_feed_forward2_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_13_norm_feed_forward2_weight_to_fp16, x = input_739_cast_fp16)[name = string("input_741_cast_fp16")]; + tensor encoder_layers_13_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(279888384))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(283034176))))[name = string("encoder_layers_13_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_13_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_13_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(283034368)))]; + tensor linear_125_cast_fp16 = linear(bias = encoder_layers_13_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_13_feed_forward2_linear1_weight_to_fp16_palettized, x = input_741_cast_fp16)[name = string("linear_125_cast_fp16")]; + tensor input_745_cast_fp16 = silu(x = linear_125_cast_fp16)[name = string("input_745_cast_fp16")]; + tensor encoder_layers_13_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(283042624))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286188416))))[name = string("encoder_layers_13_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_13_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_13_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286188608)))]; + tensor linear_126_cast_fp16 = linear(bias = encoder_layers_13_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_13_feed_forward2_linear2_weight_to_fp16_palettized, x = input_745_cast_fp16)[name = string("linear_126_cast_fp16")]; + fp16 var_3399_to_fp16 = const()[name = string("op_3399_to_fp16"), val = fp16(0x1p-1)]; + tensor var_3400_cast_fp16 = mul(x = linear_126_cast_fp16, y = var_3399_to_fp16)[name = string("op_3400_cast_fp16")]; + tensor input_751_cast_fp16 = add(x = input_739_cast_fp16, y = var_3400_cast_fp16)[name = string("input_751_cast_fp16")]; + tensor input_753_axes_0 = const()[name = string("input_753_axes_0"), val = tensor([-1])]; + tensor encoder_layers_13_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_13_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286190720)))]; + tensor encoder_layers_13_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_13_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286192832)))]; + tensor input_753_cast_fp16 = layer_norm(axes = input_753_axes_0, beta = encoder_layers_13_norm_out_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_13_norm_out_weight_to_fp16, x = input_751_cast_fp16)[name = string("input_753_cast_fp16")]; + tensor cache_57_begin_0 = const()[name = string("cache_57_begin_0"), val = tensor([14, 0, 0, 0])]; + tensor cache_57_end_0 = const()[name = string("cache_57_end_0"), val = tensor([15, 1, 42, 1024])]; + tensor cache_57_end_mask_0 = const()[name = string("cache_57_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_57_squeeze_mask_0 = const()[name = string("cache_57_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_57_cast_fp16 = slice_by_index(begin = cache_57_begin_0, end = cache_57_end_0, end_mask = cache_57_end_mask_0, squeeze_mask = cache_57_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_57_cast_fp16")]; + tensor cache_59_begin_0 = const()[name = string("cache_59_begin_0"), val = tensor([14, 0, 0, 0])]; + tensor cache_59_end_0 = const()[name = string("cache_59_end_0"), val = tensor([15, 1, 1024, 8])]; + tensor cache_59_end_mask_0 = const()[name = string("cache_59_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_59_squeeze_mask_0 = const()[name = string("cache_59_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_59_cast_fp16 = slice_by_index(begin = cache_59_begin_0, end = cache_59_end_0, end_mask = cache_59_end_mask_0, squeeze_mask = cache_59_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_59_cast_fp16")]; + tensor input_755_axes_0 = const()[name = string("input_755_axes_0"), val = tensor([-1])]; + tensor encoder_layers_14_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_14_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286194944)))]; + tensor encoder_layers_14_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_14_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286197056)))]; + tensor input_755_cast_fp16 = layer_norm(axes = input_755_axes_0, beta = encoder_layers_14_norm_feed_forward1_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_14_norm_feed_forward1_weight_to_fp16, x = input_753_cast_fp16)[name = string("input_755_cast_fp16")]; + tensor encoder_layers_14_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286199168))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(289344960))))[name = string("encoder_layers_14_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_14_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_14_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(289345152)))]; + tensor linear_127_cast_fp16 = linear(bias = encoder_layers_14_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_14_feed_forward1_linear1_weight_to_fp16_palettized, x = input_755_cast_fp16)[name = string("linear_127_cast_fp16")]; + tensor input_759_cast_fp16 = silu(x = linear_127_cast_fp16)[name = string("input_759_cast_fp16")]; + tensor encoder_layers_14_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(289353408))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(292499200))))[name = string("encoder_layers_14_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_14_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_14_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(292499392)))]; + tensor linear_128_cast_fp16 = linear(bias = encoder_layers_14_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_14_feed_forward1_linear2_weight_to_fp16_palettized, x = input_759_cast_fp16)[name = string("linear_128_cast_fp16")]; + fp16 var_3436_to_fp16 = const()[name = string("op_3436_to_fp16"), val = fp16(0x1p-1)]; + tensor var_3437_cast_fp16 = mul(x = linear_128_cast_fp16, y = var_3436_to_fp16)[name = string("op_3437_cast_fp16")]; + tensor input_765_cast_fp16 = add(x = input_753_cast_fp16, y = var_3437_cast_fp16)[name = string("input_765_cast_fp16")]; + tensor key_29_axes_0 = const()[name = string("key_29_axes_0"), val = tensor([-1])]; + tensor encoder_layers_14_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_14_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(292501504)))]; + tensor encoder_layers_14_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_14_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(292503616)))]; + tensor key_29_cast_fp16 = layer_norm(axes = key_29_axes_0, beta = encoder_layers_14_norm_self_att_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_14_norm_self_att_weight_to_fp16, x = input_765_cast_fp16)[name = string("key_29_cast_fp16")]; + bool input_767_interleave_0 = const()[name = string("input_767_interleave_0"), val = bool(false)]; + tensor input_767_cast_fp16 = concat(axis = var_67, interleave = input_767_interleave_0, values = (cache_57_cast_fp16, key_29_cast_fp16))[name = string("input_767_cast_fp16")]; + bool var_3465_interleave_0 = const()[name = string("op_3465_interleave_0"), val = bool(false)]; + tensor var_3465_cast_fp16 = concat(axis = var_67, interleave = var_3465_interleave_0, values = key_29_cast_fp16)[name = string("op_3465_cast_fp16")]; + tensor encoder_layers_14_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(292505728))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293292224))))[name = string("encoder_layers_14_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_14_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_14_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293292416)))]; + tensor linear_129_cast_fp16 = linear(bias = encoder_layers_14_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_14_self_attn_linear_q_weight_to_fp16_palettized, x = key_29_cast_fp16)[name = string("linear_129_cast_fp16")]; + tensor var_3470 = const()[name = string("op_3470"), val = tensor([1, -1, 8, 128])]; + tensor q_85_cast_fp16 = reshape(shape = var_3470, x = linear_129_cast_fp16)[name = string("q_85_cast_fp16")]; + tensor encoder_layers_14_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293294528))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294081024))))[name = string("encoder_layers_14_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_14_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_14_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294081216)))]; + tensor linear_130_cast_fp16 = linear(bias = encoder_layers_14_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_14_self_attn_linear_k_weight_to_fp16_palettized, x = input_767_cast_fp16)[name = string("linear_130_cast_fp16")]; + tensor var_3475 = const()[name = string("op_3475"), val = tensor([1, -1, 8, 128])]; + tensor k_57_cast_fp16 = reshape(shape = var_3475, x = linear_130_cast_fp16)[name = string("k_57_cast_fp16")]; + tensor encoder_layers_14_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294083328))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294869824))))[name = string("encoder_layers_14_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_14_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_14_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294870016)))]; + tensor linear_131_cast_fp16 = linear(bias = encoder_layers_14_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_14_self_attn_linear_v_weight_to_fp16_palettized, x = input_767_cast_fp16)[name = string("linear_131_cast_fp16")]; + tensor var_3480 = const()[name = string("op_3480"), val = tensor([1, -1, 8, 128])]; + tensor v_29_cast_fp16 = reshape(shape = var_3480, x = linear_131_cast_fp16)[name = string("v_29_cast_fp16")]; + tensor value_37_perm_0 = const()[name = string("value_37_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_14_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_14_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294872128)))]; + tensor var_3493_cast_fp16 = add(x = q_85_cast_fp16, y = encoder_layers_14_self_attn_pos_bias_u_to_fp16)[name = string("op_3493_cast_fp16")]; + tensor encoder_layers_14_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_14_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294874240)))]; + tensor var_3495_cast_fp16 = add(x = q_85_cast_fp16, y = encoder_layers_14_self_attn_pos_bias_v_to_fp16)[name = string("op_3495_cast_fp16")]; + tensor q_with_bias_v_29_perm_0 = const()[name = string("q_with_bias_v_29_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_371_transpose_x_0 = const()[name = string("x_371_transpose_x_0"), val = bool(false)]; + bool x_371_transpose_y_0 = const()[name = string("x_371_transpose_y_0"), val = bool(false)]; + tensor op_3497_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294876352))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(295076096))))[name = string("op_3497_to_fp16_quantized")]; + tensor q_with_bias_v_29_cast_fp16 = transpose(perm = q_with_bias_v_29_perm_0, x = var_3495_cast_fp16)[name = string("transpose_236")]; + tensor x_371_cast_fp16 = matmul(transpose_x = x_371_transpose_x_0, transpose_y = x_371_transpose_y_0, x = q_with_bias_v_29_cast_fp16, y = op_3497_to_fp16_quantized)[name = string("x_371_cast_fp16")]; + tensor x_373_pad_0 = const()[name = string("x_373_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_373_mode_0 = const()[name = string("x_373_mode_0"), val = string("constant")]; + fp16 const_261_to_fp16 = const()[name = string("const_261_to_fp16"), val = fp16(0x0p+0)]; + tensor x_373_cast_fp16 = pad(constant_val = const_261_to_fp16, mode = x_373_mode_0, pad = x_373_pad_0, x = x_371_cast_fp16)[name = string("x_373_cast_fp16")]; + tensor var_3505 = const()[name = string("op_3505"), val = tensor([1, 8, -1, 56])]; + tensor x_375_cast_fp16 = reshape(shape = var_3505, x = x_373_cast_fp16)[name = string("x_375_cast_fp16")]; + tensor var_3509_begin_0 = const()[name = string("op_3509_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_3509_end_0 = const()[name = string("op_3509_end_0"), val = tensor([1, 8, 196, 56])]; + tensor var_3509_end_mask_0 = const()[name = string("op_3509_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_3509_cast_fp16 = slice_by_index(begin = var_3509_begin_0, end = var_3509_end_0, end_mask = var_3509_end_mask_0, x = x_375_cast_fp16)[name = string("op_3509_cast_fp16")]; + tensor var_3510 = const()[name = string("op_3510"), val = tensor([1, 8, 56, 195])]; + tensor matrix_bd_57_cast_fp16 = reshape(shape = var_3510, x = var_3509_cast_fp16)[name = string("matrix_bd_57_cast_fp16")]; + bool matrix_ac_29_transpose_x_0 = const()[name = string("matrix_ac_29_transpose_x_0"), val = bool(false)]; + bool matrix_ac_29_transpose_y_0 = const()[name = string("matrix_ac_29_transpose_y_0"), val = bool(false)]; + tensor transpose_124_perm_0 = const()[name = string("transpose_124_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_125_perm_0 = const()[name = string("transpose_125_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_125 = transpose(perm = transpose_125_perm_0, x = k_57_cast_fp16)[name = string("transpose_234")]; + tensor transpose_124 = transpose(perm = transpose_124_perm_0, x = var_3493_cast_fp16)[name = string("transpose_235")]; + tensor matrix_ac_29_cast_fp16 = matmul(transpose_x = matrix_ac_29_transpose_x_0, transpose_y = matrix_ac_29_transpose_y_0, x = transpose_124, y = transpose_125)[name = string("matrix_ac_29_cast_fp16")]; + tensor matrix_bd_59_begin_0 = const()[name = string("matrix_bd_59_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_59_end_0 = const()[name = string("matrix_bd_59_end_0"), val = tensor([1, 8, 56, 98])]; + tensor matrix_bd_59_end_mask_0 = const()[name = string("matrix_bd_59_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_59_cast_fp16 = slice_by_index(begin = matrix_bd_59_begin_0, end = matrix_bd_59_end_0, end_mask = matrix_bd_59_end_mask_0, x = matrix_bd_57_cast_fp16)[name = string("matrix_bd_59_cast_fp16")]; + tensor var_3519_cast_fp16 = add(x = matrix_ac_29_cast_fp16, y = matrix_bd_59_cast_fp16)[name = string("op_3519_cast_fp16")]; + fp16 _inversed_scores_57_y_0_to_fp16 = const()[name = string("_inversed_scores_57_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_57_cast_fp16 = mul(x = var_3519_cast_fp16, y = _inversed_scores_57_y_0_to_fp16)[name = string("_inversed_scores_57_cast_fp16")]; + tensor scores_59_cast_fp16 = select(a = var_44_to_fp16, b = _inversed_scores_57_cast_fp16, cond = mask_11)[name = string("scores_59_cast_fp16")]; + tensor var_3525_cast_fp16 = softmax(axis = var_58, x = scores_59_cast_fp16)[name = string("op_3525_cast_fp16")]; + tensor input_769_cast_fp16 = select(a = var_43_to_fp16, b = var_3525_cast_fp16, cond = mask_11)[name = string("input_769_cast_fp16")]; + bool x_377_transpose_x_0 = const()[name = string("x_377_transpose_x_0"), val = bool(false)]; + bool x_377_transpose_y_0 = const()[name = string("x_377_transpose_y_0"), val = bool(false)]; + tensor value_37_cast_fp16 = transpose(perm = value_37_perm_0, x = v_29_cast_fp16)[name = string("transpose_233")]; + tensor x_377_cast_fp16 = matmul(transpose_x = x_377_transpose_x_0, transpose_y = x_377_transpose_y_0, x = input_769_cast_fp16, y = value_37_cast_fp16)[name = string("x_377_cast_fp16")]; + tensor var_3529_perm_0 = const()[name = string("op_3529_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_3530 = const()[name = string("op_3530"), val = tensor([1, -1, 1024])]; + tensor var_3529_cast_fp16 = transpose(perm = var_3529_perm_0, x = x_377_cast_fp16)[name = string("transpose_232")]; + tensor input_771_cast_fp16 = reshape(shape = var_3530, x = var_3529_cast_fp16)[name = string("input_771_cast_fp16")]; + tensor encoder_layers_14_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(295076608))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(295863104))))[name = string("encoder_layers_14_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_14_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_14_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(295863296)))]; + tensor linear_133_cast_fp16 = linear(bias = encoder_layers_14_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_14_self_attn_linear_out_weight_to_fp16_palettized, x = input_771_cast_fp16)[name = string("linear_133_cast_fp16")]; + tensor input_775_cast_fp16 = add(x = input_765_cast_fp16, y = linear_133_cast_fp16)[name = string("input_775_cast_fp16")]; + tensor x_381_axes_0 = const()[name = string("x_381_axes_0"), val = tensor([-1])]; + tensor encoder_layers_14_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_14_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(295865408)))]; + tensor encoder_layers_14_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_14_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(295867520)))]; + tensor x_381_cast_fp16 = layer_norm(axes = x_381_axes_0, beta = encoder_layers_14_norm_conv_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_14_norm_conv_weight_to_fp16, x = input_775_cast_fp16)[name = string("x_381_cast_fp16")]; + tensor input_777_perm_0 = const()[name = string("input_777_perm_0"), val = tensor([0, 2, 1])]; + string input_779_pad_type_0 = const()[name = string("input_779_pad_type_0"), val = string("valid")]; + tensor input_779_strides_0 = const()[name = string("input_779_strides_0"), val = tensor([1])]; + tensor input_779_pad_0 = const()[name = string("input_779_pad_0"), val = tensor([0, 0])]; + tensor input_779_dilations_0 = const()[name = string("input_779_dilations_0"), val = tensor([1])]; + int32 input_779_groups_0 = const()[name = string("input_779_groups_0"), val = int32(1)]; + tensor encoder_layers_14_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(295869632))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297966848))))[name = string("encoder_layers_14_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_777_cast_fp16 = transpose(perm = input_777_perm_0, x = x_381_cast_fp16)[name = string("transpose_231")]; + tensor input_779_cast_fp16 = conv(dilations = input_779_dilations_0, groups = input_779_groups_0, pad = input_779_pad_0, pad_type = input_779_pad_type_0, strides = input_779_strides_0, weight = encoder_layers_14_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_777_cast_fp16)[name = string("input_779_cast_fp16")]; + int32 x_383_split_num_splits_0 = const()[name = string("x_383_split_num_splits_0"), val = int32(2)]; + int32 x_383_split_axis_0 = const()[name = string("x_383_split_axis_0"), val = int32(1)]; + tensor x_383_split_cast_fp16_0, tensor x_383_split_cast_fp16_1 = split(axis = x_383_split_axis_0, num_splits = x_383_split_num_splits_0, x = input_779_cast_fp16)[name = string("x_383_split_cast_fp16")]; + tensor x_383_split_1_sigmoid_cast_fp16 = sigmoid(x = x_383_split_cast_fp16_1)[name = string("x_383_split_1_sigmoid_cast_fp16")]; + tensor x_383_cast_fp16 = mul(x = x_383_split_cast_fp16_0, y = x_383_split_1_sigmoid_cast_fp16)[name = string("x_383_cast_fp16")]; + tensor input_781_cast_fp16 = select(a = var_43_to_fp16, b = x_383_cast_fp16, cond = var_574)[name = string("input_781_cast_fp16")]; + bool new_x_59_interleave_0 = const()[name = string("new_x_59_interleave_0"), val = bool(false)]; + tensor new_x_59_cast_fp16 = concat(axis = var_58, interleave = new_x_59_interleave_0, values = (cache_59_cast_fp16, input_781_cast_fp16))[name = string("new_x_59_cast_fp16")]; + tensor var_3569_begin_0 = const()[name = string("op_3569_begin_0"), val = tensor([0, 0, 56])]; + tensor var_3569_end_0 = const()[name = string("op_3569_end_0"), val = tensor([1, 1024, 64])]; + tensor var_3569_end_mask_0 = const()[name = string("op_3569_end_mask_0"), val = tensor([true, true, true])]; + tensor var_3569_cast_fp16 = slice_by_index(begin = var_3569_begin_0, end = var_3569_end_0, end_mask = var_3569_end_mask_0, x = new_x_59_cast_fp16)[name = string("op_3569_cast_fp16")]; + string x_385_pad_type_0 = const()[name = string("x_385_pad_type_0"), val = string("valid")]; + int32 x_385_groups_0 = const()[name = string("x_385_groups_0"), val = int32(1024)]; + tensor x_385_strides_0 = const()[name = string("x_385_strides_0"), val = tensor([1])]; + tensor x_385_pad_0 = const()[name = string("x_385_pad_0"), val = tensor([0, 0])]; + tensor x_385_dilations_0 = const()[name = string("x_385_dilations_0"), val = tensor([1])]; + tensor encoder_layers_14_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297971008))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297980288))))[name = string("encoder_layers_14_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_385_cast_fp16 = conv(dilations = x_385_dilations_0, groups = x_385_groups_0, pad = x_385_pad_0, pad_type = x_385_pad_type_0, strides = x_385_strides_0, weight = encoder_layers_14_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_59_cast_fp16)[name = string("x_385_cast_fp16")]; + tensor input_783_perm_0 = const()[name = string("input_783_perm_0"), val = tensor([0, 2, 1])]; + tensor x_387_axes_0 = const()[name = string("x_387_axes_0"), val = tensor([-1])]; + tensor encoder_layers_14_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_14_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297982400)))]; + tensor encoder_layers_14_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_14_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297984512)))]; + tensor input_783_cast_fp16 = transpose(perm = input_783_perm_0, x = x_385_cast_fp16)[name = string("transpose_230")]; + tensor x_387_cast_fp16 = layer_norm(axes = x_387_axes_0, beta = encoder_layers_14_conv_batch_norm_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_14_conv_batch_norm_weight_to_fp16, x = input_783_cast_fp16)[name = string("x_387_cast_fp16")]; + tensor input_785_perm_0 = const()[name = string("input_785_perm_0"), val = tensor([0, 2, 1])]; + tensor input_785_cast_fp16 = transpose(perm = input_785_perm_0, x = x_387_cast_fp16)[name = string("transpose_229")]; + tensor input_787_cast_fp16 = silu(x = input_785_cast_fp16)[name = string("input_787_cast_fp16")]; + string x_389_pad_type_0 = const()[name = string("x_389_pad_type_0"), val = string("valid")]; + tensor x_389_strides_0 = const()[name = string("x_389_strides_0"), val = tensor([1])]; + tensor x_389_pad_0 = const()[name = string("x_389_pad_0"), val = tensor([0, 0])]; + tensor x_389_dilations_0 = const()[name = string("x_389_dilations_0"), val = tensor([1])]; + int32 x_389_groups_0 = const()[name = string("x_389_groups_0"), val = int32(1)]; + tensor encoder_layers_14_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297986624))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(299035264))))[name = string("encoder_layers_14_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_389_cast_fp16 = conv(dilations = x_389_dilations_0, groups = x_389_groups_0, pad = x_389_pad_0, pad_type = x_389_pad_type_0, strides = x_389_strides_0, weight = encoder_layers_14_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_787_cast_fp16)[name = string("x_389_cast_fp16")]; + tensor input_789_perm_0 = const()[name = string("input_789_perm_0"), val = tensor([0, 2, 1])]; + tensor input_789_cast_fp16 = transpose(perm = input_789_perm_0, x = x_389_cast_fp16)[name = string("transpose_228")]; + tensor input_791_cast_fp16 = add(x = input_775_cast_fp16, y = input_789_cast_fp16)[name = string("input_791_cast_fp16")]; + tensor input_793_axes_0 = const()[name = string("input_793_axes_0"), val = tensor([-1])]; + tensor encoder_layers_14_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_14_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(299037376)))]; + tensor encoder_layers_14_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_14_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(299039488)))]; + tensor input_793_cast_fp16 = layer_norm(axes = input_793_axes_0, beta = encoder_layers_14_norm_feed_forward2_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_14_norm_feed_forward2_weight_to_fp16, x = input_791_cast_fp16)[name = string("input_793_cast_fp16")]; + tensor encoder_layers_14_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(299041600))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(302187392))))[name = string("encoder_layers_14_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_14_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_14_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(302187584)))]; + tensor linear_134_cast_fp16 = linear(bias = encoder_layers_14_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_14_feed_forward2_linear1_weight_to_fp16_palettized, x = input_793_cast_fp16)[name = string("linear_134_cast_fp16")]; + tensor input_797_cast_fp16 = silu(x = linear_134_cast_fp16)[name = string("input_797_cast_fp16")]; + tensor encoder_layers_14_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(302195840))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(305341632))))[name = string("encoder_layers_14_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_14_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_14_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(305341824)))]; + tensor linear_135_cast_fp16 = linear(bias = encoder_layers_14_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_14_feed_forward2_linear2_weight_to_fp16_palettized, x = input_797_cast_fp16)[name = string("linear_135_cast_fp16")]; + fp16 var_3612_to_fp16 = const()[name = string("op_3612_to_fp16"), val = fp16(0x1p-1)]; + tensor var_3613_cast_fp16 = mul(x = linear_135_cast_fp16, y = var_3612_to_fp16)[name = string("op_3613_cast_fp16")]; + tensor input_803_cast_fp16 = add(x = input_791_cast_fp16, y = var_3613_cast_fp16)[name = string("input_803_cast_fp16")]; + tensor input_805_axes_0 = const()[name = string("input_805_axes_0"), val = tensor([-1])]; + tensor encoder_layers_14_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_14_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(305343936)))]; + tensor encoder_layers_14_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_14_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(305346048)))]; + tensor input_805_cast_fp16 = layer_norm(axes = input_805_axes_0, beta = encoder_layers_14_norm_out_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_14_norm_out_weight_to_fp16, x = input_803_cast_fp16)[name = string("input_805_cast_fp16")]; + tensor cache_61_begin_0 = const()[name = string("cache_61_begin_0"), val = tensor([15, 0, 0, 0])]; + tensor cache_61_end_0 = const()[name = string("cache_61_end_0"), val = tensor([16, 1, 42, 1024])]; + tensor cache_61_end_mask_0 = const()[name = string("cache_61_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_61_squeeze_mask_0 = const()[name = string("cache_61_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_61_cast_fp16 = slice_by_index(begin = cache_61_begin_0, end = cache_61_end_0, end_mask = cache_61_end_mask_0, squeeze_mask = cache_61_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_61_cast_fp16")]; + tensor cache_63_begin_0 = const()[name = string("cache_63_begin_0"), val = tensor([15, 0, 0, 0])]; + tensor cache_63_end_0 = const()[name = string("cache_63_end_0"), val = tensor([16, 1, 1024, 8])]; + tensor cache_63_end_mask_0 = const()[name = string("cache_63_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_63_squeeze_mask_0 = const()[name = string("cache_63_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_63_cast_fp16 = slice_by_index(begin = cache_63_begin_0, end = cache_63_end_0, end_mask = cache_63_end_mask_0, squeeze_mask = cache_63_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_63_cast_fp16")]; + tensor input_807_axes_0 = const()[name = string("input_807_axes_0"), val = tensor([-1])]; + tensor encoder_layers_15_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_15_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(305348160)))]; + tensor encoder_layers_15_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_15_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(305350272)))]; + tensor input_807_cast_fp16 = layer_norm(axes = input_807_axes_0, beta = encoder_layers_15_norm_feed_forward1_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_15_norm_feed_forward1_weight_to_fp16, x = input_805_cast_fp16)[name = string("input_807_cast_fp16")]; + tensor encoder_layers_15_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(305352384))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308498176))))[name = string("encoder_layers_15_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_15_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_15_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308498368)))]; + tensor linear_136_cast_fp16 = linear(bias = encoder_layers_15_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_15_feed_forward1_linear1_weight_to_fp16_palettized, x = input_807_cast_fp16)[name = string("linear_136_cast_fp16")]; + tensor input_811_cast_fp16 = silu(x = linear_136_cast_fp16)[name = string("input_811_cast_fp16")]; + tensor encoder_layers_15_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308506624))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311652416))))[name = string("encoder_layers_15_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_15_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_15_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311652608)))]; + tensor linear_137_cast_fp16 = linear(bias = encoder_layers_15_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_15_feed_forward1_linear2_weight_to_fp16_palettized, x = input_811_cast_fp16)[name = string("linear_137_cast_fp16")]; + fp16 var_3649_to_fp16 = const()[name = string("op_3649_to_fp16"), val = fp16(0x1p-1)]; + tensor var_3650_cast_fp16 = mul(x = linear_137_cast_fp16, y = var_3649_to_fp16)[name = string("op_3650_cast_fp16")]; + tensor input_817_cast_fp16 = add(x = input_805_cast_fp16, y = var_3650_cast_fp16)[name = string("input_817_cast_fp16")]; + tensor key_31_axes_0 = const()[name = string("key_31_axes_0"), val = tensor([-1])]; + tensor encoder_layers_15_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_15_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311654720)))]; + tensor encoder_layers_15_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_15_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311656832)))]; + tensor key_31_cast_fp16 = layer_norm(axes = key_31_axes_0, beta = encoder_layers_15_norm_self_att_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_15_norm_self_att_weight_to_fp16, x = input_817_cast_fp16)[name = string("key_31_cast_fp16")]; + bool input_819_interleave_0 = const()[name = string("input_819_interleave_0"), val = bool(false)]; + tensor input_819_cast_fp16 = concat(axis = var_67, interleave = input_819_interleave_0, values = (cache_61_cast_fp16, key_31_cast_fp16))[name = string("input_819_cast_fp16")]; + bool var_3678_interleave_0 = const()[name = string("op_3678_interleave_0"), val = bool(false)]; + tensor var_3678_cast_fp16 = concat(axis = var_67, interleave = var_3678_interleave_0, values = key_31_cast_fp16)[name = string("op_3678_cast_fp16")]; + tensor encoder_layers_15_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311658944))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(312445440))))[name = string("encoder_layers_15_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_15_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_15_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(312445632)))]; + tensor linear_138_cast_fp16 = linear(bias = encoder_layers_15_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_15_self_attn_linear_q_weight_to_fp16_palettized, x = key_31_cast_fp16)[name = string("linear_138_cast_fp16")]; + tensor var_3683 = const()[name = string("op_3683"), val = tensor([1, -1, 8, 128])]; + tensor q_91_cast_fp16 = reshape(shape = var_3683, x = linear_138_cast_fp16)[name = string("q_91_cast_fp16")]; + tensor encoder_layers_15_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(312447744))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(313234240))))[name = string("encoder_layers_15_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_15_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_15_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(313234432)))]; + tensor linear_139_cast_fp16 = linear(bias = encoder_layers_15_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_15_self_attn_linear_k_weight_to_fp16_palettized, x = input_819_cast_fp16)[name = string("linear_139_cast_fp16")]; + tensor var_3688 = const()[name = string("op_3688"), val = tensor([1, -1, 8, 128])]; + tensor k_61_cast_fp16 = reshape(shape = var_3688, x = linear_139_cast_fp16)[name = string("k_61_cast_fp16")]; + tensor encoder_layers_15_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(313236544))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(314023040))))[name = string("encoder_layers_15_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_15_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_15_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(314023232)))]; + tensor linear_140_cast_fp16 = linear(bias = encoder_layers_15_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_15_self_attn_linear_v_weight_to_fp16_palettized, x = input_819_cast_fp16)[name = string("linear_140_cast_fp16")]; + tensor var_3693 = const()[name = string("op_3693"), val = tensor([1, -1, 8, 128])]; + tensor v_31_cast_fp16 = reshape(shape = var_3693, x = linear_140_cast_fp16)[name = string("v_31_cast_fp16")]; + tensor value_39_perm_0 = const()[name = string("value_39_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_15_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_15_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(314025344)))]; + tensor var_3706_cast_fp16 = add(x = q_91_cast_fp16, y = encoder_layers_15_self_attn_pos_bias_u_to_fp16)[name = string("op_3706_cast_fp16")]; + tensor encoder_layers_15_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_15_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(314027456)))]; + tensor var_3708_cast_fp16 = add(x = q_91_cast_fp16, y = encoder_layers_15_self_attn_pos_bias_v_to_fp16)[name = string("op_3708_cast_fp16")]; + tensor q_with_bias_v_31_perm_0 = const()[name = string("q_with_bias_v_31_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_397_transpose_x_0 = const()[name = string("x_397_transpose_x_0"), val = bool(false)]; + bool x_397_transpose_y_0 = const()[name = string("x_397_transpose_y_0"), val = bool(false)]; + tensor op_3710_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(314029568))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(314229312))))[name = string("op_3710_to_fp16_quantized")]; + tensor q_with_bias_v_31_cast_fp16 = transpose(perm = q_with_bias_v_31_perm_0, x = var_3708_cast_fp16)[name = string("transpose_227")]; + tensor x_397_cast_fp16 = matmul(transpose_x = x_397_transpose_x_0, transpose_y = x_397_transpose_y_0, x = q_with_bias_v_31_cast_fp16, y = op_3710_to_fp16_quantized)[name = string("x_397_cast_fp16")]; + tensor x_399_pad_0 = const()[name = string("x_399_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_399_mode_0 = const()[name = string("x_399_mode_0"), val = string("constant")]; + fp16 const_274_to_fp16 = const()[name = string("const_274_to_fp16"), val = fp16(0x0p+0)]; + tensor x_399_cast_fp16 = pad(constant_val = const_274_to_fp16, mode = x_399_mode_0, pad = x_399_pad_0, x = x_397_cast_fp16)[name = string("x_399_cast_fp16")]; + tensor var_3718 = const()[name = string("op_3718"), val = tensor([1, 8, -1, 56])]; + tensor x_401_cast_fp16 = reshape(shape = var_3718, x = x_399_cast_fp16)[name = string("x_401_cast_fp16")]; + tensor var_3722_begin_0 = const()[name = string("op_3722_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_3722_end_0 = const()[name = string("op_3722_end_0"), val = tensor([1, 8, 196, 56])]; + tensor var_3722_end_mask_0 = const()[name = string("op_3722_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_3722_cast_fp16 = slice_by_index(begin = var_3722_begin_0, end = var_3722_end_0, end_mask = var_3722_end_mask_0, x = x_401_cast_fp16)[name = string("op_3722_cast_fp16")]; + tensor var_3723 = const()[name = string("op_3723"), val = tensor([1, 8, 56, 195])]; + tensor matrix_bd_61_cast_fp16 = reshape(shape = var_3723, x = var_3722_cast_fp16)[name = string("matrix_bd_61_cast_fp16")]; + bool matrix_ac_31_transpose_x_0 = const()[name = string("matrix_ac_31_transpose_x_0"), val = bool(false)]; + bool matrix_ac_31_transpose_y_0 = const()[name = string("matrix_ac_31_transpose_y_0"), val = bool(false)]; + tensor transpose_126_perm_0 = const()[name = string("transpose_126_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_127_perm_0 = const()[name = string("transpose_127_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_127 = transpose(perm = transpose_127_perm_0, x = k_61_cast_fp16)[name = string("transpose_225")]; + tensor transpose_126 = transpose(perm = transpose_126_perm_0, x = var_3706_cast_fp16)[name = string("transpose_226")]; + tensor matrix_ac_31_cast_fp16 = matmul(transpose_x = matrix_ac_31_transpose_x_0, transpose_y = matrix_ac_31_transpose_y_0, x = transpose_126, y = transpose_127)[name = string("matrix_ac_31_cast_fp16")]; + tensor matrix_bd_63_begin_0 = const()[name = string("matrix_bd_63_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_63_end_0 = const()[name = string("matrix_bd_63_end_0"), val = tensor([1, 8, 56, 98])]; + tensor matrix_bd_63_end_mask_0 = const()[name = string("matrix_bd_63_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_63_cast_fp16 = slice_by_index(begin = matrix_bd_63_begin_0, end = matrix_bd_63_end_0, end_mask = matrix_bd_63_end_mask_0, x = matrix_bd_61_cast_fp16)[name = string("matrix_bd_63_cast_fp16")]; + tensor var_3732_cast_fp16 = add(x = matrix_ac_31_cast_fp16, y = matrix_bd_63_cast_fp16)[name = string("op_3732_cast_fp16")]; + fp16 _inversed_scores_61_y_0_to_fp16 = const()[name = string("_inversed_scores_61_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_61_cast_fp16 = mul(x = var_3732_cast_fp16, y = _inversed_scores_61_y_0_to_fp16)[name = string("_inversed_scores_61_cast_fp16")]; + tensor scores_63_cast_fp16 = select(a = var_44_to_fp16, b = _inversed_scores_61_cast_fp16, cond = mask_11)[name = string("scores_63_cast_fp16")]; + tensor var_3738_cast_fp16 = softmax(axis = var_58, x = scores_63_cast_fp16)[name = string("op_3738_cast_fp16")]; + tensor input_821_cast_fp16 = select(a = var_43_to_fp16, b = var_3738_cast_fp16, cond = mask_11)[name = string("input_821_cast_fp16")]; + bool x_403_transpose_x_0 = const()[name = string("x_403_transpose_x_0"), val = bool(false)]; + bool x_403_transpose_y_0 = const()[name = string("x_403_transpose_y_0"), val = bool(false)]; + tensor value_39_cast_fp16 = transpose(perm = value_39_perm_0, x = v_31_cast_fp16)[name = string("transpose_224")]; + tensor x_403_cast_fp16 = matmul(transpose_x = x_403_transpose_x_0, transpose_y = x_403_transpose_y_0, x = input_821_cast_fp16, y = value_39_cast_fp16)[name = string("x_403_cast_fp16")]; + tensor var_3742_perm_0 = const()[name = string("op_3742_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_3743 = const()[name = string("op_3743"), val = tensor([1, -1, 1024])]; + tensor var_3742_cast_fp16 = transpose(perm = var_3742_perm_0, x = x_403_cast_fp16)[name = string("transpose_223")]; + tensor input_823_cast_fp16 = reshape(shape = var_3743, x = var_3742_cast_fp16)[name = string("input_823_cast_fp16")]; + tensor encoder_layers_15_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(314229824))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315016320))))[name = string("encoder_layers_15_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_15_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_15_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315016512)))]; + tensor linear_142_cast_fp16 = linear(bias = encoder_layers_15_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_15_self_attn_linear_out_weight_to_fp16_palettized, x = input_823_cast_fp16)[name = string("linear_142_cast_fp16")]; + tensor input_827_cast_fp16 = add(x = input_817_cast_fp16, y = linear_142_cast_fp16)[name = string("input_827_cast_fp16")]; + tensor x_407_axes_0 = const()[name = string("x_407_axes_0"), val = tensor([-1])]; + tensor encoder_layers_15_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_15_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315018624)))]; + tensor encoder_layers_15_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_15_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315020736)))]; + tensor x_407_cast_fp16 = layer_norm(axes = x_407_axes_0, beta = encoder_layers_15_norm_conv_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_15_norm_conv_weight_to_fp16, x = input_827_cast_fp16)[name = string("x_407_cast_fp16")]; + tensor input_829_perm_0 = const()[name = string("input_829_perm_0"), val = tensor([0, 2, 1])]; + string input_831_pad_type_0 = const()[name = string("input_831_pad_type_0"), val = string("valid")]; + tensor input_831_strides_0 = const()[name = string("input_831_strides_0"), val = tensor([1])]; + tensor input_831_pad_0 = const()[name = string("input_831_pad_0"), val = tensor([0, 0])]; + tensor input_831_dilations_0 = const()[name = string("input_831_dilations_0"), val = tensor([1])]; + int32 input_831_groups_0 = const()[name = string("input_831_groups_0"), val = int32(1)]; + tensor encoder_layers_15_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315022848))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(317120064))))[name = string("encoder_layers_15_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_829_cast_fp16 = transpose(perm = input_829_perm_0, x = x_407_cast_fp16)[name = string("transpose_222")]; + tensor input_831_cast_fp16 = conv(dilations = input_831_dilations_0, groups = input_831_groups_0, pad = input_831_pad_0, pad_type = input_831_pad_type_0, strides = input_831_strides_0, weight = encoder_layers_15_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_829_cast_fp16)[name = string("input_831_cast_fp16")]; + int32 x_409_split_num_splits_0 = const()[name = string("x_409_split_num_splits_0"), val = int32(2)]; + int32 x_409_split_axis_0 = const()[name = string("x_409_split_axis_0"), val = int32(1)]; + tensor x_409_split_cast_fp16_0, tensor x_409_split_cast_fp16_1 = split(axis = x_409_split_axis_0, num_splits = x_409_split_num_splits_0, x = input_831_cast_fp16)[name = string("x_409_split_cast_fp16")]; + tensor x_409_split_1_sigmoid_cast_fp16 = sigmoid(x = x_409_split_cast_fp16_1)[name = string("x_409_split_1_sigmoid_cast_fp16")]; + tensor x_409_cast_fp16 = mul(x = x_409_split_cast_fp16_0, y = x_409_split_1_sigmoid_cast_fp16)[name = string("x_409_cast_fp16")]; + tensor input_833_cast_fp16 = select(a = var_43_to_fp16, b = x_409_cast_fp16, cond = var_574)[name = string("input_833_cast_fp16")]; + bool new_x_63_interleave_0 = const()[name = string("new_x_63_interleave_0"), val = bool(false)]; + tensor new_x_63_cast_fp16 = concat(axis = var_58, interleave = new_x_63_interleave_0, values = (cache_63_cast_fp16, input_833_cast_fp16))[name = string("new_x_63_cast_fp16")]; + tensor var_3782_begin_0 = const()[name = string("op_3782_begin_0"), val = tensor([0, 0, 56])]; + tensor var_3782_end_0 = const()[name = string("op_3782_end_0"), val = tensor([1, 1024, 64])]; + tensor var_3782_end_mask_0 = const()[name = string("op_3782_end_mask_0"), val = tensor([true, true, true])]; + tensor var_3782_cast_fp16 = slice_by_index(begin = var_3782_begin_0, end = var_3782_end_0, end_mask = var_3782_end_mask_0, x = new_x_63_cast_fp16)[name = string("op_3782_cast_fp16")]; + string x_411_pad_type_0 = const()[name = string("x_411_pad_type_0"), val = string("valid")]; + int32 x_411_groups_0 = const()[name = string("x_411_groups_0"), val = int32(1024)]; + tensor x_411_strides_0 = const()[name = string("x_411_strides_0"), val = tensor([1])]; + tensor x_411_pad_0 = const()[name = string("x_411_pad_0"), val = tensor([0, 0])]; + tensor x_411_dilations_0 = const()[name = string("x_411_dilations_0"), val = tensor([1])]; + tensor encoder_layers_15_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(317124224))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(317133504))))[name = string("encoder_layers_15_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_411_cast_fp16 = conv(dilations = x_411_dilations_0, groups = x_411_groups_0, pad = x_411_pad_0, pad_type = x_411_pad_type_0, strides = x_411_strides_0, weight = encoder_layers_15_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_63_cast_fp16)[name = string("x_411_cast_fp16")]; + tensor input_835_perm_0 = const()[name = string("input_835_perm_0"), val = tensor([0, 2, 1])]; + tensor x_413_axes_0 = const()[name = string("x_413_axes_0"), val = tensor([-1])]; + tensor encoder_layers_15_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_15_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(317135616)))]; + tensor encoder_layers_15_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_15_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(317137728)))]; + tensor input_835_cast_fp16 = transpose(perm = input_835_perm_0, x = x_411_cast_fp16)[name = string("transpose_221")]; + tensor x_413_cast_fp16 = layer_norm(axes = x_413_axes_0, beta = encoder_layers_15_conv_batch_norm_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_15_conv_batch_norm_weight_to_fp16, x = input_835_cast_fp16)[name = string("x_413_cast_fp16")]; + tensor input_837_perm_0 = const()[name = string("input_837_perm_0"), val = tensor([0, 2, 1])]; + tensor input_837_cast_fp16 = transpose(perm = input_837_perm_0, x = x_413_cast_fp16)[name = string("transpose_220")]; + tensor input_839_cast_fp16 = silu(x = input_837_cast_fp16)[name = string("input_839_cast_fp16")]; + string x_415_pad_type_0 = const()[name = string("x_415_pad_type_0"), val = string("valid")]; + tensor x_415_strides_0 = const()[name = string("x_415_strides_0"), val = tensor([1])]; + tensor x_415_pad_0 = const()[name = string("x_415_pad_0"), val = tensor([0, 0])]; + tensor x_415_dilations_0 = const()[name = string("x_415_dilations_0"), val = tensor([1])]; + int32 x_415_groups_0 = const()[name = string("x_415_groups_0"), val = int32(1)]; + tensor encoder_layers_15_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(317139840))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(318188480))))[name = string("encoder_layers_15_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_415_cast_fp16 = conv(dilations = x_415_dilations_0, groups = x_415_groups_0, pad = x_415_pad_0, pad_type = x_415_pad_type_0, strides = x_415_strides_0, weight = encoder_layers_15_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_839_cast_fp16)[name = string("x_415_cast_fp16")]; + tensor input_841_perm_0 = const()[name = string("input_841_perm_0"), val = tensor([0, 2, 1])]; + tensor input_841_cast_fp16 = transpose(perm = input_841_perm_0, x = x_415_cast_fp16)[name = string("transpose_219")]; + tensor input_843_cast_fp16 = add(x = input_827_cast_fp16, y = input_841_cast_fp16)[name = string("input_843_cast_fp16")]; + tensor input_845_axes_0 = const()[name = string("input_845_axes_0"), val = tensor([-1])]; + tensor encoder_layers_15_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_15_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(318190592)))]; + tensor encoder_layers_15_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_15_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(318192704)))]; + tensor input_845_cast_fp16 = layer_norm(axes = input_845_axes_0, beta = encoder_layers_15_norm_feed_forward2_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_15_norm_feed_forward2_weight_to_fp16, x = input_843_cast_fp16)[name = string("input_845_cast_fp16")]; + tensor encoder_layers_15_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(318194816))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(321340608))))[name = string("encoder_layers_15_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_15_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_15_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(321340800)))]; + tensor linear_143_cast_fp16 = linear(bias = encoder_layers_15_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_15_feed_forward2_linear1_weight_to_fp16_palettized, x = input_845_cast_fp16)[name = string("linear_143_cast_fp16")]; + tensor input_849_cast_fp16 = silu(x = linear_143_cast_fp16)[name = string("input_849_cast_fp16")]; + tensor encoder_layers_15_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(321349056))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(324494848))))[name = string("encoder_layers_15_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_15_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_15_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(324495040)))]; + tensor linear_144_cast_fp16 = linear(bias = encoder_layers_15_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_15_feed_forward2_linear2_weight_to_fp16_palettized, x = input_849_cast_fp16)[name = string("linear_144_cast_fp16")]; + fp16 var_3825_to_fp16 = const()[name = string("op_3825_to_fp16"), val = fp16(0x1p-1)]; + tensor var_3826_cast_fp16 = mul(x = linear_144_cast_fp16, y = var_3825_to_fp16)[name = string("op_3826_cast_fp16")]; + tensor input_855_cast_fp16 = add(x = input_843_cast_fp16, y = var_3826_cast_fp16)[name = string("input_855_cast_fp16")]; + tensor input_857_axes_0 = const()[name = string("input_857_axes_0"), val = tensor([-1])]; + tensor encoder_layers_15_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_15_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(324497152)))]; + tensor encoder_layers_15_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_15_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(324499264)))]; + tensor input_857_cast_fp16 = layer_norm(axes = input_857_axes_0, beta = encoder_layers_15_norm_out_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_15_norm_out_weight_to_fp16, x = input_855_cast_fp16)[name = string("input_857_cast_fp16")]; + tensor cache_65_begin_0 = const()[name = string("cache_65_begin_0"), val = tensor([16, 0, 0, 0])]; + tensor cache_65_end_0 = const()[name = string("cache_65_end_0"), val = tensor([17, 1, 42, 1024])]; + tensor cache_65_end_mask_0 = const()[name = string("cache_65_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_65_squeeze_mask_0 = const()[name = string("cache_65_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_65_cast_fp16 = slice_by_index(begin = cache_65_begin_0, end = cache_65_end_0, end_mask = cache_65_end_mask_0, squeeze_mask = cache_65_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_65_cast_fp16")]; + tensor cache_67_begin_0 = const()[name = string("cache_67_begin_0"), val = tensor([16, 0, 0, 0])]; + tensor cache_67_end_0 = const()[name = string("cache_67_end_0"), val = tensor([17, 1, 1024, 8])]; + tensor cache_67_end_mask_0 = const()[name = string("cache_67_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_67_squeeze_mask_0 = const()[name = string("cache_67_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_67_cast_fp16 = slice_by_index(begin = cache_67_begin_0, end = cache_67_end_0, end_mask = cache_67_end_mask_0, squeeze_mask = cache_67_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_67_cast_fp16")]; + tensor input_859_axes_0 = const()[name = string("input_859_axes_0"), val = tensor([-1])]; + tensor encoder_layers_16_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_16_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(324501376)))]; + tensor encoder_layers_16_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_16_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(324503488)))]; + tensor input_859_cast_fp16 = layer_norm(axes = input_859_axes_0, beta = encoder_layers_16_norm_feed_forward1_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_16_norm_feed_forward1_weight_to_fp16, x = input_857_cast_fp16)[name = string("input_859_cast_fp16")]; + tensor encoder_layers_16_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(324505600))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(327651392))))[name = string("encoder_layers_16_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_16_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_16_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(327651584)))]; + tensor linear_145_cast_fp16 = linear(bias = encoder_layers_16_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_16_feed_forward1_linear1_weight_to_fp16_palettized, x = input_859_cast_fp16)[name = string("linear_145_cast_fp16")]; + tensor input_863_cast_fp16 = silu(x = linear_145_cast_fp16)[name = string("input_863_cast_fp16")]; + tensor encoder_layers_16_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(327659840))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(330805632))))[name = string("encoder_layers_16_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_16_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_16_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(330805824)))]; + tensor linear_146_cast_fp16 = linear(bias = encoder_layers_16_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_16_feed_forward1_linear2_weight_to_fp16_palettized, x = input_863_cast_fp16)[name = string("linear_146_cast_fp16")]; + fp16 var_3862_to_fp16 = const()[name = string("op_3862_to_fp16"), val = fp16(0x1p-1)]; + tensor var_3863_cast_fp16 = mul(x = linear_146_cast_fp16, y = var_3862_to_fp16)[name = string("op_3863_cast_fp16")]; + tensor input_869_cast_fp16 = add(x = input_857_cast_fp16, y = var_3863_cast_fp16)[name = string("input_869_cast_fp16")]; + tensor key_33_axes_0 = const()[name = string("key_33_axes_0"), val = tensor([-1])]; + tensor encoder_layers_16_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_16_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(330807936)))]; + tensor encoder_layers_16_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_16_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(330810048)))]; + tensor key_33_cast_fp16 = layer_norm(axes = key_33_axes_0, beta = encoder_layers_16_norm_self_att_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_16_norm_self_att_weight_to_fp16, x = input_869_cast_fp16)[name = string("key_33_cast_fp16")]; + bool input_871_interleave_0 = const()[name = string("input_871_interleave_0"), val = bool(false)]; + tensor input_871_cast_fp16 = concat(axis = var_67, interleave = input_871_interleave_0, values = (cache_65_cast_fp16, key_33_cast_fp16))[name = string("input_871_cast_fp16")]; + bool var_3891_interleave_0 = const()[name = string("op_3891_interleave_0"), val = bool(false)]; + tensor var_3891_cast_fp16 = concat(axis = var_67, interleave = var_3891_interleave_0, values = key_33_cast_fp16)[name = string("op_3891_cast_fp16")]; + tensor encoder_layers_16_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(330812160))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331598656))))[name = string("encoder_layers_16_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_16_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_16_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331598848)))]; + tensor linear_147_cast_fp16 = linear(bias = encoder_layers_16_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_16_self_attn_linear_q_weight_to_fp16_palettized, x = key_33_cast_fp16)[name = string("linear_147_cast_fp16")]; + tensor var_3896 = const()[name = string("op_3896"), val = tensor([1, -1, 8, 128])]; + tensor q_97_cast_fp16 = reshape(shape = var_3896, x = linear_147_cast_fp16)[name = string("q_97_cast_fp16")]; + tensor encoder_layers_16_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331600960))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332387456))))[name = string("encoder_layers_16_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_16_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_16_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332387648)))]; + tensor linear_148_cast_fp16 = linear(bias = encoder_layers_16_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_16_self_attn_linear_k_weight_to_fp16_palettized, x = input_871_cast_fp16)[name = string("linear_148_cast_fp16")]; + tensor var_3901 = const()[name = string("op_3901"), val = tensor([1, -1, 8, 128])]; + tensor k_65_cast_fp16 = reshape(shape = var_3901, x = linear_148_cast_fp16)[name = string("k_65_cast_fp16")]; + tensor encoder_layers_16_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332389760))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(333176256))))[name = string("encoder_layers_16_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_16_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_16_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(333176448)))]; + tensor linear_149_cast_fp16 = linear(bias = encoder_layers_16_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_16_self_attn_linear_v_weight_to_fp16_palettized, x = input_871_cast_fp16)[name = string("linear_149_cast_fp16")]; + tensor var_3906 = const()[name = string("op_3906"), val = tensor([1, -1, 8, 128])]; + tensor v_33_cast_fp16 = reshape(shape = var_3906, x = linear_149_cast_fp16)[name = string("v_33_cast_fp16")]; + tensor value_41_perm_0 = const()[name = string("value_41_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_16_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_16_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(333178560)))]; + tensor var_3919_cast_fp16 = add(x = q_97_cast_fp16, y = encoder_layers_16_self_attn_pos_bias_u_to_fp16)[name = string("op_3919_cast_fp16")]; + tensor encoder_layers_16_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_16_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(333180672)))]; + tensor var_3921_cast_fp16 = add(x = q_97_cast_fp16, y = encoder_layers_16_self_attn_pos_bias_v_to_fp16)[name = string("op_3921_cast_fp16")]; + tensor q_with_bias_v_33_perm_0 = const()[name = string("q_with_bias_v_33_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_423_transpose_x_0 = const()[name = string("x_423_transpose_x_0"), val = bool(false)]; + bool x_423_transpose_y_0 = const()[name = string("x_423_transpose_y_0"), val = bool(false)]; + tensor op_3923_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(333182784))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(333382528))))[name = string("op_3923_to_fp16_quantized")]; + tensor q_with_bias_v_33_cast_fp16 = transpose(perm = q_with_bias_v_33_perm_0, x = var_3921_cast_fp16)[name = string("transpose_218")]; + tensor x_423_cast_fp16 = matmul(transpose_x = x_423_transpose_x_0, transpose_y = x_423_transpose_y_0, x = q_with_bias_v_33_cast_fp16, y = op_3923_to_fp16_quantized)[name = string("x_423_cast_fp16")]; + tensor x_425_pad_0 = const()[name = string("x_425_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_425_mode_0 = const()[name = string("x_425_mode_0"), val = string("constant")]; + fp16 const_287_to_fp16 = const()[name = string("const_287_to_fp16"), val = fp16(0x0p+0)]; + tensor x_425_cast_fp16 = pad(constant_val = const_287_to_fp16, mode = x_425_mode_0, pad = x_425_pad_0, x = x_423_cast_fp16)[name = string("x_425_cast_fp16")]; + tensor var_3931 = const()[name = string("op_3931"), val = tensor([1, 8, -1, 56])]; + tensor x_427_cast_fp16 = reshape(shape = var_3931, x = x_425_cast_fp16)[name = string("x_427_cast_fp16")]; + tensor var_3935_begin_0 = const()[name = string("op_3935_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_3935_end_0 = const()[name = string("op_3935_end_0"), val = tensor([1, 8, 196, 56])]; + tensor var_3935_end_mask_0 = const()[name = string("op_3935_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_3935_cast_fp16 = slice_by_index(begin = var_3935_begin_0, end = var_3935_end_0, end_mask = var_3935_end_mask_0, x = x_427_cast_fp16)[name = string("op_3935_cast_fp16")]; + tensor var_3936 = const()[name = string("op_3936"), val = tensor([1, 8, 56, 195])]; + tensor matrix_bd_65_cast_fp16 = reshape(shape = var_3936, x = var_3935_cast_fp16)[name = string("matrix_bd_65_cast_fp16")]; + bool matrix_ac_33_transpose_x_0 = const()[name = string("matrix_ac_33_transpose_x_0"), val = bool(false)]; + bool matrix_ac_33_transpose_y_0 = const()[name = string("matrix_ac_33_transpose_y_0"), val = bool(false)]; + tensor transpose_128_perm_0 = const()[name = string("transpose_128_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_129_perm_0 = const()[name = string("transpose_129_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_129 = transpose(perm = transpose_129_perm_0, x = k_65_cast_fp16)[name = string("transpose_216")]; + tensor transpose_128 = transpose(perm = transpose_128_perm_0, x = var_3919_cast_fp16)[name = string("transpose_217")]; + tensor matrix_ac_33_cast_fp16 = matmul(transpose_x = matrix_ac_33_transpose_x_0, transpose_y = matrix_ac_33_transpose_y_0, x = transpose_128, y = transpose_129)[name = string("matrix_ac_33_cast_fp16")]; + tensor matrix_bd_67_begin_0 = const()[name = string("matrix_bd_67_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_67_end_0 = const()[name = string("matrix_bd_67_end_0"), val = tensor([1, 8, 56, 98])]; + tensor matrix_bd_67_end_mask_0 = const()[name = string("matrix_bd_67_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_67_cast_fp16 = slice_by_index(begin = matrix_bd_67_begin_0, end = matrix_bd_67_end_0, end_mask = matrix_bd_67_end_mask_0, x = matrix_bd_65_cast_fp16)[name = string("matrix_bd_67_cast_fp16")]; + tensor var_3945_cast_fp16 = add(x = matrix_ac_33_cast_fp16, y = matrix_bd_67_cast_fp16)[name = string("op_3945_cast_fp16")]; + fp16 _inversed_scores_65_y_0_to_fp16 = const()[name = string("_inversed_scores_65_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_65_cast_fp16 = mul(x = var_3945_cast_fp16, y = _inversed_scores_65_y_0_to_fp16)[name = string("_inversed_scores_65_cast_fp16")]; + tensor scores_67_cast_fp16 = select(a = var_44_to_fp16, b = _inversed_scores_65_cast_fp16, cond = mask_11)[name = string("scores_67_cast_fp16")]; + tensor var_3951_cast_fp16 = softmax(axis = var_58, x = scores_67_cast_fp16)[name = string("op_3951_cast_fp16")]; + tensor input_873_cast_fp16 = select(a = var_43_to_fp16, b = var_3951_cast_fp16, cond = mask_11)[name = string("input_873_cast_fp16")]; + bool x_429_transpose_x_0 = const()[name = string("x_429_transpose_x_0"), val = bool(false)]; + bool x_429_transpose_y_0 = const()[name = string("x_429_transpose_y_0"), val = bool(false)]; + tensor value_41_cast_fp16 = transpose(perm = value_41_perm_0, x = v_33_cast_fp16)[name = string("transpose_215")]; + tensor x_429_cast_fp16 = matmul(transpose_x = x_429_transpose_x_0, transpose_y = x_429_transpose_y_0, x = input_873_cast_fp16, y = value_41_cast_fp16)[name = string("x_429_cast_fp16")]; + tensor var_3955_perm_0 = const()[name = string("op_3955_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_3956 = const()[name = string("op_3956"), val = tensor([1, -1, 1024])]; + tensor var_3955_cast_fp16 = transpose(perm = var_3955_perm_0, x = x_429_cast_fp16)[name = string("transpose_214")]; + tensor input_875_cast_fp16 = reshape(shape = var_3956, x = var_3955_cast_fp16)[name = string("input_875_cast_fp16")]; + tensor encoder_layers_16_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(333383040))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(334169536))))[name = string("encoder_layers_16_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_16_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_16_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(334169728)))]; + tensor linear_151_cast_fp16 = linear(bias = encoder_layers_16_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_16_self_attn_linear_out_weight_to_fp16_palettized, x = input_875_cast_fp16)[name = string("linear_151_cast_fp16")]; + tensor input_879_cast_fp16 = add(x = input_869_cast_fp16, y = linear_151_cast_fp16)[name = string("input_879_cast_fp16")]; + tensor x_433_axes_0 = const()[name = string("x_433_axes_0"), val = tensor([-1])]; + tensor encoder_layers_16_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_16_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(334171840)))]; + tensor encoder_layers_16_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_16_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(334173952)))]; + tensor x_433_cast_fp16 = layer_norm(axes = x_433_axes_0, beta = encoder_layers_16_norm_conv_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_16_norm_conv_weight_to_fp16, x = input_879_cast_fp16)[name = string("x_433_cast_fp16")]; + tensor input_881_perm_0 = const()[name = string("input_881_perm_0"), val = tensor([0, 2, 1])]; + string input_883_pad_type_0 = const()[name = string("input_883_pad_type_0"), val = string("valid")]; + tensor input_883_strides_0 = const()[name = string("input_883_strides_0"), val = tensor([1])]; + tensor input_883_pad_0 = const()[name = string("input_883_pad_0"), val = tensor([0, 0])]; + tensor input_883_dilations_0 = const()[name = string("input_883_dilations_0"), val = tensor([1])]; + int32 input_883_groups_0 = const()[name = string("input_883_groups_0"), val = int32(1)]; + tensor encoder_layers_16_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(334176064))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(336273280))))[name = string("encoder_layers_16_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_881_cast_fp16 = transpose(perm = input_881_perm_0, x = x_433_cast_fp16)[name = string("transpose_213")]; + tensor input_883_cast_fp16 = conv(dilations = input_883_dilations_0, groups = input_883_groups_0, pad = input_883_pad_0, pad_type = input_883_pad_type_0, strides = input_883_strides_0, weight = encoder_layers_16_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_881_cast_fp16)[name = string("input_883_cast_fp16")]; + int32 x_435_split_num_splits_0 = const()[name = string("x_435_split_num_splits_0"), val = int32(2)]; + int32 x_435_split_axis_0 = const()[name = string("x_435_split_axis_0"), val = int32(1)]; + tensor x_435_split_cast_fp16_0, tensor x_435_split_cast_fp16_1 = split(axis = x_435_split_axis_0, num_splits = x_435_split_num_splits_0, x = input_883_cast_fp16)[name = string("x_435_split_cast_fp16")]; + tensor x_435_split_1_sigmoid_cast_fp16 = sigmoid(x = x_435_split_cast_fp16_1)[name = string("x_435_split_1_sigmoid_cast_fp16")]; + tensor x_435_cast_fp16 = mul(x = x_435_split_cast_fp16_0, y = x_435_split_1_sigmoid_cast_fp16)[name = string("x_435_cast_fp16")]; + tensor input_885_cast_fp16 = select(a = var_43_to_fp16, b = x_435_cast_fp16, cond = var_574)[name = string("input_885_cast_fp16")]; + bool new_x_67_interleave_0 = const()[name = string("new_x_67_interleave_0"), val = bool(false)]; + tensor new_x_67_cast_fp16 = concat(axis = var_58, interleave = new_x_67_interleave_0, values = (cache_67_cast_fp16, input_885_cast_fp16))[name = string("new_x_67_cast_fp16")]; + tensor var_3995_begin_0 = const()[name = string("op_3995_begin_0"), val = tensor([0, 0, 56])]; + tensor var_3995_end_0 = const()[name = string("op_3995_end_0"), val = tensor([1, 1024, 64])]; + tensor var_3995_end_mask_0 = const()[name = string("op_3995_end_mask_0"), val = tensor([true, true, true])]; + tensor var_3995_cast_fp16 = slice_by_index(begin = var_3995_begin_0, end = var_3995_end_0, end_mask = var_3995_end_mask_0, x = new_x_67_cast_fp16)[name = string("op_3995_cast_fp16")]; + string x_437_pad_type_0 = const()[name = string("x_437_pad_type_0"), val = string("valid")]; + int32 x_437_groups_0 = const()[name = string("x_437_groups_0"), val = int32(1024)]; + tensor x_437_strides_0 = const()[name = string("x_437_strides_0"), val = tensor([1])]; + tensor x_437_pad_0 = const()[name = string("x_437_pad_0"), val = tensor([0, 0])]; + tensor x_437_dilations_0 = const()[name = string("x_437_dilations_0"), val = tensor([1])]; + tensor encoder_layers_16_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(336277440))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(336286720))))[name = string("encoder_layers_16_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_437_cast_fp16 = conv(dilations = x_437_dilations_0, groups = x_437_groups_0, pad = x_437_pad_0, pad_type = x_437_pad_type_0, strides = x_437_strides_0, weight = encoder_layers_16_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_67_cast_fp16)[name = string("x_437_cast_fp16")]; + tensor input_887_perm_0 = const()[name = string("input_887_perm_0"), val = tensor([0, 2, 1])]; + tensor x_439_axes_0 = const()[name = string("x_439_axes_0"), val = tensor([-1])]; + tensor encoder_layers_16_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_16_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(336288832)))]; + tensor encoder_layers_16_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_16_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(336290944)))]; + tensor input_887_cast_fp16 = transpose(perm = input_887_perm_0, x = x_437_cast_fp16)[name = string("transpose_212")]; + tensor x_439_cast_fp16 = layer_norm(axes = x_439_axes_0, beta = encoder_layers_16_conv_batch_norm_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_16_conv_batch_norm_weight_to_fp16, x = input_887_cast_fp16)[name = string("x_439_cast_fp16")]; + tensor input_889_perm_0 = const()[name = string("input_889_perm_0"), val = tensor([0, 2, 1])]; + tensor input_889_cast_fp16 = transpose(perm = input_889_perm_0, x = x_439_cast_fp16)[name = string("transpose_211")]; + tensor input_891_cast_fp16 = silu(x = input_889_cast_fp16)[name = string("input_891_cast_fp16")]; + string x_441_pad_type_0 = const()[name = string("x_441_pad_type_0"), val = string("valid")]; + tensor x_441_strides_0 = const()[name = string("x_441_strides_0"), val = tensor([1])]; + tensor x_441_pad_0 = const()[name = string("x_441_pad_0"), val = tensor([0, 0])]; + tensor x_441_dilations_0 = const()[name = string("x_441_dilations_0"), val = tensor([1])]; + int32 x_441_groups_0 = const()[name = string("x_441_groups_0"), val = int32(1)]; + tensor encoder_layers_16_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(336293056))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(337341696))))[name = string("encoder_layers_16_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_441_cast_fp16 = conv(dilations = x_441_dilations_0, groups = x_441_groups_0, pad = x_441_pad_0, pad_type = x_441_pad_type_0, strides = x_441_strides_0, weight = encoder_layers_16_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_891_cast_fp16)[name = string("x_441_cast_fp16")]; + tensor input_893_perm_0 = const()[name = string("input_893_perm_0"), val = tensor([0, 2, 1])]; + tensor input_893_cast_fp16 = transpose(perm = input_893_perm_0, x = x_441_cast_fp16)[name = string("transpose_210")]; + tensor input_895_cast_fp16 = add(x = input_879_cast_fp16, y = input_893_cast_fp16)[name = string("input_895_cast_fp16")]; + tensor input_897_axes_0 = const()[name = string("input_897_axes_0"), val = tensor([-1])]; + tensor encoder_layers_16_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_16_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(337343808)))]; + tensor encoder_layers_16_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_16_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(337345920)))]; + tensor input_897_cast_fp16 = layer_norm(axes = input_897_axes_0, beta = encoder_layers_16_norm_feed_forward2_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_16_norm_feed_forward2_weight_to_fp16, x = input_895_cast_fp16)[name = string("input_897_cast_fp16")]; + tensor encoder_layers_16_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(337348032))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(340493824))))[name = string("encoder_layers_16_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_16_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_16_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(340494016)))]; + tensor linear_152_cast_fp16 = linear(bias = encoder_layers_16_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_16_feed_forward2_linear1_weight_to_fp16_palettized, x = input_897_cast_fp16)[name = string("linear_152_cast_fp16")]; + tensor input_901_cast_fp16 = silu(x = linear_152_cast_fp16)[name = string("input_901_cast_fp16")]; + tensor encoder_layers_16_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(340502272))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(343648064))))[name = string("encoder_layers_16_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_16_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_16_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(343648256)))]; + tensor linear_153_cast_fp16 = linear(bias = encoder_layers_16_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_16_feed_forward2_linear2_weight_to_fp16_palettized, x = input_901_cast_fp16)[name = string("linear_153_cast_fp16")]; + fp16 var_4038_to_fp16 = const()[name = string("op_4038_to_fp16"), val = fp16(0x1p-1)]; + tensor var_4039_cast_fp16 = mul(x = linear_153_cast_fp16, y = var_4038_to_fp16)[name = string("op_4039_cast_fp16")]; + tensor input_907_cast_fp16 = add(x = input_895_cast_fp16, y = var_4039_cast_fp16)[name = string("input_907_cast_fp16")]; + tensor input_909_axes_0 = const()[name = string("input_909_axes_0"), val = tensor([-1])]; + tensor encoder_layers_16_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_16_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(343650368)))]; + tensor encoder_layers_16_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_16_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(343652480)))]; + tensor input_909_cast_fp16 = layer_norm(axes = input_909_axes_0, beta = encoder_layers_16_norm_out_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_16_norm_out_weight_to_fp16, x = input_907_cast_fp16)[name = string("input_909_cast_fp16")]; + tensor cache_69_begin_0 = const()[name = string("cache_69_begin_0"), val = tensor([17, 0, 0, 0])]; + tensor cache_69_end_0 = const()[name = string("cache_69_end_0"), val = tensor([18, 1, 42, 1024])]; + tensor cache_69_end_mask_0 = const()[name = string("cache_69_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_69_squeeze_mask_0 = const()[name = string("cache_69_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_69_cast_fp16 = slice_by_index(begin = cache_69_begin_0, end = cache_69_end_0, end_mask = cache_69_end_mask_0, squeeze_mask = cache_69_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_69_cast_fp16")]; + tensor cache_71_begin_0 = const()[name = string("cache_71_begin_0"), val = tensor([17, 0, 0, 0])]; + tensor cache_71_end_0 = const()[name = string("cache_71_end_0"), val = tensor([18, 1, 1024, 8])]; + tensor cache_71_end_mask_0 = const()[name = string("cache_71_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_71_squeeze_mask_0 = const()[name = string("cache_71_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_71_cast_fp16 = slice_by_index(begin = cache_71_begin_0, end = cache_71_end_0, end_mask = cache_71_end_mask_0, squeeze_mask = cache_71_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_71_cast_fp16")]; + tensor input_911_axes_0 = const()[name = string("input_911_axes_0"), val = tensor([-1])]; + tensor encoder_layers_17_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_17_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(343654592)))]; + tensor encoder_layers_17_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_17_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(343656704)))]; + tensor input_911_cast_fp16 = layer_norm(axes = input_911_axes_0, beta = encoder_layers_17_norm_feed_forward1_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_17_norm_feed_forward1_weight_to_fp16, x = input_909_cast_fp16)[name = string("input_911_cast_fp16")]; + tensor encoder_layers_17_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(343658816))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346804608))))[name = string("encoder_layers_17_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_17_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_17_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346804800)))]; + tensor linear_154_cast_fp16 = linear(bias = encoder_layers_17_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_17_feed_forward1_linear1_weight_to_fp16_palettized, x = input_911_cast_fp16)[name = string("linear_154_cast_fp16")]; + tensor input_915_cast_fp16 = silu(x = linear_154_cast_fp16)[name = string("input_915_cast_fp16")]; + tensor encoder_layers_17_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346813056))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(349958848))))[name = string("encoder_layers_17_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_17_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_17_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(349959040)))]; + tensor linear_155_cast_fp16 = linear(bias = encoder_layers_17_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_17_feed_forward1_linear2_weight_to_fp16_palettized, x = input_915_cast_fp16)[name = string("linear_155_cast_fp16")]; + fp16 var_4075_to_fp16 = const()[name = string("op_4075_to_fp16"), val = fp16(0x1p-1)]; + tensor var_4076_cast_fp16 = mul(x = linear_155_cast_fp16, y = var_4075_to_fp16)[name = string("op_4076_cast_fp16")]; + tensor input_921_cast_fp16 = add(x = input_909_cast_fp16, y = var_4076_cast_fp16)[name = string("input_921_cast_fp16")]; + tensor key_35_axes_0 = const()[name = string("key_35_axes_0"), val = tensor([-1])]; + tensor encoder_layers_17_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_17_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(349961152)))]; + tensor encoder_layers_17_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_17_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(349963264)))]; + tensor key_35_cast_fp16 = layer_norm(axes = key_35_axes_0, beta = encoder_layers_17_norm_self_att_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_17_norm_self_att_weight_to_fp16, x = input_921_cast_fp16)[name = string("key_35_cast_fp16")]; + bool input_923_interleave_0 = const()[name = string("input_923_interleave_0"), val = bool(false)]; + tensor input_923_cast_fp16 = concat(axis = var_67, interleave = input_923_interleave_0, values = (cache_69_cast_fp16, key_35_cast_fp16))[name = string("input_923_cast_fp16")]; + bool var_4104_interleave_0 = const()[name = string("op_4104_interleave_0"), val = bool(false)]; + tensor var_4104_cast_fp16 = concat(axis = var_67, interleave = var_4104_interleave_0, values = key_35_cast_fp16)[name = string("op_4104_cast_fp16")]; + tensor encoder_layers_17_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(349965376))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350751872))))[name = string("encoder_layers_17_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_17_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_17_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350752064)))]; + tensor linear_156_cast_fp16 = linear(bias = encoder_layers_17_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_17_self_attn_linear_q_weight_to_fp16_palettized, x = key_35_cast_fp16)[name = string("linear_156_cast_fp16")]; + tensor var_4109 = const()[name = string("op_4109"), val = tensor([1, -1, 8, 128])]; + tensor q_103_cast_fp16 = reshape(shape = var_4109, x = linear_156_cast_fp16)[name = string("q_103_cast_fp16")]; + tensor encoder_layers_17_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350754176))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351540672))))[name = string("encoder_layers_17_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_17_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_17_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351540864)))]; + tensor linear_157_cast_fp16 = linear(bias = encoder_layers_17_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_17_self_attn_linear_k_weight_to_fp16_palettized, x = input_923_cast_fp16)[name = string("linear_157_cast_fp16")]; + tensor var_4114 = const()[name = string("op_4114"), val = tensor([1, -1, 8, 128])]; + tensor k_69_cast_fp16 = reshape(shape = var_4114, x = linear_157_cast_fp16)[name = string("k_69_cast_fp16")]; + tensor encoder_layers_17_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351542976))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(352329472))))[name = string("encoder_layers_17_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_17_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_17_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(352329664)))]; + tensor linear_158_cast_fp16 = linear(bias = encoder_layers_17_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_17_self_attn_linear_v_weight_to_fp16_palettized, x = input_923_cast_fp16)[name = string("linear_158_cast_fp16")]; + tensor var_4119 = const()[name = string("op_4119"), val = tensor([1, -1, 8, 128])]; + tensor v_35_cast_fp16 = reshape(shape = var_4119, x = linear_158_cast_fp16)[name = string("v_35_cast_fp16")]; + tensor value_43_perm_0 = const()[name = string("value_43_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_17_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_17_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(352331776)))]; + tensor var_4132_cast_fp16 = add(x = q_103_cast_fp16, y = encoder_layers_17_self_attn_pos_bias_u_to_fp16)[name = string("op_4132_cast_fp16")]; + tensor encoder_layers_17_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_17_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(352333888)))]; + tensor var_4134_cast_fp16 = add(x = q_103_cast_fp16, y = encoder_layers_17_self_attn_pos_bias_v_to_fp16)[name = string("op_4134_cast_fp16")]; + tensor q_with_bias_v_35_perm_0 = const()[name = string("q_with_bias_v_35_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_449_transpose_x_0 = const()[name = string("x_449_transpose_x_0"), val = bool(false)]; + bool x_449_transpose_y_0 = const()[name = string("x_449_transpose_y_0"), val = bool(false)]; + tensor op_4136_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(352336000))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(352535744))))[name = string("op_4136_to_fp16_quantized")]; + tensor q_with_bias_v_35_cast_fp16 = transpose(perm = q_with_bias_v_35_perm_0, x = var_4134_cast_fp16)[name = string("transpose_209")]; + tensor x_449_cast_fp16 = matmul(transpose_x = x_449_transpose_x_0, transpose_y = x_449_transpose_y_0, x = q_with_bias_v_35_cast_fp16, y = op_4136_to_fp16_quantized)[name = string("x_449_cast_fp16")]; + tensor x_451_pad_0 = const()[name = string("x_451_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_451_mode_0 = const()[name = string("x_451_mode_0"), val = string("constant")]; + fp16 const_300_to_fp16 = const()[name = string("const_300_to_fp16"), val = fp16(0x0p+0)]; + tensor x_451_cast_fp16 = pad(constant_val = const_300_to_fp16, mode = x_451_mode_0, pad = x_451_pad_0, x = x_449_cast_fp16)[name = string("x_451_cast_fp16")]; + tensor var_4144 = const()[name = string("op_4144"), val = tensor([1, 8, -1, 56])]; + tensor x_453_cast_fp16 = reshape(shape = var_4144, x = x_451_cast_fp16)[name = string("x_453_cast_fp16")]; + tensor var_4148_begin_0 = const()[name = string("op_4148_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_4148_end_0 = const()[name = string("op_4148_end_0"), val = tensor([1, 8, 196, 56])]; + tensor var_4148_end_mask_0 = const()[name = string("op_4148_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_4148_cast_fp16 = slice_by_index(begin = var_4148_begin_0, end = var_4148_end_0, end_mask = var_4148_end_mask_0, x = x_453_cast_fp16)[name = string("op_4148_cast_fp16")]; + tensor var_4149 = const()[name = string("op_4149"), val = tensor([1, 8, 56, 195])]; + tensor matrix_bd_69_cast_fp16 = reshape(shape = var_4149, x = var_4148_cast_fp16)[name = string("matrix_bd_69_cast_fp16")]; + bool matrix_ac_35_transpose_x_0 = const()[name = string("matrix_ac_35_transpose_x_0"), val = bool(false)]; + bool matrix_ac_35_transpose_y_0 = const()[name = string("matrix_ac_35_transpose_y_0"), val = bool(false)]; + tensor transpose_130_perm_0 = const()[name = string("transpose_130_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_131_perm_0 = const()[name = string("transpose_131_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_131 = transpose(perm = transpose_131_perm_0, x = k_69_cast_fp16)[name = string("transpose_207")]; + tensor transpose_130 = transpose(perm = transpose_130_perm_0, x = var_4132_cast_fp16)[name = string("transpose_208")]; + tensor matrix_ac_35_cast_fp16 = matmul(transpose_x = matrix_ac_35_transpose_x_0, transpose_y = matrix_ac_35_transpose_y_0, x = transpose_130, y = transpose_131)[name = string("matrix_ac_35_cast_fp16")]; + tensor matrix_bd_71_begin_0 = const()[name = string("matrix_bd_71_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_71_end_0 = const()[name = string("matrix_bd_71_end_0"), val = tensor([1, 8, 56, 98])]; + tensor matrix_bd_71_end_mask_0 = const()[name = string("matrix_bd_71_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_71_cast_fp16 = slice_by_index(begin = matrix_bd_71_begin_0, end = matrix_bd_71_end_0, end_mask = matrix_bd_71_end_mask_0, x = matrix_bd_69_cast_fp16)[name = string("matrix_bd_71_cast_fp16")]; + tensor var_4158_cast_fp16 = add(x = matrix_ac_35_cast_fp16, y = matrix_bd_71_cast_fp16)[name = string("op_4158_cast_fp16")]; + fp16 _inversed_scores_69_y_0_to_fp16 = const()[name = string("_inversed_scores_69_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_69_cast_fp16 = mul(x = var_4158_cast_fp16, y = _inversed_scores_69_y_0_to_fp16)[name = string("_inversed_scores_69_cast_fp16")]; + tensor scores_71_cast_fp16 = select(a = var_44_to_fp16, b = _inversed_scores_69_cast_fp16, cond = mask_11)[name = string("scores_71_cast_fp16")]; + tensor var_4164_cast_fp16 = softmax(axis = var_58, x = scores_71_cast_fp16)[name = string("op_4164_cast_fp16")]; + tensor input_925_cast_fp16 = select(a = var_43_to_fp16, b = var_4164_cast_fp16, cond = mask_11)[name = string("input_925_cast_fp16")]; + bool x_455_transpose_x_0 = const()[name = string("x_455_transpose_x_0"), val = bool(false)]; + bool x_455_transpose_y_0 = const()[name = string("x_455_transpose_y_0"), val = bool(false)]; + tensor value_43_cast_fp16 = transpose(perm = value_43_perm_0, x = v_35_cast_fp16)[name = string("transpose_206")]; + tensor x_455_cast_fp16 = matmul(transpose_x = x_455_transpose_x_0, transpose_y = x_455_transpose_y_0, x = input_925_cast_fp16, y = value_43_cast_fp16)[name = string("x_455_cast_fp16")]; + tensor var_4168_perm_0 = const()[name = string("op_4168_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_4169 = const()[name = string("op_4169"), val = tensor([1, -1, 1024])]; + tensor var_4168_cast_fp16 = transpose(perm = var_4168_perm_0, x = x_455_cast_fp16)[name = string("transpose_205")]; + tensor input_927_cast_fp16 = reshape(shape = var_4169, x = var_4168_cast_fp16)[name = string("input_927_cast_fp16")]; + tensor encoder_layers_17_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(352536256))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(353322752))))[name = string("encoder_layers_17_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_17_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_17_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(353322944)))]; + tensor linear_160_cast_fp16 = linear(bias = encoder_layers_17_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_17_self_attn_linear_out_weight_to_fp16_palettized, x = input_927_cast_fp16)[name = string("linear_160_cast_fp16")]; + tensor input_931_cast_fp16 = add(x = input_921_cast_fp16, y = linear_160_cast_fp16)[name = string("input_931_cast_fp16")]; + tensor x_459_axes_0 = const()[name = string("x_459_axes_0"), val = tensor([-1])]; + tensor encoder_layers_17_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_17_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(353325056)))]; + tensor encoder_layers_17_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_17_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(353327168)))]; + tensor x_459_cast_fp16 = layer_norm(axes = x_459_axes_0, beta = encoder_layers_17_norm_conv_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_17_norm_conv_weight_to_fp16, x = input_931_cast_fp16)[name = string("x_459_cast_fp16")]; + tensor input_933_perm_0 = const()[name = string("input_933_perm_0"), val = tensor([0, 2, 1])]; + string input_935_pad_type_0 = const()[name = string("input_935_pad_type_0"), val = string("valid")]; + tensor input_935_strides_0 = const()[name = string("input_935_strides_0"), val = tensor([1])]; + tensor input_935_pad_0 = const()[name = string("input_935_pad_0"), val = tensor([0, 0])]; + tensor input_935_dilations_0 = const()[name = string("input_935_dilations_0"), val = tensor([1])]; + int32 input_935_groups_0 = const()[name = string("input_935_groups_0"), val = int32(1)]; + tensor encoder_layers_17_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(353329280))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(355426496))))[name = string("encoder_layers_17_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_933_cast_fp16 = transpose(perm = input_933_perm_0, x = x_459_cast_fp16)[name = string("transpose_204")]; + tensor input_935_cast_fp16 = conv(dilations = input_935_dilations_0, groups = input_935_groups_0, pad = input_935_pad_0, pad_type = input_935_pad_type_0, strides = input_935_strides_0, weight = encoder_layers_17_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_933_cast_fp16)[name = string("input_935_cast_fp16")]; + int32 x_461_split_num_splits_0 = const()[name = string("x_461_split_num_splits_0"), val = int32(2)]; + int32 x_461_split_axis_0 = const()[name = string("x_461_split_axis_0"), val = int32(1)]; + tensor x_461_split_cast_fp16_0, tensor x_461_split_cast_fp16_1 = split(axis = x_461_split_axis_0, num_splits = x_461_split_num_splits_0, x = input_935_cast_fp16)[name = string("x_461_split_cast_fp16")]; + tensor x_461_split_1_sigmoid_cast_fp16 = sigmoid(x = x_461_split_cast_fp16_1)[name = string("x_461_split_1_sigmoid_cast_fp16")]; + tensor x_461_cast_fp16 = mul(x = x_461_split_cast_fp16_0, y = x_461_split_1_sigmoid_cast_fp16)[name = string("x_461_cast_fp16")]; + tensor input_937_cast_fp16 = select(a = var_43_to_fp16, b = x_461_cast_fp16, cond = var_574)[name = string("input_937_cast_fp16")]; + bool new_x_71_interleave_0 = const()[name = string("new_x_71_interleave_0"), val = bool(false)]; + tensor new_x_71_cast_fp16 = concat(axis = var_58, interleave = new_x_71_interleave_0, values = (cache_71_cast_fp16, input_937_cast_fp16))[name = string("new_x_71_cast_fp16")]; + tensor var_4208_begin_0 = const()[name = string("op_4208_begin_0"), val = tensor([0, 0, 56])]; + tensor var_4208_end_0 = const()[name = string("op_4208_end_0"), val = tensor([1, 1024, 64])]; + tensor var_4208_end_mask_0 = const()[name = string("op_4208_end_mask_0"), val = tensor([true, true, true])]; + tensor var_4208_cast_fp16 = slice_by_index(begin = var_4208_begin_0, end = var_4208_end_0, end_mask = var_4208_end_mask_0, x = new_x_71_cast_fp16)[name = string("op_4208_cast_fp16")]; + string x_463_pad_type_0 = const()[name = string("x_463_pad_type_0"), val = string("valid")]; + int32 x_463_groups_0 = const()[name = string("x_463_groups_0"), val = int32(1024)]; + tensor x_463_strides_0 = const()[name = string("x_463_strides_0"), val = tensor([1])]; + tensor x_463_pad_0 = const()[name = string("x_463_pad_0"), val = tensor([0, 0])]; + tensor x_463_dilations_0 = const()[name = string("x_463_dilations_0"), val = tensor([1])]; + tensor encoder_layers_17_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(355430656))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(355439936))))[name = string("encoder_layers_17_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_463_cast_fp16 = conv(dilations = x_463_dilations_0, groups = x_463_groups_0, pad = x_463_pad_0, pad_type = x_463_pad_type_0, strides = x_463_strides_0, weight = encoder_layers_17_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_71_cast_fp16)[name = string("x_463_cast_fp16")]; + tensor input_939_perm_0 = const()[name = string("input_939_perm_0"), val = tensor([0, 2, 1])]; + tensor x_465_axes_0 = const()[name = string("x_465_axes_0"), val = tensor([-1])]; + tensor encoder_layers_17_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_17_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(355442048)))]; + tensor encoder_layers_17_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_17_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(355444160)))]; + tensor input_939_cast_fp16 = transpose(perm = input_939_perm_0, x = x_463_cast_fp16)[name = string("transpose_203")]; + tensor x_465_cast_fp16 = layer_norm(axes = x_465_axes_0, beta = encoder_layers_17_conv_batch_norm_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_17_conv_batch_norm_weight_to_fp16, x = input_939_cast_fp16)[name = string("x_465_cast_fp16")]; + tensor input_941_perm_0 = const()[name = string("input_941_perm_0"), val = tensor([0, 2, 1])]; + tensor input_941_cast_fp16 = transpose(perm = input_941_perm_0, x = x_465_cast_fp16)[name = string("transpose_202")]; + tensor input_943_cast_fp16 = silu(x = input_941_cast_fp16)[name = string("input_943_cast_fp16")]; + string x_467_pad_type_0 = const()[name = string("x_467_pad_type_0"), val = string("valid")]; + tensor x_467_strides_0 = const()[name = string("x_467_strides_0"), val = tensor([1])]; + tensor x_467_pad_0 = const()[name = string("x_467_pad_0"), val = tensor([0, 0])]; + tensor x_467_dilations_0 = const()[name = string("x_467_dilations_0"), val = tensor([1])]; + int32 x_467_groups_0 = const()[name = string("x_467_groups_0"), val = int32(1)]; + tensor encoder_layers_17_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(355446272))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(356494912))))[name = string("encoder_layers_17_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_467_cast_fp16 = conv(dilations = x_467_dilations_0, groups = x_467_groups_0, pad = x_467_pad_0, pad_type = x_467_pad_type_0, strides = x_467_strides_0, weight = encoder_layers_17_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_943_cast_fp16)[name = string("x_467_cast_fp16")]; + tensor input_945_perm_0 = const()[name = string("input_945_perm_0"), val = tensor([0, 2, 1])]; + tensor input_945_cast_fp16 = transpose(perm = input_945_perm_0, x = x_467_cast_fp16)[name = string("transpose_201")]; + tensor input_947_cast_fp16 = add(x = input_931_cast_fp16, y = input_945_cast_fp16)[name = string("input_947_cast_fp16")]; + tensor input_949_axes_0 = const()[name = string("input_949_axes_0"), val = tensor([-1])]; + tensor encoder_layers_17_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_17_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(356497024)))]; + tensor encoder_layers_17_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_17_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(356499136)))]; + tensor input_949_cast_fp16 = layer_norm(axes = input_949_axes_0, beta = encoder_layers_17_norm_feed_forward2_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_17_norm_feed_forward2_weight_to_fp16, x = input_947_cast_fp16)[name = string("input_949_cast_fp16")]; + tensor encoder_layers_17_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(356501248))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(359647040))))[name = string("encoder_layers_17_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_17_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_17_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(359647232)))]; + tensor linear_161_cast_fp16 = linear(bias = encoder_layers_17_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_17_feed_forward2_linear1_weight_to_fp16_palettized, x = input_949_cast_fp16)[name = string("linear_161_cast_fp16")]; + tensor input_953_cast_fp16 = silu(x = linear_161_cast_fp16)[name = string("input_953_cast_fp16")]; + tensor encoder_layers_17_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(359655488))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(362801280))))[name = string("encoder_layers_17_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_17_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_17_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(362801472)))]; + tensor linear_162_cast_fp16 = linear(bias = encoder_layers_17_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_17_feed_forward2_linear2_weight_to_fp16_palettized, x = input_953_cast_fp16)[name = string("linear_162_cast_fp16")]; + fp16 var_4251_to_fp16 = const()[name = string("op_4251_to_fp16"), val = fp16(0x1p-1)]; + tensor var_4252_cast_fp16 = mul(x = linear_162_cast_fp16, y = var_4251_to_fp16)[name = string("op_4252_cast_fp16")]; + tensor input_959_cast_fp16 = add(x = input_947_cast_fp16, y = var_4252_cast_fp16)[name = string("input_959_cast_fp16")]; + tensor input_961_axes_0 = const()[name = string("input_961_axes_0"), val = tensor([-1])]; + tensor encoder_layers_17_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_17_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(362803584)))]; + tensor encoder_layers_17_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_17_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(362805696)))]; + tensor input_961_cast_fp16 = layer_norm(axes = input_961_axes_0, beta = encoder_layers_17_norm_out_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_17_norm_out_weight_to_fp16, x = input_959_cast_fp16)[name = string("input_961_cast_fp16")]; + tensor cache_73_begin_0 = const()[name = string("cache_73_begin_0"), val = tensor([18, 0, 0, 0])]; + tensor cache_73_end_0 = const()[name = string("cache_73_end_0"), val = tensor([19, 1, 42, 1024])]; + tensor cache_73_end_mask_0 = const()[name = string("cache_73_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_73_squeeze_mask_0 = const()[name = string("cache_73_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_73_cast_fp16 = slice_by_index(begin = cache_73_begin_0, end = cache_73_end_0, end_mask = cache_73_end_mask_0, squeeze_mask = cache_73_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_73_cast_fp16")]; + tensor cache_75_begin_0 = const()[name = string("cache_75_begin_0"), val = tensor([18, 0, 0, 0])]; + tensor cache_75_end_0 = const()[name = string("cache_75_end_0"), val = tensor([19, 1, 1024, 8])]; + tensor cache_75_end_mask_0 = const()[name = string("cache_75_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_75_squeeze_mask_0 = const()[name = string("cache_75_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_75_cast_fp16 = slice_by_index(begin = cache_75_begin_0, end = cache_75_end_0, end_mask = cache_75_end_mask_0, squeeze_mask = cache_75_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_75_cast_fp16")]; + tensor input_963_axes_0 = const()[name = string("input_963_axes_0"), val = tensor([-1])]; + tensor encoder_layers_18_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_18_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(362807808)))]; + tensor encoder_layers_18_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_18_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(362809920)))]; + tensor input_963_cast_fp16 = layer_norm(axes = input_963_axes_0, beta = encoder_layers_18_norm_feed_forward1_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_18_norm_feed_forward1_weight_to_fp16, x = input_961_cast_fp16)[name = string("input_963_cast_fp16")]; + tensor encoder_layers_18_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(362812032))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(365957824))))[name = string("encoder_layers_18_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_18_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_18_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(365958016)))]; + tensor linear_163_cast_fp16 = linear(bias = encoder_layers_18_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_18_feed_forward1_linear1_weight_to_fp16_palettized, x = input_963_cast_fp16)[name = string("linear_163_cast_fp16")]; + tensor input_967_cast_fp16 = silu(x = linear_163_cast_fp16)[name = string("input_967_cast_fp16")]; + tensor encoder_layers_18_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(365966272))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369112064))))[name = string("encoder_layers_18_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_18_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_18_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369112256)))]; + tensor linear_164_cast_fp16 = linear(bias = encoder_layers_18_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_18_feed_forward1_linear2_weight_to_fp16_palettized, x = input_967_cast_fp16)[name = string("linear_164_cast_fp16")]; + fp16 var_4288_to_fp16 = const()[name = string("op_4288_to_fp16"), val = fp16(0x1p-1)]; + tensor var_4289_cast_fp16 = mul(x = linear_164_cast_fp16, y = var_4288_to_fp16)[name = string("op_4289_cast_fp16")]; + tensor input_973_cast_fp16 = add(x = input_961_cast_fp16, y = var_4289_cast_fp16)[name = string("input_973_cast_fp16")]; + tensor key_37_axes_0 = const()[name = string("key_37_axes_0"), val = tensor([-1])]; + tensor encoder_layers_18_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_18_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369114368)))]; + tensor encoder_layers_18_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_18_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369116480)))]; + tensor key_37_cast_fp16 = layer_norm(axes = key_37_axes_0, beta = encoder_layers_18_norm_self_att_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_18_norm_self_att_weight_to_fp16, x = input_973_cast_fp16)[name = string("key_37_cast_fp16")]; + bool input_975_interleave_0 = const()[name = string("input_975_interleave_0"), val = bool(false)]; + tensor input_975_cast_fp16 = concat(axis = var_67, interleave = input_975_interleave_0, values = (cache_73_cast_fp16, key_37_cast_fp16))[name = string("input_975_cast_fp16")]; + bool var_4317_interleave_0 = const()[name = string("op_4317_interleave_0"), val = bool(false)]; + tensor var_4317_cast_fp16 = concat(axis = var_67, interleave = var_4317_interleave_0, values = key_37_cast_fp16)[name = string("op_4317_cast_fp16")]; + tensor encoder_layers_18_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369118592))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369905088))))[name = string("encoder_layers_18_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_18_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_18_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369905280)))]; + tensor linear_165_cast_fp16 = linear(bias = encoder_layers_18_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_18_self_attn_linear_q_weight_to_fp16_palettized, x = key_37_cast_fp16)[name = string("linear_165_cast_fp16")]; + tensor var_4322 = const()[name = string("op_4322"), val = tensor([1, -1, 8, 128])]; + tensor q_109_cast_fp16 = reshape(shape = var_4322, x = linear_165_cast_fp16)[name = string("q_109_cast_fp16")]; + tensor encoder_layers_18_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369907392))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(370693888))))[name = string("encoder_layers_18_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_18_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_18_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(370694080)))]; + tensor linear_166_cast_fp16 = linear(bias = encoder_layers_18_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_18_self_attn_linear_k_weight_to_fp16_palettized, x = input_975_cast_fp16)[name = string("linear_166_cast_fp16")]; + tensor var_4327 = const()[name = string("op_4327"), val = tensor([1, -1, 8, 128])]; + tensor k_73_cast_fp16 = reshape(shape = var_4327, x = linear_166_cast_fp16)[name = string("k_73_cast_fp16")]; + tensor encoder_layers_18_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(370696192))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371482688))))[name = string("encoder_layers_18_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_18_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_18_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371482880)))]; + tensor linear_167_cast_fp16 = linear(bias = encoder_layers_18_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_18_self_attn_linear_v_weight_to_fp16_palettized, x = input_975_cast_fp16)[name = string("linear_167_cast_fp16")]; + tensor var_4332 = const()[name = string("op_4332"), val = tensor([1, -1, 8, 128])]; + tensor v_37_cast_fp16 = reshape(shape = var_4332, x = linear_167_cast_fp16)[name = string("v_37_cast_fp16")]; + tensor value_45_perm_0 = const()[name = string("value_45_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_18_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_18_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371484992)))]; + tensor var_4345_cast_fp16 = add(x = q_109_cast_fp16, y = encoder_layers_18_self_attn_pos_bias_u_to_fp16)[name = string("op_4345_cast_fp16")]; + tensor encoder_layers_18_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_18_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371487104)))]; + tensor var_4347_cast_fp16 = add(x = q_109_cast_fp16, y = encoder_layers_18_self_attn_pos_bias_v_to_fp16)[name = string("op_4347_cast_fp16")]; + tensor q_with_bias_v_37_perm_0 = const()[name = string("q_with_bias_v_37_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_475_transpose_x_0 = const()[name = string("x_475_transpose_x_0"), val = bool(false)]; + bool x_475_transpose_y_0 = const()[name = string("x_475_transpose_y_0"), val = bool(false)]; + tensor op_4349_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371489216))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371688960))))[name = string("op_4349_to_fp16_quantized")]; + tensor q_with_bias_v_37_cast_fp16 = transpose(perm = q_with_bias_v_37_perm_0, x = var_4347_cast_fp16)[name = string("transpose_200")]; + tensor x_475_cast_fp16 = matmul(transpose_x = x_475_transpose_x_0, transpose_y = x_475_transpose_y_0, x = q_with_bias_v_37_cast_fp16, y = op_4349_to_fp16_quantized)[name = string("x_475_cast_fp16")]; + tensor x_477_pad_0 = const()[name = string("x_477_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_477_mode_0 = const()[name = string("x_477_mode_0"), val = string("constant")]; + fp16 const_313_to_fp16 = const()[name = string("const_313_to_fp16"), val = fp16(0x0p+0)]; + tensor x_477_cast_fp16 = pad(constant_val = const_313_to_fp16, mode = x_477_mode_0, pad = x_477_pad_0, x = x_475_cast_fp16)[name = string("x_477_cast_fp16")]; + tensor var_4357 = const()[name = string("op_4357"), val = tensor([1, 8, -1, 56])]; + tensor x_479_cast_fp16 = reshape(shape = var_4357, x = x_477_cast_fp16)[name = string("x_479_cast_fp16")]; + tensor var_4361_begin_0 = const()[name = string("op_4361_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_4361_end_0 = const()[name = string("op_4361_end_0"), val = tensor([1, 8, 196, 56])]; + tensor var_4361_end_mask_0 = const()[name = string("op_4361_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_4361_cast_fp16 = slice_by_index(begin = var_4361_begin_0, end = var_4361_end_0, end_mask = var_4361_end_mask_0, x = x_479_cast_fp16)[name = string("op_4361_cast_fp16")]; + tensor var_4362 = const()[name = string("op_4362"), val = tensor([1, 8, 56, 195])]; + tensor matrix_bd_73_cast_fp16 = reshape(shape = var_4362, x = var_4361_cast_fp16)[name = string("matrix_bd_73_cast_fp16")]; + bool matrix_ac_37_transpose_x_0 = const()[name = string("matrix_ac_37_transpose_x_0"), val = bool(false)]; + bool matrix_ac_37_transpose_y_0 = const()[name = string("matrix_ac_37_transpose_y_0"), val = bool(false)]; + tensor transpose_132_perm_0 = const()[name = string("transpose_132_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_133_perm_0 = const()[name = string("transpose_133_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_133 = transpose(perm = transpose_133_perm_0, x = k_73_cast_fp16)[name = string("transpose_198")]; + tensor transpose_132 = transpose(perm = transpose_132_perm_0, x = var_4345_cast_fp16)[name = string("transpose_199")]; + tensor matrix_ac_37_cast_fp16 = matmul(transpose_x = matrix_ac_37_transpose_x_0, transpose_y = matrix_ac_37_transpose_y_0, x = transpose_132, y = transpose_133)[name = string("matrix_ac_37_cast_fp16")]; + tensor matrix_bd_75_begin_0 = const()[name = string("matrix_bd_75_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_75_end_0 = const()[name = string("matrix_bd_75_end_0"), val = tensor([1, 8, 56, 98])]; + tensor matrix_bd_75_end_mask_0 = const()[name = string("matrix_bd_75_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_75_cast_fp16 = slice_by_index(begin = matrix_bd_75_begin_0, end = matrix_bd_75_end_0, end_mask = matrix_bd_75_end_mask_0, x = matrix_bd_73_cast_fp16)[name = string("matrix_bd_75_cast_fp16")]; + tensor var_4371_cast_fp16 = add(x = matrix_ac_37_cast_fp16, y = matrix_bd_75_cast_fp16)[name = string("op_4371_cast_fp16")]; + fp16 _inversed_scores_73_y_0_to_fp16 = const()[name = string("_inversed_scores_73_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_73_cast_fp16 = mul(x = var_4371_cast_fp16, y = _inversed_scores_73_y_0_to_fp16)[name = string("_inversed_scores_73_cast_fp16")]; + tensor scores_75_cast_fp16 = select(a = var_44_to_fp16, b = _inversed_scores_73_cast_fp16, cond = mask_11)[name = string("scores_75_cast_fp16")]; + tensor var_4377_cast_fp16 = softmax(axis = var_58, x = scores_75_cast_fp16)[name = string("op_4377_cast_fp16")]; + tensor input_977_cast_fp16 = select(a = var_43_to_fp16, b = var_4377_cast_fp16, cond = mask_11)[name = string("input_977_cast_fp16")]; + bool x_481_transpose_x_0 = const()[name = string("x_481_transpose_x_0"), val = bool(false)]; + bool x_481_transpose_y_0 = const()[name = string("x_481_transpose_y_0"), val = bool(false)]; + tensor value_45_cast_fp16 = transpose(perm = value_45_perm_0, x = v_37_cast_fp16)[name = string("transpose_197")]; + tensor x_481_cast_fp16 = matmul(transpose_x = x_481_transpose_x_0, transpose_y = x_481_transpose_y_0, x = input_977_cast_fp16, y = value_45_cast_fp16)[name = string("x_481_cast_fp16")]; + tensor var_4381_perm_0 = const()[name = string("op_4381_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_4382 = const()[name = string("op_4382"), val = tensor([1, -1, 1024])]; + tensor var_4381_cast_fp16 = transpose(perm = var_4381_perm_0, x = x_481_cast_fp16)[name = string("transpose_196")]; + tensor input_979_cast_fp16 = reshape(shape = var_4382, x = var_4381_cast_fp16)[name = string("input_979_cast_fp16")]; + tensor encoder_layers_18_self_attn_linear_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371689472))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(372738112))))[name = string("encoder_layers_18_self_attn_linear_out_weight_to_fp16_quantized")]; + tensor encoder_layers_18_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_18_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(372740224)))]; + tensor linear_169_cast_fp16 = linear(bias = encoder_layers_18_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_18_self_attn_linear_out_weight_to_fp16_quantized, x = input_979_cast_fp16)[name = string("linear_169_cast_fp16")]; + tensor input_983_cast_fp16 = add(x = input_973_cast_fp16, y = linear_169_cast_fp16)[name = string("input_983_cast_fp16")]; + tensor x_485_axes_0 = const()[name = string("x_485_axes_0"), val = tensor([-1])]; + tensor encoder_layers_18_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_18_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(372742336)))]; + tensor encoder_layers_18_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_18_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(372744448)))]; + tensor x_485_cast_fp16 = layer_norm(axes = x_485_axes_0, beta = encoder_layers_18_norm_conv_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_18_norm_conv_weight_to_fp16, x = input_983_cast_fp16)[name = string("x_485_cast_fp16")]; + tensor input_985_perm_0 = const()[name = string("input_985_perm_0"), val = tensor([0, 2, 1])]; + string input_987_pad_type_0 = const()[name = string("input_987_pad_type_0"), val = string("valid")]; + tensor input_987_strides_0 = const()[name = string("input_987_strides_0"), val = tensor([1])]; + tensor input_987_pad_0 = const()[name = string("input_987_pad_0"), val = tensor([0, 0])]; + tensor input_987_dilations_0 = const()[name = string("input_987_dilations_0"), val = tensor([1])]; + int32 input_987_groups_0 = const()[name = string("input_987_groups_0"), val = int32(1)]; + tensor encoder_layers_18_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(372746560))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(374843776))))[name = string("encoder_layers_18_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_985_cast_fp16 = transpose(perm = input_985_perm_0, x = x_485_cast_fp16)[name = string("transpose_195")]; + tensor input_987_cast_fp16 = conv(dilations = input_987_dilations_0, groups = input_987_groups_0, pad = input_987_pad_0, pad_type = input_987_pad_type_0, strides = input_987_strides_0, weight = encoder_layers_18_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_985_cast_fp16)[name = string("input_987_cast_fp16")]; + int32 x_487_split_num_splits_0 = const()[name = string("x_487_split_num_splits_0"), val = int32(2)]; + int32 x_487_split_axis_0 = const()[name = string("x_487_split_axis_0"), val = int32(1)]; + tensor x_487_split_cast_fp16_0, tensor x_487_split_cast_fp16_1 = split(axis = x_487_split_axis_0, num_splits = x_487_split_num_splits_0, x = input_987_cast_fp16)[name = string("x_487_split_cast_fp16")]; + tensor x_487_split_1_sigmoid_cast_fp16 = sigmoid(x = x_487_split_cast_fp16_1)[name = string("x_487_split_1_sigmoid_cast_fp16")]; + tensor x_487_cast_fp16 = mul(x = x_487_split_cast_fp16_0, y = x_487_split_1_sigmoid_cast_fp16)[name = string("x_487_cast_fp16")]; + tensor input_989_cast_fp16 = select(a = var_43_to_fp16, b = x_487_cast_fp16, cond = var_574)[name = string("input_989_cast_fp16")]; + bool new_x_75_interleave_0 = const()[name = string("new_x_75_interleave_0"), val = bool(false)]; + tensor new_x_75_cast_fp16 = concat(axis = var_58, interleave = new_x_75_interleave_0, values = (cache_75_cast_fp16, input_989_cast_fp16))[name = string("new_x_75_cast_fp16")]; + tensor var_4421_begin_0 = const()[name = string("op_4421_begin_0"), val = tensor([0, 0, 56])]; + tensor var_4421_end_0 = const()[name = string("op_4421_end_0"), val = tensor([1, 1024, 64])]; + tensor var_4421_end_mask_0 = const()[name = string("op_4421_end_mask_0"), val = tensor([true, true, true])]; + tensor var_4421_cast_fp16 = slice_by_index(begin = var_4421_begin_0, end = var_4421_end_0, end_mask = var_4421_end_mask_0, x = new_x_75_cast_fp16)[name = string("op_4421_cast_fp16")]; + string x_489_pad_type_0 = const()[name = string("x_489_pad_type_0"), val = string("valid")]; + int32 x_489_groups_0 = const()[name = string("x_489_groups_0"), val = int32(1024)]; + tensor x_489_strides_0 = const()[name = string("x_489_strides_0"), val = tensor([1])]; + tensor x_489_pad_0 = const()[name = string("x_489_pad_0"), val = tensor([0, 0])]; + tensor x_489_dilations_0 = const()[name = string("x_489_dilations_0"), val = tensor([1])]; + tensor encoder_layers_18_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(374847936))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(374857216))))[name = string("encoder_layers_18_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_489_cast_fp16 = conv(dilations = x_489_dilations_0, groups = x_489_groups_0, pad = x_489_pad_0, pad_type = x_489_pad_type_0, strides = x_489_strides_0, weight = encoder_layers_18_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_75_cast_fp16)[name = string("x_489_cast_fp16")]; + tensor input_991_perm_0 = const()[name = string("input_991_perm_0"), val = tensor([0, 2, 1])]; + tensor x_491_axes_0 = const()[name = string("x_491_axes_0"), val = tensor([-1])]; + tensor encoder_layers_18_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_18_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(374859328)))]; + tensor encoder_layers_18_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_18_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(374861440)))]; + tensor input_991_cast_fp16 = transpose(perm = input_991_perm_0, x = x_489_cast_fp16)[name = string("transpose_194")]; + tensor x_491_cast_fp16 = layer_norm(axes = x_491_axes_0, beta = encoder_layers_18_conv_batch_norm_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_18_conv_batch_norm_weight_to_fp16, x = input_991_cast_fp16)[name = string("x_491_cast_fp16")]; + tensor input_993_perm_0 = const()[name = string("input_993_perm_0"), val = tensor([0, 2, 1])]; + tensor input_993_cast_fp16 = transpose(perm = input_993_perm_0, x = x_491_cast_fp16)[name = string("transpose_193")]; + tensor input_995_cast_fp16 = silu(x = input_993_cast_fp16)[name = string("input_995_cast_fp16")]; + string x_493_pad_type_0 = const()[name = string("x_493_pad_type_0"), val = string("valid")]; + tensor x_493_strides_0 = const()[name = string("x_493_strides_0"), val = tensor([1])]; + tensor x_493_pad_0 = const()[name = string("x_493_pad_0"), val = tensor([0, 0])]; + tensor x_493_dilations_0 = const()[name = string("x_493_dilations_0"), val = tensor([1])]; + int32 x_493_groups_0 = const()[name = string("x_493_groups_0"), val = int32(1)]; + tensor encoder_layers_18_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(374863552))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(375912192))))[name = string("encoder_layers_18_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_493_cast_fp16 = conv(dilations = x_493_dilations_0, groups = x_493_groups_0, pad = x_493_pad_0, pad_type = x_493_pad_type_0, strides = x_493_strides_0, weight = encoder_layers_18_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_995_cast_fp16)[name = string("x_493_cast_fp16")]; + tensor input_997_perm_0 = const()[name = string("input_997_perm_0"), val = tensor([0, 2, 1])]; + tensor input_997_cast_fp16 = transpose(perm = input_997_perm_0, x = x_493_cast_fp16)[name = string("transpose_192")]; + tensor input_999_cast_fp16 = add(x = input_983_cast_fp16, y = input_997_cast_fp16)[name = string("input_999_cast_fp16")]; + tensor input_1001_axes_0 = const()[name = string("input_1001_axes_0"), val = tensor([-1])]; + tensor encoder_layers_18_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_18_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(375914304)))]; + tensor encoder_layers_18_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_18_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(375916416)))]; + tensor input_1001_cast_fp16 = layer_norm(axes = input_1001_axes_0, beta = encoder_layers_18_norm_feed_forward2_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_18_norm_feed_forward2_weight_to_fp16, x = input_999_cast_fp16)[name = string("input_1001_cast_fp16")]; + tensor encoder_layers_18_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(375918528))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(380112896))))[name = string("encoder_layers_18_feed_forward2_linear1_weight_to_fp16_quantized")]; + tensor encoder_layers_18_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_18_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(380121152)))]; + tensor linear_170_cast_fp16 = linear(bias = encoder_layers_18_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_18_feed_forward2_linear1_weight_to_fp16_quantized, x = input_1001_cast_fp16)[name = string("linear_170_cast_fp16")]; + tensor input_1005_cast_fp16 = silu(x = linear_170_cast_fp16)[name = string("input_1005_cast_fp16")]; + tensor encoder_layers_18_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(380129408))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(384323776))))[name = string("encoder_layers_18_feed_forward2_linear2_weight_to_fp16_quantized")]; + tensor encoder_layers_18_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_18_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(384325888)))]; + tensor linear_171_cast_fp16 = linear(bias = encoder_layers_18_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_18_feed_forward2_linear2_weight_to_fp16_quantized, x = input_1005_cast_fp16)[name = string("linear_171_cast_fp16")]; + fp16 var_4464_to_fp16 = const()[name = string("op_4464_to_fp16"), val = fp16(0x1p-1)]; + tensor var_4465_cast_fp16 = mul(x = linear_171_cast_fp16, y = var_4464_to_fp16)[name = string("op_4465_cast_fp16")]; + tensor input_1011_cast_fp16 = add(x = input_999_cast_fp16, y = var_4465_cast_fp16)[name = string("input_1011_cast_fp16")]; + tensor input_1013_axes_0 = const()[name = string("input_1013_axes_0"), val = tensor([-1])]; + tensor encoder_layers_18_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_18_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(384328000)))]; + tensor encoder_layers_18_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_18_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(384330112)))]; + tensor input_1013_cast_fp16 = layer_norm(axes = input_1013_axes_0, beta = encoder_layers_18_norm_out_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_18_norm_out_weight_to_fp16, x = input_1011_cast_fp16)[name = string("input_1013_cast_fp16")]; + tensor cache_77_begin_0 = const()[name = string("cache_77_begin_0"), val = tensor([19, 0, 0, 0])]; + tensor cache_77_end_0 = const()[name = string("cache_77_end_0"), val = tensor([20, 1, 42, 1024])]; + tensor cache_77_end_mask_0 = const()[name = string("cache_77_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_77_squeeze_mask_0 = const()[name = string("cache_77_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_77_cast_fp16 = slice_by_index(begin = cache_77_begin_0, end = cache_77_end_0, end_mask = cache_77_end_mask_0, squeeze_mask = cache_77_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_77_cast_fp16")]; + tensor cache_79_begin_0 = const()[name = string("cache_79_begin_0"), val = tensor([19, 0, 0, 0])]; + tensor cache_79_end_0 = const()[name = string("cache_79_end_0"), val = tensor([20, 1, 1024, 8])]; + tensor cache_79_end_mask_0 = const()[name = string("cache_79_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_79_squeeze_mask_0 = const()[name = string("cache_79_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_79_cast_fp16 = slice_by_index(begin = cache_79_begin_0, end = cache_79_end_0, end_mask = cache_79_end_mask_0, squeeze_mask = cache_79_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_79_cast_fp16")]; + tensor input_1015_axes_0 = const()[name = string("input_1015_axes_0"), val = tensor([-1])]; + tensor encoder_layers_19_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_19_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(384332224)))]; + tensor encoder_layers_19_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_19_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(384334336)))]; + tensor input_1015_cast_fp16 = layer_norm(axes = input_1015_axes_0, beta = encoder_layers_19_norm_feed_forward1_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_19_norm_feed_forward1_weight_to_fp16, x = input_1013_cast_fp16)[name = string("input_1015_cast_fp16")]; + tensor encoder_layers_19_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(384336448))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(388530816))))[name = string("encoder_layers_19_feed_forward1_linear1_weight_to_fp16_quantized")]; + tensor encoder_layers_19_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_19_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(388539072)))]; + tensor linear_172_cast_fp16 = linear(bias = encoder_layers_19_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_19_feed_forward1_linear1_weight_to_fp16_quantized, x = input_1015_cast_fp16)[name = string("linear_172_cast_fp16")]; + tensor input_1019_cast_fp16 = silu(x = linear_172_cast_fp16)[name = string("input_1019_cast_fp16")]; + tensor encoder_layers_19_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(388547328))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(392741696))))[name = string("encoder_layers_19_feed_forward1_linear2_weight_to_fp16_quantized")]; + tensor encoder_layers_19_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_19_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(392743808)))]; + tensor linear_173_cast_fp16 = linear(bias = encoder_layers_19_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_19_feed_forward1_linear2_weight_to_fp16_quantized, x = input_1019_cast_fp16)[name = string("linear_173_cast_fp16")]; + fp16 var_4501_to_fp16 = const()[name = string("op_4501_to_fp16"), val = fp16(0x1p-1)]; + tensor var_4502_cast_fp16 = mul(x = linear_173_cast_fp16, y = var_4501_to_fp16)[name = string("op_4502_cast_fp16")]; + tensor input_1025_cast_fp16 = add(x = input_1013_cast_fp16, y = var_4502_cast_fp16)[name = string("input_1025_cast_fp16")]; + tensor key_39_axes_0 = const()[name = string("key_39_axes_0"), val = tensor([-1])]; + tensor encoder_layers_19_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_19_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(392745920)))]; + tensor encoder_layers_19_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_19_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(392748032)))]; + tensor key_39_cast_fp16 = layer_norm(axes = key_39_axes_0, beta = encoder_layers_19_norm_self_att_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_19_norm_self_att_weight_to_fp16, x = input_1025_cast_fp16)[name = string("key_39_cast_fp16")]; + bool input_1027_interleave_0 = const()[name = string("input_1027_interleave_0"), val = bool(false)]; + tensor input_1027_cast_fp16 = concat(axis = var_67, interleave = input_1027_interleave_0, values = (cache_77_cast_fp16, key_39_cast_fp16))[name = string("input_1027_cast_fp16")]; + bool var_4530_interleave_0 = const()[name = string("op_4530_interleave_0"), val = bool(false)]; + tensor var_4530_cast_fp16 = concat(axis = var_67, interleave = var_4530_interleave_0, values = key_39_cast_fp16)[name = string("op_4530_cast_fp16")]; + tensor encoder_layers_19_self_attn_linear_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(392750144))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(393798784))))[name = string("encoder_layers_19_self_attn_linear_q_weight_to_fp16_quantized")]; + tensor encoder_layers_19_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_19_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(393800896)))]; + tensor linear_174_cast_fp16 = linear(bias = encoder_layers_19_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_19_self_attn_linear_q_weight_to_fp16_quantized, x = key_39_cast_fp16)[name = string("linear_174_cast_fp16")]; + tensor var_4535 = const()[name = string("op_4535"), val = tensor([1, -1, 8, 128])]; + tensor q_115_cast_fp16 = reshape(shape = var_4535, x = linear_174_cast_fp16)[name = string("q_115_cast_fp16")]; + tensor encoder_layers_19_self_attn_linear_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(393803008))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(394851648))))[name = string("encoder_layers_19_self_attn_linear_k_weight_to_fp16_quantized")]; + tensor encoder_layers_19_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_19_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(394853760)))]; + tensor linear_175_cast_fp16 = linear(bias = encoder_layers_19_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_19_self_attn_linear_k_weight_to_fp16_quantized, x = input_1027_cast_fp16)[name = string("linear_175_cast_fp16")]; + tensor var_4540 = const()[name = string("op_4540"), val = tensor([1, -1, 8, 128])]; + tensor k_77_cast_fp16 = reshape(shape = var_4540, x = linear_175_cast_fp16)[name = string("k_77_cast_fp16")]; + tensor encoder_layers_19_self_attn_linear_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(394855872))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(395904512))))[name = string("encoder_layers_19_self_attn_linear_v_weight_to_fp16_quantized")]; + tensor encoder_layers_19_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_19_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(395906624)))]; + tensor linear_176_cast_fp16 = linear(bias = encoder_layers_19_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_19_self_attn_linear_v_weight_to_fp16_quantized, x = input_1027_cast_fp16)[name = string("linear_176_cast_fp16")]; + tensor var_4545 = const()[name = string("op_4545"), val = tensor([1, -1, 8, 128])]; + tensor v_39_cast_fp16 = reshape(shape = var_4545, x = linear_176_cast_fp16)[name = string("v_39_cast_fp16")]; + tensor value_47_perm_0 = const()[name = string("value_47_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_19_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_19_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(395908736)))]; + tensor var_4558_cast_fp16 = add(x = q_115_cast_fp16, y = encoder_layers_19_self_attn_pos_bias_u_to_fp16)[name = string("op_4558_cast_fp16")]; + tensor encoder_layers_19_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_19_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(395910848)))]; + tensor var_4560_cast_fp16 = add(x = q_115_cast_fp16, y = encoder_layers_19_self_attn_pos_bias_v_to_fp16)[name = string("op_4560_cast_fp16")]; + tensor q_with_bias_v_39_perm_0 = const()[name = string("q_with_bias_v_39_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_501_transpose_x_0 = const()[name = string("x_501_transpose_x_0"), val = bool(false)]; + bool x_501_transpose_y_0 = const()[name = string("x_501_transpose_y_0"), val = bool(false)]; + tensor op_4562_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(395912960))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(396112704))))[name = string("op_4562_to_fp16_quantized")]; + tensor q_with_bias_v_39_cast_fp16 = transpose(perm = q_with_bias_v_39_perm_0, x = var_4560_cast_fp16)[name = string("transpose_191")]; + tensor x_501_cast_fp16 = matmul(transpose_x = x_501_transpose_x_0, transpose_y = x_501_transpose_y_0, x = q_with_bias_v_39_cast_fp16, y = op_4562_to_fp16_quantized)[name = string("x_501_cast_fp16")]; + tensor x_503_pad_0 = const()[name = string("x_503_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_503_mode_0 = const()[name = string("x_503_mode_0"), val = string("constant")]; + fp16 const_326_to_fp16 = const()[name = string("const_326_to_fp16"), val = fp16(0x0p+0)]; + tensor x_503_cast_fp16 = pad(constant_val = const_326_to_fp16, mode = x_503_mode_0, pad = x_503_pad_0, x = x_501_cast_fp16)[name = string("x_503_cast_fp16")]; + tensor var_4570 = const()[name = string("op_4570"), val = tensor([1, 8, -1, 56])]; + tensor x_505_cast_fp16 = reshape(shape = var_4570, x = x_503_cast_fp16)[name = string("x_505_cast_fp16")]; + tensor var_4574_begin_0 = const()[name = string("op_4574_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_4574_end_0 = const()[name = string("op_4574_end_0"), val = tensor([1, 8, 196, 56])]; + tensor var_4574_end_mask_0 = const()[name = string("op_4574_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_4574_cast_fp16 = slice_by_index(begin = var_4574_begin_0, end = var_4574_end_0, end_mask = var_4574_end_mask_0, x = x_505_cast_fp16)[name = string("op_4574_cast_fp16")]; + tensor var_4575 = const()[name = string("op_4575"), val = tensor([1, 8, 56, 195])]; + tensor matrix_bd_77_cast_fp16 = reshape(shape = var_4575, x = var_4574_cast_fp16)[name = string("matrix_bd_77_cast_fp16")]; + bool matrix_ac_39_transpose_x_0 = const()[name = string("matrix_ac_39_transpose_x_0"), val = bool(false)]; + bool matrix_ac_39_transpose_y_0 = const()[name = string("matrix_ac_39_transpose_y_0"), val = bool(false)]; + tensor transpose_134_perm_0 = const()[name = string("transpose_134_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_135_perm_0 = const()[name = string("transpose_135_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_135 = transpose(perm = transpose_135_perm_0, x = k_77_cast_fp16)[name = string("transpose_189")]; + tensor transpose_134 = transpose(perm = transpose_134_perm_0, x = var_4558_cast_fp16)[name = string("transpose_190")]; + tensor matrix_ac_39_cast_fp16 = matmul(transpose_x = matrix_ac_39_transpose_x_0, transpose_y = matrix_ac_39_transpose_y_0, x = transpose_134, y = transpose_135)[name = string("matrix_ac_39_cast_fp16")]; + tensor matrix_bd_79_begin_0 = const()[name = string("matrix_bd_79_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_79_end_0 = const()[name = string("matrix_bd_79_end_0"), val = tensor([1, 8, 56, 98])]; + tensor matrix_bd_79_end_mask_0 = const()[name = string("matrix_bd_79_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_79_cast_fp16 = slice_by_index(begin = matrix_bd_79_begin_0, end = matrix_bd_79_end_0, end_mask = matrix_bd_79_end_mask_0, x = matrix_bd_77_cast_fp16)[name = string("matrix_bd_79_cast_fp16")]; + tensor var_4584_cast_fp16 = add(x = matrix_ac_39_cast_fp16, y = matrix_bd_79_cast_fp16)[name = string("op_4584_cast_fp16")]; + fp16 _inversed_scores_77_y_0_to_fp16 = const()[name = string("_inversed_scores_77_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_77_cast_fp16 = mul(x = var_4584_cast_fp16, y = _inversed_scores_77_y_0_to_fp16)[name = string("_inversed_scores_77_cast_fp16")]; + tensor scores_79_cast_fp16 = select(a = var_44_to_fp16, b = _inversed_scores_77_cast_fp16, cond = mask_11)[name = string("scores_79_cast_fp16")]; + tensor var_4590_cast_fp16 = softmax(axis = var_58, x = scores_79_cast_fp16)[name = string("op_4590_cast_fp16")]; + tensor input_1029_cast_fp16 = select(a = var_43_to_fp16, b = var_4590_cast_fp16, cond = mask_11)[name = string("input_1029_cast_fp16")]; + bool x_507_transpose_x_0 = const()[name = string("x_507_transpose_x_0"), val = bool(false)]; + bool x_507_transpose_y_0 = const()[name = string("x_507_transpose_y_0"), val = bool(false)]; + tensor value_47_cast_fp16 = transpose(perm = value_47_perm_0, x = v_39_cast_fp16)[name = string("transpose_188")]; + tensor x_507_cast_fp16 = matmul(transpose_x = x_507_transpose_x_0, transpose_y = x_507_transpose_y_0, x = input_1029_cast_fp16, y = value_47_cast_fp16)[name = string("x_507_cast_fp16")]; + tensor var_4594_perm_0 = const()[name = string("op_4594_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_4595 = const()[name = string("op_4595"), val = tensor([1, -1, 1024])]; + tensor var_4594_cast_fp16 = transpose(perm = var_4594_perm_0, x = x_507_cast_fp16)[name = string("transpose_187")]; + tensor input_1031_cast_fp16 = reshape(shape = var_4595, x = var_4594_cast_fp16)[name = string("input_1031_cast_fp16")]; + tensor encoder_layers_19_self_attn_linear_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(396113216))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397161856))))[name = string("encoder_layers_19_self_attn_linear_out_weight_to_fp16_quantized")]; + tensor encoder_layers_19_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_19_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397163968)))]; + tensor linear_178_cast_fp16 = linear(bias = encoder_layers_19_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_19_self_attn_linear_out_weight_to_fp16_quantized, x = input_1031_cast_fp16)[name = string("linear_178_cast_fp16")]; + tensor input_1035_cast_fp16 = add(x = input_1025_cast_fp16, y = linear_178_cast_fp16)[name = string("input_1035_cast_fp16")]; + tensor x_511_axes_0 = const()[name = string("x_511_axes_0"), val = tensor([-1])]; + tensor encoder_layers_19_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_19_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397166080)))]; + tensor encoder_layers_19_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_19_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397168192)))]; + tensor x_511_cast_fp16 = layer_norm(axes = x_511_axes_0, beta = encoder_layers_19_norm_conv_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_19_norm_conv_weight_to_fp16, x = input_1035_cast_fp16)[name = string("x_511_cast_fp16")]; + tensor input_1037_perm_0 = const()[name = string("input_1037_perm_0"), val = tensor([0, 2, 1])]; + string input_1039_pad_type_0 = const()[name = string("input_1039_pad_type_0"), val = string("valid")]; + tensor input_1039_strides_0 = const()[name = string("input_1039_strides_0"), val = tensor([1])]; + tensor input_1039_pad_0 = const()[name = string("input_1039_pad_0"), val = tensor([0, 0])]; + tensor input_1039_dilations_0 = const()[name = string("input_1039_dilations_0"), val = tensor([1])]; + int32 input_1039_groups_0 = const()[name = string("input_1039_groups_0"), val = int32(1)]; + tensor encoder_layers_19_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397170304))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(399267520))))[name = string("encoder_layers_19_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_1037_cast_fp16 = transpose(perm = input_1037_perm_0, x = x_511_cast_fp16)[name = string("transpose_186")]; + tensor input_1039_cast_fp16 = conv(dilations = input_1039_dilations_0, groups = input_1039_groups_0, pad = input_1039_pad_0, pad_type = input_1039_pad_type_0, strides = input_1039_strides_0, weight = encoder_layers_19_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_1037_cast_fp16)[name = string("input_1039_cast_fp16")]; + int32 x_513_split_num_splits_0 = const()[name = string("x_513_split_num_splits_0"), val = int32(2)]; + int32 x_513_split_axis_0 = const()[name = string("x_513_split_axis_0"), val = int32(1)]; + tensor x_513_split_cast_fp16_0, tensor x_513_split_cast_fp16_1 = split(axis = x_513_split_axis_0, num_splits = x_513_split_num_splits_0, x = input_1039_cast_fp16)[name = string("x_513_split_cast_fp16")]; + tensor x_513_split_1_sigmoid_cast_fp16 = sigmoid(x = x_513_split_cast_fp16_1)[name = string("x_513_split_1_sigmoid_cast_fp16")]; + tensor x_513_cast_fp16 = mul(x = x_513_split_cast_fp16_0, y = x_513_split_1_sigmoid_cast_fp16)[name = string("x_513_cast_fp16")]; + tensor input_1041_cast_fp16 = select(a = var_43_to_fp16, b = x_513_cast_fp16, cond = var_574)[name = string("input_1041_cast_fp16")]; + bool new_x_79_interleave_0 = const()[name = string("new_x_79_interleave_0"), val = bool(false)]; + tensor new_x_79_cast_fp16 = concat(axis = var_58, interleave = new_x_79_interleave_0, values = (cache_79_cast_fp16, input_1041_cast_fp16))[name = string("new_x_79_cast_fp16")]; + tensor var_4634_begin_0 = const()[name = string("op_4634_begin_0"), val = tensor([0, 0, 56])]; + tensor var_4634_end_0 = const()[name = string("op_4634_end_0"), val = tensor([1, 1024, 64])]; + tensor var_4634_end_mask_0 = const()[name = string("op_4634_end_mask_0"), val = tensor([true, true, true])]; + tensor var_4634_cast_fp16 = slice_by_index(begin = var_4634_begin_0, end = var_4634_end_0, end_mask = var_4634_end_mask_0, x = new_x_79_cast_fp16)[name = string("op_4634_cast_fp16")]; + string x_515_pad_type_0 = const()[name = string("x_515_pad_type_0"), val = string("valid")]; + int32 x_515_groups_0 = const()[name = string("x_515_groups_0"), val = int32(1024)]; + tensor x_515_strides_0 = const()[name = string("x_515_strides_0"), val = tensor([1])]; + tensor x_515_pad_0 = const()[name = string("x_515_pad_0"), val = tensor([0, 0])]; + tensor x_515_dilations_0 = const()[name = string("x_515_dilations_0"), val = tensor([1])]; + tensor encoder_layers_19_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(399271680))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(399280960))))[name = string("encoder_layers_19_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_515_cast_fp16 = conv(dilations = x_515_dilations_0, groups = x_515_groups_0, pad = x_515_pad_0, pad_type = x_515_pad_type_0, strides = x_515_strides_0, weight = encoder_layers_19_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_79_cast_fp16)[name = string("x_515_cast_fp16")]; + tensor input_1043_perm_0 = const()[name = string("input_1043_perm_0"), val = tensor([0, 2, 1])]; + tensor x_517_axes_0 = const()[name = string("x_517_axes_0"), val = tensor([-1])]; + tensor encoder_layers_19_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_19_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(399283072)))]; + tensor encoder_layers_19_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_19_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(399285184)))]; + tensor input_1043_cast_fp16 = transpose(perm = input_1043_perm_0, x = x_515_cast_fp16)[name = string("transpose_185")]; + tensor x_517_cast_fp16 = layer_norm(axes = x_517_axes_0, beta = encoder_layers_19_conv_batch_norm_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_19_conv_batch_norm_weight_to_fp16, x = input_1043_cast_fp16)[name = string("x_517_cast_fp16")]; + tensor input_1045_perm_0 = const()[name = string("input_1045_perm_0"), val = tensor([0, 2, 1])]; + tensor input_1045_cast_fp16 = transpose(perm = input_1045_perm_0, x = x_517_cast_fp16)[name = string("transpose_184")]; + tensor input_1047_cast_fp16 = silu(x = input_1045_cast_fp16)[name = string("input_1047_cast_fp16")]; + string x_519_pad_type_0 = const()[name = string("x_519_pad_type_0"), val = string("valid")]; + tensor x_519_strides_0 = const()[name = string("x_519_strides_0"), val = tensor([1])]; + tensor x_519_pad_0 = const()[name = string("x_519_pad_0"), val = tensor([0, 0])]; + tensor x_519_dilations_0 = const()[name = string("x_519_dilations_0"), val = tensor([1])]; + int32 x_519_groups_0 = const()[name = string("x_519_groups_0"), val = int32(1)]; + tensor encoder_layers_19_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(399287296))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(400335936))))[name = string("encoder_layers_19_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_519_cast_fp16 = conv(dilations = x_519_dilations_0, groups = x_519_groups_0, pad = x_519_pad_0, pad_type = x_519_pad_type_0, strides = x_519_strides_0, weight = encoder_layers_19_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_1047_cast_fp16)[name = string("x_519_cast_fp16")]; + tensor input_1049_perm_0 = const()[name = string("input_1049_perm_0"), val = tensor([0, 2, 1])]; + tensor input_1049_cast_fp16 = transpose(perm = input_1049_perm_0, x = x_519_cast_fp16)[name = string("transpose_183")]; + tensor input_1051_cast_fp16 = add(x = input_1035_cast_fp16, y = input_1049_cast_fp16)[name = string("input_1051_cast_fp16")]; + tensor input_1053_axes_0 = const()[name = string("input_1053_axes_0"), val = tensor([-1])]; + tensor encoder_layers_19_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_19_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(400338048)))]; + tensor encoder_layers_19_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_19_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(400340160)))]; + tensor input_1053_cast_fp16 = layer_norm(axes = input_1053_axes_0, beta = encoder_layers_19_norm_feed_forward2_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_19_norm_feed_forward2_weight_to_fp16, x = input_1051_cast_fp16)[name = string("input_1053_cast_fp16")]; + tensor encoder_layers_19_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(400342272))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(404536640))))[name = string("encoder_layers_19_feed_forward2_linear1_weight_to_fp16_quantized")]; + tensor encoder_layers_19_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_19_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(404544896)))]; + tensor linear_179_cast_fp16 = linear(bias = encoder_layers_19_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_19_feed_forward2_linear1_weight_to_fp16_quantized, x = input_1053_cast_fp16)[name = string("linear_179_cast_fp16")]; + tensor input_1057_cast_fp16 = silu(x = linear_179_cast_fp16)[name = string("input_1057_cast_fp16")]; + tensor encoder_layers_19_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(404553152))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(408747520))))[name = string("encoder_layers_19_feed_forward2_linear2_weight_to_fp16_quantized")]; + tensor encoder_layers_19_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_19_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(408749632)))]; + tensor linear_180_cast_fp16 = linear(bias = encoder_layers_19_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_19_feed_forward2_linear2_weight_to_fp16_quantized, x = input_1057_cast_fp16)[name = string("linear_180_cast_fp16")]; + fp16 var_4677_to_fp16 = const()[name = string("op_4677_to_fp16"), val = fp16(0x1p-1)]; + tensor var_4678_cast_fp16 = mul(x = linear_180_cast_fp16, y = var_4677_to_fp16)[name = string("op_4678_cast_fp16")]; + tensor input_1063_cast_fp16 = add(x = input_1051_cast_fp16, y = var_4678_cast_fp16)[name = string("input_1063_cast_fp16")]; + tensor input_1065_axes_0 = const()[name = string("input_1065_axes_0"), val = tensor([-1])]; + tensor encoder_layers_19_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_19_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(408751744)))]; + tensor encoder_layers_19_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_19_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(408753856)))]; + tensor input_1065_cast_fp16 = layer_norm(axes = input_1065_axes_0, beta = encoder_layers_19_norm_out_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_19_norm_out_weight_to_fp16, x = input_1063_cast_fp16)[name = string("input_1065_cast_fp16")]; + tensor cache_81_begin_0 = const()[name = string("cache_81_begin_0"), val = tensor([20, 0, 0, 0])]; + tensor cache_81_end_0 = const()[name = string("cache_81_end_0"), val = tensor([21, 1, 42, 1024])]; + tensor cache_81_end_mask_0 = const()[name = string("cache_81_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_81_squeeze_mask_0 = const()[name = string("cache_81_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_81_cast_fp16 = slice_by_index(begin = cache_81_begin_0, end = cache_81_end_0, end_mask = cache_81_end_mask_0, squeeze_mask = cache_81_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_81_cast_fp16")]; + tensor cache_83_begin_0 = const()[name = string("cache_83_begin_0"), val = tensor([20, 0, 0, 0])]; + tensor cache_83_end_0 = const()[name = string("cache_83_end_0"), val = tensor([21, 1, 1024, 8])]; + tensor cache_83_end_mask_0 = const()[name = string("cache_83_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_83_squeeze_mask_0 = const()[name = string("cache_83_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_83_cast_fp16 = slice_by_index(begin = cache_83_begin_0, end = cache_83_end_0, end_mask = cache_83_end_mask_0, squeeze_mask = cache_83_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_83_cast_fp16")]; + tensor input_1067_axes_0 = const()[name = string("input_1067_axes_0"), val = tensor([-1])]; + tensor encoder_layers_20_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_20_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(408755968)))]; + tensor encoder_layers_20_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_20_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(408758080)))]; + tensor input_1067_cast_fp16 = layer_norm(axes = input_1067_axes_0, beta = encoder_layers_20_norm_feed_forward1_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_20_norm_feed_forward1_weight_to_fp16, x = input_1065_cast_fp16)[name = string("input_1067_cast_fp16")]; + tensor encoder_layers_20_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(408760192))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(412954560))))[name = string("encoder_layers_20_feed_forward1_linear1_weight_to_fp16_quantized")]; + tensor encoder_layers_20_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_20_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(412962816)))]; + tensor linear_181_cast_fp16 = linear(bias = encoder_layers_20_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_20_feed_forward1_linear1_weight_to_fp16_quantized, x = input_1067_cast_fp16)[name = string("linear_181_cast_fp16")]; + tensor input_1071_cast_fp16 = silu(x = linear_181_cast_fp16)[name = string("input_1071_cast_fp16")]; + tensor encoder_layers_20_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(412971072))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(417165440))))[name = string("encoder_layers_20_feed_forward1_linear2_weight_to_fp16_quantized")]; + tensor encoder_layers_20_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_20_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(417167552)))]; + tensor linear_182_cast_fp16 = linear(bias = encoder_layers_20_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_20_feed_forward1_linear2_weight_to_fp16_quantized, x = input_1071_cast_fp16)[name = string("linear_182_cast_fp16")]; + fp16 var_4714_to_fp16 = const()[name = string("op_4714_to_fp16"), val = fp16(0x1p-1)]; + tensor var_4715_cast_fp16 = mul(x = linear_182_cast_fp16, y = var_4714_to_fp16)[name = string("op_4715_cast_fp16")]; + tensor input_1077_cast_fp16 = add(x = input_1065_cast_fp16, y = var_4715_cast_fp16)[name = string("input_1077_cast_fp16")]; + tensor key_41_axes_0 = const()[name = string("key_41_axes_0"), val = tensor([-1])]; + tensor encoder_layers_20_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_20_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(417169664)))]; + tensor encoder_layers_20_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_20_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(417171776)))]; + tensor key_41_cast_fp16 = layer_norm(axes = key_41_axes_0, beta = encoder_layers_20_norm_self_att_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_20_norm_self_att_weight_to_fp16, x = input_1077_cast_fp16)[name = string("key_41_cast_fp16")]; + bool input_1079_interleave_0 = const()[name = string("input_1079_interleave_0"), val = bool(false)]; + tensor input_1079_cast_fp16 = concat(axis = var_67, interleave = input_1079_interleave_0, values = (cache_81_cast_fp16, key_41_cast_fp16))[name = string("input_1079_cast_fp16")]; + bool var_4743_interleave_0 = const()[name = string("op_4743_interleave_0"), val = bool(false)]; + tensor var_4743_cast_fp16 = concat(axis = var_67, interleave = var_4743_interleave_0, values = key_41_cast_fp16)[name = string("op_4743_cast_fp16")]; + tensor encoder_layers_20_self_attn_linear_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(417173888))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(418222528))))[name = string("encoder_layers_20_self_attn_linear_q_weight_to_fp16_quantized")]; + tensor encoder_layers_20_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_20_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(418224640)))]; + tensor linear_183_cast_fp16 = linear(bias = encoder_layers_20_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_20_self_attn_linear_q_weight_to_fp16_quantized, x = key_41_cast_fp16)[name = string("linear_183_cast_fp16")]; + tensor var_4748 = const()[name = string("op_4748"), val = tensor([1, -1, 8, 128])]; + tensor q_121_cast_fp16 = reshape(shape = var_4748, x = linear_183_cast_fp16)[name = string("q_121_cast_fp16")]; + tensor encoder_layers_20_self_attn_linear_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(418226752))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(419275392))))[name = string("encoder_layers_20_self_attn_linear_k_weight_to_fp16_quantized")]; + tensor encoder_layers_20_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_20_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(419277504)))]; + tensor linear_184_cast_fp16 = linear(bias = encoder_layers_20_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_20_self_attn_linear_k_weight_to_fp16_quantized, x = input_1079_cast_fp16)[name = string("linear_184_cast_fp16")]; + tensor var_4753 = const()[name = string("op_4753"), val = tensor([1, -1, 8, 128])]; + tensor k_81_cast_fp16 = reshape(shape = var_4753, x = linear_184_cast_fp16)[name = string("k_81_cast_fp16")]; + tensor encoder_layers_20_self_attn_linear_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(419279616))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(420328256))))[name = string("encoder_layers_20_self_attn_linear_v_weight_to_fp16_quantized")]; + tensor encoder_layers_20_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_20_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(420330368)))]; + tensor linear_185_cast_fp16 = linear(bias = encoder_layers_20_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_20_self_attn_linear_v_weight_to_fp16_quantized, x = input_1079_cast_fp16)[name = string("linear_185_cast_fp16")]; + tensor var_4758 = const()[name = string("op_4758"), val = tensor([1, -1, 8, 128])]; + tensor v_41_cast_fp16 = reshape(shape = var_4758, x = linear_185_cast_fp16)[name = string("v_41_cast_fp16")]; + tensor value_49_perm_0 = const()[name = string("value_49_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_20_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_20_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(420332480)))]; + tensor var_4771_cast_fp16 = add(x = q_121_cast_fp16, y = encoder_layers_20_self_attn_pos_bias_u_to_fp16)[name = string("op_4771_cast_fp16")]; + tensor encoder_layers_20_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_20_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(420334592)))]; + tensor var_4773_cast_fp16 = add(x = q_121_cast_fp16, y = encoder_layers_20_self_attn_pos_bias_v_to_fp16)[name = string("op_4773_cast_fp16")]; + tensor q_with_bias_v_41_perm_0 = const()[name = string("q_with_bias_v_41_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_527_transpose_x_0 = const()[name = string("x_527_transpose_x_0"), val = bool(false)]; + bool x_527_transpose_y_0 = const()[name = string("x_527_transpose_y_0"), val = bool(false)]; + tensor op_4775_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(420336704))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(420536448))))[name = string("op_4775_to_fp16_quantized")]; + tensor q_with_bias_v_41_cast_fp16 = transpose(perm = q_with_bias_v_41_perm_0, x = var_4773_cast_fp16)[name = string("transpose_182")]; + tensor x_527_cast_fp16 = matmul(transpose_x = x_527_transpose_x_0, transpose_y = x_527_transpose_y_0, x = q_with_bias_v_41_cast_fp16, y = op_4775_to_fp16_quantized)[name = string("x_527_cast_fp16")]; + tensor x_529_pad_0 = const()[name = string("x_529_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_529_mode_0 = const()[name = string("x_529_mode_0"), val = string("constant")]; + fp16 const_339_to_fp16 = const()[name = string("const_339_to_fp16"), val = fp16(0x0p+0)]; + tensor x_529_cast_fp16 = pad(constant_val = const_339_to_fp16, mode = x_529_mode_0, pad = x_529_pad_0, x = x_527_cast_fp16)[name = string("x_529_cast_fp16")]; + tensor var_4783 = const()[name = string("op_4783"), val = tensor([1, 8, -1, 56])]; + tensor x_531_cast_fp16 = reshape(shape = var_4783, x = x_529_cast_fp16)[name = string("x_531_cast_fp16")]; + tensor var_4787_begin_0 = const()[name = string("op_4787_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_4787_end_0 = const()[name = string("op_4787_end_0"), val = tensor([1, 8, 196, 56])]; + tensor var_4787_end_mask_0 = const()[name = string("op_4787_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_4787_cast_fp16 = slice_by_index(begin = var_4787_begin_0, end = var_4787_end_0, end_mask = var_4787_end_mask_0, x = x_531_cast_fp16)[name = string("op_4787_cast_fp16")]; + tensor var_4788 = const()[name = string("op_4788"), val = tensor([1, 8, 56, 195])]; + tensor matrix_bd_81_cast_fp16 = reshape(shape = var_4788, x = var_4787_cast_fp16)[name = string("matrix_bd_81_cast_fp16")]; + bool matrix_ac_41_transpose_x_0 = const()[name = string("matrix_ac_41_transpose_x_0"), val = bool(false)]; + bool matrix_ac_41_transpose_y_0 = const()[name = string("matrix_ac_41_transpose_y_0"), val = bool(false)]; + tensor transpose_136_perm_0 = const()[name = string("transpose_136_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_137_perm_0 = const()[name = string("transpose_137_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_137 = transpose(perm = transpose_137_perm_0, x = k_81_cast_fp16)[name = string("transpose_180")]; + tensor transpose_136 = transpose(perm = transpose_136_perm_0, x = var_4771_cast_fp16)[name = string("transpose_181")]; + tensor matrix_ac_41_cast_fp16 = matmul(transpose_x = matrix_ac_41_transpose_x_0, transpose_y = matrix_ac_41_transpose_y_0, x = transpose_136, y = transpose_137)[name = string("matrix_ac_41_cast_fp16")]; + tensor matrix_bd_83_begin_0 = const()[name = string("matrix_bd_83_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_83_end_0 = const()[name = string("matrix_bd_83_end_0"), val = tensor([1, 8, 56, 98])]; + tensor matrix_bd_83_end_mask_0 = const()[name = string("matrix_bd_83_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_83_cast_fp16 = slice_by_index(begin = matrix_bd_83_begin_0, end = matrix_bd_83_end_0, end_mask = matrix_bd_83_end_mask_0, x = matrix_bd_81_cast_fp16)[name = string("matrix_bd_83_cast_fp16")]; + tensor var_4797_cast_fp16 = add(x = matrix_ac_41_cast_fp16, y = matrix_bd_83_cast_fp16)[name = string("op_4797_cast_fp16")]; + fp16 _inversed_scores_81_y_0_to_fp16 = const()[name = string("_inversed_scores_81_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_81_cast_fp16 = mul(x = var_4797_cast_fp16, y = _inversed_scores_81_y_0_to_fp16)[name = string("_inversed_scores_81_cast_fp16")]; + tensor scores_83_cast_fp16 = select(a = var_44_to_fp16, b = _inversed_scores_81_cast_fp16, cond = mask_11)[name = string("scores_83_cast_fp16")]; + tensor var_4803_cast_fp16 = softmax(axis = var_58, x = scores_83_cast_fp16)[name = string("op_4803_cast_fp16")]; + tensor input_1081_cast_fp16 = select(a = var_43_to_fp16, b = var_4803_cast_fp16, cond = mask_11)[name = string("input_1081_cast_fp16")]; + bool x_533_transpose_x_0 = const()[name = string("x_533_transpose_x_0"), val = bool(false)]; + bool x_533_transpose_y_0 = const()[name = string("x_533_transpose_y_0"), val = bool(false)]; + tensor value_49_cast_fp16 = transpose(perm = value_49_perm_0, x = v_41_cast_fp16)[name = string("transpose_179")]; + tensor x_533_cast_fp16 = matmul(transpose_x = x_533_transpose_x_0, transpose_y = x_533_transpose_y_0, x = input_1081_cast_fp16, y = value_49_cast_fp16)[name = string("x_533_cast_fp16")]; + tensor var_4807_perm_0 = const()[name = string("op_4807_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_4808 = const()[name = string("op_4808"), val = tensor([1, -1, 1024])]; + tensor var_4807_cast_fp16 = transpose(perm = var_4807_perm_0, x = x_533_cast_fp16)[name = string("transpose_178")]; + tensor input_1083_cast_fp16 = reshape(shape = var_4808, x = var_4807_cast_fp16)[name = string("input_1083_cast_fp16")]; + tensor encoder_layers_20_self_attn_linear_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(420536960))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421585600))))[name = string("encoder_layers_20_self_attn_linear_out_weight_to_fp16_quantized")]; + tensor encoder_layers_20_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_20_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421587712)))]; + tensor linear_187_cast_fp16 = linear(bias = encoder_layers_20_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_20_self_attn_linear_out_weight_to_fp16_quantized, x = input_1083_cast_fp16)[name = string("linear_187_cast_fp16")]; + tensor input_1087_cast_fp16 = add(x = input_1077_cast_fp16, y = linear_187_cast_fp16)[name = string("input_1087_cast_fp16")]; + tensor x_537_axes_0 = const()[name = string("x_537_axes_0"), val = tensor([-1])]; + tensor encoder_layers_20_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_20_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421589824)))]; + tensor encoder_layers_20_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_20_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421591936)))]; + tensor x_537_cast_fp16 = layer_norm(axes = x_537_axes_0, beta = encoder_layers_20_norm_conv_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_20_norm_conv_weight_to_fp16, x = input_1087_cast_fp16)[name = string("x_537_cast_fp16")]; + tensor input_1089_perm_0 = const()[name = string("input_1089_perm_0"), val = tensor([0, 2, 1])]; + string input_1091_pad_type_0 = const()[name = string("input_1091_pad_type_0"), val = string("valid")]; + tensor input_1091_strides_0 = const()[name = string("input_1091_strides_0"), val = tensor([1])]; + tensor input_1091_pad_0 = const()[name = string("input_1091_pad_0"), val = tensor([0, 0])]; + tensor input_1091_dilations_0 = const()[name = string("input_1091_dilations_0"), val = tensor([1])]; + int32 input_1091_groups_0 = const()[name = string("input_1091_groups_0"), val = int32(1)]; + tensor encoder_layers_20_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421594048))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(423691264))))[name = string("encoder_layers_20_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_1089_cast_fp16 = transpose(perm = input_1089_perm_0, x = x_537_cast_fp16)[name = string("transpose_177")]; + tensor input_1091_cast_fp16 = conv(dilations = input_1091_dilations_0, groups = input_1091_groups_0, pad = input_1091_pad_0, pad_type = input_1091_pad_type_0, strides = input_1091_strides_0, weight = encoder_layers_20_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_1089_cast_fp16)[name = string("input_1091_cast_fp16")]; + int32 x_539_split_num_splits_0 = const()[name = string("x_539_split_num_splits_0"), val = int32(2)]; + int32 x_539_split_axis_0 = const()[name = string("x_539_split_axis_0"), val = int32(1)]; + tensor x_539_split_cast_fp16_0, tensor x_539_split_cast_fp16_1 = split(axis = x_539_split_axis_0, num_splits = x_539_split_num_splits_0, x = input_1091_cast_fp16)[name = string("x_539_split_cast_fp16")]; + tensor x_539_split_1_sigmoid_cast_fp16 = sigmoid(x = x_539_split_cast_fp16_1)[name = string("x_539_split_1_sigmoid_cast_fp16")]; + tensor x_539_cast_fp16 = mul(x = x_539_split_cast_fp16_0, y = x_539_split_1_sigmoid_cast_fp16)[name = string("x_539_cast_fp16")]; + tensor input_1093_cast_fp16 = select(a = var_43_to_fp16, b = x_539_cast_fp16, cond = var_574)[name = string("input_1093_cast_fp16")]; + bool new_x_83_interleave_0 = const()[name = string("new_x_83_interleave_0"), val = bool(false)]; + tensor new_x_83_cast_fp16 = concat(axis = var_58, interleave = new_x_83_interleave_0, values = (cache_83_cast_fp16, input_1093_cast_fp16))[name = string("new_x_83_cast_fp16")]; + tensor var_4847_begin_0 = const()[name = string("op_4847_begin_0"), val = tensor([0, 0, 56])]; + tensor var_4847_end_0 = const()[name = string("op_4847_end_0"), val = tensor([1, 1024, 64])]; + tensor var_4847_end_mask_0 = const()[name = string("op_4847_end_mask_0"), val = tensor([true, true, true])]; + tensor var_4847_cast_fp16 = slice_by_index(begin = var_4847_begin_0, end = var_4847_end_0, end_mask = var_4847_end_mask_0, x = new_x_83_cast_fp16)[name = string("op_4847_cast_fp16")]; + string x_541_pad_type_0 = const()[name = string("x_541_pad_type_0"), val = string("valid")]; + int32 x_541_groups_0 = const()[name = string("x_541_groups_0"), val = int32(1024)]; + tensor x_541_strides_0 = const()[name = string("x_541_strides_0"), val = tensor([1])]; + tensor x_541_pad_0 = const()[name = string("x_541_pad_0"), val = tensor([0, 0])]; + tensor x_541_dilations_0 = const()[name = string("x_541_dilations_0"), val = tensor([1])]; + tensor encoder_layers_20_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(423695424))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(423704704))))[name = string("encoder_layers_20_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_541_cast_fp16 = conv(dilations = x_541_dilations_0, groups = x_541_groups_0, pad = x_541_pad_0, pad_type = x_541_pad_type_0, strides = x_541_strides_0, weight = encoder_layers_20_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_83_cast_fp16)[name = string("x_541_cast_fp16")]; + tensor input_1095_perm_0 = const()[name = string("input_1095_perm_0"), val = tensor([0, 2, 1])]; + tensor x_543_axes_0 = const()[name = string("x_543_axes_0"), val = tensor([-1])]; + tensor encoder_layers_20_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_20_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(423706816)))]; + tensor encoder_layers_20_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_20_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(423708928)))]; + tensor input_1095_cast_fp16 = transpose(perm = input_1095_perm_0, x = x_541_cast_fp16)[name = string("transpose_176")]; + tensor x_543_cast_fp16 = layer_norm(axes = x_543_axes_0, beta = encoder_layers_20_conv_batch_norm_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_20_conv_batch_norm_weight_to_fp16, x = input_1095_cast_fp16)[name = string("x_543_cast_fp16")]; + tensor input_1097_perm_0 = const()[name = string("input_1097_perm_0"), val = tensor([0, 2, 1])]; + tensor input_1097_cast_fp16 = transpose(perm = input_1097_perm_0, x = x_543_cast_fp16)[name = string("transpose_175")]; + tensor input_1099_cast_fp16 = silu(x = input_1097_cast_fp16)[name = string("input_1099_cast_fp16")]; + string x_545_pad_type_0 = const()[name = string("x_545_pad_type_0"), val = string("valid")]; + tensor x_545_strides_0 = const()[name = string("x_545_strides_0"), val = tensor([1])]; + tensor x_545_pad_0 = const()[name = string("x_545_pad_0"), val = tensor([0, 0])]; + tensor x_545_dilations_0 = const()[name = string("x_545_dilations_0"), val = tensor([1])]; + int32 x_545_groups_0 = const()[name = string("x_545_groups_0"), val = int32(1)]; + tensor encoder_layers_20_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(423711040))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(424759680))))[name = string("encoder_layers_20_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_545_cast_fp16 = conv(dilations = x_545_dilations_0, groups = x_545_groups_0, pad = x_545_pad_0, pad_type = x_545_pad_type_0, strides = x_545_strides_0, weight = encoder_layers_20_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_1099_cast_fp16)[name = string("x_545_cast_fp16")]; + tensor input_1101_perm_0 = const()[name = string("input_1101_perm_0"), val = tensor([0, 2, 1])]; + tensor input_1101_cast_fp16 = transpose(perm = input_1101_perm_0, x = x_545_cast_fp16)[name = string("transpose_174")]; + tensor input_1103_cast_fp16 = add(x = input_1087_cast_fp16, y = input_1101_cast_fp16)[name = string("input_1103_cast_fp16")]; + tensor input_1105_axes_0 = const()[name = string("input_1105_axes_0"), val = tensor([-1])]; + tensor encoder_layers_20_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_20_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(424761792)))]; + tensor encoder_layers_20_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_20_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(424763904)))]; + tensor input_1105_cast_fp16 = layer_norm(axes = input_1105_axes_0, beta = encoder_layers_20_norm_feed_forward2_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_20_norm_feed_forward2_weight_to_fp16, x = input_1103_cast_fp16)[name = string("input_1105_cast_fp16")]; + tensor encoder_layers_20_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(424766016))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(428960384))))[name = string("encoder_layers_20_feed_forward2_linear1_weight_to_fp16_quantized")]; + tensor encoder_layers_20_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_20_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(428968640)))]; + tensor linear_188_cast_fp16 = linear(bias = encoder_layers_20_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_20_feed_forward2_linear1_weight_to_fp16_quantized, x = input_1105_cast_fp16)[name = string("linear_188_cast_fp16")]; + tensor input_1109_cast_fp16 = silu(x = linear_188_cast_fp16)[name = string("input_1109_cast_fp16")]; + tensor encoder_layers_20_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(428976896))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(433171264))))[name = string("encoder_layers_20_feed_forward2_linear2_weight_to_fp16_quantized")]; + tensor encoder_layers_20_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_20_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(433173376)))]; + tensor linear_189_cast_fp16 = linear(bias = encoder_layers_20_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_20_feed_forward2_linear2_weight_to_fp16_quantized, x = input_1109_cast_fp16)[name = string("linear_189_cast_fp16")]; + fp16 var_4890_to_fp16 = const()[name = string("op_4890_to_fp16"), val = fp16(0x1p-1)]; + tensor var_4891_cast_fp16 = mul(x = linear_189_cast_fp16, y = var_4890_to_fp16)[name = string("op_4891_cast_fp16")]; + tensor input_1115_cast_fp16 = add(x = input_1103_cast_fp16, y = var_4891_cast_fp16)[name = string("input_1115_cast_fp16")]; + tensor input_1117_axes_0 = const()[name = string("input_1117_axes_0"), val = tensor([-1])]; + tensor encoder_layers_20_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_20_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(433175488)))]; + tensor encoder_layers_20_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_20_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(433177600)))]; + tensor input_1117_cast_fp16 = layer_norm(axes = input_1117_axes_0, beta = encoder_layers_20_norm_out_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_20_norm_out_weight_to_fp16, x = input_1115_cast_fp16)[name = string("input_1117_cast_fp16")]; + tensor cache_85_begin_0 = const()[name = string("cache_85_begin_0"), val = tensor([21, 0, 0, 0])]; + tensor cache_85_end_0 = const()[name = string("cache_85_end_0"), val = tensor([22, 1, 42, 1024])]; + tensor cache_85_end_mask_0 = const()[name = string("cache_85_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_85_squeeze_mask_0 = const()[name = string("cache_85_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_85_cast_fp16 = slice_by_index(begin = cache_85_begin_0, end = cache_85_end_0, end_mask = cache_85_end_mask_0, squeeze_mask = cache_85_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_85_cast_fp16")]; + tensor cache_87_begin_0 = const()[name = string("cache_87_begin_0"), val = tensor([21, 0, 0, 0])]; + tensor cache_87_end_0 = const()[name = string("cache_87_end_0"), val = tensor([22, 1, 1024, 8])]; + tensor cache_87_end_mask_0 = const()[name = string("cache_87_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_87_squeeze_mask_0 = const()[name = string("cache_87_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_87_cast_fp16 = slice_by_index(begin = cache_87_begin_0, end = cache_87_end_0, end_mask = cache_87_end_mask_0, squeeze_mask = cache_87_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_87_cast_fp16")]; + tensor input_1119_axes_0 = const()[name = string("input_1119_axes_0"), val = tensor([-1])]; + tensor encoder_layers_21_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_21_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(433179712)))]; + tensor encoder_layers_21_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_21_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(433181824)))]; + tensor input_1119_cast_fp16 = layer_norm(axes = input_1119_axes_0, beta = encoder_layers_21_norm_feed_forward1_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_21_norm_feed_forward1_weight_to_fp16, x = input_1117_cast_fp16)[name = string("input_1119_cast_fp16")]; + tensor encoder_layers_21_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(433183936))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(437378304))))[name = string("encoder_layers_21_feed_forward1_linear1_weight_to_fp16_quantized")]; + tensor encoder_layers_21_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_21_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(437386560)))]; + tensor linear_190_cast_fp16 = linear(bias = encoder_layers_21_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_21_feed_forward1_linear1_weight_to_fp16_quantized, x = input_1119_cast_fp16)[name = string("linear_190_cast_fp16")]; + tensor input_1123_cast_fp16 = silu(x = linear_190_cast_fp16)[name = string("input_1123_cast_fp16")]; + tensor encoder_layers_21_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(437394816))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(441589184))))[name = string("encoder_layers_21_feed_forward1_linear2_weight_to_fp16_quantized")]; + tensor encoder_layers_21_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_21_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(441591296)))]; + tensor linear_191_cast_fp16 = linear(bias = encoder_layers_21_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_21_feed_forward1_linear2_weight_to_fp16_quantized, x = input_1123_cast_fp16)[name = string("linear_191_cast_fp16")]; + fp16 var_4927_to_fp16 = const()[name = string("op_4927_to_fp16"), val = fp16(0x1p-1)]; + tensor var_4928_cast_fp16 = mul(x = linear_191_cast_fp16, y = var_4927_to_fp16)[name = string("op_4928_cast_fp16")]; + tensor input_1129_cast_fp16 = add(x = input_1117_cast_fp16, y = var_4928_cast_fp16)[name = string("input_1129_cast_fp16")]; + tensor key_43_axes_0 = const()[name = string("key_43_axes_0"), val = tensor([-1])]; + tensor encoder_layers_21_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_21_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(441593408)))]; + tensor encoder_layers_21_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_21_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(441595520)))]; + tensor key_43_cast_fp16 = layer_norm(axes = key_43_axes_0, beta = encoder_layers_21_norm_self_att_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_21_norm_self_att_weight_to_fp16, x = input_1129_cast_fp16)[name = string("key_43_cast_fp16")]; + bool input_1131_interleave_0 = const()[name = string("input_1131_interleave_0"), val = bool(false)]; + tensor input_1131_cast_fp16 = concat(axis = var_67, interleave = input_1131_interleave_0, values = (cache_85_cast_fp16, key_43_cast_fp16))[name = string("input_1131_cast_fp16")]; + bool var_4956_interleave_0 = const()[name = string("op_4956_interleave_0"), val = bool(false)]; + tensor var_4956_cast_fp16 = concat(axis = var_67, interleave = var_4956_interleave_0, values = key_43_cast_fp16)[name = string("op_4956_cast_fp16")]; + tensor encoder_layers_21_self_attn_linear_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(441597632))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(442646272))))[name = string("encoder_layers_21_self_attn_linear_q_weight_to_fp16_quantized")]; + tensor encoder_layers_21_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_21_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(442648384)))]; + tensor linear_192_cast_fp16 = linear(bias = encoder_layers_21_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_21_self_attn_linear_q_weight_to_fp16_quantized, x = key_43_cast_fp16)[name = string("linear_192_cast_fp16")]; + tensor var_4961 = const()[name = string("op_4961"), val = tensor([1, -1, 8, 128])]; + tensor q_127_cast_fp16 = reshape(shape = var_4961, x = linear_192_cast_fp16)[name = string("q_127_cast_fp16")]; + tensor encoder_layers_21_self_attn_linear_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(442650496))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(443699136))))[name = string("encoder_layers_21_self_attn_linear_k_weight_to_fp16_quantized")]; + tensor encoder_layers_21_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_21_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(443701248)))]; + tensor linear_193_cast_fp16 = linear(bias = encoder_layers_21_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_21_self_attn_linear_k_weight_to_fp16_quantized, x = input_1131_cast_fp16)[name = string("linear_193_cast_fp16")]; + tensor var_4966 = const()[name = string("op_4966"), val = tensor([1, -1, 8, 128])]; + tensor k_85_cast_fp16 = reshape(shape = var_4966, x = linear_193_cast_fp16)[name = string("k_85_cast_fp16")]; + tensor encoder_layers_21_self_attn_linear_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(443703360))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(444752000))))[name = string("encoder_layers_21_self_attn_linear_v_weight_to_fp16_quantized")]; + tensor encoder_layers_21_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_21_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(444754112)))]; + tensor linear_194_cast_fp16 = linear(bias = encoder_layers_21_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_21_self_attn_linear_v_weight_to_fp16_quantized, x = input_1131_cast_fp16)[name = string("linear_194_cast_fp16")]; + tensor var_4971 = const()[name = string("op_4971"), val = tensor([1, -1, 8, 128])]; + tensor v_43_cast_fp16 = reshape(shape = var_4971, x = linear_194_cast_fp16)[name = string("v_43_cast_fp16")]; + tensor value_51_perm_0 = const()[name = string("value_51_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_21_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_21_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(444756224)))]; + tensor var_4984_cast_fp16 = add(x = q_127_cast_fp16, y = encoder_layers_21_self_attn_pos_bias_u_to_fp16)[name = string("op_4984_cast_fp16")]; + tensor encoder_layers_21_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_21_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(444758336)))]; + tensor var_4986_cast_fp16 = add(x = q_127_cast_fp16, y = encoder_layers_21_self_attn_pos_bias_v_to_fp16)[name = string("op_4986_cast_fp16")]; + tensor q_with_bias_v_43_perm_0 = const()[name = string("q_with_bias_v_43_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_553_transpose_x_0 = const()[name = string("x_553_transpose_x_0"), val = bool(false)]; + bool x_553_transpose_y_0 = const()[name = string("x_553_transpose_y_0"), val = bool(false)]; + tensor op_4988_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(444760448))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(444960192))))[name = string("op_4988_to_fp16_quantized")]; + tensor q_with_bias_v_43_cast_fp16 = transpose(perm = q_with_bias_v_43_perm_0, x = var_4986_cast_fp16)[name = string("transpose_173")]; + tensor x_553_cast_fp16 = matmul(transpose_x = x_553_transpose_x_0, transpose_y = x_553_transpose_y_0, x = q_with_bias_v_43_cast_fp16, y = op_4988_to_fp16_quantized)[name = string("x_553_cast_fp16")]; + tensor x_555_pad_0 = const()[name = string("x_555_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_555_mode_0 = const()[name = string("x_555_mode_0"), val = string("constant")]; + fp16 const_352_to_fp16 = const()[name = string("const_352_to_fp16"), val = fp16(0x0p+0)]; + tensor x_555_cast_fp16 = pad(constant_val = const_352_to_fp16, mode = x_555_mode_0, pad = x_555_pad_0, x = x_553_cast_fp16)[name = string("x_555_cast_fp16")]; + tensor var_4996 = const()[name = string("op_4996"), val = tensor([1, 8, -1, 56])]; + tensor x_557_cast_fp16 = reshape(shape = var_4996, x = x_555_cast_fp16)[name = string("x_557_cast_fp16")]; + tensor var_5000_begin_0 = const()[name = string("op_5000_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_5000_end_0 = const()[name = string("op_5000_end_0"), val = tensor([1, 8, 196, 56])]; + tensor var_5000_end_mask_0 = const()[name = string("op_5000_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_5000_cast_fp16 = slice_by_index(begin = var_5000_begin_0, end = var_5000_end_0, end_mask = var_5000_end_mask_0, x = x_557_cast_fp16)[name = string("op_5000_cast_fp16")]; + tensor var_5001 = const()[name = string("op_5001"), val = tensor([1, 8, 56, 195])]; + tensor matrix_bd_85_cast_fp16 = reshape(shape = var_5001, x = var_5000_cast_fp16)[name = string("matrix_bd_85_cast_fp16")]; + bool matrix_ac_43_transpose_x_0 = const()[name = string("matrix_ac_43_transpose_x_0"), val = bool(false)]; + bool matrix_ac_43_transpose_y_0 = const()[name = string("matrix_ac_43_transpose_y_0"), val = bool(false)]; + tensor transpose_138_perm_0 = const()[name = string("transpose_138_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_139_perm_0 = const()[name = string("transpose_139_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_139 = transpose(perm = transpose_139_perm_0, x = k_85_cast_fp16)[name = string("transpose_171")]; + tensor transpose_138 = transpose(perm = transpose_138_perm_0, x = var_4984_cast_fp16)[name = string("transpose_172")]; + tensor matrix_ac_43_cast_fp16 = matmul(transpose_x = matrix_ac_43_transpose_x_0, transpose_y = matrix_ac_43_transpose_y_0, x = transpose_138, y = transpose_139)[name = string("matrix_ac_43_cast_fp16")]; + tensor matrix_bd_87_begin_0 = const()[name = string("matrix_bd_87_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_87_end_0 = const()[name = string("matrix_bd_87_end_0"), val = tensor([1, 8, 56, 98])]; + tensor matrix_bd_87_end_mask_0 = const()[name = string("matrix_bd_87_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_87_cast_fp16 = slice_by_index(begin = matrix_bd_87_begin_0, end = matrix_bd_87_end_0, end_mask = matrix_bd_87_end_mask_0, x = matrix_bd_85_cast_fp16)[name = string("matrix_bd_87_cast_fp16")]; + tensor var_5010_cast_fp16 = add(x = matrix_ac_43_cast_fp16, y = matrix_bd_87_cast_fp16)[name = string("op_5010_cast_fp16")]; + fp16 _inversed_scores_85_y_0_to_fp16 = const()[name = string("_inversed_scores_85_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_85_cast_fp16 = mul(x = var_5010_cast_fp16, y = _inversed_scores_85_y_0_to_fp16)[name = string("_inversed_scores_85_cast_fp16")]; + tensor scores_87_cast_fp16 = select(a = var_44_to_fp16, b = _inversed_scores_85_cast_fp16, cond = mask_11)[name = string("scores_87_cast_fp16")]; + tensor var_5016_cast_fp16 = softmax(axis = var_58, x = scores_87_cast_fp16)[name = string("op_5016_cast_fp16")]; + tensor input_1133_cast_fp16 = select(a = var_43_to_fp16, b = var_5016_cast_fp16, cond = mask_11)[name = string("input_1133_cast_fp16")]; + bool x_559_transpose_x_0 = const()[name = string("x_559_transpose_x_0"), val = bool(false)]; + bool x_559_transpose_y_0 = const()[name = string("x_559_transpose_y_0"), val = bool(false)]; + tensor value_51_cast_fp16 = transpose(perm = value_51_perm_0, x = v_43_cast_fp16)[name = string("transpose_170")]; + tensor x_559_cast_fp16 = matmul(transpose_x = x_559_transpose_x_0, transpose_y = x_559_transpose_y_0, x = input_1133_cast_fp16, y = value_51_cast_fp16)[name = string("x_559_cast_fp16")]; + tensor var_5020_perm_0 = const()[name = string("op_5020_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_5021 = const()[name = string("op_5021"), val = tensor([1, -1, 1024])]; + tensor var_5020_cast_fp16 = transpose(perm = var_5020_perm_0, x = x_559_cast_fp16)[name = string("transpose_169")]; + tensor input_1135_cast_fp16 = reshape(shape = var_5021, x = var_5020_cast_fp16)[name = string("input_1135_cast_fp16")]; + tensor encoder_layers_21_self_attn_linear_out_weight_to_fp16 = const()[name = string("encoder_layers_21_self_attn_linear_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(444960704)))]; + tensor encoder_layers_21_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_21_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(447057920)))]; + tensor linear_196_cast_fp16 = linear(bias = encoder_layers_21_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_21_self_attn_linear_out_weight_to_fp16, x = input_1135_cast_fp16)[name = string("linear_196_cast_fp16")]; + tensor input_1139_cast_fp16 = add(x = input_1129_cast_fp16, y = linear_196_cast_fp16)[name = string("input_1139_cast_fp16")]; + tensor x_563_axes_0 = const()[name = string("x_563_axes_0"), val = tensor([-1])]; + tensor encoder_layers_21_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_21_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(447060032)))]; + tensor encoder_layers_21_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_21_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(447062144)))]; + tensor x_563_cast_fp16 = layer_norm(axes = x_563_axes_0, beta = encoder_layers_21_norm_conv_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_21_norm_conv_weight_to_fp16, x = input_1139_cast_fp16)[name = string("x_563_cast_fp16")]; + tensor input_1141_perm_0 = const()[name = string("input_1141_perm_0"), val = tensor([0, 2, 1])]; + string input_1143_pad_type_0 = const()[name = string("input_1143_pad_type_0"), val = string("valid")]; + tensor input_1143_strides_0 = const()[name = string("input_1143_strides_0"), val = tensor([1])]; + tensor input_1143_pad_0 = const()[name = string("input_1143_pad_0"), val = tensor([0, 0])]; + tensor input_1143_dilations_0 = const()[name = string("input_1143_dilations_0"), val = tensor([1])]; + int32 input_1143_groups_0 = const()[name = string("input_1143_groups_0"), val = int32(1)]; + tensor encoder_layers_21_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(447064256))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(449161472))))[name = string("encoder_layers_21_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_1141_cast_fp16 = transpose(perm = input_1141_perm_0, x = x_563_cast_fp16)[name = string("transpose_168")]; + tensor input_1143_cast_fp16 = conv(dilations = input_1143_dilations_0, groups = input_1143_groups_0, pad = input_1143_pad_0, pad_type = input_1143_pad_type_0, strides = input_1143_strides_0, weight = encoder_layers_21_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_1141_cast_fp16)[name = string("input_1143_cast_fp16")]; + int32 x_565_split_num_splits_0 = const()[name = string("x_565_split_num_splits_0"), val = int32(2)]; + int32 x_565_split_axis_0 = const()[name = string("x_565_split_axis_0"), val = int32(1)]; + tensor x_565_split_cast_fp16_0, tensor x_565_split_cast_fp16_1 = split(axis = x_565_split_axis_0, num_splits = x_565_split_num_splits_0, x = input_1143_cast_fp16)[name = string("x_565_split_cast_fp16")]; + tensor x_565_split_1_sigmoid_cast_fp16 = sigmoid(x = x_565_split_cast_fp16_1)[name = string("x_565_split_1_sigmoid_cast_fp16")]; + tensor x_565_cast_fp16 = mul(x = x_565_split_cast_fp16_0, y = x_565_split_1_sigmoid_cast_fp16)[name = string("x_565_cast_fp16")]; + tensor input_1145_cast_fp16 = select(a = var_43_to_fp16, b = x_565_cast_fp16, cond = var_574)[name = string("input_1145_cast_fp16")]; + bool new_x_87_interleave_0 = const()[name = string("new_x_87_interleave_0"), val = bool(false)]; + tensor new_x_87_cast_fp16 = concat(axis = var_58, interleave = new_x_87_interleave_0, values = (cache_87_cast_fp16, input_1145_cast_fp16))[name = string("new_x_87_cast_fp16")]; + tensor var_5060_begin_0 = const()[name = string("op_5060_begin_0"), val = tensor([0, 0, 56])]; + tensor var_5060_end_0 = const()[name = string("op_5060_end_0"), val = tensor([1, 1024, 64])]; + tensor var_5060_end_mask_0 = const()[name = string("op_5060_end_mask_0"), val = tensor([true, true, true])]; + tensor var_5060_cast_fp16 = slice_by_index(begin = var_5060_begin_0, end = var_5060_end_0, end_mask = var_5060_end_mask_0, x = new_x_87_cast_fp16)[name = string("op_5060_cast_fp16")]; + string x_567_pad_type_0 = const()[name = string("x_567_pad_type_0"), val = string("valid")]; + int32 x_567_groups_0 = const()[name = string("x_567_groups_0"), val = int32(1024)]; + tensor x_567_strides_0 = const()[name = string("x_567_strides_0"), val = tensor([1])]; + tensor x_567_pad_0 = const()[name = string("x_567_pad_0"), val = tensor([0, 0])]; + tensor x_567_dilations_0 = const()[name = string("x_567_dilations_0"), val = tensor([1])]; + tensor encoder_layers_21_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(449165632))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(449174912))))[name = string("encoder_layers_21_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_567_cast_fp16 = conv(dilations = x_567_dilations_0, groups = x_567_groups_0, pad = x_567_pad_0, pad_type = x_567_pad_type_0, strides = x_567_strides_0, weight = encoder_layers_21_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_87_cast_fp16)[name = string("x_567_cast_fp16")]; + tensor input_1147_perm_0 = const()[name = string("input_1147_perm_0"), val = tensor([0, 2, 1])]; + tensor x_569_axes_0 = const()[name = string("x_569_axes_0"), val = tensor([-1])]; + tensor encoder_layers_21_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_21_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(449177024)))]; + tensor encoder_layers_21_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_21_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(449179136)))]; + tensor input_1147_cast_fp16 = transpose(perm = input_1147_perm_0, x = x_567_cast_fp16)[name = string("transpose_167")]; + tensor x_569_cast_fp16 = layer_norm(axes = x_569_axes_0, beta = encoder_layers_21_conv_batch_norm_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_21_conv_batch_norm_weight_to_fp16, x = input_1147_cast_fp16)[name = string("x_569_cast_fp16")]; + tensor input_1149_perm_0 = const()[name = string("input_1149_perm_0"), val = tensor([0, 2, 1])]; + tensor input_1149_cast_fp16 = transpose(perm = input_1149_perm_0, x = x_569_cast_fp16)[name = string("transpose_166")]; + tensor input_1151_cast_fp16 = silu(x = input_1149_cast_fp16)[name = string("input_1151_cast_fp16")]; + string x_571_pad_type_0 = const()[name = string("x_571_pad_type_0"), val = string("valid")]; + tensor x_571_strides_0 = const()[name = string("x_571_strides_0"), val = tensor([1])]; + tensor x_571_pad_0 = const()[name = string("x_571_pad_0"), val = tensor([0, 0])]; + tensor x_571_dilations_0 = const()[name = string("x_571_dilations_0"), val = tensor([1])]; + int32 x_571_groups_0 = const()[name = string("x_571_groups_0"), val = int32(1)]; + tensor encoder_layers_21_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(449181248))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(450229888))))[name = string("encoder_layers_21_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_571_cast_fp16 = conv(dilations = x_571_dilations_0, groups = x_571_groups_0, pad = x_571_pad_0, pad_type = x_571_pad_type_0, strides = x_571_strides_0, weight = encoder_layers_21_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_1151_cast_fp16)[name = string("x_571_cast_fp16")]; + tensor input_1153_perm_0 = const()[name = string("input_1153_perm_0"), val = tensor([0, 2, 1])]; + tensor input_1153_cast_fp16 = transpose(perm = input_1153_perm_0, x = x_571_cast_fp16)[name = string("transpose_165")]; + tensor input_1155_cast_fp16 = add(x = input_1139_cast_fp16, y = input_1153_cast_fp16)[name = string("input_1155_cast_fp16")]; + tensor input_1157_axes_0 = const()[name = string("input_1157_axes_0"), val = tensor([-1])]; + tensor encoder_layers_21_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_21_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(450232000)))]; + tensor encoder_layers_21_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_21_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(450234112)))]; + tensor input_1157_cast_fp16 = layer_norm(axes = input_1157_axes_0, beta = encoder_layers_21_norm_feed_forward2_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_21_norm_feed_forward2_weight_to_fp16, x = input_1155_cast_fp16)[name = string("input_1157_cast_fp16")]; + tensor encoder_layers_21_feed_forward2_linear1_weight_to_fp16 = const()[name = string("encoder_layers_21_feed_forward2_linear1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(450236224)))]; + tensor encoder_layers_21_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_21_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(458624896)))]; + tensor linear_197_cast_fp16 = linear(bias = encoder_layers_21_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_21_feed_forward2_linear1_weight_to_fp16, x = input_1157_cast_fp16)[name = string("linear_197_cast_fp16")]; + tensor input_1161_cast_fp16 = silu(x = linear_197_cast_fp16)[name = string("input_1161_cast_fp16")]; + tensor encoder_layers_21_feed_forward2_linear2_weight_to_fp16 = const()[name = string("encoder_layers_21_feed_forward2_linear2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(458633152)))]; + tensor encoder_layers_21_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_21_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(467021824)))]; + tensor linear_198_cast_fp16 = linear(bias = encoder_layers_21_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_21_feed_forward2_linear2_weight_to_fp16, x = input_1161_cast_fp16)[name = string("linear_198_cast_fp16")]; + fp16 var_5103_to_fp16 = const()[name = string("op_5103_to_fp16"), val = fp16(0x1p-1)]; + tensor var_5104_cast_fp16 = mul(x = linear_198_cast_fp16, y = var_5103_to_fp16)[name = string("op_5104_cast_fp16")]; + tensor input_1167_cast_fp16 = add(x = input_1155_cast_fp16, y = var_5104_cast_fp16)[name = string("input_1167_cast_fp16")]; + tensor input_1169_axes_0 = const()[name = string("input_1169_axes_0"), val = tensor([-1])]; + tensor encoder_layers_21_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_21_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(467023936)))]; + tensor encoder_layers_21_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_21_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(467026048)))]; + tensor input_1169_cast_fp16 = layer_norm(axes = input_1169_axes_0, beta = encoder_layers_21_norm_out_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_21_norm_out_weight_to_fp16, x = input_1167_cast_fp16)[name = string("input_1169_cast_fp16")]; + tensor cache_89_begin_0 = const()[name = string("cache_89_begin_0"), val = tensor([22, 0, 0, 0])]; + tensor cache_89_end_0 = const()[name = string("cache_89_end_0"), val = tensor([23, 1, 42, 1024])]; + tensor cache_89_end_mask_0 = const()[name = string("cache_89_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_89_squeeze_mask_0 = const()[name = string("cache_89_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_89_cast_fp16 = slice_by_index(begin = cache_89_begin_0, end = cache_89_end_0, end_mask = cache_89_end_mask_0, squeeze_mask = cache_89_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_89_cast_fp16")]; + tensor cache_91_begin_0 = const()[name = string("cache_91_begin_0"), val = tensor([22, 0, 0, 0])]; + tensor cache_91_end_0 = const()[name = string("cache_91_end_0"), val = tensor([23, 1, 1024, 8])]; + tensor cache_91_end_mask_0 = const()[name = string("cache_91_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_91_squeeze_mask_0 = const()[name = string("cache_91_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_91_cast_fp16 = slice_by_index(begin = cache_91_begin_0, end = cache_91_end_0, end_mask = cache_91_end_mask_0, squeeze_mask = cache_91_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_91_cast_fp16")]; + tensor input_1171_axes_0 = const()[name = string("input_1171_axes_0"), val = tensor([-1])]; + tensor encoder_layers_22_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_22_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(467028160)))]; + tensor encoder_layers_22_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_22_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(467030272)))]; + tensor input_1171_cast_fp16 = layer_norm(axes = input_1171_axes_0, beta = encoder_layers_22_norm_feed_forward1_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_22_norm_feed_forward1_weight_to_fp16, x = input_1169_cast_fp16)[name = string("input_1171_cast_fp16")]; + tensor encoder_layers_22_feed_forward1_linear1_weight_to_fp16 = const()[name = string("encoder_layers_22_feed_forward1_linear1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(467032384)))]; + tensor encoder_layers_22_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_22_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(475421056)))]; + tensor linear_199_cast_fp16 = linear(bias = encoder_layers_22_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_22_feed_forward1_linear1_weight_to_fp16, x = input_1171_cast_fp16)[name = string("linear_199_cast_fp16")]; + tensor input_1175_cast_fp16 = silu(x = linear_199_cast_fp16)[name = string("input_1175_cast_fp16")]; + tensor encoder_layers_22_feed_forward1_linear2_weight_to_fp16 = const()[name = string("encoder_layers_22_feed_forward1_linear2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(475429312)))]; + tensor encoder_layers_22_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_22_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(483817984)))]; + tensor linear_200_cast_fp16 = linear(bias = encoder_layers_22_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_22_feed_forward1_linear2_weight_to_fp16, x = input_1175_cast_fp16)[name = string("linear_200_cast_fp16")]; + fp16 var_5140_to_fp16 = const()[name = string("op_5140_to_fp16"), val = fp16(0x1p-1)]; + tensor var_5141_cast_fp16 = mul(x = linear_200_cast_fp16, y = var_5140_to_fp16)[name = string("op_5141_cast_fp16")]; + tensor input_1181_cast_fp16 = add(x = input_1169_cast_fp16, y = var_5141_cast_fp16)[name = string("input_1181_cast_fp16")]; + tensor key_45_axes_0 = const()[name = string("key_45_axes_0"), val = tensor([-1])]; + tensor encoder_layers_22_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_22_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(483820096)))]; + tensor encoder_layers_22_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_22_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(483822208)))]; + tensor key_45_cast_fp16 = layer_norm(axes = key_45_axes_0, beta = encoder_layers_22_norm_self_att_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_22_norm_self_att_weight_to_fp16, x = input_1181_cast_fp16)[name = string("key_45_cast_fp16")]; + bool input_1183_interleave_0 = const()[name = string("input_1183_interleave_0"), val = bool(false)]; + tensor input_1183_cast_fp16 = concat(axis = var_67, interleave = input_1183_interleave_0, values = (cache_89_cast_fp16, key_45_cast_fp16))[name = string("input_1183_cast_fp16")]; + bool var_5169_interleave_0 = const()[name = string("op_5169_interleave_0"), val = bool(false)]; + tensor var_5169_cast_fp16 = concat(axis = var_67, interleave = var_5169_interleave_0, values = key_45_cast_fp16)[name = string("op_5169_cast_fp16")]; + tensor encoder_layers_22_self_attn_linear_q_weight_to_fp16 = const()[name = string("encoder_layers_22_self_attn_linear_q_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(483824320)))]; + tensor encoder_layers_22_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_22_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(485921536)))]; + tensor linear_201_cast_fp16 = linear(bias = encoder_layers_22_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_22_self_attn_linear_q_weight_to_fp16, x = key_45_cast_fp16)[name = string("linear_201_cast_fp16")]; + tensor var_5174 = const()[name = string("op_5174"), val = tensor([1, -1, 8, 128])]; + tensor q_133_cast_fp16 = reshape(shape = var_5174, x = linear_201_cast_fp16)[name = string("q_133_cast_fp16")]; + tensor encoder_layers_22_self_attn_linear_k_weight_to_fp16 = const()[name = string("encoder_layers_22_self_attn_linear_k_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(485923648)))]; + tensor encoder_layers_22_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_22_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(488020864)))]; + tensor linear_202_cast_fp16 = linear(bias = encoder_layers_22_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_22_self_attn_linear_k_weight_to_fp16, x = input_1183_cast_fp16)[name = string("linear_202_cast_fp16")]; + tensor var_5179 = const()[name = string("op_5179"), val = tensor([1, -1, 8, 128])]; + tensor k_89_cast_fp16 = reshape(shape = var_5179, x = linear_202_cast_fp16)[name = string("k_89_cast_fp16")]; + tensor encoder_layers_22_self_attn_linear_v_weight_to_fp16 = const()[name = string("encoder_layers_22_self_attn_linear_v_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(488022976)))]; + tensor encoder_layers_22_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_22_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(490120192)))]; + tensor linear_203_cast_fp16 = linear(bias = encoder_layers_22_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_22_self_attn_linear_v_weight_to_fp16, x = input_1183_cast_fp16)[name = string("linear_203_cast_fp16")]; + tensor var_5184 = const()[name = string("op_5184"), val = tensor([1, -1, 8, 128])]; + tensor v_45_cast_fp16 = reshape(shape = var_5184, x = linear_203_cast_fp16)[name = string("v_45_cast_fp16")]; + tensor value_53_perm_0 = const()[name = string("value_53_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_22_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_22_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(490122304)))]; + tensor var_5197_cast_fp16 = add(x = q_133_cast_fp16, y = encoder_layers_22_self_attn_pos_bias_u_to_fp16)[name = string("op_5197_cast_fp16")]; + tensor encoder_layers_22_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_22_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(490124416)))]; + tensor var_5199_cast_fp16 = add(x = q_133_cast_fp16, y = encoder_layers_22_self_attn_pos_bias_v_to_fp16)[name = string("op_5199_cast_fp16")]; + tensor q_with_bias_v_45_perm_0 = const()[name = string("q_with_bias_v_45_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_579_transpose_x_0 = const()[name = string("x_579_transpose_x_0"), val = bool(false)]; + bool x_579_transpose_y_0 = const()[name = string("x_579_transpose_y_0"), val = bool(false)]; + tensor op_5201_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(490126528))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(490326272))))[name = string("op_5201_to_fp16_quantized")]; + tensor q_with_bias_v_45_cast_fp16 = transpose(perm = q_with_bias_v_45_perm_0, x = var_5199_cast_fp16)[name = string("transpose_164")]; + tensor x_579_cast_fp16 = matmul(transpose_x = x_579_transpose_x_0, transpose_y = x_579_transpose_y_0, x = q_with_bias_v_45_cast_fp16, y = op_5201_to_fp16_quantized)[name = string("x_579_cast_fp16")]; + tensor x_581_pad_0 = const()[name = string("x_581_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_581_mode_0 = const()[name = string("x_581_mode_0"), val = string("constant")]; + fp16 const_365_to_fp16 = const()[name = string("const_365_to_fp16"), val = fp16(0x0p+0)]; + tensor x_581_cast_fp16 = pad(constant_val = const_365_to_fp16, mode = x_581_mode_0, pad = x_581_pad_0, x = x_579_cast_fp16)[name = string("x_581_cast_fp16")]; + tensor var_5209 = const()[name = string("op_5209"), val = tensor([1, 8, -1, 56])]; + tensor x_583_cast_fp16 = reshape(shape = var_5209, x = x_581_cast_fp16)[name = string("x_583_cast_fp16")]; + tensor var_5213_begin_0 = const()[name = string("op_5213_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_5213_end_0 = const()[name = string("op_5213_end_0"), val = tensor([1, 8, 196, 56])]; + tensor var_5213_end_mask_0 = const()[name = string("op_5213_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_5213_cast_fp16 = slice_by_index(begin = var_5213_begin_0, end = var_5213_end_0, end_mask = var_5213_end_mask_0, x = x_583_cast_fp16)[name = string("op_5213_cast_fp16")]; + tensor var_5214 = const()[name = string("op_5214"), val = tensor([1, 8, 56, 195])]; + tensor matrix_bd_89_cast_fp16 = reshape(shape = var_5214, x = var_5213_cast_fp16)[name = string("matrix_bd_89_cast_fp16")]; + bool matrix_ac_45_transpose_x_0 = const()[name = string("matrix_ac_45_transpose_x_0"), val = bool(false)]; + bool matrix_ac_45_transpose_y_0 = const()[name = string("matrix_ac_45_transpose_y_0"), val = bool(false)]; + tensor transpose_140_perm_0 = const()[name = string("transpose_140_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_141_perm_0 = const()[name = string("transpose_141_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_141 = transpose(perm = transpose_141_perm_0, x = k_89_cast_fp16)[name = string("transpose_162")]; + tensor transpose_140 = transpose(perm = transpose_140_perm_0, x = var_5197_cast_fp16)[name = string("transpose_163")]; + tensor matrix_ac_45_cast_fp16 = matmul(transpose_x = matrix_ac_45_transpose_x_0, transpose_y = matrix_ac_45_transpose_y_0, x = transpose_140, y = transpose_141)[name = string("matrix_ac_45_cast_fp16")]; + tensor matrix_bd_91_begin_0 = const()[name = string("matrix_bd_91_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_91_end_0 = const()[name = string("matrix_bd_91_end_0"), val = tensor([1, 8, 56, 98])]; + tensor matrix_bd_91_end_mask_0 = const()[name = string("matrix_bd_91_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_91_cast_fp16 = slice_by_index(begin = matrix_bd_91_begin_0, end = matrix_bd_91_end_0, end_mask = matrix_bd_91_end_mask_0, x = matrix_bd_89_cast_fp16)[name = string("matrix_bd_91_cast_fp16")]; + tensor var_5223_cast_fp16 = add(x = matrix_ac_45_cast_fp16, y = matrix_bd_91_cast_fp16)[name = string("op_5223_cast_fp16")]; + fp16 _inversed_scores_89_y_0_to_fp16 = const()[name = string("_inversed_scores_89_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_89_cast_fp16 = mul(x = var_5223_cast_fp16, y = _inversed_scores_89_y_0_to_fp16)[name = string("_inversed_scores_89_cast_fp16")]; + tensor scores_91_cast_fp16 = select(a = var_44_to_fp16, b = _inversed_scores_89_cast_fp16, cond = mask_11)[name = string("scores_91_cast_fp16")]; + tensor var_5229_cast_fp16 = softmax(axis = var_58, x = scores_91_cast_fp16)[name = string("op_5229_cast_fp16")]; + tensor input_1185_cast_fp16 = select(a = var_43_to_fp16, b = var_5229_cast_fp16, cond = mask_11)[name = string("input_1185_cast_fp16")]; + bool x_585_transpose_x_0 = const()[name = string("x_585_transpose_x_0"), val = bool(false)]; + bool x_585_transpose_y_0 = const()[name = string("x_585_transpose_y_0"), val = bool(false)]; + tensor value_53_cast_fp16 = transpose(perm = value_53_perm_0, x = v_45_cast_fp16)[name = string("transpose_161")]; + tensor x_585_cast_fp16 = matmul(transpose_x = x_585_transpose_x_0, transpose_y = x_585_transpose_y_0, x = input_1185_cast_fp16, y = value_53_cast_fp16)[name = string("x_585_cast_fp16")]; + tensor var_5233_perm_0 = const()[name = string("op_5233_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_5234 = const()[name = string("op_5234"), val = tensor([1, -1, 1024])]; + tensor var_5233_cast_fp16 = transpose(perm = var_5233_perm_0, x = x_585_cast_fp16)[name = string("transpose_160")]; + tensor input_1187_cast_fp16 = reshape(shape = var_5234, x = var_5233_cast_fp16)[name = string("input_1187_cast_fp16")]; + tensor encoder_layers_22_self_attn_linear_out_weight_to_fp16 = const()[name = string("encoder_layers_22_self_attn_linear_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(490326784)))]; + tensor encoder_layers_22_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_22_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(492424000)))]; + tensor linear_205_cast_fp16 = linear(bias = encoder_layers_22_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_22_self_attn_linear_out_weight_to_fp16, x = input_1187_cast_fp16)[name = string("linear_205_cast_fp16")]; + tensor input_1191_cast_fp16 = add(x = input_1181_cast_fp16, y = linear_205_cast_fp16)[name = string("input_1191_cast_fp16")]; + tensor x_589_axes_0 = const()[name = string("x_589_axes_0"), val = tensor([-1])]; + tensor encoder_layers_22_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_22_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(492426112)))]; + tensor encoder_layers_22_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_22_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(492428224)))]; + tensor x_589_cast_fp16 = layer_norm(axes = x_589_axes_0, beta = encoder_layers_22_norm_conv_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_22_norm_conv_weight_to_fp16, x = input_1191_cast_fp16)[name = string("x_589_cast_fp16")]; + tensor input_1193_perm_0 = const()[name = string("input_1193_perm_0"), val = tensor([0, 2, 1])]; + string input_1195_pad_type_0 = const()[name = string("input_1195_pad_type_0"), val = string("valid")]; + tensor input_1195_strides_0 = const()[name = string("input_1195_strides_0"), val = tensor([1])]; + tensor input_1195_pad_0 = const()[name = string("input_1195_pad_0"), val = tensor([0, 0])]; + tensor input_1195_dilations_0 = const()[name = string("input_1195_dilations_0"), val = tensor([1])]; + int32 input_1195_groups_0 = const()[name = string("input_1195_groups_0"), val = int32(1)]; + tensor encoder_layers_22_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(492430336))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(494527552))))[name = string("encoder_layers_22_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_1193_cast_fp16 = transpose(perm = input_1193_perm_0, x = x_589_cast_fp16)[name = string("transpose_159")]; + tensor input_1195_cast_fp16 = conv(dilations = input_1195_dilations_0, groups = input_1195_groups_0, pad = input_1195_pad_0, pad_type = input_1195_pad_type_0, strides = input_1195_strides_0, weight = encoder_layers_22_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_1193_cast_fp16)[name = string("input_1195_cast_fp16")]; + int32 x_591_split_num_splits_0 = const()[name = string("x_591_split_num_splits_0"), val = int32(2)]; + int32 x_591_split_axis_0 = const()[name = string("x_591_split_axis_0"), val = int32(1)]; + tensor x_591_split_cast_fp16_0, tensor x_591_split_cast_fp16_1 = split(axis = x_591_split_axis_0, num_splits = x_591_split_num_splits_0, x = input_1195_cast_fp16)[name = string("x_591_split_cast_fp16")]; + tensor x_591_split_1_sigmoid_cast_fp16 = sigmoid(x = x_591_split_cast_fp16_1)[name = string("x_591_split_1_sigmoid_cast_fp16")]; + tensor x_591_cast_fp16 = mul(x = x_591_split_cast_fp16_0, y = x_591_split_1_sigmoid_cast_fp16)[name = string("x_591_cast_fp16")]; + tensor input_1197_cast_fp16 = select(a = var_43_to_fp16, b = x_591_cast_fp16, cond = var_574)[name = string("input_1197_cast_fp16")]; + bool new_x_91_interleave_0 = const()[name = string("new_x_91_interleave_0"), val = bool(false)]; + tensor new_x_91_cast_fp16 = concat(axis = var_58, interleave = new_x_91_interleave_0, values = (cache_91_cast_fp16, input_1197_cast_fp16))[name = string("new_x_91_cast_fp16")]; + tensor var_5273_begin_0 = const()[name = string("op_5273_begin_0"), val = tensor([0, 0, 56])]; + tensor var_5273_end_0 = const()[name = string("op_5273_end_0"), val = tensor([1, 1024, 64])]; + tensor var_5273_end_mask_0 = const()[name = string("op_5273_end_mask_0"), val = tensor([true, true, true])]; + tensor var_5273_cast_fp16 = slice_by_index(begin = var_5273_begin_0, end = var_5273_end_0, end_mask = var_5273_end_mask_0, x = new_x_91_cast_fp16)[name = string("op_5273_cast_fp16")]; + string x_593_pad_type_0 = const()[name = string("x_593_pad_type_0"), val = string("valid")]; + int32 x_593_groups_0 = const()[name = string("x_593_groups_0"), val = int32(1024)]; + tensor x_593_strides_0 = const()[name = string("x_593_strides_0"), val = tensor([1])]; + tensor x_593_pad_0 = const()[name = string("x_593_pad_0"), val = tensor([0, 0])]; + tensor x_593_dilations_0 = const()[name = string("x_593_dilations_0"), val = tensor([1])]; + tensor encoder_layers_22_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(494531712))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(494540992))))[name = string("encoder_layers_22_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_593_cast_fp16 = conv(dilations = x_593_dilations_0, groups = x_593_groups_0, pad = x_593_pad_0, pad_type = x_593_pad_type_0, strides = x_593_strides_0, weight = encoder_layers_22_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_91_cast_fp16)[name = string("x_593_cast_fp16")]; + tensor input_1199_perm_0 = const()[name = string("input_1199_perm_0"), val = tensor([0, 2, 1])]; + tensor x_595_axes_0 = const()[name = string("x_595_axes_0"), val = tensor([-1])]; + tensor encoder_layers_22_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_22_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(494543104)))]; + tensor encoder_layers_22_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_22_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(494545216)))]; + tensor input_1199_cast_fp16 = transpose(perm = input_1199_perm_0, x = x_593_cast_fp16)[name = string("transpose_158")]; + tensor x_595_cast_fp16 = layer_norm(axes = x_595_axes_0, beta = encoder_layers_22_conv_batch_norm_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_22_conv_batch_norm_weight_to_fp16, x = input_1199_cast_fp16)[name = string("x_595_cast_fp16")]; + tensor input_1201_perm_0 = const()[name = string("input_1201_perm_0"), val = tensor([0, 2, 1])]; + tensor input_1201_cast_fp16 = transpose(perm = input_1201_perm_0, x = x_595_cast_fp16)[name = string("transpose_157")]; + tensor input_1203_cast_fp16 = silu(x = input_1201_cast_fp16)[name = string("input_1203_cast_fp16")]; + string x_597_pad_type_0 = const()[name = string("x_597_pad_type_0"), val = string("valid")]; + tensor x_597_strides_0 = const()[name = string("x_597_strides_0"), val = tensor([1])]; + tensor x_597_pad_0 = const()[name = string("x_597_pad_0"), val = tensor([0, 0])]; + tensor x_597_dilations_0 = const()[name = string("x_597_dilations_0"), val = tensor([1])]; + int32 x_597_groups_0 = const()[name = string("x_597_groups_0"), val = int32(1)]; + tensor encoder_layers_22_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(494547328))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(495595968))))[name = string("encoder_layers_22_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_597_cast_fp16 = conv(dilations = x_597_dilations_0, groups = x_597_groups_0, pad = x_597_pad_0, pad_type = x_597_pad_type_0, strides = x_597_strides_0, weight = encoder_layers_22_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_1203_cast_fp16)[name = string("x_597_cast_fp16")]; + tensor input_1205_perm_0 = const()[name = string("input_1205_perm_0"), val = tensor([0, 2, 1])]; + tensor input_1205_cast_fp16 = transpose(perm = input_1205_perm_0, x = x_597_cast_fp16)[name = string("transpose_156")]; + tensor input_1207_cast_fp16 = add(x = input_1191_cast_fp16, y = input_1205_cast_fp16)[name = string("input_1207_cast_fp16")]; + tensor input_1209_axes_0 = const()[name = string("input_1209_axes_0"), val = tensor([-1])]; + tensor encoder_layers_22_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_22_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(495598080)))]; + tensor encoder_layers_22_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_22_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(495600192)))]; + tensor input_1209_cast_fp16 = layer_norm(axes = input_1209_axes_0, beta = encoder_layers_22_norm_feed_forward2_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_22_norm_feed_forward2_weight_to_fp16, x = input_1207_cast_fp16)[name = string("input_1209_cast_fp16")]; + tensor encoder_layers_22_feed_forward2_linear1_weight_to_fp16 = const()[name = string("encoder_layers_22_feed_forward2_linear1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(495602304)))]; + tensor encoder_layers_22_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_22_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(503990976)))]; + tensor linear_206_cast_fp16 = linear(bias = encoder_layers_22_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_22_feed_forward2_linear1_weight_to_fp16, x = input_1209_cast_fp16)[name = string("linear_206_cast_fp16")]; + tensor input_1213_cast_fp16 = silu(x = linear_206_cast_fp16)[name = string("input_1213_cast_fp16")]; + tensor encoder_layers_22_feed_forward2_linear2_weight_to_fp16 = const()[name = string("encoder_layers_22_feed_forward2_linear2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(503999232)))]; + tensor encoder_layers_22_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_22_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(512387904)))]; + tensor linear_207_cast_fp16 = linear(bias = encoder_layers_22_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_22_feed_forward2_linear2_weight_to_fp16, x = input_1213_cast_fp16)[name = string("linear_207_cast_fp16")]; + fp16 var_5316_to_fp16 = const()[name = string("op_5316_to_fp16"), val = fp16(0x1p-1)]; + tensor var_5317_cast_fp16 = mul(x = linear_207_cast_fp16, y = var_5316_to_fp16)[name = string("op_5317_cast_fp16")]; + tensor input_1219_cast_fp16 = add(x = input_1207_cast_fp16, y = var_5317_cast_fp16)[name = string("input_1219_cast_fp16")]; + tensor input_1221_axes_0 = const()[name = string("input_1221_axes_0"), val = tensor([-1])]; + tensor encoder_layers_22_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_22_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(512390016)))]; + tensor encoder_layers_22_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_22_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(512392128)))]; + tensor input_1221_cast_fp16 = layer_norm(axes = input_1221_axes_0, beta = encoder_layers_22_norm_out_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_22_norm_out_weight_to_fp16, x = input_1219_cast_fp16)[name = string("input_1221_cast_fp16")]; + tensor cache_93_begin_0 = const()[name = string("cache_93_begin_0"), val = tensor([23, 0, 0, 0])]; + tensor cache_93_end_0 = const()[name = string("cache_93_end_0"), val = tensor([24, 1, 42, 1024])]; + tensor cache_93_end_mask_0 = const()[name = string("cache_93_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_93_squeeze_mask_0 = const()[name = string("cache_93_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_93_cast_fp16 = slice_by_index(begin = cache_93_begin_0, end = cache_93_end_0, end_mask = cache_93_end_mask_0, squeeze_mask = cache_93_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_93_cast_fp16")]; + tensor cache_begin_0 = const()[name = string("cache_begin_0"), val = tensor([23, 0, 0, 0])]; + tensor cache_end_0 = const()[name = string("cache_end_0"), val = tensor([24, 1, 1024, 8])]; + tensor cache_end_mask_0 = const()[name = string("cache_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_squeeze_mask_0 = const()[name = string("cache_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_cast_fp16 = slice_by_index(begin = cache_begin_0, end = cache_end_0, end_mask = cache_end_mask_0, squeeze_mask = cache_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_cast_fp16")]; + tensor input_1223_axes_0 = const()[name = string("input_1223_axes_0"), val = tensor([-1])]; + tensor encoder_layers_23_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_23_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(512394240)))]; + tensor encoder_layers_23_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_23_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(512396352)))]; + tensor input_1223_cast_fp16 = layer_norm(axes = input_1223_axes_0, beta = encoder_layers_23_norm_feed_forward1_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_23_norm_feed_forward1_weight_to_fp16, x = input_1221_cast_fp16)[name = string("input_1223_cast_fp16")]; + tensor encoder_layers_23_feed_forward1_linear1_weight_to_fp16 = const()[name = string("encoder_layers_23_feed_forward1_linear1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(512398464)))]; + tensor encoder_layers_23_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_23_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(520787136)))]; + tensor linear_208_cast_fp16 = linear(bias = encoder_layers_23_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_23_feed_forward1_linear1_weight_to_fp16, x = input_1223_cast_fp16)[name = string("linear_208_cast_fp16")]; + tensor input_1227_cast_fp16 = silu(x = linear_208_cast_fp16)[name = string("input_1227_cast_fp16")]; + tensor encoder_layers_23_feed_forward1_linear2_weight_to_fp16 = const()[name = string("encoder_layers_23_feed_forward1_linear2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(520795392)))]; + tensor encoder_layers_23_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_23_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(529184064)))]; + tensor linear_209_cast_fp16 = linear(bias = encoder_layers_23_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_23_feed_forward1_linear2_weight_to_fp16, x = input_1227_cast_fp16)[name = string("linear_209_cast_fp16")]; + fp16 var_5353_to_fp16 = const()[name = string("op_5353_to_fp16"), val = fp16(0x1p-1)]; + tensor var_5354_cast_fp16 = mul(x = linear_209_cast_fp16, y = var_5353_to_fp16)[name = string("op_5354_cast_fp16")]; + tensor input_1233_cast_fp16 = add(x = input_1221_cast_fp16, y = var_5354_cast_fp16)[name = string("input_1233_cast_fp16")]; + tensor key_axes_0 = const()[name = string("key_axes_0"), val = tensor([-1])]; + tensor encoder_layers_23_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_23_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(529186176)))]; + tensor encoder_layers_23_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_23_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(529188288)))]; + tensor key_cast_fp16 = layer_norm(axes = key_axes_0, beta = encoder_layers_23_norm_self_att_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_23_norm_self_att_weight_to_fp16, x = input_1233_cast_fp16)[name = string("key_cast_fp16")]; + bool input_1235_interleave_0 = const()[name = string("input_1235_interleave_0"), val = bool(false)]; + tensor input_1235_cast_fp16 = concat(axis = var_67, interleave = input_1235_interleave_0, values = (cache_93_cast_fp16, key_cast_fp16))[name = string("input_1235_cast_fp16")]; + bool cache_last_channel_cur_interleave_0 = const()[name = string("cache_last_channel_cur_interleave_0"), val = bool(false)]; + tensor cache_last_channel_cur_cast_fp16 = concat(axis = var_67, interleave = cache_last_channel_cur_interleave_0, values = key_cast_fp16)[name = string("cache_last_channel_cur_cast_fp16")]; + tensor encoder_layers_23_self_attn_linear_q_weight_to_fp16 = const()[name = string("encoder_layers_23_self_attn_linear_q_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(529190400)))]; + tensor encoder_layers_23_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_23_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(531287616)))]; + tensor linear_210_cast_fp16 = linear(bias = encoder_layers_23_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_23_self_attn_linear_q_weight_to_fp16, x = key_cast_fp16)[name = string("linear_210_cast_fp16")]; + tensor var_5387 = const()[name = string("op_5387"), val = tensor([1, -1, 8, 128])]; + tensor q_139_cast_fp16 = reshape(shape = var_5387, x = linear_210_cast_fp16)[name = string("q_139_cast_fp16")]; + tensor encoder_layers_23_self_attn_linear_k_weight_to_fp16 = const()[name = string("encoder_layers_23_self_attn_linear_k_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(531289728)))]; + tensor encoder_layers_23_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_23_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(533386944)))]; + tensor linear_211_cast_fp16 = linear(bias = encoder_layers_23_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_23_self_attn_linear_k_weight_to_fp16, x = input_1235_cast_fp16)[name = string("linear_211_cast_fp16")]; + tensor var_5392 = const()[name = string("op_5392"), val = tensor([1, -1, 8, 128])]; + tensor k_93_cast_fp16 = reshape(shape = var_5392, x = linear_211_cast_fp16)[name = string("k_93_cast_fp16")]; + tensor encoder_layers_23_self_attn_linear_v_weight_to_fp16 = const()[name = string("encoder_layers_23_self_attn_linear_v_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(533389056)))]; + tensor encoder_layers_23_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_23_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(535486272)))]; + tensor linear_212_cast_fp16 = linear(bias = encoder_layers_23_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_23_self_attn_linear_v_weight_to_fp16, x = input_1235_cast_fp16)[name = string("linear_212_cast_fp16")]; + tensor var_5397 = const()[name = string("op_5397"), val = tensor([1, -1, 8, 128])]; + tensor v_cast_fp16 = reshape(shape = var_5397, x = linear_212_cast_fp16)[name = string("v_cast_fp16")]; + tensor value_perm_0 = const()[name = string("value_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_23_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_23_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(535488384)))]; + tensor var_5410_cast_fp16 = add(x = q_139_cast_fp16, y = encoder_layers_23_self_attn_pos_bias_u_to_fp16)[name = string("op_5410_cast_fp16")]; + tensor encoder_layers_23_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_23_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(535490496)))]; + tensor var_5412_cast_fp16 = add(x = q_139_cast_fp16, y = encoder_layers_23_self_attn_pos_bias_v_to_fp16)[name = string("op_5412_cast_fp16")]; + tensor q_with_bias_v_perm_0 = const()[name = string("q_with_bias_v_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_605_transpose_x_0 = const()[name = string("x_605_transpose_x_0"), val = bool(false)]; + bool x_605_transpose_y_0 = const()[name = string("x_605_transpose_y_0"), val = bool(false)]; + tensor op_5414_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(535492608))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(535692352))))[name = string("op_5414_to_fp16_quantized")]; + tensor q_with_bias_v_cast_fp16 = transpose(perm = q_with_bias_v_perm_0, x = var_5412_cast_fp16)[name = string("transpose_155")]; + tensor x_605_cast_fp16 = matmul(transpose_x = x_605_transpose_x_0, transpose_y = x_605_transpose_y_0, x = q_with_bias_v_cast_fp16, y = op_5414_to_fp16_quantized)[name = string("x_605_cast_fp16")]; + tensor x_607_pad_0 = const()[name = string("x_607_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_607_mode_0 = const()[name = string("x_607_mode_0"), val = string("constant")]; + fp16 const_378_to_fp16 = const()[name = string("const_378_to_fp16"), val = fp16(0x0p+0)]; + tensor x_607_cast_fp16 = pad(constant_val = const_378_to_fp16, mode = x_607_mode_0, pad = x_607_pad_0, x = x_605_cast_fp16)[name = string("x_607_cast_fp16")]; + tensor var_5422 = const()[name = string("op_5422"), val = tensor([1, 8, -1, 56])]; + tensor x_609_cast_fp16 = reshape(shape = var_5422, x = x_607_cast_fp16)[name = string("x_609_cast_fp16")]; + tensor var_5426_begin_0 = const()[name = string("op_5426_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_5426_end_0 = const()[name = string("op_5426_end_0"), val = tensor([1, 8, 196, 56])]; + tensor var_5426_end_mask_0 = const()[name = string("op_5426_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_5426_cast_fp16 = slice_by_index(begin = var_5426_begin_0, end = var_5426_end_0, end_mask = var_5426_end_mask_0, x = x_609_cast_fp16)[name = string("op_5426_cast_fp16")]; + tensor var_5427 = const()[name = string("op_5427"), val = tensor([1, 8, 56, 195])]; + tensor matrix_bd_93_cast_fp16 = reshape(shape = var_5427, x = var_5426_cast_fp16)[name = string("matrix_bd_93_cast_fp16")]; + bool matrix_ac_transpose_x_0 = const()[name = string("matrix_ac_transpose_x_0"), val = bool(false)]; + bool matrix_ac_transpose_y_0 = const()[name = string("matrix_ac_transpose_y_0"), val = bool(false)]; + tensor transpose_142_perm_0 = const()[name = string("transpose_142_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_143_perm_0 = const()[name = string("transpose_143_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_143 = transpose(perm = transpose_143_perm_0, x = k_93_cast_fp16)[name = string("transpose_153")]; + tensor transpose_142 = transpose(perm = transpose_142_perm_0, x = var_5410_cast_fp16)[name = string("transpose_154")]; + tensor matrix_ac_cast_fp16 = matmul(transpose_x = matrix_ac_transpose_x_0, transpose_y = matrix_ac_transpose_y_0, x = transpose_142, y = transpose_143)[name = string("matrix_ac_cast_fp16")]; + tensor matrix_bd_begin_0 = const()[name = string("matrix_bd_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_end_0 = const()[name = string("matrix_bd_end_0"), val = tensor([1, 8, 56, 98])]; + tensor matrix_bd_end_mask_0 = const()[name = string("matrix_bd_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_cast_fp16 = slice_by_index(begin = matrix_bd_begin_0, end = matrix_bd_end_0, end_mask = matrix_bd_end_mask_0, x = matrix_bd_93_cast_fp16)[name = string("matrix_bd_cast_fp16")]; + tensor var_5436_cast_fp16 = add(x = matrix_ac_cast_fp16, y = matrix_bd_cast_fp16)[name = string("op_5436_cast_fp16")]; + fp16 _inversed_scores_93_y_0_to_fp16 = const()[name = string("_inversed_scores_93_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_93_cast_fp16 = mul(x = var_5436_cast_fp16, y = _inversed_scores_93_y_0_to_fp16)[name = string("_inversed_scores_93_cast_fp16")]; + tensor scores_cast_fp16 = select(a = var_44_to_fp16, b = _inversed_scores_93_cast_fp16, cond = mask_11)[name = string("scores_cast_fp16")]; + tensor var_5442_cast_fp16 = softmax(axis = var_58, x = scores_cast_fp16)[name = string("op_5442_cast_fp16")]; + tensor input_1237_cast_fp16 = select(a = var_43_to_fp16, b = var_5442_cast_fp16, cond = mask_11)[name = string("input_1237_cast_fp16")]; + bool x_611_transpose_x_0 = const()[name = string("x_611_transpose_x_0"), val = bool(false)]; + bool x_611_transpose_y_0 = const()[name = string("x_611_transpose_y_0"), val = bool(false)]; + tensor value_cast_fp16 = transpose(perm = value_perm_0, x = v_cast_fp16)[name = string("transpose_152")]; + tensor x_611_cast_fp16 = matmul(transpose_x = x_611_transpose_x_0, transpose_y = x_611_transpose_y_0, x = input_1237_cast_fp16, y = value_cast_fp16)[name = string("x_611_cast_fp16")]; + tensor var_5446_perm_0 = const()[name = string("op_5446_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_5447 = const()[name = string("op_5447"), val = tensor([1, -1, 1024])]; + tensor var_5446_cast_fp16 = transpose(perm = var_5446_perm_0, x = x_611_cast_fp16)[name = string("transpose_151")]; + tensor input_1239_cast_fp16 = reshape(shape = var_5447, x = var_5446_cast_fp16)[name = string("input_1239_cast_fp16")]; + tensor encoder_layers_23_self_attn_linear_out_weight_to_fp16 = const()[name = string("encoder_layers_23_self_attn_linear_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(535692864)))]; + tensor encoder_layers_23_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_23_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(537790080)))]; + tensor linear_214_cast_fp16 = linear(bias = encoder_layers_23_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_23_self_attn_linear_out_weight_to_fp16, x = input_1239_cast_fp16)[name = string("linear_214_cast_fp16")]; + tensor input_1243_cast_fp16 = add(x = input_1233_cast_fp16, y = linear_214_cast_fp16)[name = string("input_1243_cast_fp16")]; + tensor x_615_axes_0 = const()[name = string("x_615_axes_0"), val = tensor([-1])]; + tensor encoder_layers_23_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_23_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(537792192)))]; + tensor encoder_layers_23_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_23_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(537794304)))]; + tensor x_615_cast_fp16 = layer_norm(axes = x_615_axes_0, beta = encoder_layers_23_norm_conv_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_23_norm_conv_weight_to_fp16, x = input_1243_cast_fp16)[name = string("x_615_cast_fp16")]; + tensor input_1245_perm_0 = const()[name = string("input_1245_perm_0"), val = tensor([0, 2, 1])]; + string input_1247_pad_type_0 = const()[name = string("input_1247_pad_type_0"), val = string("valid")]; + tensor input_1247_strides_0 = const()[name = string("input_1247_strides_0"), val = tensor([1])]; + tensor input_1247_pad_0 = const()[name = string("input_1247_pad_0"), val = tensor([0, 0])]; + tensor input_1247_dilations_0 = const()[name = string("input_1247_dilations_0"), val = tensor([1])]; + int32 input_1247_groups_0 = const()[name = string("input_1247_groups_0"), val = int32(1)]; + tensor encoder_layers_23_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(537796416))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(539893632))))[name = string("encoder_layers_23_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_1245_cast_fp16 = transpose(perm = input_1245_perm_0, x = x_615_cast_fp16)[name = string("transpose_150")]; + tensor input_1247_cast_fp16 = conv(dilations = input_1247_dilations_0, groups = input_1247_groups_0, pad = input_1247_pad_0, pad_type = input_1247_pad_type_0, strides = input_1247_strides_0, weight = encoder_layers_23_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_1245_cast_fp16)[name = string("input_1247_cast_fp16")]; + int32 x_617_split_num_splits_0 = const()[name = string("x_617_split_num_splits_0"), val = int32(2)]; + int32 x_617_split_axis_0 = const()[name = string("x_617_split_axis_0"), val = int32(1)]; + tensor x_617_split_cast_fp16_0, tensor x_617_split_cast_fp16_1 = split(axis = x_617_split_axis_0, num_splits = x_617_split_num_splits_0, x = input_1247_cast_fp16)[name = string("x_617_split_cast_fp16")]; + tensor x_617_split_1_sigmoid_cast_fp16 = sigmoid(x = x_617_split_cast_fp16_1)[name = string("x_617_split_1_sigmoid_cast_fp16")]; + tensor x_617_cast_fp16 = mul(x = x_617_split_cast_fp16_0, y = x_617_split_1_sigmoid_cast_fp16)[name = string("x_617_cast_fp16")]; + tensor input_1249_cast_fp16 = select(a = var_43_to_fp16, b = x_617_cast_fp16, cond = var_574)[name = string("input_1249_cast_fp16")]; + bool new_x_interleave_0 = const()[name = string("new_x_interleave_0"), val = bool(false)]; + tensor new_x_cast_fp16 = concat(axis = var_58, interleave = new_x_interleave_0, values = (cache_cast_fp16, input_1249_cast_fp16))[name = string("new_x_cast_fp16")]; + tensor cache_last_time_cur_begin_0 = const()[name = string("cache_last_time_cur_begin_0"), val = tensor([0, 0, 56])]; + tensor cache_last_time_cur_end_0 = const()[name = string("cache_last_time_cur_end_0"), val = tensor([1, 1024, 64])]; + tensor cache_last_time_cur_end_mask_0 = const()[name = string("cache_last_time_cur_end_mask_0"), val = tensor([true, true, true])]; + tensor cache_last_time_cur_cast_fp16 = slice_by_index(begin = cache_last_time_cur_begin_0, end = cache_last_time_cur_end_0, end_mask = cache_last_time_cur_end_mask_0, x = new_x_cast_fp16)[name = string("cache_last_time_cur_cast_fp16")]; + string x_619_pad_type_0 = const()[name = string("x_619_pad_type_0"), val = string("valid")]; + int32 x_619_groups_0 = const()[name = string("x_619_groups_0"), val = int32(1024)]; + tensor x_619_strides_0 = const()[name = string("x_619_strides_0"), val = tensor([1])]; + tensor x_619_pad_0 = const()[name = string("x_619_pad_0"), val = tensor([0, 0])]; + tensor x_619_dilations_0 = const()[name = string("x_619_dilations_0"), val = tensor([1])]; + tensor encoder_layers_23_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(539897792))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(539907072))))[name = string("encoder_layers_23_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_619_cast_fp16 = conv(dilations = x_619_dilations_0, groups = x_619_groups_0, pad = x_619_pad_0, pad_type = x_619_pad_type_0, strides = x_619_strides_0, weight = encoder_layers_23_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_cast_fp16)[name = string("x_619_cast_fp16")]; + tensor input_1251_perm_0 = const()[name = string("input_1251_perm_0"), val = tensor([0, 2, 1])]; + tensor x_621_axes_0 = const()[name = string("x_621_axes_0"), val = tensor([-1])]; + tensor encoder_layers_23_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_23_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(539909184)))]; + tensor encoder_layers_23_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_23_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(539911296)))]; + tensor input_1251_cast_fp16 = transpose(perm = input_1251_perm_0, x = x_619_cast_fp16)[name = string("transpose_149")]; + tensor x_621_cast_fp16 = layer_norm(axes = x_621_axes_0, beta = encoder_layers_23_conv_batch_norm_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_23_conv_batch_norm_weight_to_fp16, x = input_1251_cast_fp16)[name = string("x_621_cast_fp16")]; + tensor input_1253_perm_0 = const()[name = string("input_1253_perm_0"), val = tensor([0, 2, 1])]; + tensor input_1253_cast_fp16 = transpose(perm = input_1253_perm_0, x = x_621_cast_fp16)[name = string("transpose_148")]; + tensor input_1255_cast_fp16 = silu(x = input_1253_cast_fp16)[name = string("input_1255_cast_fp16")]; + string x_623_pad_type_0 = const()[name = string("x_623_pad_type_0"), val = string("valid")]; + tensor x_623_strides_0 = const()[name = string("x_623_strides_0"), val = tensor([1])]; + tensor x_623_pad_0 = const()[name = string("x_623_pad_0"), val = tensor([0, 0])]; + tensor x_623_dilations_0 = const()[name = string("x_623_dilations_0"), val = tensor([1])]; + int32 x_623_groups_0 = const()[name = string("x_623_groups_0"), val = int32(1)]; + tensor encoder_layers_23_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(539913408))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(540962048))))[name = string("encoder_layers_23_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_623_cast_fp16 = conv(dilations = x_623_dilations_0, groups = x_623_groups_0, pad = x_623_pad_0, pad_type = x_623_pad_type_0, strides = x_623_strides_0, weight = encoder_layers_23_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_1255_cast_fp16)[name = string("x_623_cast_fp16")]; + tensor input_1257_perm_0 = const()[name = string("input_1257_perm_0"), val = tensor([0, 2, 1])]; + tensor input_1257_cast_fp16 = transpose(perm = input_1257_perm_0, x = x_623_cast_fp16)[name = string("transpose_147")]; + tensor input_1259_cast_fp16 = add(x = input_1243_cast_fp16, y = input_1257_cast_fp16)[name = string("input_1259_cast_fp16")]; + tensor input_1261_axes_0 = const()[name = string("input_1261_axes_0"), val = tensor([-1])]; + tensor encoder_layers_23_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_23_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(540964160)))]; + tensor encoder_layers_23_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_23_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(540966272)))]; + tensor input_1261_cast_fp16 = layer_norm(axes = input_1261_axes_0, beta = encoder_layers_23_norm_feed_forward2_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_23_norm_feed_forward2_weight_to_fp16, x = input_1259_cast_fp16)[name = string("input_1261_cast_fp16")]; + tensor encoder_layers_23_feed_forward2_linear1_weight_to_fp16 = const()[name = string("encoder_layers_23_feed_forward2_linear1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(540968384)))]; + tensor encoder_layers_23_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_23_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(549357056)))]; + tensor linear_215_cast_fp16 = linear(bias = encoder_layers_23_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_23_feed_forward2_linear1_weight_to_fp16, x = input_1261_cast_fp16)[name = string("linear_215_cast_fp16")]; + tensor input_1265_cast_fp16 = silu(x = linear_215_cast_fp16)[name = string("input_1265_cast_fp16")]; + tensor encoder_layers_23_feed_forward2_linear2_weight_to_fp16 = const()[name = string("encoder_layers_23_feed_forward2_linear2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(549365312)))]; + tensor encoder_layers_23_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_23_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(557753984)))]; + tensor linear_216_cast_fp16 = linear(bias = encoder_layers_23_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_23_feed_forward2_linear2_weight_to_fp16, x = input_1265_cast_fp16)[name = string("linear_216_cast_fp16")]; + fp16 var_5529_to_fp16 = const()[name = string("op_5529_to_fp16"), val = fp16(0x1p-1)]; + tensor var_5530_cast_fp16 = mul(x = linear_216_cast_fp16, y = var_5529_to_fp16)[name = string("op_5530_cast_fp16")]; + tensor input_1271_cast_fp16 = add(x = input_1259_cast_fp16, y = var_5530_cast_fp16)[name = string("input_1271_cast_fp16")]; + tensor audio_signal_axes_0 = const()[name = string("audio_signal_axes_0"), val = tensor([-1])]; + tensor encoder_layers_23_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_23_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(557756096)))]; + tensor encoder_layers_23_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_23_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(557758208)))]; + tensor audio_signal_cast_fp16 = layer_norm(axes = audio_signal_axes_0, beta = encoder_layers_23_norm_out_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_23_norm_out_weight_to_fp16, x = input_1271_cast_fp16)[name = string("audio_signal_cast_fp16")]; + int32 obj_5_axis_0 = const()[name = string("obj_5_axis_0"), val = int32(0)]; + tensor obj_5_cast_fp16 = stack(axis = obj_5_axis_0, values = (var_483_cast_fp16, var_696_cast_fp16, var_909_cast_fp16, var_1122_cast_fp16, var_1335_cast_fp16, var_1548_cast_fp16, var_1761_cast_fp16, var_1974_cast_fp16, var_2187_cast_fp16, var_2400_cast_fp16, var_2613_cast_fp16, var_2826_cast_fp16, var_3039_cast_fp16, var_3252_cast_fp16, var_3465_cast_fp16, var_3678_cast_fp16, var_3891_cast_fp16, var_4104_cast_fp16, var_4317_cast_fp16, var_4530_cast_fp16, var_4743_cast_fp16, var_4956_cast_fp16, var_5169_cast_fp16, cache_last_channel_cur_cast_fp16))[name = string("obj_5_cast_fp16")]; + int32 obj_7_axis_0 = const()[name = string("obj_7_axis_0"), val = int32(0)]; + tensor obj_7_cast_fp16 = stack(axis = obj_7_axis_0, values = (var_587_cast_fp16, var_800_cast_fp16, var_1013_cast_fp16, var_1226_cast_fp16, var_1439_cast_fp16, var_1652_cast_fp16, var_1865_cast_fp16, var_2078_cast_fp16, var_2291_cast_fp16, var_2504_cast_fp16, var_2717_cast_fp16, var_2930_cast_fp16, var_3143_cast_fp16, var_3356_cast_fp16, var_3569_cast_fp16, var_3782_cast_fp16, var_3995_cast_fp16, var_4208_cast_fp16, var_4421_cast_fp16, var_4634_cast_fp16, var_4847_cast_fp16, var_5060_cast_fp16, var_5273_cast_fp16, cache_last_time_cur_cast_fp16))[name = string("obj_7_cast_fp16")]; + tensor var_5546 = add(x = cache_len, y = max_audio_length_1)[name = string("op_5546")]; + string var_5546_promoted_to_fp16_dtype_0 = const()[name = string("op_5546_promoted_to_fp16_dtype_0"), val = string("fp16")]; + fp16 const_384_to_fp16 = const()[name = string("const_384_to_fp16"), val = fp16(-inf)]; + fp16 var_48_promoted_to_fp16 = const()[name = string("op_48_promoted_to_fp16"), val = fp16(0x1.5p+5)]; + tensor var_5546_to_fp16 = cast(dtype = var_5546_promoted_to_fp16_dtype_0, x = var_5546)[name = string("cast_6")]; + tensor clip_1_cast_fp16 = clip(alpha = const_384_to_fp16, beta = var_48_promoted_to_fp16, x = var_5546_to_fp16)[name = string("clip_1_cast_fp16")]; + tensor var_5568_begin_0 = const()[name = string("op_5568_begin_0"), val = tensor([0, 0, 14, 0])]; + tensor var_5568_end_0 = const()[name = string("op_5568_end_0"), val = tensor([24, 1, 56, 1024])]; + tensor var_5568_end_mask_0 = const()[name = string("op_5568_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_5568_cast_fp16 = slice_by_index(begin = var_5568_begin_0, end = var_5568_end_0, end_mask = var_5568_end_mask_0, x = obj_5_cast_fp16)[name = string("op_5568_cast_fp16")]; + int32 var_5588_one_hot_vector_size_0 = const()[name = string("op_5588_one_hot_vector_size_0"), val = int32(128)]; + int32 var_5588_axis_0 = const()[name = string("op_5588_axis_0"), val = int32(-1)]; + int32 var_5588_on_value_0 = const()[name = string("op_5588_on_value_0"), val = int32(1)]; + int32 var_5588_off_value_0 = const()[name = string("op_5588_off_value_0"), val = int32(0)]; + tensor var_5588 = one_hot(axis = var_5588_axis_0, indices = prompt_id, off_value = var_5588_off_value_0, on_value = var_5588_on_value_0, one_hot_vector_size = var_5588_one_hot_vector_size_0)[name = string("op_5588")]; + tensor var_5591_axes_0 = const()[name = string("op_5591_axes_0"), val = tensor([1])]; + string cast_245_to_fp16_dtype_0 = const()[name = string("cast_245_to_fp16_dtype_0"), val = string("fp16")]; + tensor var_5588_to_fp16 = cast(dtype = cast_245_to_fp16_dtype_0, x = var_5588)[name = string("cast_5")]; + tensor var_5591_cast_fp16 = expand_dims(axes = var_5591_axes_0, x = var_5588_to_fp16)[name = string("op_5591_cast_fp16")]; + tensor one_hot_reps_0 = const()[name = string("one_hot_reps_0"), val = tensor([1, 56, 1])]; + tensor one_hot_cast_fp16 = tile(reps = one_hot_reps_0, x = var_5591_cast_fp16)[name = string("one_hot_cast_fp16")]; + int32 var_5600 = const()[name = string("op_5600"), val = int32(-1)]; + bool input_1273_interleave_0 = const()[name = string("input_1273_interleave_0"), val = bool(false)]; + tensor input_1273_cast_fp16 = concat(axis = var_5600, interleave = input_1273_interleave_0, values = (audio_signal_cast_fp16, one_hot_cast_fp16))[name = string("input_1273_cast_fp16")]; + tensor prompt_kernel_0_weight_to_fp16 = const()[name = string("prompt_kernel_0_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(557760320)))]; + tensor prompt_kernel_0_bias_to_fp16 = const()[name = string("prompt_kernel_0_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(562478976)))]; + tensor linear_217_cast_fp16 = linear(bias = prompt_kernel_0_bias_to_fp16, weight = prompt_kernel_0_weight_to_fp16, x = input_1273_cast_fp16)[name = string("linear_217_cast_fp16")]; + tensor input_cast_fp16 = relu(x = linear_217_cast_fp16)[name = string("input_cast_fp16")]; + tensor prompt_kernel_2_weight_to_fp16 = const()[name = string("prompt_kernel_2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(562483136)))]; + tensor prompt_kernel_2_bias_to_fp16 = const()[name = string("prompt_kernel_2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(566677504)))]; + tensor linear_218_cast_fp16 = linear(bias = prompt_kernel_2_bias_to_fp16, weight = prompt_kernel_2_weight_to_fp16, x = input_cast_fp16)[name = string("linear_218_cast_fp16")]; + tensor var_5613_perm_0 = const()[name = string("op_5613_perm_0"), val = tensor([0, 2, 1])]; + string var_5613_cast_fp16_to_fp32_dtype_0 = const()[name = string("op_5613_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + string cast_246_dtype_0 = const()[name = string("cast_246_dtype_0"), val = string("int32")]; + tensor var_5621_perm_0 = const()[name = string("op_5621_perm_0"), val = tensor([1, 0, 2, 3])]; + string var_5621_cast_fp16_to_fp32_dtype_0 = const()[name = string("op_5621_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + tensor var_5624_perm_0 = const()[name = string("op_5624_perm_0"), val = tensor([1, 0, 2, 3])]; + string var_5624_cast_fp16_to_fp32_dtype_0 = const()[name = string("op_5624_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + string cast_247_dtype_0 = const()[name = string("cast_247_dtype_0"), val = string("int32")]; + tensor cache_len_out = cast(dtype = cast_247_dtype_0, x = clip_1_cast_fp16)[name = string("cast_0")]; + tensor var_5624_cast_fp16 = transpose(perm = var_5624_perm_0, x = obj_7_cast_fp16)[name = string("transpose_144")]; + tensor cache_time_out = cast(dtype = var_5624_cast_fp16_to_fp32_dtype_0, x = var_5624_cast_fp16)[name = string("cast_1")]; + tensor var_5621_cast_fp16 = transpose(perm = var_5621_perm_0, x = var_5568_cast_fp16)[name = string("transpose_145")]; + tensor cache_channel_out = cast(dtype = var_5621_cast_fp16_to_fp32_dtype_0, x = var_5621_cast_fp16)[name = string("cast_2")]; + tensor encoded_length = cast(dtype = cast_246_dtype_0, x = clip_0_cast_fp16)[name = string("cast_3")]; + tensor var_5613_cast_fp16 = transpose(perm = var_5613_perm_0, x = linear_218_cast_fp16)[name = string("transpose_146")]; + tensor encoded = cast(dtype = var_5613_cast_fp16_to_fp32_dtype_0, x = var_5613_cast_fp16)[name = string("cast_4")]; + } -> (encoded, encoded_length, cache_channel_out, cache_time_out, cache_len_out); +} \ No newline at end of file diff --git a/ja/4480ms/encoder.mlmodelc/weights/weight.bin b/ja/4480ms/encoder.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..e770743e8f69de769852fc42baf0304b71ade176 --- /dev/null +++ b/ja/4480ms/encoder.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6a88ef3a70bab53d01290a8c9d3f29c171843a29356f93213384700ab3525ab0 +size 566679616 diff --git a/ja/4480ms/encoder.mlpackage/Data/com.apple.CoreML/model.mlmodel b/ja/4480ms/encoder.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..c61e9d5dbed4e53cf445b5183683f7437ec0c084 --- /dev/null +++ b/ja/4480ms/encoder.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dd9cb26cdd3e8b27c2966692ada2b9b8cc7bb79a48770213713a1fb1a08e5aed +size 793485 diff --git a/ja/4480ms/encoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/ja/4480ms/encoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..e770743e8f69de769852fc42baf0304b71ade176 --- /dev/null +++ b/ja/4480ms/encoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6a88ef3a70bab53d01290a8c9d3f29c171843a29356f93213384700ab3525ab0 +size 566679616 diff --git a/ja/4480ms/encoder.mlpackage/Manifest.json b/ja/4480ms/encoder.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..648bba6d2c3cd23bf1cfd59d14d3c64b9f1237dc --- /dev/null +++ b/ja/4480ms/encoder.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "0A4614E6-2B65-4F67-A3BD-DA028BB6FB28": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + }, + "18FBBE10-AC02-470A-A266-74B2AD0F575E": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + } + }, + "rootModelIdentifier": "0A4614E6-2B65-4F67-A3BD-DA028BB6FB28" +} diff --git a/ja/4480ms/joint.mlmodelc/analytics/coremldata.bin b/ja/4480ms/joint.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..a7ba5a90671da5c40e03362f44f23df528bc6d93 --- /dev/null +++ b/ja/4480ms/joint.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e342ce20383866520d2c6c860c2bf14d887b9e7fef53606661b41a23ad09472e +size 243 diff --git a/ja/4480ms/joint.mlmodelc/coremldata.bin b/ja/4480ms/joint.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..219c9bac9ed82b5626e34705f215643647b64d90 --- /dev/null +++ b/ja/4480ms/joint.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1df3d770386f24c28a1a187cd341e4f2527d0b1c7f2959e5f606383e2ba9ddc6 +size 401 diff --git a/ja/4480ms/joint.mlmodelc/model.mil b/ja/4480ms/joint.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..827f6cd71b5910ea07d4f6ba43462967d8b86410 --- /dev/null +++ b/ja/4480ms/joint.mlmodelc/model.mil @@ -0,0 +1,31 @@ +program(1.3) +[buildInfo = dict({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.10.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor decoder, tensor encoder) { + tensor input_1_perm_0 = const()[name = string("input_1_perm_0"), val = tensor([0, 2, 1])]; + string encoder_to_fp16_dtype_0 = const()[name = string("encoder_to_fp16_dtype_0"), val = string("fp16")]; + tensor input_3_perm_0 = const()[name = string("input_3_perm_0"), val = tensor([0, 2, 1])]; + string decoder_to_fp16_dtype_0 = const()[name = string("decoder_to_fp16_dtype_0"), val = string("fp16")]; + tensor module_enc_weight_to_fp16 = const()[name = string("module_enc_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))]; + tensor module_enc_bias_to_fp16 = const()[name = string("module_enc_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1310848)))]; + tensor encoder_to_fp16 = cast(dtype = encoder_to_fp16_dtype_0, x = encoder)[name = string("cast_2")]; + tensor input_1_cast_fp16 = transpose(perm = input_1_perm_0, x = encoder_to_fp16)[name = string("transpose_1")]; + tensor linear_0_cast_fp16 = linear(bias = module_enc_bias_to_fp16, weight = module_enc_weight_to_fp16, x = input_1_cast_fp16)[name = string("linear_0_cast_fp16")]; + tensor module_pred_weight_to_fp16 = const()[name = string("module_pred_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1312192)))]; + tensor module_pred_bias_to_fp16 = const()[name = string("module_pred_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2131456)))]; + tensor decoder_to_fp16 = cast(dtype = decoder_to_fp16_dtype_0, x = decoder)[name = string("cast_1")]; + tensor input_3_cast_fp16 = transpose(perm = input_3_perm_0, x = decoder_to_fp16)[name = string("transpose_0")]; + tensor linear_1_cast_fp16 = linear(bias = module_pred_bias_to_fp16, weight = module_pred_weight_to_fp16, x = input_3_cast_fp16)[name = string("linear_1_cast_fp16")]; + tensor var_23_axes_0 = const()[name = string("op_23_axes_0"), val = tensor([2])]; + tensor var_23_cast_fp16 = expand_dims(axes = var_23_axes_0, x = linear_0_cast_fp16)[name = string("op_23_cast_fp16")]; + tensor var_25_axes_0 = const()[name = string("op_25_axes_0"), val = tensor([1])]; + tensor var_25_cast_fp16 = expand_dims(axes = var_25_axes_0, x = linear_1_cast_fp16)[name = string("op_25_cast_fp16")]; + tensor input_5_cast_fp16 = add(x = var_23_cast_fp16, y = var_25_cast_fp16)[name = string("input_5_cast_fp16")]; + tensor input_7_cast_fp16 = relu(x = input_5_cast_fp16)[name = string("input_7_cast_fp16")]; + tensor module_joint_net_2_weight_to_fp16 = const()[name = string("module_joint_net_2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2132800)))]; + tensor module_joint_net_2_bias_to_fp16 = const()[name = string("module_joint_net_2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3929984)))]; + tensor linear_2_cast_fp16 = linear(bias = module_joint_net_2_bias_to_fp16, weight = module_joint_net_2_weight_to_fp16, x = input_7_cast_fp16)[name = string("linear_2_cast_fp16")]; + string linear_2_cast_fp16_to_fp32_dtype_0 = const()[name = string("linear_2_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + tensor logits = cast(dtype = linear_2_cast_fp16_to_fp32_dtype_0, x = linear_2_cast_fp16)[name = string("cast_0")]; + } -> (logits); +} \ No newline at end of file diff --git a/ja/4480ms/joint.mlmodelc/weights/weight.bin b/ja/4480ms/joint.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..31da4412e1214e0b52e77c023e0490150a12e242 --- /dev/null +++ b/ja/4480ms/joint.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:44c15c8be3c89e5d531258b840e560e467084d55ed400453a9d21325757fb111 +size 3932856 diff --git a/ja/4480ms/joint.mlpackage/Data/com.apple.CoreML/model.mlmodel b/ja/4480ms/joint.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..1bc98711fb995d49c835ce43242e33bc518a943d --- /dev/null +++ b/ja/4480ms/joint.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1df49e8dd61d86e19d95b935020748db2eda7ae1273f39988001a30535c5be45 +size 4545 diff --git a/ja/4480ms/joint.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/ja/4480ms/joint.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..31da4412e1214e0b52e77c023e0490150a12e242 --- /dev/null +++ b/ja/4480ms/joint.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:44c15c8be3c89e5d531258b840e560e467084d55ed400453a9d21325757fb111 +size 3932856 diff --git a/ja/4480ms/joint.mlpackage/Manifest.json b/ja/4480ms/joint.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..ef5aca0b538b9de8e541f27a29fff913203df8c7 --- /dev/null +++ b/ja/4480ms/joint.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "8348189F-7CB9-4961-A35F-4049C53D63B6": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + }, + "AA6A8B4F-747E-4EC1-87E1-2B387F1149D8": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + } + }, + "rootModelIdentifier": "8348189F-7CB9-4961-A35F-4049C53D63B6" +} diff --git a/ja/4480ms/metadata.json b/ja/4480ms/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..7a24cbcb01c58da153481428d2166bedfe1a250f --- /dev/null +++ b/ja/4480ms/metadata.json @@ -0,0 +1,199 @@ +{ + "model": "nvidia/nemotron-asr-streaming-multilingual-0.6b", + "model_class": "nemo.collections.asr.models.rnnt_bpe_models_prompt.EncDecRNNTBPEModelWithPrompt", + "sample_rate": 16000, + "mel_features": 128, + "chunk_mel_frames": 448, + "pre_encode_cache": 9, + "total_mel_frames": 457, + "att_context_size": [ + 42, + 13 + ], + "vocab_size": 1403, + "blank_idx": 1403, + "vocab_pruned": true, + "vocab_pruned_original_size": 13087, + "cache_channel_shape": [ + 1, + 24, + 42, + 1024 + ], + "cache_time_shape": [ + 1, + 24, + 1024, + 8 + ], + "decoder_hidden": 640, + "decoder_layers": 2, + "encoder_dim": 1024, + "num_prompts": 128, + "prompt_dictionary": { + "af-ZA": 54, + "am-ET": 49, + "ar": 7, + "ar-AR": 7, + "auto": 101, + "ay-BO": 81, + "az-AZ": 66, + "bg": 30, + "bg-BG": 30, + "bn-IN": 36, + "cs": 22, + "cs-CZ": 22, + "da": 25, + "da-DK": 25, + "de": 9, + "de-DE": 9, + "el": 21, + "el-GR": 21, + "en": 0, + "en-GB": 1, + "en-US": 0, + "enGB": 1, + "es": 3, + "es-ES": 2, + "es-US": 3, + "esES": 2, + "et": 60, + "et-EE": 60, + "fa-IR": 38, + "fi": 26, + "fi-FI": 26, + "fr": 8, + "fr-CA": 100, + "fr-FR": 8, + "gn-PY": 82, + "gu-IN": 42, + "ha-NG": 50, + "haw-US": 97, + "he-IL": 64, + "hi": 6, + "hi-HI": 6, + "hi-IN": 6, + "hr": 29, + "hr-HR": 29, + "hu": 23, + "hu-HU": 23, + "hy-AM": 68, + "id-ID": 34, + "ig-NG": 53, + "it": 15, + "it-IT": 15, + "ja-JA": 10, + "ja-JP": 10, + "ka-GE": 67, + "km-KH": 47, + "kn-IN": 43, + "ko": 14, + "ko-KO": 14, + "ko-KR": 14, + "ku-TR": 65, + "ky-KG": 71, + "ln-CD": 58, + "lt": 31, + "lt-LT": 31, + "lv": 61, + "lv-LV": 61, + "mi-NZ": 96, + "ml-IN": 44, + "mr-IN": 41, + "ms-MY": 35, + "mt-MT": 102, + "nah-MX": 83, + "nb": 103, + "nb-NO": 103, + "ne-NP": 46, + "nl": 16, + "nl-NL": 16, + "nn": 104, + "nn-NO": 104, + "no": 27, + "no-NO": 27, + "ny-MW": 57, + "or-KE": 59, + "pl": 17, + "pl-PL": 17, + "pt": 13, + "pt-BR": 12, + "pt-PT": 13, + "qu-PE": 80, + "ro": 20, + "ro-RO": 20, + "ru": 11, + "ru-RU": 11, + "rw-RW": 55, + "si-LK": 45, + "sk": 28, + "sk-SK": 28, + "sl": 62, + "sl-SI": 62, + "sm-WS": 98, + "so-SO": 56, + "sv": 24, + "sv-SE": 24, + "sw-KE": 48, + "ta-IN": 39, + "te-IN": 40, + "tg-TJ": 70, + "th-TH": 32, + "to-TO": 99, + "tr": 18, + "tr-TR": 18, + "uk": 19, + "uk-UA": 19, + "ur-PK": 37, + "uz-UZ": 69, + "vi-VN": 33, + "yo-NG": 52, + "zh-CN": 4, + "zh-TW": 5, + "zh-ZH": 4, + "zu-ZA": 51 + }, + "default_prompt_id": 101, + "lang_tag_token_ids": [ + 1, + 52, + 62, + 66, + 69, + 70, + 75, + 76, + 77, + 79, + 81, + 83, + 86, + 88, + 89, + 90, + 92, + 94, + 95, + 96, + 97, + 99, + 100, + 103, + 107, + 109, + 111, + 112, + 114, + 115, + 117, + 1389, + 1390, + 1391, + 1392, + 1393, + 1394, + 1395, + 1402 + ], + "chunk_ms": 4480 +} \ No newline at end of file diff --git a/ja/4480ms/preprocessor.mlmodelc/analytics/coremldata.bin b/ja/4480ms/preprocessor.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..e87594f68006bef8db4fcfe2a2379a3c1197ba56 --- /dev/null +++ b/ja/4480ms/preprocessor.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:422a77fc3b64b27260a8ae2031d287abaa1630d1ebc70343e9dcd280dd4c7e5c +size 243 diff --git a/ja/4480ms/preprocessor.mlmodelc/coremldata.bin b/ja/4480ms/preprocessor.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..9f88d2976124fbdeae6f6c4c492443eb5f32c97e --- /dev/null +++ b/ja/4480ms/preprocessor.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:41ec4b4c1059ff1f2ac7c71d90b1da0caf9244499d06f6c9de175a1ef992bec1 +size 371 diff --git a/ja/4480ms/preprocessor.mlmodelc/model.mil b/ja/4480ms/preprocessor.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..0b8261362f9cbf465b530a0d2d0ee9a2b2f462cd --- /dev/null +++ b/ja/4480ms/preprocessor.mlmodelc/model.mil @@ -0,0 +1,122 @@ +program(1.3) +[buildInfo = dict({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.5.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.3.0"}})] +{ + func main(tensor audio, tensor audio_length) [FlexibleShapeInformation = tuple>>, tuple, ?>>>>((("DefaultShapes", {{"audio", [1, 1]}}), ("RangeDims", {{"audio", [[1, 1], [1, 480000]]}})))] { + int32 var_9 = const()[name = string("op_9"), val = int32(1)]; + int32 var_10 = const()[name = string("op_10"), val = int32(160)]; + int32 var_12 = const()[name = string("op_12"), val = int32(0)]; + int32 var_33 = const()[name = string("op_33"), val = int32(512)]; + tensor var_34 = add(x = audio_length, y = var_33)[name = string("op_34")]; + int32 var_35 = const()[name = string("op_35"), val = int32(512)]; + tensor var_36 = sub(x = var_34, y = var_35)[name = string("op_36")]; + tensor floor_div_0 = floor_div(x = var_36, y = var_10)[name = string("floor_div_0")]; + tensor var_39 = equal(x = audio_length, y = var_12)[name = string("op_39")]; + tensor var_40 = const()[name = string("op_40"), val = tensor([0])]; + tensor mel_length = select(a = var_40, b = floor_div_0, cond = var_39)[name = string("seq_len")]; + string audio_to_fp16_dtype_0 = const()[name = string("audio_to_fp16_dtype_0"), val = string("fp16")]; + tensor audio_to_fp16 = cast(dtype = audio_to_fp16_dtype_0, x = audio)[name = string("cast_14")]; + tensor var_42_shape_cast_fp16 = shape(x = audio_to_fp16)[name = string("op_42_shape_cast_fp16")]; + int32 gather_0_axis_0 = const()[name = string("gather_0_axis_0"), val = int32(0)]; + int32 gather_0_batch_dims_0 = const()[name = string("gather_0_batch_dims_0"), val = int32(0)]; + bool gather_0_validate_indices_0 = const()[name = string("gather_0_validate_indices_0"), val = bool(false)]; + string var_42_shape_cast_fp16_to_int16_dtype_0 = const()[name = string("op_42_shape_cast_fp16_to_int16_dtype_0"), val = string("int16")]; + uint16 select_0_to_uint16 = const()[name = string("select_0_to_uint16"), val = uint16(1)]; + tensor var_42_shape_cast_fp16_to_int16 = cast(dtype = var_42_shape_cast_fp16_to_int16_dtype_0, x = var_42_shape_cast_fp16)[name = string("cast_13")]; + int16 gather_0_cast_uint16 = gather(axis = gather_0_axis_0, batch_dims = gather_0_batch_dims_0, indices = select_0_to_uint16, validate_indices = gather_0_validate_indices_0, x = var_42_shape_cast_fp16_to_int16)[name = string("gather_0_cast_uint16")]; + string gather_0_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_0_cast_uint16_to_int32_dtype_0"), val = string("int32")]; + int32 const_0 = const()[name = string("const_0"), val = int32(0)]; + int32 const_1 = const()[name = string("const_1"), val = int32(1)]; + int32 gather_0_cast_uint16_to_int32 = cast(dtype = gather_0_cast_uint16_to_int32_dtype_0, x = gather_0_cast_uint16)[name = string("cast_12")]; + tensor var_43 = range_1d(end = gather_0_cast_uint16_to_int32, start = const_0, step = const_1)[name = string("op_43")]; + tensor var_44_axes_0 = const()[name = string("op_44_axes_0"), val = tensor([0])]; + tensor var_44 = expand_dims(axes = var_44_axes_0, x = var_43)[name = string("op_44")]; + tensor var_45_axes_0 = const()[name = string("op_45_axes_0"), val = tensor([1])]; + tensor var_45 = expand_dims(axes = var_45_axes_0, x = audio_length)[name = string("op_45")]; + tensor timemask = less(x = var_44, y = var_45)[name = string("timemask")]; + tensor var_48_begin_0 = const()[name = string("op_48_begin_0"), val = tensor([0, 0])]; + tensor var_48_end_0 = const()[name = string("op_48_end_0"), val = tensor([1, 1])]; + tensor var_48_end_mask_0 = const()[name = string("op_48_end_mask_0"), val = tensor([true, false])]; + tensor var_48_squeeze_mask_0 = const()[name = string("op_48_squeeze_mask_0"), val = tensor([false, true])]; + tensor var_48_cast_fp16 = slice_by_index(begin = var_48_begin_0, end = var_48_end_0, end_mask = var_48_end_mask_0, squeeze_mask = var_48_squeeze_mask_0, x = audio_to_fp16)[name = string("op_48_cast_fp16")]; + tensor var_49_axes_0 = const()[name = string("op_49_axes_0"), val = tensor([1])]; + tensor var_49_cast_fp16 = expand_dims(axes = var_49_axes_0, x = var_48_cast_fp16)[name = string("op_49_cast_fp16")]; + tensor var_51_begin_0 = const()[name = string("op_51_begin_0"), val = tensor([0, 1])]; + tensor var_51_end_0 = const()[name = string("op_51_end_0"), val = tensor([1, 0])]; + tensor var_51_end_mask_0 = const()[name = string("op_51_end_mask_0"), val = tensor([true, true])]; + tensor var_51_cast_fp16 = slice_by_index(begin = var_51_begin_0, end = var_51_end_0, end_mask = var_51_end_mask_0, x = audio_to_fp16)[name = string("op_51_cast_fp16")]; + tensor var_53_begin_0 = const()[name = string("op_53_begin_0"), val = tensor([0, 0])]; + tensor var_53_end_0 = const()[name = string("op_53_end_0"), val = tensor([1, -1])]; + tensor var_53_end_mask_0 = const()[name = string("op_53_end_mask_0"), val = tensor([true, false])]; + tensor var_53_cast_fp16 = slice_by_index(begin = var_53_begin_0, end = var_53_end_0, end_mask = var_53_end_mask_0, x = audio_to_fp16)[name = string("op_53_cast_fp16")]; + fp16 var_54_to_fp16 = const()[name = string("op_54_to_fp16"), val = fp16(0x1.f0cp-1)]; + tensor var_55_cast_fp16 = mul(x = var_53_cast_fp16, y = var_54_to_fp16)[name = string("op_55_cast_fp16")]; + tensor var_56_cast_fp16 = sub(x = var_51_cast_fp16, y = var_55_cast_fp16)[name = string("op_56_cast_fp16")]; + bool x_3_interleave_0 = const()[name = string("x_3_interleave_0"), val = bool(false)]; + tensor x_3_cast_fp16 = concat(axis = var_9, interleave = x_3_interleave_0, values = (var_49_cast_fp16, var_56_cast_fp16))[name = string("x_3_cast_fp16")]; + tensor var_59 = logical_not(x = timemask)[name = string("op_59")]; + fp16 var_16_to_fp16 = const()[name = string("op_16_to_fp16"), val = fp16(0x0p+0)]; + tensor input_1_cast_fp16 = select(a = var_16_to_fp16, b = x_3_cast_fp16, cond = var_59)[name = string("input_1_cast_fp16")]; + tensor concat_1x = const()[name = string("concat_1x"), val = tensor([1, 1, -1])]; + tensor input_3_cast_fp16 = reshape(shape = concat_1x, x = input_1_cast_fp16)[name = string("input_3_cast_fp16")]; + tensor input_5_pad_0 = const()[name = string("input_5_pad_0"), val = tensor([0, 0, 0, 0, 256, 256])]; + string input_5_mode_0 = const()[name = string("input_5_mode_0"), val = string("constant")]; + fp16 const_3_to_fp16 = const()[name = string("const_3_to_fp16"), val = fp16(0x0p+0)]; + tensor input_5_cast_fp16 = pad(constant_val = const_3_to_fp16, mode = input_5_mode_0, pad = input_5_pad_0, x = input_3_cast_fp16)[name = string("input_5_cast_fp16")]; + tensor concat_2x = const()[name = string("concat_2x"), val = tensor([1, -1])]; + tensor input_cast_fp16 = reshape(shape = concat_2x, x = input_5_cast_fp16)[name = string("input_cast_fp16")]; + tensor expand_dims_3 = const()[name = string("expand_dims_3"), val = tensor([160])]; + tensor expand_dims_4_axes_0 = const()[name = string("expand_dims_4_axes_0"), val = tensor([1])]; + tensor expand_dims_4_cast_fp16 = expand_dims(axes = expand_dims_4_axes_0, x = input_cast_fp16)[name = string("expand_dims_4_cast_fp16")]; + string conv_0_pad_type_0 = const()[name = string("conv_0_pad_type_0"), val = string("valid")]; + tensor conv_0_pad_0 = const()[name = string("conv_0_pad_0"), val = tensor([0, 0])]; + tensor conv_0_dilations_0 = const()[name = string("conv_0_dilations_0"), val = tensor([1])]; + int32 conv_0_groups_0 = const()[name = string("conv_0_groups_0"), val = int32(1)]; + tensor expand_dims_1_to_fp16 = const()[name = string("expand_dims_1_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))]; + tensor conv_0_cast_fp16 = conv(dilations = conv_0_dilations_0, groups = conv_0_groups_0, pad = conv_0_pad_0, pad_type = conv_0_pad_type_0, strides = expand_dims_3, weight = expand_dims_1_to_fp16, x = expand_dims_4_cast_fp16)[name = string("conv_0_cast_fp16")]; + string conv_1_pad_type_0 = const()[name = string("conv_1_pad_type_0"), val = string("valid")]; + tensor conv_1_pad_0 = const()[name = string("conv_1_pad_0"), val = tensor([0, 0])]; + tensor conv_1_dilations_0 = const()[name = string("conv_1_dilations_0"), val = tensor([1])]; + int32 conv_1_groups_0 = const()[name = string("conv_1_groups_0"), val = int32(1)]; + tensor expand_dims_2_to_fp16 = const()[name = string("expand_dims_2_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(263296)))]; + tensor conv_1_cast_fp16 = conv(dilations = conv_1_dilations_0, groups = conv_1_groups_0, pad = conv_1_pad_0, pad_type = conv_1_pad_type_0, strides = expand_dims_3, weight = expand_dims_2_to_fp16, x = expand_dims_4_cast_fp16)[name = string("conv_1_cast_fp16")]; + int32 stack_0_axis_0 = const()[name = string("stack_0_axis_0"), val = int32(-1)]; + tensor stack_0_cast_fp16 = stack(axis = stack_0_axis_0, values = (conv_0_cast_fp16, conv_1_cast_fp16))[name = string("stack_0_cast_fp16")]; + fp16 var_19_promoted_to_fp16 = const()[name = string("op_19_promoted_to_fp16"), val = fp16(0x1p+1)]; + tensor var_74_cast_fp16 = pow(x = stack_0_cast_fp16, y = var_19_promoted_to_fp16)[name = string("op_74_cast_fp16")]; + tensor var_76_axes_0 = const()[name = string("op_76_axes_0"), val = tensor([-1])]; + bool var_76_keep_dims_0 = const()[name = string("op_76_keep_dims_0"), val = bool(false)]; + tensor var_76_cast_fp16 = reduce_sum(axes = var_76_axes_0, keep_dims = var_76_keep_dims_0, x = var_74_cast_fp16)[name = string("op_76_cast_fp16")]; + tensor x_11_cast_fp16 = identity(x = var_76_cast_fp16)[name = string("x_11_cast_fp16")]; + bool x_13_transpose_x_0 = const()[name = string("x_13_transpose_x_0"), val = bool(false)]; + bool x_13_transpose_y_0 = const()[name = string("x_13_transpose_y_0"), val = bool(false)]; + tensor const_4_to_fp16 = const()[name = string("const_4_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(526528)))]; + tensor x_13_cast_fp16 = matmul(transpose_x = x_13_transpose_x_0, transpose_y = x_13_transpose_y_0, x = const_4_to_fp16, y = x_11_cast_fp16)[name = string("x_13_cast_fp16")]; + fp16 var_83_to_fp16 = const()[name = string("op_83_to_fp16"), val = fp16(0x1p-24)]; + tensor var_84_cast_fp16 = add(x = x_13_cast_fp16, y = var_83_to_fp16)[name = string("op_84_cast_fp16")]; + fp32 x_epsilon_0 = const()[name = string("x_epsilon_0"), val = fp32(0x1p-149)]; + tensor x_cast_fp16 = log(epsilon = x_epsilon_0, x = var_84_cast_fp16)[name = string("x_cast_fp16")]; + tensor var_86_shape_cast_fp16 = shape(x = x_cast_fp16)[name = string("op_86_shape_cast_fp16")]; + int32 gather_5_axis_0 = const()[name = string("gather_5_axis_0"), val = int32(0)]; + int32 gather_5_batch_dims_0 = const()[name = string("gather_5_batch_dims_0"), val = int32(0)]; + bool gather_5_validate_indices_0 = const()[name = string("gather_5_validate_indices_0"), val = bool(false)]; + string var_86_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_86_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")]; + uint16 select_5_to_uint16 = const()[name = string("select_5_to_uint16"), val = uint16(2)]; + tensor var_86_shape_cast_fp16_to_uint16 = cast(dtype = var_86_shape_cast_fp16_to_uint16_dtype_0, x = var_86_shape_cast_fp16)[name = string("cast_11")]; + uint16 gather_5_cast_uint16 = gather(axis = gather_5_axis_0, batch_dims = gather_5_batch_dims_0, indices = select_5_to_uint16, validate_indices = gather_5_validate_indices_0, x = var_86_shape_cast_fp16_to_uint16)[name = string("gather_5_cast_uint16")]; + string gather_5_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_5_cast_uint16_to_int32_dtype_0"), val = string("int32")]; + int32 const_5 = const()[name = string("const_5"), val = int32(0)]; + int32 const_6 = const()[name = string("const_6"), val = int32(1)]; + int32 gather_5_cast_uint16_to_int32 = cast(dtype = gather_5_cast_uint16_to_int32_dtype_0, x = gather_5_cast_uint16)[name = string("cast_10")]; + tensor mask_1 = range_1d(end = gather_5_cast_uint16_to_int32, start = const_5, step = const_6)[name = string("mask_1")]; + tensor expand_dims_0_axes_0 = const()[name = string("expand_dims_0_axes_0"), val = tensor([0])]; + tensor expand_dims_0 = expand_dims(axes = expand_dims_0_axes_0, x = mask_1)[name = string("expand_dims_0")]; + tensor var_91_axes_0 = const()[name = string("op_91_axes_0"), val = tensor([1])]; + tensor var_91 = expand_dims(axes = var_91_axes_0, x = mel_length)[name = string("op_91")]; + tensor mask = greater_equal(x = expand_dims_0, y = var_91)[name = string("mask")]; + tensor var_93_axes_0 = const()[name = string("op_93_axes_0"), val = tensor([1])]; + tensor var_93 = expand_dims(axes = var_93_axes_0, x = mask)[name = string("op_93")]; + tensor processed_signal_cast_fp16 = select(a = var_16_to_fp16, b = x_cast_fp16, cond = var_93)[name = string("processed_signal_cast_fp16")]; + string processed_signal_cast_fp16_to_fp32_dtype_0 = const()[name = string("processed_signal_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + tensor mel = cast(dtype = processed_signal_cast_fp16_to_fp32_dtype_0, x = processed_signal_cast_fp16)[name = string("cast_9")]; + } -> (mel, mel_length); +} \ No newline at end of file diff --git a/ja/4480ms/preprocessor.mlmodelc/weights/weight.bin b/ja/4480ms/preprocessor.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..86dd375f6649d262d58c9cf8b89006ceab216411 --- /dev/null +++ b/ja/4480ms/preprocessor.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:297514e2b211d14b0e53cb97193d679bb89ead98d28e578f3f1d049ddbcc36b3 +size 592384 diff --git a/ja/4480ms/preprocessor.mlpackage/Data/com.apple.CoreML/model.mlmodel b/ja/4480ms/preprocessor.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..961c08f6554bdf851377e1e59750c9ee4a63de19 --- /dev/null +++ b/ja/4480ms/preprocessor.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:912e39c12c0ddeb4d3716693621d0140bb7de556056ab8dc9e5cd9336ebf1baa +size 15878 diff --git a/ja/4480ms/preprocessor.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/ja/4480ms/preprocessor.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..86dd375f6649d262d58c9cf8b89006ceab216411 --- /dev/null +++ b/ja/4480ms/preprocessor.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:297514e2b211d14b0e53cb97193d679bb89ead98d28e578f3f1d049ddbcc36b3 +size 592384 diff --git a/ja/4480ms/preprocessor.mlpackage/Manifest.json b/ja/4480ms/preprocessor.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..3bfc4c45d1b99a3d127a61aff3d1d0ab4d95963c --- /dev/null +++ b/ja/4480ms/preprocessor.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "470B2EDA-A2F1-4777-8F1D-55E0D6A275A8": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + }, + "D83A1966-0388-45B9-9CA7-1DACB258C26C": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + } + }, + "rootModelIdentifier": "470B2EDA-A2F1-4777-8F1D-55E0D6A275A8" +} diff --git a/ja/4480ms/tokenizer.json b/ja/4480ms/tokenizer.json new file mode 100644 index 0000000000000000000000000000000000000000..b05c35042625729d82c27359c8b8cf9ae4e2e4c9 --- /dev/null +++ b/ja/4480ms/tokenizer.json @@ -0,0 +1,1406 @@ +{ + "0": "", + "1": "", + "2": "▁", + "3": ".", + "4": ",", + "5": "e", + "6": "t", + "7": "a", + "8": "s", + "9": "o", + "10": "i", + "11": "r", + "12": "h", + "13": "m", + "14": "p", + "15": "n", + "16": "g", + "17": "en", + "18": "on", + "19": "y", + "20": "0", + "21": "b", + "22": "2", + "23": "1", + "24": "k", + "25": "3", + "26": "5", + "27": "I", + "28": "4", + "29": "z", + "30": "A", + "31": "6", + "32": "E", + "33": "7", + "34": "T", + "35": "8", + "36": "9", + "37": "S", + "38": "C", + "39": "M", + "40": "P", + "41": "D", + "42": "B", + "43": "U", + "44": "N", + "45": "O", + "46": "G", + "47": "F", + "48": "L", + "49": "R", + "50": "H", + "51": "K", + "52": "", + "53": "le", + "54": "li", + "55": "to", + "56": "na", + "57": "ci", + "58": "▁A", + "59": "vi", + "60": "Z", + "61": "X", + "62": "", + "63": "et", + "64": "ar", + "65": "el", + "66": "", + "67": "ie", + "68": "▁G", + "69": "", + "70": "", + "71": "ta", + "72": "ks", + "73": "gi", + "74": "di", + "75": "", + "76": "", + "77": "", + "78": "cs", + "79": "", + "80": "sti", + "81": "", + "82": "pe", + "83": "", + "84": "as", + "85": "▁Tai", + "86": "", + "87": "▁20", + "88": "", + "89": "", + "90": "", + "91": "-", + "92": "", + "93": "▁M", + "94": "", + "95": "", + "96": "", + "97": "", + "98": "pp", + "99": "", + "100": "", + "101": "▁N", + "102": "mm", + "103": "", + "104": "▁and", + "105": "th", + "106": "ex", + "107": "", + "108": "ng", + "109": "", + "110": "▁La", + "111": "", + "112": "", + "113": "avec", + "114": "", + "115": "", + "116": "!", + "117": "", + "118": "▁あ", + "119": "▁す", + "120": "▁し", + "121": "▁そ", + "122": "▁ま", + "123": "▁こ", + "124": "以", + "125": "仕", + "126": "▁文", + "127": "▁強", + "128": "▁月", + "129": "▁連", + "130": "▁ペ", + "131": "▁海", + "132": "▁科", + "133": "▁ゴ", + "134": "▁原", + "135": "▁ギ", + "136": "▁ヤ", + "137": "▁特", + "138": "▁参", + "139": "▁男", + "140": "普", + "141": "受", + "142": "▁都", + "143": "▁水", + "144": "▁終", + "145": "▁南", + "146": "▁例", + "147": "▁利", + "148": "▁現", + "149": "▁真", + "150": "▁古", + "151": "▁彼", + "152": "▁交", + "153": "専", + "154": "▁悪", + "155": "▁旅", + "156": "▁技", + "157": "▁再", + "158": "企", + "159": "登", + "160": "総", + "161": "改", + "162": "探", + "163": "異", + "164": "完", + "165": "頑", + "166": "深", + "167": "推", + "168": "昨", + "169": "危", + "170": "詳", + "171": "余", + "172": "描", + "173": "被", + "174": "固", + "175": "副", + "176": "混", + "177": "救", + "178": "互", + "179": "閉", + "180": "亜", + "181": "植", + "182": "責", + "183": "仮", + "184": "脱", + "185": "恐", + "186": "契", + "187": "隠", + "188": "麻", + "189": "哲", + "190": "宣", + "191": "浮", + "192": "諸", + "193": "巨", + "194": "午", + "195": "占", + "196": "距", + "197": "汚", + "198": "延", + "199": "訪", + "200": "沿", + "201": "唯", + "202": "批", + "203": "傾", + "204": "邦", + "205": "購", + "206": "維", + "207": "燃", + "208": "既", + "209": "緊", + "210": "獲", + "211": "炭", + "212": "搭", + "213": "努", + "214": "顧", + "215": "狭", + "216": "雰", + "217": "塗", + "218": "排", + "219": "奪", + "220": "埋", + "221": "履", + "222": "侵", + "223": "該", + "224": "頻", + "225": "勧", + "226": "魅", + "227": "抑", + "228": "驚", + "229": "衝", + "230": "耐", + "231": "披", + "232": "透", + "233": "促", + "234": "撤", + "235": "潜", + "236": "吐", + "237": "睡", + "238": "孤", + "239": "昆", + "240": "遭", + "241": "溶", + "242": "貿", + "243": "疾", + "244": "拒", + "245": "脅", + "246": "挟", + "247": "漏", + "248": "覆", + "249": "畜", + "250": "紛", + "251": "絞", + "252": "擬", + "253": "疎", + "254": "跳", + "255": "廷", + "256": "郊", + "257": "克", + "258": "措", + "259": "虐", + "260": "陥", + "261": "徐", + "262": "漠", + "263": "愚", + "264": "噴", + "265": "卿", + "266": "胡", + "267": "溢", + "268": "妨", + "269": "沸", + "270": "阻", + "271": "妃", + "272": "舶", + "273": "濡", + "274": "剥", + "275": "洪", + "276": "鋭", + "277": "娯", + "278": "腎", + "279": "訂", + "280": "肖", + "281": "枢", + "282": "髄", + "283": "傑", + "284": "砕", + "285": "幣", + "286": "勾", + "287": "臆", + "288": "歪", + "289": "喝", + "290": "彗", + "291": "礁", + "292": "蚊", + "293": "焉", + "294": "塵", + "295": "脊", + "296": "鱗", + "297": "瞑", + "298": "峻", + "299": "徊", + "300": "徘", + "301": "猥", + "302": "竣", + "303": "穣", + "304": "隻", + "305": "睨", + "306": "瀕", + "307": "褻", + "308": "踵", + "309": "奢", + "310": "董", + "311": "毅", + "312": "醒", + "313": "墜", + "314": "窟", + "315": "綿", + "316": "遮", + "317": "騰", + "318": "顎", + "319": "遍", + "320": "蓋", + "321": "澄", + "322": "棲", + "323": "虎", + "324": "粒", + "325": "胎", + "326": "爪", + "327": "殻", + "328": "鉛", + "329": "枕", + "330": "堤", + "331": "膝", + "332": "峠", + "333": "搬", + "334": "晋", + "335": "餌", + "336": "狼", + "337": "衰", + "338": "閲", + "339": "筒", + "340": "軸", + "341": "緯", + "342": "暦", + "343": "刃", + "344": "潔", + "345": "裂", + "346": "叫", + "347": "殖", + "348": "媒", + "349": "癌", + "350": "鎖", + "351": "磁", + "352": "洞", + "353": "遇", + "354": "枝", + "355": "紫", + "356": "縦", + "357": "胞", + "358": "釈", + "359": "威", + "360": "晶", + "361": "砲", + "362": "焦", + "363": "尿", + "364": "魂", + "365": "潮", + "366": "旬", + "367": "慎", + "368": "噂", + "369": "隔", + "370": "穴", + "371": "慮", + "372": "即", + "373": "滑", + "374": "雷", + "375": "摘", + "376": "鏡", + "377": "棒", + "378": "悟", + "379": "葬", + "380": "序", + "381": "貫", + "382": "氷", + "383": "針", + "384": "煮", + "385": "棄", + "386": "銃", + "387": "汁", + "388": "封", + "389": "湿", + "390": "靴", + "391": "豚", + "392": "締", + "393": "豪", + "394": "票", + "395": "皮", + "396": "縮", + "397": "徹", + "398": "較", + "399": "忍", + "400": "核", + "401": "儀", + "402": "到", + "403": "削", + "404": "駆", + "405": "繁", + "406": "陰", + "407": "浄", + "408": "脈", + "409": "滞", + "410": "至", + "411": "枚", + "412": "偉", + "413": "致", + "414": "貨", + "415": "漢", + "416": "己", + "417": "握", + "418": "欧", + "419": "薄", + "420": "献", + "421": "預", + "422": "龍", + "423": "快", + "424": "句", + "425": "縁", + "426": "微", + "427": "妙", + "428": "晩", + "429": "粉", + "430": "卓", + "431": "圏", + "432": "兼", + "433": "脳", + "434": "竜", + "435": "鳴", + "436": "騒", + "437": "請", + "438": "卵", + "439": "唱", + "440": "嵐", + "441": "臓", + "442": "箱", + "443": "祖", + "444": "浴", + "445": "壁", + "446": "析", + "447": "厚", + "448": "筆", + "449": "承", + "450": "均", + "451": "律", + "452": "否", + "453": "脚", + "454": "湖", + "455": "乳", + "456": "揮", + "457": "滅", + "458": "乾", + "459": "羽", + "460": "候", + "461": "拡", + "462": "貸", + "463": "砂", + "464": "敬", + "465": "庁", + "466": "煙", + "467": "底", + "468": "露", + "469": "骨", + "470": "倍", + "471": "殿", + "472": "易", + "473": "層", + "474": "幕", + "475": "毛", + "476": "爆", + "477": "暇", + "478": "械", + "479": "隣", + "480": "輸", + "481": "柄", + "482": "範", + "483": "掲", + "484": "嘘", + "485": "剤", + "486": "墓", + "487": "衣", + "488": "射", + "489": "菓", + "490": "募", + "491": "乱", + "492": "迎", + "493": "抱", + "494": "酸", + "495": "雄", + "496": "虫", + "497": "複", + "498": "為", + "499": "泳", + "500": "宝", + "501": "激", + "502": "暑", + "503": "疑", + "504": "誘", + "505": "暴", + "506": "聖", + "507": "捨", + "508": "破", + "509": "革", + "510": "希", + "511": "折", + "512": "惑", + "513": "測", + "514": "紀", + "515": "舎", + "516": "署", + "517": "患", + "518": "岸", + "519": "秀", + "520": "免", + "521": "禁", + "522": "躍", + "523": "聴", + "524": "抗", + "525": "税", + "526": "奏", + "527": "弾", + "528": "礼", + "529": "童", + "530": "裏", + "531": "吹", + "532": "眠", + "533": "歯", + "534": "拠", + "535": "慣", + "536": "触", + "537": "飼", + "538": "群", + "539": "宗", + "540": "傷", + "541": "額", + "542": "塩", + "543": "静", + "544": "留", + "545": "罪", + "546": "純", + "547": "壊", + "548": "闘", + "549": "弱", + "550": "刻", + "551": "航", + "552": "栄", + "553": "姿", + "554": "亡", + "555": "織", + "556": "敗", + "557": "章", + "558": "吸", + "559": "令", + "560": "捜", + "561": "模", + "562": "絵", + "563": "申", + "564": "盤", + "565": "積", + "566": "標", + "567": "階", + "568": "省", + "569": "項", + "570": "猫", + "571": "従", + "572": "非", + "573": "季", + "574": "捕", + "575": "党", + "576": "圧", + "577": "香", + "578": "操", + "579": "暗", + "580": "症", + "581": "散", + "582": "突", + "583": "適", + "584": "雑", + "585": "跡", + "586": "厳", + "587": "鳥", + "588": "逃", + "589": "講", + "590": "晴", + "591": "徴", + "592": "困", + "593": "短", + "594": "婦", + "595": "略", + "596": "齢", + "597": "震", + "598": "敵", + "599": "博", + "600": "血", + "601": "満", + "602": "舗", + "603": "宙", + "604": "寿", + "605": "遺", + "606": "極", + "607": "里", + "608": "因", + "609": "典", + "610": "染", + "611": "徒", + "612": "巻", + "613": "頂", + "614": "超", + "615": "河", + "616": "盛", + "617": "犬", + "618": "豊", + "619": "端", + "620": "紹", + "621": "首", + "622": "陽", + "623": "歳", + "624": "印", + "625": "紙", + "626": "払", + "627": "求", + "628": "障", + "629": "簡", + "630": "途", + "631": "創", + "632": "船", + "633": "菜", + "634": "ゥ", + "635": "勤", + "636": "痛", + "637": "並", + "638": "景", + "639": "雪", + "640": "節", + "641": "浜", + "642": "清", + "643": "抜", + "644": "勢", + "645": "暮", + "646": "銀", + "647": "盟", + "648": "魚", + "649": "率", + "650": "洋", + "651": "渡", + "652": "順", + "653": "況", + "654": "談", + "655": "舞", + "656": "案", + "657": "岩", + "658": "負", + "659": "旧", + "660": "財", + "661": "故", + "662": "冬", + "663": "横", + "664": "奥", + "665": "比", + "666": "囲", + "667": "停", + "668": "築", + "669": "波", + "670": "林", + "671": "暖", + "672": "索", + "673": "赤", + "674": "給", + "675": "末", + "676": "催", + "677": "遅", + "678": "述", + "679": "黒", + "680": "細", + "681": "与", + "682": "減", + "683": "級", + "684": "費", + "685": "越", + "686": "差", + "687": "領", + "688": "衛", + "689": "隊", + "690": "薬", + "691": "氏", + "692": "望", + "693": "似", + "694": "就", + "695": "条", + "696": "処", + "697": "谷", + "698": "策", + "699": "効", + "700": "熱", + "701": "復", + "702": "ヌ", + "703": "振", + "704": "規", + "705": "港", + "706": "注", + "707": "森", + "708": "防", + "709": "継", + "710": "退", + "711": "火", + "712": "陸", + "713": "去", + "714": "視", + "715": "整", + "716": "準", + "717": "庭", + "718": "ゾ", + "719": "独", + "720": "撃", + "721": "児", + "722": "橋", + "723": "換", + "724": "念", + "725": "識", + "726": "打", + "727": "津", + "728": "雨", + "729": "幸", + "730": "含", + "731": "響", + "732": "労", + "733": "官", + "734": "追", + "735": "遠", + "736": "未", + "737": "販", + "738": "街", + "739": "曜", + "740": "程", + "741": "提", + "742": "玉", + "743": "判", + "744": "移", + "745": "攻", + "746": "低", + "747": "装", + "748": "断", + "749": "及", + "750": "証", + "751": "象", + "752": "守", + "753": "戻", + "754": "詞", + "755": "投", + "756": "載", + "757": "具", + "758": "除", + "759": "環", + "760": "展", + "761": "争", + "762": "失", + "763": "春", + "764": "挙", + "765": "返", + "766": "馬", + "767": "欲", + "768": "材", + "769": "図", + "770": "養", + "771": "焼", + "772": "導", + "773": "夢", + "774": "米", + "775": "冷", + "776": "息", + "777": "兵", + "778": "済", + "779": "劇", + "780": "央", + "781": "険", + "782": "服", + "783": "態", + "784": "走", + "785": "評", + "786": "権", + "787": "論", + "788": "ゅ", + "789": "境", + "790": "察", + "791": "授", + "792": "頼", + "793": "派", + "794": "撮", + "795": "素", + "796": "修", + "797": "第", + "798": "質", + "799": "告", + "800": "興", + "801": "秒", + "802": "宇", + "803": "肉", + "804": "像", + "805": "称", + "806": "値", + "807": "頭", + "808": "週", + "809": "督", + "810": "消", + "811": "芸", + "812": "顔", + "813": "読", + "814": "仲", + "815": "遊", + "816": "試", + "817": "酒", + "818": "離", + "819": "増", + "820": "殺", + "821": "鉄", + "822": "害", + "823": "割", + "824": "石", + "825": "夏", + "826": "助", + "827": "英", + "828": "想", + "829": "管", + "830": "急", + "831": "頃", + "832": "づ", + "833": "造", + "834": "史", + "835": "量", + "836": "製", + "837": "府", + "838": "足", + "839": "王", + "840": "委", + "841": "両", + "842": "辺", + "843": "残", + "844": "逆", + "845": "備", + "846": "軍", + "847": "警", + "848": "査", + "849": "列", + "850": "編", + "851": "段", + "852": "反", + "853": "ゼ", + "854": "携", + "855": "歩", + "856": "座", + "857": "飛", + "858": "丈", + "859": "価", + "860": "監", + "861": "ヘ", + "862": "周", + "863": "毎", + "864": "統", + "865": "収", + "866": "落", + "867": "星", + "868": "降", + "869": "側", + "870": "療", + "871": "師", + "872": "写", + "873": "類", + "874": "命", + "875": "介", + "876": "護", + "877": "死", + "878": "果", + "879": "任", + "880": "更", + "881": "常", + "882": "検", + "883": "過", + "884": "資", + "885": "働", + "886": "認", + "887": "般", + "888": "示", + "889": "客", + "890": "習", + "891": "究", + "892": "半", + "893": "録", + "894": "字", + "895": "昔", + "896": "影", + "897": "覚", + "898": "型", + "899": "声", + "900": "件", + "901": "義", + "902": "施", + "903": "容", + "904": "路", + "905": "呼", + "906": "役", + "907": "単", + "908": "状", + "909": "建", + "910": "由", + "911": "属", + "912": "土", + "913": "葉", + "914": "起", + "915": "覧", + "916": "配", + "917": "張", + "918": "接", + "919": "込", + "920": "待", + "921": "室", + "922": "病", + "923": "帯", + "924": "婚", + "925": "光", + "926": "個", + "927": "職", + "928": "営", + "929": "ぼ", + "930": "研", + "931": "計", + "932": "直", + "933": "難", + "934": "絶", + "935": "ヨ", + "936": "照", + "937": "西", + "938": "約", + "939": "存", + "940": "験", + "941": "治", + "942": "解", + "943": "転", + "944": "商", + "945": "進", + "946": "係", + "947": "説", + "948": "観", + "949": "球", + "950": "乗", + "951": "支", + "952": "得", + "953": "議", + "954": "門", + "955": "止", + "956": "重", + "957": "温", + "958": "着", + "959": "飲", + "960": "母", + "961": "士", + "962": "ざ", + "963": "集", + "964": "万", + "965": "太", + "966": "続", + "967": "線", + "968": "種", + "969": "格", + "970": "位", + "971": "ユ", + "972": "歌", + "973": "夜", + "974": "共", + "975": "正", + "976": "必", + "977": "ヒ", + "978": "色", + "979": "問", + "980": "再", + "981": "域", + "982": "ゆ", + "983": "勝", + "984": "台", + "985": "技", + "986": "旅", + "987": "引", + "988": "系", + "989": "院", + "990": "悪", + "991": "基", + "992": "神", + "993": "産", + "994": "決", + "995": "民", + "996": "交", + "997": "政", + "998": "賞", + "999": "空", + "1000": "医", + "1001": "彼", + "1002": "夫", + "1003": "可", + "1004": "誰", + "1005": "古", + "1006": "帰", + "1007": "術", + "1008": "相", + "1009": "団", + "1010": "伝", + "1011": "住", + "1012": "題", + "1013": "平", + "1014": "予", + "1015": "音", + "1016": "朝", + "1017": "指", + "1018": "真", + "1019": "ヴ", + "1020": "務", + "1021": "点", + "1022": "各", + "1023": "館", + "1024": "応", + "1025": "現", + "1026": "利", + "1027": "天", + "1028": "等", + "1029": "木", + "1030": "白", + "1031": "形", + "1032": "供", + "1033": "経", + "1034": "族", + "1035": "早", + "1036": "例", + "1037": "不", + "1038": "切", + "1039": "南", + "1040": "加", + "1041": "際", + "1042": "終", + "1043": "様", + "1044": "放", + "1045": "和", + "1046": "州", + "1047": "水", + "1048": "協", + "1049": "在", + "1050": "組", + "1051": "向", + "1052": "広", + "1053": "身", + "1054": "界", + "1055": "工", + "1056": "選", + "1057": "始", + "1058": "元", + "1059": "々", + "1060": "親", + "1061": "美", + "1062": "信", + "1063": "都", + "1064": "置", + "1065": "局", + "1066": "運", + "1067": "送", + "1068": "風", + "1069": "口", + "1070": "演", + "1071": "調", + "1072": "ぎ", + "1073": "優", + "1074": "次", + "1075": "ォ", + "1076": "他", + "1077": "園", + "1078": "保", + "1079": "男", + "1080": "参", + "1081": "少", + "1082": "百", + "1083": "特", + "1084": "考", + "1085": "無", + "1086": "七", + "1087": "ヤ", + "1088": "ギ", + "1089": "良", + "1090": "ザ", + "1091": "制", + "1092": "売", + "1093": "能", + "1094": "原", + "1095": "ゲ", + "1096": "有", + "1097": "安", + "1098": "ゴ", + "1099": "育", + "1100": "科", + "1101": "要", + "1102": "料", + "1103": "書", + "1104": "語", + "1105": "設", + "1106": "海", + "1107": "期", + "1108": "流", + "1109": "確", + "1110": "ペ", + "1111": "区", + "1112": "む", + "1113": "連", + "1114": "買", + "1115": "ひ", + "1116": "ふ", + "1117": "付", + "1118": "町", + "1119": "活", + "1120": "情", + "1121": "月", + "1122": "表", + "1123": "曲", + "1124": "強", + "1125": "世", + "1126": "明", + "1127": "成", + "1128": "ノ", + "1129": "ァ", + "1130": "文", + "1131": "違", + "1132": "東", + "1133": "友", + "1134": "意", + "1135": "力", + "1136": "式", + "1137": "法", + "1138": "報", + "1139": "員", + "1140": "心", + "1141": "屋", + "1142": "品", + "1143": "北", + "1144": "先", + "1145": "島", + "1146": "味", + "1147": "川", + "1148": "開", + "1149": "千", + "1150": "関", + "1151": "電", + "1152": "然", + "1153": "度", + "1154": "達", + "1155": "面", + "1156": "九", + "1157": "数", + "1158": "取", + "1159": "楽", + "1160": "金", + "1161": "性", + "1162": "野", + "1163": "別", + "1164": "戦", + "1165": "公", + "1166": "機", + "1167": "道", + "1168": "目", + "1169": "記", + "1170": "び", + "1171": "発", + "1172": "対", + "1173": "立", + "1174": "初", + "1175": "化", + "1176": "ソ", + "1177": "ワ", + "1178": "田", + "1179": "持", + "1180": "ガ", + "1181": "車", + "1182": "番", + "1183": "ピ", + "1184": "聞", + "1185": "回", + "1186": "ぶ", + "1187": "ベ", + "1188": "げ", + "1189": "実", + "1190": "ボ", + "1191": "店", + "1192": "小", + "1193": "定", + "1194": "モ", + "1195": "長", + "1196": "新", + "1197": "ハ", + "1198": "ケ", + "1199": "外", + "1200": "ポ", + "1201": "近", + "1202": "所", + "1203": "へ", + "1204": "同", + "1205": "ネ", + "1206": "内", + "1207": "女", + "1208": "ホ", + "1209": "体", + "1210": "好", + "1211": "ツ", + "1212": "セ", + "1213": "知", + "1214": "山", + "1215": "来", + "1216": "ェ", + "1217": "使", + "1218": "ョ", + "1219": "ズ", + "1220": "主", + "1221": "動", + "1222": "理", + "1223": "物", + "1224": "映", + "1225": "者", + "1226": "ぐ", + "1227": "的", + "1228": "代", + "1229": "変", + "1230": "教", + "1231": "社", + "1232": "用", + "1233": "話", + "1234": "名", + "1235": "構", + "1236": "高", + "1237": "最", + "1238": "ず", + "1239": "ミ", + "1240": "校", + "1241": "ダ", + "1242": "食", + "1243": "後", + "1244": "手", + "1245": "三", + "1246": "通", + "1247": "感", + "1248": "合", + "1249": "多", + "1250": "業", + "1251": "入", + "1252": "エ", + "1253": "場", + "1254": "べ", + "1255": "上", + "1256": "家", + "1257": "私", + "1258": "年", + "1259": "間", + "1260": "画", + "1261": "前", + "1262": "下", + "1263": "ャ", + "1264": "地", + "1265": "二", + "1266": "ウ", + "1267": "ナ", + "1268": "ビ", + "1269": "自", + "1270": "全", + "1271": "パ", + "1272": "結", + "1273": "ブ", + "1274": "ュ", + "1275": "市", + "1276": "サ", + "1277": "気", + "1278": "方", + "1279": "デ", + "1280": "十", + "1281": "キ", + "1282": "当", + "1283": "国", + "1284": "作", + "1285": "ィ", + "1286": "部", + "1287": "オ", + "1288": "ニ", + "1289": "チ", + "1290": "ム", + "1291": "グ", + "1292": "メ", + "1293": "ご", + "1294": "子", + "1295": "ば", + "1296": "生", + "1297": "ほ", + "1298": "せ", + "1299": "何", + "1300": "出", + "1301": "言", + "1302": "今", + "1303": "バ", + "1304": "事", + "1305": "中", + "1306": "プ", + "1307": "時", + "1308": "コ", + "1309": "見", + "1310": "テ", + "1311": "会", + "1312": "マ", + "1313": "カ", + "1314": "思", + "1315": "ロ", + "1316": "ジ", + "1317": "フ", + "1318": "シ", + "1319": "め", + "1320": "レ", + "1321": "ド", + "1322": "分", + "1323": "ょ", + "1324": "ろ", + "1325": "学", + "1326": "行", + "1327": "タ", + "1328": "大", + "1329": "つ", + "1330": "本", + "1331": "日", + "1332": "わ", + "1333": "一", + "1334": "ク", + "1335": "み", + "1336": "リ", + "1337": "ア", + "1338": "ッ", + "1339": "人", + "1340": "ラ", + "1341": "お", + "1342": "じ", + "1343": "イ", + "1344": "ル", + "1345": "ト", + "1346": "ゃ", + "1347": "き", + "1348": "さ", + "1349": "ち", + "1350": "や", + "1351": "ス", + "1352": "ど", + "1353": "け", + "1354": "く", + "1355": "え", + "1356": "を", + "1357": "り", + "1358": "よ", + "1359": "こ", + "1360": "ン", + "1361": "だ", + "1362": "れ", + "1363": "ら", + "1364": "ね", + "1365": "が", + "1366": "ま", + "1367": "ー", + "1368": "も", + "1369": "そ", + "1370": "し", + "1371": "に", + "1372": "は", + "1373": "る", + "1374": "す", + "1375": "と", + "1376": "た", + "1377": "あ", + "1378": "て", + "1379": "っ", + "1380": "で", + "1381": "か", + "1382": "な", + "1383": "ん", + "1384": "う", + "1385": "の", + "1386": "、", + "1387": "。", + "1388": "い", + "1389": "", + "1390": "", + "1391": "", + "1392": "", + "1393": "", + "1394": "", + "1395": "", + "1396": "▁香", + "1397": "▁群", + "1398": "▁米", + "1399": "咆", + "1400": "哮", + "1401": "翅", + "1402": "", + "1403": "" +} \ No newline at end of file diff --git a/ja/560ms/decoder.mlmodelc/analytics/coremldata.bin b/ja/560ms/decoder.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..a23f14dd8e4d2bccc2844d3d81c6c9ca86ea3cba --- /dev/null +++ b/ja/560ms/decoder.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2fcae710f3db79230f47be6daadc8af085539067285a96f89b2a4c0fd0cb3808 +size 243 diff --git a/ja/560ms/decoder.mlmodelc/coremldata.bin b/ja/560ms/decoder.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..89ec734bb4645199981d835aff20eb64bd9e3c4e --- /dev/null +++ b/ja/560ms/decoder.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:298eeaf999bc86b5914efa85450328efc8cf08459f1e8edaefd676f7d5a8410c +size 493 diff --git a/ja/560ms/decoder.mlmodelc/model.mil b/ja/560ms/decoder.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..9f4b6b4cebc16f759164ca05a77b06fb57dedbce --- /dev/null +++ b/ja/560ms/decoder.mlmodelc/model.mil @@ -0,0 +1,73 @@ +program(1.3) +[buildInfo = dict({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.10.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor c_in, tensor h_in, tensor token, tensor token_length) { + int32 y_batch_dims_0 = const()[name = string("y_batch_dims_0"), val = int32(0)]; + bool y_validate_indices_0 = const()[name = string("y_validate_indices_0"), val = bool(false)]; + tensor module_prediction_embed_weight_to_fp16 = const()[name = string("module_prediction_embed_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))]; + string token_to_int16_dtype_0 = const()[name = string("token_to_int16_dtype_0"), val = string("int16")]; + string cast_1_dtype_0 = const()[name = string("cast_1_dtype_0"), val = string("int32")]; + int32 greater_equal_0_y_0 = const()[name = string("greater_equal_0_y_0"), val = int32(0)]; + tensor token_to_int16 = cast(dtype = token_to_int16_dtype_0, x = token)[name = string("cast_9")]; + tensor cast_1 = cast(dtype = cast_1_dtype_0, x = token_to_int16)[name = string("cast_8")]; + tensor greater_equal_0 = greater_equal(x = cast_1, y = greater_equal_0_y_0)[name = string("greater_equal_0")]; + int32 slice_by_index_0 = const()[name = string("slice_by_index_0"), val = int32(1404)]; + tensor add_2 = add(x = cast_1, y = slice_by_index_0)[name = string("add_2")]; + tensor select_0 = select(a = cast_1, b = add_2, cond = greater_equal_0)[name = string("select_0")]; + int32 y_cast_fp16_cast_uint16_axis_0 = const()[name = string("y_cast_fp16_cast_uint16_axis_0"), val = int32(0)]; + string select_0_to_int16_dtype_0 = const()[name = string("select_0_to_int16_dtype_0"), val = string("int16")]; + tensor select_0_to_int16 = cast(dtype = select_0_to_int16_dtype_0, x = select_0)[name = string("cast_7")]; + tensor y_cast_fp16_cast_uint16_cast_uint16 = gather(axis = y_cast_fp16_cast_uint16_axis_0, batch_dims = y_batch_dims_0, indices = select_0_to_int16, validate_indices = y_validate_indices_0, x = module_prediction_embed_weight_to_fp16)[name = string("y_cast_fp16_cast_uint16_cast_uint16")]; + tensor input_3_perm_0 = const()[name = string("input_3_perm_0"), val = tensor([1, 0, 2])]; + int32 split_0_num_splits_0 = const()[name = string("split_0_num_splits_0"), val = int32(2)]; + int32 split_0_axis_0 = const()[name = string("split_0_axis_0"), val = int32(0)]; + string h_in_to_fp16_dtype_0 = const()[name = string("h_in_to_fp16_dtype_0"), val = string("fp16")]; + tensor h_in_to_fp16 = cast(dtype = h_in_to_fp16_dtype_0, x = h_in)[name = string("cast_6")]; + tensor split_0_cast_fp16_0, tensor split_0_cast_fp16_1 = split(axis = split_0_axis_0, num_splits = split_0_num_splits_0, x = h_in_to_fp16)[name = string("split_0_cast_fp16")]; + int32 split_1_num_splits_0 = const()[name = string("split_1_num_splits_0"), val = int32(2)]; + int32 split_1_axis_0 = const()[name = string("split_1_axis_0"), val = int32(0)]; + string c_in_to_fp16_dtype_0 = const()[name = string("c_in_to_fp16_dtype_0"), val = string("fp16")]; + tensor c_in_to_fp16 = cast(dtype = c_in_to_fp16_dtype_0, x = c_in)[name = string("cast_5")]; + tensor split_1_cast_fp16_0, tensor split_1_cast_fp16_1 = split(axis = split_1_axis_0, num_splits = split_1_num_splits_0, x = c_in_to_fp16)[name = string("split_1_cast_fp16")]; + tensor input_lstm_layer_0_lstm_h0_squeeze_axes_0 = const()[name = string("input_lstm_layer_0_lstm_h0_squeeze_axes_0"), val = tensor([0])]; + tensor input_lstm_layer_0_lstm_h0_squeeze_cast_fp16 = squeeze(axes = input_lstm_layer_0_lstm_h0_squeeze_axes_0, x = split_0_cast_fp16_0)[name = string("input_lstm_layer_0_lstm_h0_squeeze_cast_fp16")]; + tensor input_lstm_layer_0_lstm_c0_squeeze_axes_0 = const()[name = string("input_lstm_layer_0_lstm_c0_squeeze_axes_0"), val = tensor([0])]; + tensor input_lstm_layer_0_lstm_c0_squeeze_cast_fp16 = squeeze(axes = input_lstm_layer_0_lstm_c0_squeeze_axes_0, x = split_1_cast_fp16_0)[name = string("input_lstm_layer_0_lstm_c0_squeeze_cast_fp16")]; + string input_lstm_layer_0_direction_0 = const()[name = string("input_lstm_layer_0_direction_0"), val = string("forward")]; + bool input_lstm_layer_0_output_sequence_0 = const()[name = string("input_lstm_layer_0_output_sequence_0"), val = bool(true)]; + string input_lstm_layer_0_recurrent_activation_0 = const()[name = string("input_lstm_layer_0_recurrent_activation_0"), val = string("sigmoid")]; + string input_lstm_layer_0_cell_activation_0 = const()[name = string("input_lstm_layer_0_cell_activation_0"), val = string("tanh")]; + string input_lstm_layer_0_activation_0 = const()[name = string("input_lstm_layer_0_activation_0"), val = string("tanh")]; + tensor concat_1_to_fp16 = const()[name = string("concat_1_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1797248)))]; + tensor concat_2_to_fp16 = const()[name = string("concat_2_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5074112)))]; + tensor concat_0_to_fp16 = const()[name = string("concat_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8350976)))]; + tensor input_3_cast_fp16 = transpose(perm = input_3_perm_0, x = y_cast_fp16_cast_uint16_cast_uint16)[name = string("transpose_2")]; + tensor input_lstm_layer_0_cast_fp16_0, tensor input_lstm_layer_0_cast_fp16_1, tensor input_lstm_layer_0_cast_fp16_2 = lstm(activation = input_lstm_layer_0_activation_0, bias = concat_0_to_fp16, cell_activation = input_lstm_layer_0_cell_activation_0, direction = input_lstm_layer_0_direction_0, initial_c = input_lstm_layer_0_lstm_c0_squeeze_cast_fp16, initial_h = input_lstm_layer_0_lstm_h0_squeeze_cast_fp16, output_sequence = input_lstm_layer_0_output_sequence_0, recurrent_activation = input_lstm_layer_0_recurrent_activation_0, weight_hh = concat_2_to_fp16, weight_ih = concat_1_to_fp16, x = input_3_cast_fp16)[name = string("input_lstm_layer_0_cast_fp16")]; + tensor input_lstm_h0_squeeze_axes_0 = const()[name = string("input_lstm_h0_squeeze_axes_0"), val = tensor([0])]; + tensor input_lstm_h0_squeeze_cast_fp16 = squeeze(axes = input_lstm_h0_squeeze_axes_0, x = split_0_cast_fp16_1)[name = string("input_lstm_h0_squeeze_cast_fp16")]; + tensor input_lstm_c0_squeeze_axes_0 = const()[name = string("input_lstm_c0_squeeze_axes_0"), val = tensor([0])]; + tensor input_lstm_c0_squeeze_cast_fp16 = squeeze(axes = input_lstm_c0_squeeze_axes_0, x = split_1_cast_fp16_1)[name = string("input_lstm_c0_squeeze_cast_fp16")]; + string input_direction_0 = const()[name = string("input_direction_0"), val = string("forward")]; + bool input_output_sequence_0 = const()[name = string("input_output_sequence_0"), val = bool(true)]; + string input_recurrent_activation_0 = const()[name = string("input_recurrent_activation_0"), val = string("sigmoid")]; + string input_cell_activation_0 = const()[name = string("input_cell_activation_0"), val = string("tanh")]; + string input_activation_0 = const()[name = string("input_activation_0"), val = string("tanh")]; + tensor concat_4_to_fp16 = const()[name = string("concat_4_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8356160)))]; + tensor concat_5_to_fp16 = const()[name = string("concat_5_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11633024)))]; + tensor concat_3_to_fp16 = const()[name = string("concat_3_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14909888)))]; + tensor input_cast_fp16_0, tensor input_cast_fp16_1, tensor input_cast_fp16_2 = lstm(activation = input_activation_0, bias = concat_3_to_fp16, cell_activation = input_cell_activation_0, direction = input_direction_0, initial_c = input_lstm_c0_squeeze_cast_fp16, initial_h = input_lstm_h0_squeeze_cast_fp16, output_sequence = input_output_sequence_0, recurrent_activation = input_recurrent_activation_0, weight_hh = concat_5_to_fp16, weight_ih = concat_4_to_fp16, x = input_lstm_layer_0_cast_fp16_0)[name = string("input_cast_fp16")]; + int32 obj_3_axis_0 = const()[name = string("obj_3_axis_0"), val = int32(0)]; + tensor obj_3_cast_fp16 = stack(axis = obj_3_axis_0, values = (input_lstm_layer_0_cast_fp16_1, input_cast_fp16_1))[name = string("obj_3_cast_fp16")]; + string obj_3_cast_fp16_to_fp32_dtype_0 = const()[name = string("obj_3_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + int32 obj_axis_0 = const()[name = string("obj_axis_0"), val = int32(0)]; + tensor obj_cast_fp16 = stack(axis = obj_axis_0, values = (input_lstm_layer_0_cast_fp16_2, input_cast_fp16_2))[name = string("obj_cast_fp16")]; + string obj_cast_fp16_to_fp32_dtype_0 = const()[name = string("obj_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + tensor transpose_0_perm_0 = const()[name = string("transpose_0_perm_0"), val = tensor([1, 2, 0])]; + string transpose_0_cast_fp16_to_fp32_dtype_0 = const()[name = string("transpose_0_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + tensor transpose_0_cast_fp16 = transpose(perm = transpose_0_perm_0, x = input_cast_fp16_0)[name = string("transpose_1")]; + tensor decoder_out = cast(dtype = transpose_0_cast_fp16_to_fp32_dtype_0, x = transpose_0_cast_fp16)[name = string("cast_2")]; + tensor c_out = cast(dtype = obj_cast_fp16_to_fp32_dtype_0, x = obj_cast_fp16)[name = string("cast_3")]; + tensor h_out = cast(dtype = obj_3_cast_fp16_to_fp32_dtype_0, x = obj_3_cast_fp16)[name = string("cast_4")]; + tensor token_length_tmp = identity(x = token_length)[name = string("token_length_tmp")]; + } -> (decoder_out, h_out, c_out); +} \ No newline at end of file diff --git a/ja/560ms/decoder.mlmodelc/weights/weight.bin b/ja/560ms/decoder.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..9bcbce3e617e1135c0460941dfc723b71230d030 --- /dev/null +++ b/ja/560ms/decoder.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d5eeb17b5de822e941f66ce3bd6739aae939dfa054f61b69ff6ce89d0ac9e778 +size 14915072 diff --git a/ja/560ms/decoder.mlpackage/Data/com.apple.CoreML/model.mlmodel b/ja/560ms/decoder.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..931ee2253d124627cf1b2689c6e01d5cf3746838 --- /dev/null +++ b/ja/560ms/decoder.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:25c461bd45595f33022b4ce50bf3d493d5b70ae73c50bd0a98598336bd38864a +size 11598 diff --git a/ja/560ms/decoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/ja/560ms/decoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..9bcbce3e617e1135c0460941dfc723b71230d030 --- /dev/null +++ b/ja/560ms/decoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d5eeb17b5de822e941f66ce3bd6739aae939dfa054f61b69ff6ce89d0ac9e778 +size 14915072 diff --git a/ja/560ms/decoder.mlpackage/Manifest.json b/ja/560ms/decoder.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..b4431e1d0753f3273eeff30f89de1232349486a7 --- /dev/null +++ b/ja/560ms/decoder.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "8C20B369-4E12-4E4E-B3E8-A79B91D9CAFC": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + }, + "9356FC01-CF91-4D74-A142-118AF15703DD": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + } + }, + "rootModelIdentifier": "9356FC01-CF91-4D74-A142-118AF15703DD" +} diff --git a/ja/560ms/decoder_joint.mlmodelc/analytics/coremldata.bin b/ja/560ms/decoder_joint.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..30ee1bc4e73ed57bede1d9e6315c983146d06e8c --- /dev/null +++ b/ja/560ms/decoder_joint.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:adba69d0e8e1547064d062072f64e8a9f1da383a6d09e2986a28268dd78cb23c +size 243 diff --git a/ja/560ms/decoder_joint.mlmodelc/coremldata.bin b/ja/560ms/decoder_joint.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..ff735d9d794bd3717f0344022771f29df72a633d --- /dev/null +++ b/ja/560ms/decoder_joint.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:85c85ecec674a9bee9777b7cf93682fd4cb5ea9bed388a030224a6f90dd72cde +size 514 diff --git a/ja/560ms/decoder_joint.mlmodelc/model.mil b/ja/560ms/decoder_joint.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..2cb099751b91d7bd911aeac28092cc495bcaf315 --- /dev/null +++ b/ja/560ms/decoder_joint.mlmodelc/model.mil @@ -0,0 +1,92 @@ +program(1.3) +[buildInfo = dict({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.10.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor c_in, tensor encoder, tensor h_in, tensor token, tensor token_length) { + int32 y_batch_dims_0 = const()[name = string("y_batch_dims_0"), val = int32(0)]; + bool y_validate_indices_0 = const()[name = string("y_validate_indices_0"), val = bool(false)]; + tensor decoder_module_prediction_embed_weight_to_fp16 = const()[name = string("decoder_module_prediction_embed_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))]; + string token_to_int16_dtype_0 = const()[name = string("token_to_int16_dtype_0"), val = string("int16")]; + string cast_1_dtype_0 = const()[name = string("cast_1_dtype_0"), val = string("int32")]; + int32 greater_equal_0_y_0 = const()[name = string("greater_equal_0_y_0"), val = int32(0)]; + tensor token_to_int16 = cast(dtype = token_to_int16_dtype_0, x = token)[name = string("cast_10")]; + tensor cast_1 = cast(dtype = cast_1_dtype_0, x = token_to_int16)[name = string("cast_9")]; + tensor greater_equal_0 = greater_equal(x = cast_1, y = greater_equal_0_y_0)[name = string("greater_equal_0")]; + int32 slice_by_index_0 = const()[name = string("slice_by_index_0"), val = int32(1404)]; + tensor add_2 = add(x = cast_1, y = slice_by_index_0)[name = string("add_2")]; + tensor select_0 = select(a = cast_1, b = add_2, cond = greater_equal_0)[name = string("select_0")]; + int32 y_cast_fp16_cast_uint16_axis_0 = const()[name = string("y_cast_fp16_cast_uint16_axis_0"), val = int32(0)]; + string select_0_to_int16_dtype_0 = const()[name = string("select_0_to_int16_dtype_0"), val = string("int16")]; + tensor select_0_to_int16 = cast(dtype = select_0_to_int16_dtype_0, x = select_0)[name = string("cast_8")]; + tensor y_cast_fp16_cast_uint16_cast_uint16 = gather(axis = y_cast_fp16_cast_uint16_axis_0, batch_dims = y_batch_dims_0, indices = select_0_to_int16, validate_indices = y_validate_indices_0, x = decoder_module_prediction_embed_weight_to_fp16)[name = string("y_cast_fp16_cast_uint16_cast_uint16")]; + tensor input_3_perm_0 = const()[name = string("input_3_perm_0"), val = tensor([1, 0, 2])]; + int32 split_0_num_splits_0 = const()[name = string("split_0_num_splits_0"), val = int32(2)]; + int32 split_0_axis_0 = const()[name = string("split_0_axis_0"), val = int32(0)]; + string h_in_to_fp16_dtype_0 = const()[name = string("h_in_to_fp16_dtype_0"), val = string("fp16")]; + tensor h_in_to_fp16 = cast(dtype = h_in_to_fp16_dtype_0, x = h_in)[name = string("cast_7")]; + tensor split_0_cast_fp16_0, tensor split_0_cast_fp16_1 = split(axis = split_0_axis_0, num_splits = split_0_num_splits_0, x = h_in_to_fp16)[name = string("split_0_cast_fp16")]; + int32 split_1_num_splits_0 = const()[name = string("split_1_num_splits_0"), val = int32(2)]; + int32 split_1_axis_0 = const()[name = string("split_1_axis_0"), val = int32(0)]; + string c_in_to_fp16_dtype_0 = const()[name = string("c_in_to_fp16_dtype_0"), val = string("fp16")]; + tensor c_in_to_fp16 = cast(dtype = c_in_to_fp16_dtype_0, x = c_in)[name = string("cast_6")]; + tensor split_1_cast_fp16_0, tensor split_1_cast_fp16_1 = split(axis = split_1_axis_0, num_splits = split_1_num_splits_0, x = c_in_to_fp16)[name = string("split_1_cast_fp16")]; + tensor input_5_lstm_layer_0_lstm_h0_squeeze_axes_0 = const()[name = string("input_5_lstm_layer_0_lstm_h0_squeeze_axes_0"), val = tensor([0])]; + tensor input_5_lstm_layer_0_lstm_h0_squeeze_cast_fp16 = squeeze(axes = input_5_lstm_layer_0_lstm_h0_squeeze_axes_0, x = split_0_cast_fp16_0)[name = string("input_5_lstm_layer_0_lstm_h0_squeeze_cast_fp16")]; + tensor input_5_lstm_layer_0_lstm_c0_squeeze_axes_0 = const()[name = string("input_5_lstm_layer_0_lstm_c0_squeeze_axes_0"), val = tensor([0])]; + tensor input_5_lstm_layer_0_lstm_c0_squeeze_cast_fp16 = squeeze(axes = input_5_lstm_layer_0_lstm_c0_squeeze_axes_0, x = split_1_cast_fp16_0)[name = string("input_5_lstm_layer_0_lstm_c0_squeeze_cast_fp16")]; + string input_5_lstm_layer_0_direction_0 = const()[name = string("input_5_lstm_layer_0_direction_0"), val = string("forward")]; + bool input_5_lstm_layer_0_output_sequence_0 = const()[name = string("input_5_lstm_layer_0_output_sequence_0"), val = bool(true)]; + string input_5_lstm_layer_0_recurrent_activation_0 = const()[name = string("input_5_lstm_layer_0_recurrent_activation_0"), val = string("sigmoid")]; + string input_5_lstm_layer_0_cell_activation_0 = const()[name = string("input_5_lstm_layer_0_cell_activation_0"), val = string("tanh")]; + string input_5_lstm_layer_0_activation_0 = const()[name = string("input_5_lstm_layer_0_activation_0"), val = string("tanh")]; + tensor concat_1_to_fp16 = const()[name = string("concat_1_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1797248)))]; + tensor concat_2_to_fp16 = const()[name = string("concat_2_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5074112)))]; + tensor concat_0_to_fp16 = const()[name = string("concat_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8350976)))]; + tensor input_3_cast_fp16 = transpose(perm = input_3_perm_0, x = y_cast_fp16_cast_uint16_cast_uint16)[name = string("transpose_4")]; + tensor input_5_lstm_layer_0_cast_fp16_0, tensor input_5_lstm_layer_0_cast_fp16_1, tensor input_5_lstm_layer_0_cast_fp16_2 = lstm(activation = input_5_lstm_layer_0_activation_0, bias = concat_0_to_fp16, cell_activation = input_5_lstm_layer_0_cell_activation_0, direction = input_5_lstm_layer_0_direction_0, initial_c = input_5_lstm_layer_0_lstm_c0_squeeze_cast_fp16, initial_h = input_5_lstm_layer_0_lstm_h0_squeeze_cast_fp16, output_sequence = input_5_lstm_layer_0_output_sequence_0, recurrent_activation = input_5_lstm_layer_0_recurrent_activation_0, weight_hh = concat_2_to_fp16, weight_ih = concat_1_to_fp16, x = input_3_cast_fp16)[name = string("input_5_lstm_layer_0_cast_fp16")]; + tensor input_5_lstm_h0_squeeze_axes_0 = const()[name = string("input_5_lstm_h0_squeeze_axes_0"), val = tensor([0])]; + tensor input_5_lstm_h0_squeeze_cast_fp16 = squeeze(axes = input_5_lstm_h0_squeeze_axes_0, x = split_0_cast_fp16_1)[name = string("input_5_lstm_h0_squeeze_cast_fp16")]; + tensor input_5_lstm_c0_squeeze_axes_0 = const()[name = string("input_5_lstm_c0_squeeze_axes_0"), val = tensor([0])]; + tensor input_5_lstm_c0_squeeze_cast_fp16 = squeeze(axes = input_5_lstm_c0_squeeze_axes_0, x = split_1_cast_fp16_1)[name = string("input_5_lstm_c0_squeeze_cast_fp16")]; + string input_5_direction_0 = const()[name = string("input_5_direction_0"), val = string("forward")]; + bool input_5_output_sequence_0 = const()[name = string("input_5_output_sequence_0"), val = bool(true)]; + string input_5_recurrent_activation_0 = const()[name = string("input_5_recurrent_activation_0"), val = string("sigmoid")]; + string input_5_cell_activation_0 = const()[name = string("input_5_cell_activation_0"), val = string("tanh")]; + string input_5_activation_0 = const()[name = string("input_5_activation_0"), val = string("tanh")]; + tensor concat_4_to_fp16 = const()[name = string("concat_4_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8356160)))]; + tensor concat_5_to_fp16 = const()[name = string("concat_5_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11633024)))]; + tensor concat_3_to_fp16 = const()[name = string("concat_3_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14909888)))]; + tensor input_5_cast_fp16_0, tensor input_5_cast_fp16_1, tensor input_5_cast_fp16_2 = lstm(activation = input_5_activation_0, bias = concat_3_to_fp16, cell_activation = input_5_cell_activation_0, direction = input_5_direction_0, initial_c = input_5_lstm_c0_squeeze_cast_fp16, initial_h = input_5_lstm_h0_squeeze_cast_fp16, output_sequence = input_5_output_sequence_0, recurrent_activation = input_5_recurrent_activation_0, weight_hh = concat_5_to_fp16, weight_ih = concat_4_to_fp16, x = input_5_lstm_layer_0_cast_fp16_0)[name = string("input_5_cast_fp16")]; + int32 obj_3_axis_0 = const()[name = string("obj_3_axis_0"), val = int32(0)]; + tensor obj_3_cast_fp16 = stack(axis = obj_3_axis_0, values = (input_5_lstm_layer_0_cast_fp16_1, input_5_cast_fp16_1))[name = string("obj_3_cast_fp16")]; + string obj_3_cast_fp16_to_fp32_dtype_0 = const()[name = string("obj_3_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + int32 obj_axis_0 = const()[name = string("obj_axis_0"), val = int32(0)]; + tensor obj_cast_fp16 = stack(axis = obj_axis_0, values = (input_5_lstm_layer_0_cast_fp16_2, input_5_cast_fp16_2))[name = string("obj_cast_fp16")]; + string obj_cast_fp16_to_fp32_dtype_0 = const()[name = string("obj_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + tensor transpose_1_perm_0 = const()[name = string("transpose_1_perm_0"), val = tensor([1, 0, 2])]; + tensor input_7_perm_0 = const()[name = string("input_7_perm_0"), val = tensor([0, 2, 1])]; + string encoder_to_fp16_dtype_0 = const()[name = string("encoder_to_fp16_dtype_0"), val = string("fp16")]; + tensor joint_module_enc_weight_to_fp16 = const()[name = string("joint_module_enc_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14915072)))]; + tensor joint_module_enc_bias_to_fp16 = const()[name = string("joint_module_enc_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16225856)))]; + tensor encoder_to_fp16 = cast(dtype = encoder_to_fp16_dtype_0, x = encoder)[name = string("cast_3")]; + tensor input_7_cast_fp16 = transpose(perm = input_7_perm_0, x = encoder_to_fp16)[name = string("transpose_2")]; + tensor linear_0_cast_fp16 = linear(bias = joint_module_enc_bias_to_fp16, weight = joint_module_enc_weight_to_fp16, x = input_7_cast_fp16)[name = string("linear_0_cast_fp16")]; + tensor joint_module_pred_weight_to_fp16 = const()[name = string("joint_module_pred_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16227200)))]; + tensor joint_module_pred_bias_to_fp16 = const()[name = string("joint_module_pred_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17046464)))]; + tensor transpose_1_cast_fp16 = transpose(perm = transpose_1_perm_0, x = input_5_cast_fp16_0)[name = string("transpose_3")]; + tensor linear_1_cast_fp16 = linear(bias = joint_module_pred_bias_to_fp16, weight = joint_module_pred_weight_to_fp16, x = transpose_1_cast_fp16)[name = string("linear_1_cast_fp16")]; + tensor var_79_axes_0 = const()[name = string("op_79_axes_0"), val = tensor([2])]; + tensor var_79_cast_fp16 = expand_dims(axes = var_79_axes_0, x = linear_0_cast_fp16)[name = string("op_79_cast_fp16")]; + tensor var_80_axes_0 = const()[name = string("op_80_axes_0"), val = tensor([1])]; + tensor var_80_cast_fp16 = expand_dims(axes = var_80_axes_0, x = linear_1_cast_fp16)[name = string("op_80_cast_fp16")]; + tensor input_11_cast_fp16 = add(x = var_79_cast_fp16, y = var_80_cast_fp16)[name = string("input_11_cast_fp16")]; + tensor input_13_cast_fp16 = relu(x = input_11_cast_fp16)[name = string("input_13_cast_fp16")]; + tensor joint_module_joint_net_2_weight_to_fp16 = const()[name = string("joint_module_joint_net_2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17047808)))]; + tensor joint_module_joint_net_2_bias_to_fp16 = const()[name = string("joint_module_joint_net_2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18844992)))]; + tensor linear_2_cast_fp16 = linear(bias = joint_module_joint_net_2_bias_to_fp16, weight = joint_module_joint_net_2_weight_to_fp16, x = input_13_cast_fp16)[name = string("linear_2_cast_fp16")]; + string linear_2_cast_fp16_to_fp32_dtype_0 = const()[name = string("linear_2_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + tensor logits = cast(dtype = linear_2_cast_fp16_to_fp32_dtype_0, x = linear_2_cast_fp16)[name = string("cast_2")]; + tensor c_out = cast(dtype = obj_cast_fp16_to_fp32_dtype_0, x = obj_cast_fp16)[name = string("cast_4")]; + tensor h_out = cast(dtype = obj_3_cast_fp16_to_fp32_dtype_0, x = obj_3_cast_fp16)[name = string("cast_5")]; + tensor token_length_tmp = identity(x = token_length)[name = string("token_length_tmp")]; + } -> (logits, h_out, c_out); +} \ No newline at end of file diff --git a/ja/560ms/decoder_joint.mlmodelc/weights/weight.bin b/ja/560ms/decoder_joint.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..c3bb2e494d21dfd602e504fdfe76da274071d914 --- /dev/null +++ b/ja/560ms/decoder_joint.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bfc1c768dd0e0e61c0ab8806894ecc03902d2a5028e9c30f5d0a5e38d5139fd9 +size 18847864 diff --git a/ja/560ms/decoder_joint.mlpackage/Data/com.apple.CoreML/model.mlmodel b/ja/560ms/decoder_joint.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..4b6b70ea10ec445416345cff7ae90b536c38ce69 --- /dev/null +++ b/ja/560ms/decoder_joint.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dad926552d028aa609be27dde32ede7e7e0687ce48d057ee7bae6d472f15229e +size 14986 diff --git a/ja/560ms/decoder_joint.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/ja/560ms/decoder_joint.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..c3bb2e494d21dfd602e504fdfe76da274071d914 --- /dev/null +++ b/ja/560ms/decoder_joint.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bfc1c768dd0e0e61c0ab8806894ecc03902d2a5028e9c30f5d0a5e38d5139fd9 +size 18847864 diff --git a/ja/560ms/decoder_joint.mlpackage/Manifest.json b/ja/560ms/decoder_joint.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..8a2ce629ffcccdfce8e228506afd499877ee4507 --- /dev/null +++ b/ja/560ms/decoder_joint.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "4DAC4947-89EA-4167-8E25-F5740051D450": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + }, + "D23EE53F-5E4C-405E-B75A-E4553E8229D1": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + } + }, + "rootModelIdentifier": "D23EE53F-5E4C-405E-B75A-E4553E8229D1" +} diff --git a/ja/560ms/encoder.mlmodelc/analytics/coremldata.bin b/ja/560ms/encoder.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..f631009aa2b7c2a2f48501a3e27f90a0688e220e --- /dev/null +++ b/ja/560ms/encoder.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c7172b10d4d8cda6411e6356b4ba90d9d72d946859055983012fa4a9c9984deb +size 243 diff --git a/ja/560ms/encoder.mlmodelc/coremldata.bin b/ja/560ms/encoder.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..c7de42463ee4488658603df6c89c76810298287f --- /dev/null +++ b/ja/560ms/encoder.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ca2acb0ada101aeb6f17e151b3427b07c156a77b122fd68eb9cd98d73a035310 +size 632 diff --git a/ja/560ms/encoder.mlmodelc/model.mil b/ja/560ms/encoder.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..d6f0179d3761ce7d81034c775fc7550c16626d14 --- /dev/null +++ b/ja/560ms/encoder.mlmodelc/model.mil @@ -0,0 +1,4434 @@ +program(1.3) +[buildInfo = dict({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}})] +{ + func main(tensor cache_channel, tensor cache_len, tensor cache_time, tensor mel, tensor mel_length, tensor prompt_id) { + tensor value_3_perm_0 = const()[name = string("value_3_perm_0"), val = tensor([1, 0, 2, 3])]; + string cache_channel_to_fp16_dtype_0 = const()[name = string("cache_channel_to_fp16_dtype_0"), val = string("fp16")]; + tensor value_5_perm_0 = const()[name = string("value_5_perm_0"), val = tensor([1, 0, 2, 3])]; + string cache_time_to_fp16_dtype_0 = const()[name = string("cache_time_to_fp16_dtype_0"), val = string("fp16")]; + int32 var_59 = const()[name = string("op_59"), val = int32(-1)]; + int32 var_68 = const()[name = string("op_68"), val = int32(1)]; + tensor x_1_perm_0 = const()[name = string("x_1_perm_0"), val = tensor([0, 2, 1])]; + string mel_to_fp16_dtype_0 = const()[name = string("mel_to_fp16_dtype_0"), val = string("fp16")]; + tensor tensor_1_axes_0 = const()[name = string("tensor_1_axes_0"), val = tensor([1])]; + tensor mel_to_fp16 = cast(dtype = mel_to_fp16_dtype_0, x = mel)[name = string("cast_21")]; + tensor x_1_cast_fp16 = transpose(perm = x_1_perm_0, x = mel_to_fp16)[name = string("transpose_367")]; + tensor tensor_1_cast_fp16 = expand_dims(axes = tensor_1_axes_0, x = x_1_cast_fp16)[name = string("tensor_1_cast_fp16")]; + tensor expand_dims_0 = const()[name = string("expand_dims_0"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))]; + tensor var_137_axes_0 = const()[name = string("op_137_axes_0"), val = tensor([1])]; + tensor var_137 = expand_dims(axes = var_137_axes_0, x = mel_length)[name = string("op_137")]; + tensor time_mask_1 = less(x = expand_dims_0, y = var_137)[name = string("time_mask_1")]; + tensor var_139_axes_0 = const()[name = string("op_139_axes_0"), val = tensor([-1])]; + tensor var_139 = expand_dims(axes = var_139_axes_0, x = time_mask_1)[name = string("op_139")]; + tensor var_141_reps_0 = const()[name = string("op_141_reps_0"), val = tensor([1, 1, 128])]; + tensor var_141 = tile(reps = var_141_reps_0, x = var_139)[name = string("op_141")]; + tensor var_147_axes_0 = const()[name = string("op_147_axes_0"), val = tensor([1])]; + string mask_1_to_fp16_dtype_0 = const()[name = string("mask_1_to_fp16_dtype_0"), val = string("fp16")]; + tensor var_141_to_fp16 = cast(dtype = mask_1_to_fp16_dtype_0, x = var_141)[name = string("cast_20")]; + tensor var_147_cast_fp16 = expand_dims(axes = var_147_axes_0, x = var_141_to_fp16)[name = string("op_147_cast_fp16")]; + tensor input_1_cast_fp16 = mul(x = tensor_1_cast_fp16, y = var_147_cast_fp16)[name = string("input_1_cast_fp16")]; + tensor input_3_pad_0 = const()[name = string("input_3_pad_0"), val = tensor([0, 0, 0, 0, 2, 1, 2, 1])]; + string input_3_mode_0 = const()[name = string("input_3_mode_0"), val = string("constant")]; + fp16 const_9_to_fp16 = const()[name = string("const_9_to_fp16"), val = fp16(0x0p+0)]; + tensor input_3_cast_fp16 = pad(constant_val = const_9_to_fp16, mode = input_3_mode_0, pad = input_3_pad_0, x = input_1_cast_fp16)[name = string("input_3_cast_fp16")]; + string tensor_3_pad_type_0 = const()[name = string("tensor_3_pad_type_0"), val = string("valid")]; + tensor tensor_3_strides_0 = const()[name = string("tensor_3_strides_0"), val = tensor([2, 2])]; + tensor tensor_3_pad_0 = const()[name = string("tensor_3_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor tensor_3_dilations_0 = const()[name = string("tensor_3_dilations_0"), val = tensor([1, 1])]; + int32 tensor_3_groups_0 = const()[name = string("tensor_3_groups_0"), val = int32(1)]; + tensor encoder_pre_encode_conv_0_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(448))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2816))))[name = string("encoder_pre_encode_conv_0_weight_to_fp16_quantized")]; + tensor encoder_pre_encode_conv_0_bias_to_fp16 = const()[name = string("encoder_pre_encode_conv_0_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3392)))]; + tensor tensor_3_cast_fp16 = conv(bias = encoder_pre_encode_conv_0_bias_to_fp16, dilations = tensor_3_dilations_0, groups = tensor_3_groups_0, pad = tensor_3_pad_0, pad_type = tensor_3_pad_type_0, strides = tensor_3_strides_0, weight = encoder_pre_encode_conv_0_weight_to_fp16_quantized, x = input_3_cast_fp16)[name = string("tensor_3_cast_fp16")]; + string current_lengths_1_to_fp16_dtype_0 = const()[name = string("current_lengths_1_to_fp16_dtype_0"), val = string("fp16")]; + fp16 var_160_promoted_to_fp16 = const()[name = string("op_160_promoted_to_fp16"), val = fp16(0x1p+1)]; + tensor mel_length_to_fp16 = cast(dtype = current_lengths_1_to_fp16_dtype_0, x = mel_length)[name = string("cast_19")]; + tensor var_161_cast_fp16 = add(x = mel_length_to_fp16, y = var_160_promoted_to_fp16)[name = string("op_161_cast_fp16")]; + fp16 var_162_promoted_to_fp16 = const()[name = string("op_162_promoted_to_fp16"), val = fp16(0x1p+0)]; + tensor var_163_cast_fp16 = add(x = var_161_cast_fp16, y = var_162_promoted_to_fp16)[name = string("op_163_cast_fp16")]; + fp16 var_164_promoted_to_fp16 = const()[name = string("op_164_promoted_to_fp16"), val = fp16(0x1.8p+1)]; + tensor var_165_cast_fp16 = sub(x = var_163_cast_fp16, y = var_164_promoted_to_fp16)[name = string("op_165_cast_fp16")]; + fp16 var_56_promoted_to_fp16 = const()[name = string("op_56_promoted_to_fp16"), val = fp16(0x1p+1)]; + tensor floor_div_0_cast_fp16 = floor_div(x = var_165_cast_fp16, y = var_56_promoted_to_fp16)[name = string("floor_div_0_cast_fp16")]; + fp16 var_167_promoted_to_fp16 = const()[name = string("op_167_promoted_to_fp16"), val = fp16(0x1p+0)]; + tensor current_lengths_3_cast_fp16 = add(x = floor_div_0_cast_fp16, y = var_167_promoted_to_fp16)[name = string("current_lengths_3_cast_fp16")]; + string lengths_19_dtype_0 = const()[name = string("lengths_19_dtype_0"), val = string("int32")]; + tensor expand_dims_1 = const()[name = string("expand_dims_1"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3968)))]; + tensor var_176_axes_0 = const()[name = string("op_176_axes_0"), val = tensor([1])]; + tensor current_lengths_3_cast_fp16_to_int32 = cast(dtype = lengths_19_dtype_0, x = current_lengths_3_cast_fp16)[name = string("cast_18")]; + tensor var_176 = expand_dims(axes = var_176_axes_0, x = current_lengths_3_cast_fp16_to_int32)[name = string("op_176")]; + tensor time_mask_3 = less(x = expand_dims_1, y = var_176)[name = string("time_mask_3")]; + tensor var_178_axes_0 = const()[name = string("op_178_axes_0"), val = tensor([-1])]; + tensor var_178 = expand_dims(axes = var_178_axes_0, x = time_mask_3)[name = string("op_178")]; + tensor var_180_reps_0 = const()[name = string("op_180_reps_0"), val = tensor([1, 1, 65])]; + tensor var_180 = tile(reps = var_180_reps_0, x = var_178)[name = string("op_180")]; + tensor var_186_axes_0 = const()[name = string("op_186_axes_0"), val = tensor([1])]; + string mask_3_to_fp16_dtype_0 = const()[name = string("mask_3_to_fp16_dtype_0"), val = string("fp16")]; + tensor var_180_to_fp16 = cast(dtype = mask_3_to_fp16_dtype_0, x = var_180)[name = string("cast_17")]; + tensor var_186_cast_fp16 = expand_dims(axes = var_186_axes_0, x = var_180_to_fp16)[name = string("op_186_cast_fp16")]; + tensor expanded_mask_3_reps_0 = const()[name = string("expanded_mask_3_reps_0"), val = tensor([1, 256, 1, 1])]; + tensor expanded_mask_3_cast_fp16 = tile(reps = expanded_mask_3_reps_0, x = var_186_cast_fp16)[name = string("expanded_mask_3_cast_fp16")]; + tensor input_5_cast_fp16 = mul(x = tensor_3_cast_fp16, y = expanded_mask_3_cast_fp16)[name = string("input_5_cast_fp16")]; + tensor tensor_5_cast_fp16 = relu(x = input_5_cast_fp16)[name = string("tensor_5_cast_fp16")]; + tensor input_7_cast_fp16 = mul(x = tensor_5_cast_fp16, y = expanded_mask_3_cast_fp16)[name = string("input_7_cast_fp16")]; + tensor input_9_pad_0 = const()[name = string("input_9_pad_0"), val = tensor([0, 0, 0, 0, 2, 1, 2, 1])]; + string input_9_mode_0 = const()[name = string("input_9_mode_0"), val = string("constant")]; + fp16 const_23_to_fp16 = const()[name = string("const_23_to_fp16"), val = fp16(0x0p+0)]; + tensor input_9_cast_fp16 = pad(constant_val = const_23_to_fp16, mode = input_9_mode_0, pad = input_9_pad_0, x = input_7_cast_fp16)[name = string("input_9_cast_fp16")]; + string tensor_7_pad_type_0 = const()[name = string("tensor_7_pad_type_0"), val = string("valid")]; + tensor tensor_7_strides_0 = const()[name = string("tensor_7_strides_0"), val = tensor([2, 2])]; + int32 tensor_7_groups_0 = const()[name = string("tensor_7_groups_0"), val = int32(256)]; + tensor tensor_7_pad_0 = const()[name = string("tensor_7_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor tensor_7_dilations_0 = const()[name = string("tensor_7_dilations_0"), val = tensor([1, 1])]; + tensor encoder_pre_encode_conv_2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4224))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6592))))[name = string("encoder_pre_encode_conv_2_weight_to_fp16_quantized")]; + tensor encoder_pre_encode_conv_2_bias_to_fp16 = const()[name = string("encoder_pre_encode_conv_2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7168)))]; + tensor tensor_7_cast_fp16 = conv(bias = encoder_pre_encode_conv_2_bias_to_fp16, dilations = tensor_7_dilations_0, groups = tensor_7_groups_0, pad = tensor_7_pad_0, pad_type = tensor_7_pad_type_0, strides = tensor_7_strides_0, weight = encoder_pre_encode_conv_2_weight_to_fp16_quantized, x = input_9_cast_fp16)[name = string("tensor_7_cast_fp16")]; + fp16 var_208_promoted_to_fp16 = const()[name = string("op_208_promoted_to_fp16"), val = fp16(0x1p+1)]; + tensor var_209_cast_fp16 = add(x = current_lengths_3_cast_fp16, y = var_208_promoted_to_fp16)[name = string("op_209_cast_fp16")]; + fp16 var_210_promoted_to_fp16 = const()[name = string("op_210_promoted_to_fp16"), val = fp16(0x1p+0)]; + tensor var_211_cast_fp16 = add(x = var_209_cast_fp16, y = var_210_promoted_to_fp16)[name = string("op_211_cast_fp16")]; + fp16 var_212_promoted_to_fp16 = const()[name = string("op_212_promoted_to_fp16"), val = fp16(0x1.8p+1)]; + tensor var_213_cast_fp16 = sub(x = var_211_cast_fp16, y = var_212_promoted_to_fp16)[name = string("op_213_cast_fp16")]; + fp16 var_56_promoted_1_to_fp16 = const()[name = string("op_56_promoted_1_to_fp16"), val = fp16(0x1p+1)]; + tensor floor_div_1_cast_fp16 = floor_div(x = var_213_cast_fp16, y = var_56_promoted_1_to_fp16)[name = string("floor_div_1_cast_fp16")]; + fp16 var_215_promoted_to_fp16 = const()[name = string("op_215_promoted_to_fp16"), val = fp16(0x1p+0)]; + tensor current_lengths_5_cast_fp16 = add(x = floor_div_1_cast_fp16, y = var_215_promoted_to_fp16)[name = string("current_lengths_5_cast_fp16")]; + string lengths_21_dtype_0 = const()[name = string("lengths_21_dtype_0"), val = string("int32")]; + tensor expand_dims_2 = const()[name = string("expand_dims_2"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7744)))]; + tensor var_224_axes_0 = const()[name = string("op_224_axes_0"), val = tensor([1])]; + tensor current_lengths_5_cast_fp16_to_int32 = cast(dtype = lengths_21_dtype_0, x = current_lengths_5_cast_fp16)[name = string("cast_16")]; + tensor var_224 = expand_dims(axes = var_224_axes_0, x = current_lengths_5_cast_fp16_to_int32)[name = string("op_224")]; + tensor time_mask_5 = less(x = expand_dims_2, y = var_224)[name = string("time_mask_5")]; + tensor var_226_axes_0 = const()[name = string("op_226_axes_0"), val = tensor([-1])]; + tensor var_226 = expand_dims(axes = var_226_axes_0, x = time_mask_5)[name = string("op_226")]; + tensor var_228_reps_0 = const()[name = string("op_228_reps_0"), val = tensor([1, 1, 33])]; + tensor var_228 = tile(reps = var_228_reps_0, x = var_226)[name = string("op_228")]; + tensor var_234_axes_0 = const()[name = string("op_234_axes_0"), val = tensor([1])]; + string mask_5_to_fp16_dtype_0 = const()[name = string("mask_5_to_fp16_dtype_0"), val = string("fp16")]; + tensor var_228_to_fp16 = cast(dtype = mask_5_to_fp16_dtype_0, x = var_228)[name = string("cast_15")]; + tensor var_234_cast_fp16 = expand_dims(axes = var_234_axes_0, x = var_228_to_fp16)[name = string("op_234_cast_fp16")]; + tensor expanded_mask_7_reps_0 = const()[name = string("expanded_mask_7_reps_0"), val = tensor([1, 256, 1, 1])]; + tensor expanded_mask_7_cast_fp16 = tile(reps = expanded_mask_7_reps_0, x = var_234_cast_fp16)[name = string("expanded_mask_7_cast_fp16")]; + tensor input_11_cast_fp16 = mul(x = tensor_7_cast_fp16, y = expanded_mask_7_cast_fp16)[name = string("input_11_cast_fp16")]; + string tensor_9_pad_type_0 = const()[name = string("tensor_9_pad_type_0"), val = string("valid")]; + tensor tensor_9_strides_0 = const()[name = string("tensor_9_strides_0"), val = tensor([1, 1])]; + tensor tensor_9_pad_0 = const()[name = string("tensor_9_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor tensor_9_dilations_0 = const()[name = string("tensor_9_dilations_0"), val = tensor([1, 1])]; + int32 tensor_9_groups_0 = const()[name = string("tensor_9_groups_0"), val = int32(1)]; + tensor encoder_pre_encode_conv_3_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7936))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(73536))))[name = string("encoder_pre_encode_conv_3_weight_to_fp16_quantized")]; + tensor encoder_pre_encode_conv_3_bias_to_fp16 = const()[name = string("encoder_pre_encode_conv_3_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(74112)))]; + tensor tensor_9_cast_fp16 = conv(bias = encoder_pre_encode_conv_3_bias_to_fp16, dilations = tensor_9_dilations_0, groups = tensor_9_groups_0, pad = tensor_9_pad_0, pad_type = tensor_9_pad_type_0, strides = tensor_9_strides_0, weight = encoder_pre_encode_conv_3_weight_to_fp16_quantized, x = input_11_cast_fp16)[name = string("tensor_9_cast_fp16")]; + tensor input_13_cast_fp16 = mul(x = tensor_9_cast_fp16, y = expanded_mask_7_cast_fp16)[name = string("input_13_cast_fp16")]; + tensor tensor_11_cast_fp16 = relu(x = input_13_cast_fp16)[name = string("tensor_11_cast_fp16")]; + tensor input_15_cast_fp16 = mul(x = tensor_11_cast_fp16, y = expanded_mask_7_cast_fp16)[name = string("input_15_cast_fp16")]; + tensor input_17_pad_0 = const()[name = string("input_17_pad_0"), val = tensor([0, 0, 0, 0, 2, 1, 2, 1])]; + string input_17_mode_0 = const()[name = string("input_17_mode_0"), val = string("constant")]; + fp16 const_41_to_fp16 = const()[name = string("const_41_to_fp16"), val = fp16(0x0p+0)]; + tensor input_17_cast_fp16 = pad(constant_val = const_41_to_fp16, mode = input_17_mode_0, pad = input_17_pad_0, x = input_15_cast_fp16)[name = string("input_17_cast_fp16")]; + string tensor_13_pad_type_0 = const()[name = string("tensor_13_pad_type_0"), val = string("valid")]; + tensor tensor_13_strides_0 = const()[name = string("tensor_13_strides_0"), val = tensor([2, 2])]; + int32 tensor_13_groups_0 = const()[name = string("tensor_13_groups_0"), val = int32(256)]; + tensor tensor_13_pad_0 = const()[name = string("tensor_13_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor tensor_13_dilations_0 = const()[name = string("tensor_13_dilations_0"), val = tensor([1, 1])]; + tensor encoder_pre_encode_conv_5_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(74688))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77056))))[name = string("encoder_pre_encode_conv_5_weight_to_fp16_quantized")]; + tensor encoder_pre_encode_conv_5_bias_to_fp16 = const()[name = string("encoder_pre_encode_conv_5_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77632)))]; + tensor tensor_13_cast_fp16 = conv(bias = encoder_pre_encode_conv_5_bias_to_fp16, dilations = tensor_13_dilations_0, groups = tensor_13_groups_0, pad = tensor_13_pad_0, pad_type = tensor_13_pad_type_0, strides = tensor_13_strides_0, weight = encoder_pre_encode_conv_5_weight_to_fp16_quantized, x = input_17_cast_fp16)[name = string("tensor_13_cast_fp16")]; + fp16 var_271_promoted_to_fp16 = const()[name = string("op_271_promoted_to_fp16"), val = fp16(0x1p+1)]; + tensor var_272_cast_fp16 = add(x = current_lengths_5_cast_fp16, y = var_271_promoted_to_fp16)[name = string("op_272_cast_fp16")]; + fp16 var_273_promoted_to_fp16 = const()[name = string("op_273_promoted_to_fp16"), val = fp16(0x1p+0)]; + tensor var_274_cast_fp16 = add(x = var_272_cast_fp16, y = var_273_promoted_to_fp16)[name = string("op_274_cast_fp16")]; + fp16 var_275_promoted_to_fp16 = const()[name = string("op_275_promoted_to_fp16"), val = fp16(0x1.8p+1)]; + tensor var_276_cast_fp16 = sub(x = var_274_cast_fp16, y = var_275_promoted_to_fp16)[name = string("op_276_cast_fp16")]; + fp16 var_56_promoted_2_to_fp16 = const()[name = string("op_56_promoted_2_to_fp16"), val = fp16(0x1p+1)]; + tensor floor_div_2_cast_fp16 = floor_div(x = var_276_cast_fp16, y = var_56_promoted_2_to_fp16)[name = string("floor_div_2_cast_fp16")]; + fp16 var_278_promoted_to_fp16 = const()[name = string("op_278_promoted_to_fp16"), val = fp16(0x1p+0)]; + tensor current_lengths_cast_fp16 = add(x = floor_div_2_cast_fp16, y = var_278_promoted_to_fp16)[name = string("current_lengths_cast_fp16")]; + string lengths_dtype_0 = const()[name = string("lengths_dtype_0"), val = string("int32")]; + tensor expand_dims_3 = const()[name = string("expand_dims_3"), val = tensor([[0, 1, 2, 3, 4, 5, 6, 7, 8]])]; + tensor var_287_axes_0 = const()[name = string("op_287_axes_0"), val = tensor([1])]; + tensor current_lengths_cast_fp16_to_int32 = cast(dtype = lengths_dtype_0, x = current_lengths_cast_fp16)[name = string("cast_14")]; + tensor var_287 = expand_dims(axes = var_287_axes_0, x = current_lengths_cast_fp16_to_int32)[name = string("op_287")]; + tensor time_mask = less(x = expand_dims_3, y = var_287)[name = string("time_mask")]; + tensor var_289_axes_0 = const()[name = string("op_289_axes_0"), val = tensor([-1])]; + tensor var_289 = expand_dims(axes = var_289_axes_0, x = time_mask)[name = string("op_289")]; + tensor var_291_reps_0 = const()[name = string("op_291_reps_0"), val = tensor([1, 1, 17])]; + tensor var_291 = tile(reps = var_291_reps_0, x = var_289)[name = string("op_291")]; + tensor var_297_axes_0 = const()[name = string("op_297_axes_0"), val = tensor([1])]; + string mask_7_to_fp16_dtype_0 = const()[name = string("mask_7_to_fp16_dtype_0"), val = string("fp16")]; + tensor var_291_to_fp16 = cast(dtype = mask_7_to_fp16_dtype_0, x = var_291)[name = string("cast_13")]; + tensor var_297_cast_fp16 = expand_dims(axes = var_297_axes_0, x = var_291_to_fp16)[name = string("op_297_cast_fp16")]; + tensor expanded_mask_13_reps_0 = const()[name = string("expanded_mask_13_reps_0"), val = tensor([1, 256, 1, 1])]; + tensor expanded_mask_13_cast_fp16 = tile(reps = expanded_mask_13_reps_0, x = var_297_cast_fp16)[name = string("expanded_mask_13_cast_fp16")]; + tensor input_19_cast_fp16 = mul(x = tensor_13_cast_fp16, y = expanded_mask_13_cast_fp16)[name = string("input_19_cast_fp16")]; + string tensor_15_pad_type_0 = const()[name = string("tensor_15_pad_type_0"), val = string("valid")]; + tensor tensor_15_strides_0 = const()[name = string("tensor_15_strides_0"), val = tensor([1, 1])]; + tensor tensor_15_pad_0 = const()[name = string("tensor_15_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor tensor_15_dilations_0 = const()[name = string("tensor_15_dilations_0"), val = tensor([1, 1])]; + int32 tensor_15_groups_0 = const()[name = string("tensor_15_groups_0"), val = int32(1)]; + tensor encoder_pre_encode_conv_6_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78208))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(143808))))[name = string("encoder_pre_encode_conv_6_weight_to_fp16_quantized")]; + tensor encoder_pre_encode_conv_6_bias_to_fp16 = const()[name = string("encoder_pre_encode_conv_6_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144384)))]; + tensor tensor_15_cast_fp16 = conv(bias = encoder_pre_encode_conv_6_bias_to_fp16, dilations = tensor_15_dilations_0, groups = tensor_15_groups_0, pad = tensor_15_pad_0, pad_type = tensor_15_pad_type_0, strides = tensor_15_strides_0, weight = encoder_pre_encode_conv_6_weight_to_fp16_quantized, x = input_19_cast_fp16)[name = string("tensor_15_cast_fp16")]; + tensor input_21_cast_fp16 = mul(x = tensor_15_cast_fp16, y = expanded_mask_13_cast_fp16)[name = string("input_21_cast_fp16")]; + tensor tensor_cast_fp16 = relu(x = input_21_cast_fp16)[name = string("tensor_cast_fp16")]; + tensor x_3_cast_fp16 = mul(x = tensor_cast_fp16, y = expanded_mask_13_cast_fp16)[name = string("x_3_cast_fp16")]; + tensor var_331_perm_0 = const()[name = string("op_331_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_332 = const()[name = string("op_332"), val = tensor([1, 9, -1])]; + tensor var_331_cast_fp16 = transpose(perm = var_331_perm_0, x = x_3_cast_fp16)[name = string("transpose_366")]; + tensor input_23_cast_fp16 = reshape(shape = var_332, x = var_331_cast_fp16)[name = string("input_23_cast_fp16")]; + tensor encoder_pre_encode_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144960))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4601472))))[name = string("encoder_pre_encode_out_weight_to_fp16_quantized")]; + tensor encoder_pre_encode_out_bias_to_fp16 = const()[name = string("encoder_pre_encode_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4603584)))]; + tensor linear_0_cast_fp16 = linear(bias = encoder_pre_encode_out_bias_to_fp16, weight = encoder_pre_encode_out_weight_to_fp16_quantized, x = input_23_cast_fp16)[name = string("linear_0_cast_fp16")]; + tensor var_342_begin_0 = const()[name = string("op_342_begin_0"), val = tensor([0, 2, 0])]; + tensor var_342_end_0 = const()[name = string("op_342_end_0"), val = tensor([1, 9, 1024])]; + tensor var_342_end_mask_0 = const()[name = string("op_342_end_mask_0"), val = tensor([true, true, true])]; + tensor var_342_cast_fp16 = slice_by_index(begin = var_342_begin_0, end = var_342_end_0, end_mask = var_342_end_mask_0, x = linear_0_cast_fp16)[name = string("op_342_cast_fp16")]; + int32 var_344 = const()[name = string("op_344"), val = int32(2)]; + tensor var_345 = sub(x = current_lengths_cast_fp16_to_int32, y = var_344)[name = string("op_345")]; + string var_345_promoted_to_fp16_dtype_0 = const()[name = string("op_345_promoted_to_fp16_dtype_0"), val = string("fp16")]; + fp16 var_62_promoted_to_fp16 = const()[name = string("op_62_promoted_to_fp16"), val = fp16(0x0p+0)]; + fp16 const_61_to_fp16 = const()[name = string("const_61_to_fp16"), val = fp16(inf)]; + tensor var_345_to_fp16 = cast(dtype = var_345_promoted_to_fp16_dtype_0, x = var_345)[name = string("cast_12")]; + tensor clip_0_cast_fp16 = clip(alpha = var_62_promoted_to_fp16, beta = const_61_to_fp16, x = var_345_to_fp16)[name = string("clip_0_cast_fp16")]; + tensor max_audio_length_1 = const()[name = string("max_audio_length_1"), val = tensor([7])]; + fp16 var_361_promoted_to_fp16 = const()[name = string("op_361_promoted_to_fp16"), val = fp16(0x1.5p+5)]; + tensor padding_length_cast_fp16 = add(x = clip_0_cast_fp16, y = var_361_promoted_to_fp16)[name = string("padding_length_cast_fp16")]; + int32 const_63 = const()[name = string("const_63"), val = int32(-1)]; + tensor var_363 = mul(x = cache_len, y = const_63)[name = string("op_363")]; + int32 var_364 = const()[name = string("op_364"), val = int32(42)]; + tensor offset = add(x = var_363, y = var_364)[name = string("offset")]; + tensor var_404_axes_0 = const()[name = string("op_404_axes_0"), val = tensor([-1])]; + tensor var_404_cast_fp16 = expand_dims(axes = var_404_axes_0, x = padding_length_cast_fp16)[name = string("op_404_cast_fp16")]; + tensor var_403_promoted_to_fp16 = const()[name = string("op_403_promoted_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4605696)))]; + tensor pad_mask_1_cast_fp16 = less(x = var_403_promoted_to_fp16, y = var_404_cast_fp16)[name = string("pad_mask_1_cast_fp16")]; + tensor expand_dims_5 = const()[name = string("expand_dims_5"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4605888)))]; + tensor var_410_axes_0 = const()[name = string("op_410_axes_0"), val = tensor([-1])]; + tensor var_410 = expand_dims(axes = var_410_axes_0, x = offset)[name = string("op_410")]; + tensor pad_mask_off = greater_equal(x = expand_dims_5, y = var_410)[name = string("pad_mask_off")]; + tensor pad_mask_3 = logical_and(x = pad_mask_off, y = pad_mask_1_cast_fp16)[name = string("pad_mask_3")]; + tensor var_413_axes_0 = const()[name = string("op_413_axes_0"), val = tensor([1])]; + tensor var_413 = expand_dims(axes = var_413_axes_0, x = pad_mask_3)[name = string("op_413")]; + tensor var_414 = const()[name = string("op_414"), val = tensor([1, 49, 1])]; + tensor pad_mask_for_att_mask_1 = tile(reps = var_414, x = var_413)[name = string("pad_mask_for_att_mask_1")]; + tensor var_416_perm_0 = const()[name = string("op_416_perm_0"), val = tensor([0, 2, 1])]; + tensor var_416 = transpose(perm = var_416_perm_0, x = pad_mask_for_att_mask_1)[name = string("transpose_365")]; + tensor pad_mask_for_att_mask = logical_and(x = pad_mask_for_att_mask_1, y = var_416)[name = string("pad_mask_for_att_mask")]; + tensor const_71 = const()[name = string("const_71"), val = tensor([[[true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true]]])]; + tensor att_mask_9 = logical_and(x = pad_mask_for_att_mask, y = const_71)[name = string("att_mask_9")]; + tensor att_mask = logical_not(x = att_mask_9)[name = string("att_mask")]; + tensor pad_mask_5 = logical_not(x = pad_mask_3)[name = string("pad_mask_5")]; + tensor pad_mask_begin_0 = const()[name = string("pad_mask_begin_0"), val = tensor([0, 42])]; + tensor pad_mask_end_0 = const()[name = string("pad_mask_end_0"), val = tensor([1, 49])]; + tensor pad_mask_end_mask_0 = const()[name = string("pad_mask_end_mask_0"), val = tensor([true, true])]; + tensor pad_mask = slice_by_index(begin = pad_mask_begin_0, end = pad_mask_end_0, end_mask = pad_mask_end_mask_0, x = pad_mask_5)[name = string("pad_mask")]; + tensor mask_9_begin_0 = const()[name = string("mask_9_begin_0"), val = tensor([0, 42, 0])]; + tensor mask_9_end_0 = const()[name = string("mask_9_end_0"), val = tensor([1, 49, 49])]; + tensor mask_9_end_mask_0 = const()[name = string("mask_9_end_mask_0"), val = tensor([true, true, true])]; + tensor mask_9 = slice_by_index(begin = mask_9_begin_0, end = mask_9_end_0, end_mask = mask_9_end_mask_0, x = att_mask)[name = string("mask_9")]; + tensor cache_1_begin_0 = const()[name = string("cache_1_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor cache_1_end_0 = const()[name = string("cache_1_end_0"), val = tensor([1, 1, 42, 1024])]; + tensor cache_1_end_mask_0 = const()[name = string("cache_1_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_1_squeeze_mask_0 = const()[name = string("cache_1_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_channel_to_fp16 = cast(dtype = cache_channel_to_fp16_dtype_0, x = cache_channel)[name = string("cast_11")]; + tensor value_3_cast_fp16 = transpose(perm = value_3_perm_0, x = cache_channel_to_fp16)[name = string("transpose_364")]; + tensor cache_1_cast_fp16 = slice_by_index(begin = cache_1_begin_0, end = cache_1_end_0, end_mask = cache_1_end_mask_0, squeeze_mask = cache_1_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_1_cast_fp16")]; + tensor cache_3_begin_0 = const()[name = string("cache_3_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor cache_3_end_0 = const()[name = string("cache_3_end_0"), val = tensor([1, 1, 1024, 8])]; + tensor cache_3_end_mask_0 = const()[name = string("cache_3_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_3_squeeze_mask_0 = const()[name = string("cache_3_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_time_to_fp16 = cast(dtype = cache_time_to_fp16_dtype_0, x = cache_time)[name = string("cast_10")]; + tensor value_5_cast_fp16 = transpose(perm = value_5_perm_0, x = cache_time_to_fp16)[name = string("transpose_363")]; + tensor cache_3_cast_fp16 = slice_by_index(begin = cache_3_begin_0, end = cache_3_end_0, end_mask = cache_3_end_mask_0, squeeze_mask = cache_3_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_3_cast_fp16")]; + tensor input_27_axes_0 = const()[name = string("input_27_axes_0"), val = tensor([-1])]; + tensor encoder_layers_0_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_0_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4606208)))]; + tensor encoder_layers_0_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_0_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4608320)))]; + fp16 var_42_to_fp16 = const()[name = string("op_42_to_fp16"), val = fp16(0x1.5p-17)]; + tensor input_27_cast_fp16 = layer_norm(axes = input_27_axes_0, beta = encoder_layers_0_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_0_norm_feed_forward1_weight_to_fp16, x = var_342_cast_fp16)[name = string("input_27_cast_fp16")]; + tensor encoder_layers_0_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4610432))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8804800))))[name = string("encoder_layers_0_feed_forward1_linear1_weight_to_fp16_quantized")]; + tensor encoder_layers_0_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_0_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8813056)))]; + tensor linear_1_cast_fp16 = linear(bias = encoder_layers_0_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_0_feed_forward1_linear1_weight_to_fp16_quantized, x = input_27_cast_fp16)[name = string("linear_1_cast_fp16")]; + tensor input_31_cast_fp16 = silu(x = linear_1_cast_fp16)[name = string("input_31_cast_fp16")]; + tensor encoder_layers_0_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8821312))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13015680))))[name = string("encoder_layers_0_feed_forward1_linear2_weight_to_fp16_quantized")]; + tensor encoder_layers_0_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_0_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13017792)))]; + tensor linear_2_cast_fp16 = linear(bias = encoder_layers_0_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_0_feed_forward1_linear2_weight_to_fp16_quantized, x = input_31_cast_fp16)[name = string("linear_2_cast_fp16")]; + fp16 var_455_to_fp16 = const()[name = string("op_455_to_fp16"), val = fp16(0x1p-1)]; + tensor var_456_cast_fp16 = mul(x = linear_2_cast_fp16, y = var_455_to_fp16)[name = string("op_456_cast_fp16")]; + tensor input_37_cast_fp16 = add(x = var_342_cast_fp16, y = var_456_cast_fp16)[name = string("input_37_cast_fp16")]; + tensor key_1_axes_0 = const()[name = string("key_1_axes_0"), val = tensor([-1])]; + tensor encoder_layers_0_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_0_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13019904)))]; + tensor encoder_layers_0_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_0_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13022016)))]; + tensor key_1_cast_fp16 = layer_norm(axes = key_1_axes_0, beta = encoder_layers_0_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_0_norm_self_att_weight_to_fp16, x = input_37_cast_fp16)[name = string("key_1_cast_fp16")]; + bool input_39_interleave_0 = const()[name = string("input_39_interleave_0"), val = bool(false)]; + tensor input_39_cast_fp16 = concat(axis = var_68, interleave = input_39_interleave_0, values = (cache_1_cast_fp16, key_1_cast_fp16))[name = string("input_39_cast_fp16")]; + tensor var_478_begin_0 = const()[name = string("op_478_begin_0"), val = tensor([0, 7, 0])]; + tensor var_478_end_0 = const()[name = string("op_478_end_0"), val = tensor([1, 42, 1024])]; + tensor var_478_end_mask_0 = const()[name = string("op_478_end_mask_0"), val = tensor([true, true, true])]; + tensor var_478_cast_fp16 = slice_by_index(begin = var_478_begin_0, end = var_478_end_0, end_mask = var_478_end_mask_0, x = cache_1_cast_fp16)[name = string("op_478_cast_fp16")]; + bool var_484_interleave_0 = const()[name = string("op_484_interleave_0"), val = bool(false)]; + tensor var_484_cast_fp16 = concat(axis = var_68, interleave = var_484_interleave_0, values = (var_478_cast_fp16, key_1_cast_fp16))[name = string("op_484_cast_fp16")]; + tensor encoder_layers_0_self_attn_linear_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13024128))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14072768))))[name = string("encoder_layers_0_self_attn_linear_q_weight_to_fp16_quantized")]; + tensor encoder_layers_0_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_0_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14074880)))]; + tensor linear_3_cast_fp16 = linear(bias = encoder_layers_0_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_0_self_attn_linear_q_weight_to_fp16_quantized, x = key_1_cast_fp16)[name = string("linear_3_cast_fp16")]; + tensor var_489 = const()[name = string("op_489"), val = tensor([1, -1, 8, 128])]; + tensor q_1_cast_fp16 = reshape(shape = var_489, x = linear_3_cast_fp16)[name = string("q_1_cast_fp16")]; + tensor encoder_layers_0_self_attn_linear_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14076992))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15125632))))[name = string("encoder_layers_0_self_attn_linear_k_weight_to_fp16_quantized")]; + tensor encoder_layers_0_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_0_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15127744)))]; + tensor linear_4_cast_fp16 = linear(bias = encoder_layers_0_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_0_self_attn_linear_k_weight_to_fp16_quantized, x = input_39_cast_fp16)[name = string("linear_4_cast_fp16")]; + tensor var_494 = const()[name = string("op_494"), val = tensor([1, -1, 8, 128])]; + tensor k_1_cast_fp16 = reshape(shape = var_494, x = linear_4_cast_fp16)[name = string("k_1_cast_fp16")]; + tensor encoder_layers_0_self_attn_linear_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15129856))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16178496))))[name = string("encoder_layers_0_self_attn_linear_v_weight_to_fp16_quantized")]; + tensor encoder_layers_0_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_0_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16180608)))]; + tensor linear_5_cast_fp16 = linear(bias = encoder_layers_0_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_0_self_attn_linear_v_weight_to_fp16_quantized, x = input_39_cast_fp16)[name = string("linear_5_cast_fp16")]; + tensor var_499 = const()[name = string("op_499"), val = tensor([1, -1, 8, 128])]; + tensor v_1_cast_fp16 = reshape(shape = var_499, x = linear_5_cast_fp16)[name = string("v_1_cast_fp16")]; + tensor value_9_perm_0 = const()[name = string("value_9_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_0_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_0_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16182720)))]; + tensor var_512_cast_fp16 = add(x = q_1_cast_fp16, y = encoder_layers_0_self_attn_pos_bias_u_to_fp16)[name = string("op_512_cast_fp16")]; + tensor encoder_layers_0_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_0_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16184832)))]; + tensor var_514_cast_fp16 = add(x = q_1_cast_fp16, y = encoder_layers_0_self_attn_pos_bias_v_to_fp16)[name = string("op_514_cast_fp16")]; + tensor q_with_bias_v_1_perm_0 = const()[name = string("q_with_bias_v_1_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_7_transpose_x_0 = const()[name = string("x_7_transpose_x_0"), val = bool(false)]; + bool x_7_transpose_y_0 = const()[name = string("x_7_transpose_y_0"), val = bool(false)]; + tensor op_516_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16186944))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16286336))))[name = string("op_516_to_fp16_quantized")]; + tensor q_with_bias_v_1_cast_fp16 = transpose(perm = q_with_bias_v_1_perm_0, x = var_514_cast_fp16)[name = string("transpose_362")]; + tensor x_7_cast_fp16 = matmul(transpose_x = x_7_transpose_x_0, transpose_y = x_7_transpose_y_0, x = q_with_bias_v_1_cast_fp16, y = op_516_to_fp16_quantized)[name = string("x_7_cast_fp16")]; + tensor x_9_pad_0 = const()[name = string("x_9_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_9_mode_0 = const()[name = string("x_9_mode_0"), val = string("constant")]; + fp16 const_79_to_fp16 = const()[name = string("const_79_to_fp16"), val = fp16(0x0p+0)]; + tensor x_9_cast_fp16 = pad(constant_val = const_79_to_fp16, mode = x_9_mode_0, pad = x_9_pad_0, x = x_7_cast_fp16)[name = string("x_9_cast_fp16")]; + tensor var_524 = const()[name = string("op_524"), val = tensor([1, 8, -1, 7])]; + tensor x_11_cast_fp16 = reshape(shape = var_524, x = x_9_cast_fp16)[name = string("x_11_cast_fp16")]; + tensor var_528_begin_0 = const()[name = string("op_528_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_528_end_0 = const()[name = string("op_528_end_0"), val = tensor([1, 8, 98, 7])]; + tensor var_528_end_mask_0 = const()[name = string("op_528_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_528_cast_fp16 = slice_by_index(begin = var_528_begin_0, end = var_528_end_0, end_mask = var_528_end_mask_0, x = x_11_cast_fp16)[name = string("op_528_cast_fp16")]; + tensor var_529 = const()[name = string("op_529"), val = tensor([1, 8, 7, 97])]; + tensor matrix_bd_1_cast_fp16 = reshape(shape = var_529, x = var_528_cast_fp16)[name = string("matrix_bd_1_cast_fp16")]; + bool matrix_ac_1_transpose_x_0 = const()[name = string("matrix_ac_1_transpose_x_0"), val = bool(false)]; + bool matrix_ac_1_transpose_y_0 = const()[name = string("matrix_ac_1_transpose_y_0"), val = bool(false)]; + tensor transpose_96_perm_0 = const()[name = string("transpose_96_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_97_perm_0 = const()[name = string("transpose_97_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_97 = transpose(perm = transpose_97_perm_0, x = k_1_cast_fp16)[name = string("transpose_360")]; + tensor transpose_96 = transpose(perm = transpose_96_perm_0, x = var_512_cast_fp16)[name = string("transpose_361")]; + tensor matrix_ac_1_cast_fp16 = matmul(transpose_x = matrix_ac_1_transpose_x_0, transpose_y = matrix_ac_1_transpose_y_0, x = transpose_96, y = transpose_97)[name = string("matrix_ac_1_cast_fp16")]; + tensor matrix_bd_3_begin_0 = const()[name = string("matrix_bd_3_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_3_end_0 = const()[name = string("matrix_bd_3_end_0"), val = tensor([1, 8, 7, 49])]; + tensor matrix_bd_3_end_mask_0 = const()[name = string("matrix_bd_3_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_3_cast_fp16 = slice_by_index(begin = matrix_bd_3_begin_0, end = matrix_bd_3_end_0, end_mask = matrix_bd_3_end_mask_0, x = matrix_bd_1_cast_fp16)[name = string("matrix_bd_3_cast_fp16")]; + tensor var_538_cast_fp16 = add(x = matrix_ac_1_cast_fp16, y = matrix_bd_3_cast_fp16)[name = string("op_538_cast_fp16")]; + fp16 _inversed_scores_1_y_0_to_fp16 = const()[name = string("_inversed_scores_1_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_1_cast_fp16 = mul(x = var_538_cast_fp16, y = _inversed_scores_1_y_0_to_fp16)[name = string("_inversed_scores_1_cast_fp16")]; + tensor mask_11_axes_0 = const()[name = string("mask_11_axes_0"), val = tensor([1])]; + tensor mask_11 = expand_dims(axes = mask_11_axes_0, x = mask_9)[name = string("mask_11")]; + fp16 var_45_to_fp16 = const()[name = string("op_45_to_fp16"), val = fp16(-0x1.388p+13)]; + tensor scores_3_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_1_cast_fp16, cond = mask_11)[name = string("scores_3_cast_fp16")]; + tensor var_544_cast_fp16 = softmax(axis = var_59, x = scores_3_cast_fp16)[name = string("op_544_cast_fp16")]; + fp16 var_44_to_fp16 = const()[name = string("op_44_to_fp16"), val = fp16(0x0p+0)]; + tensor input_41_cast_fp16 = select(a = var_44_to_fp16, b = var_544_cast_fp16, cond = mask_11)[name = string("input_41_cast_fp16")]; + bool x_13_transpose_x_0 = const()[name = string("x_13_transpose_x_0"), val = bool(false)]; + bool x_13_transpose_y_0 = const()[name = string("x_13_transpose_y_0"), val = bool(false)]; + tensor value_9_cast_fp16 = transpose(perm = value_9_perm_0, x = v_1_cast_fp16)[name = string("transpose_359")]; + tensor x_13_cast_fp16 = matmul(transpose_x = x_13_transpose_x_0, transpose_y = x_13_transpose_y_0, x = input_41_cast_fp16, y = value_9_cast_fp16)[name = string("x_13_cast_fp16")]; + tensor var_548_perm_0 = const()[name = string("op_548_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_549 = const()[name = string("op_549"), val = tensor([1, -1, 1024])]; + tensor var_548_cast_fp16 = transpose(perm = var_548_perm_0, x = x_13_cast_fp16)[name = string("transpose_358")]; + tensor input_43_cast_fp16 = reshape(shape = var_549, x = var_548_cast_fp16)[name = string("input_43_cast_fp16")]; + tensor encoder_layers_0_self_attn_linear_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16286656))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17335296))))[name = string("encoder_layers_0_self_attn_linear_out_weight_to_fp16_quantized")]; + tensor encoder_layers_0_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_0_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17337408)))]; + tensor linear_7_cast_fp16 = linear(bias = encoder_layers_0_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_0_self_attn_linear_out_weight_to_fp16_quantized, x = input_43_cast_fp16)[name = string("linear_7_cast_fp16")]; + tensor input_47_cast_fp16 = add(x = input_37_cast_fp16, y = linear_7_cast_fp16)[name = string("input_47_cast_fp16")]; + tensor x_17_axes_0 = const()[name = string("x_17_axes_0"), val = tensor([-1])]; + tensor encoder_layers_0_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_0_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17339520)))]; + tensor encoder_layers_0_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_0_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17341632)))]; + tensor x_17_cast_fp16 = layer_norm(axes = x_17_axes_0, beta = encoder_layers_0_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_0_norm_conv_weight_to_fp16, x = input_47_cast_fp16)[name = string("x_17_cast_fp16")]; + tensor input_49_perm_0 = const()[name = string("input_49_perm_0"), val = tensor([0, 2, 1])]; + string input_51_pad_type_0 = const()[name = string("input_51_pad_type_0"), val = string("valid")]; + tensor input_51_strides_0 = const()[name = string("input_51_strides_0"), val = tensor([1])]; + tensor input_51_pad_0 = const()[name = string("input_51_pad_0"), val = tensor([0, 0])]; + tensor input_51_dilations_0 = const()[name = string("input_51_dilations_0"), val = tensor([1])]; + int32 input_51_groups_0 = const()[name = string("input_51_groups_0"), val = int32(1)]; + tensor encoder_layers_0_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17343744))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19440960))))[name = string("encoder_layers_0_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_49_cast_fp16 = transpose(perm = input_49_perm_0, x = x_17_cast_fp16)[name = string("transpose_357")]; + tensor input_51_cast_fp16 = conv(dilations = input_51_dilations_0, groups = input_51_groups_0, pad = input_51_pad_0, pad_type = input_51_pad_type_0, strides = input_51_strides_0, weight = encoder_layers_0_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_49_cast_fp16)[name = string("input_51_cast_fp16")]; + int32 x_19_split_num_splits_0 = const()[name = string("x_19_split_num_splits_0"), val = int32(2)]; + int32 x_19_split_axis_0 = const()[name = string("x_19_split_axis_0"), val = int32(1)]; + tensor x_19_split_cast_fp16_0, tensor x_19_split_cast_fp16_1 = split(axis = x_19_split_axis_0, num_splits = x_19_split_num_splits_0, x = input_51_cast_fp16)[name = string("x_19_split_cast_fp16")]; + tensor x_19_split_1_sigmoid_cast_fp16 = sigmoid(x = x_19_split_cast_fp16_1)[name = string("x_19_split_1_sigmoid_cast_fp16")]; + tensor x_19_cast_fp16 = mul(x = x_19_split_cast_fp16_0, y = x_19_split_1_sigmoid_cast_fp16)[name = string("x_19_cast_fp16")]; + tensor var_575_axes_0 = const()[name = string("op_575_axes_0"), val = tensor([1])]; + tensor var_575 = expand_dims(axes = var_575_axes_0, x = pad_mask)[name = string("op_575")]; + tensor input_53_cast_fp16 = select(a = var_44_to_fp16, b = x_19_cast_fp16, cond = var_575)[name = string("input_53_cast_fp16")]; + bool new_x_3_interleave_0 = const()[name = string("new_x_3_interleave_0"), val = bool(false)]; + tensor new_x_3_cast_fp16 = concat(axis = var_59, interleave = new_x_3_interleave_0, values = (cache_3_cast_fp16, input_53_cast_fp16))[name = string("new_x_3_cast_fp16")]; + tensor var_588_begin_0 = const()[name = string("op_588_begin_0"), val = tensor([0, 0, 7])]; + tensor var_588_end_0 = const()[name = string("op_588_end_0"), val = tensor([1, 1024, 15])]; + tensor var_588_end_mask_0 = const()[name = string("op_588_end_mask_0"), val = tensor([true, true, true])]; + tensor var_588_cast_fp16 = slice_by_index(begin = var_588_begin_0, end = var_588_end_0, end_mask = var_588_end_mask_0, x = new_x_3_cast_fp16)[name = string("op_588_cast_fp16")]; + string x_21_pad_type_0 = const()[name = string("x_21_pad_type_0"), val = string("valid")]; + int32 x_21_groups_0 = const()[name = string("x_21_groups_0"), val = int32(1024)]; + tensor x_21_strides_0 = const()[name = string("x_21_strides_0"), val = tensor([1])]; + tensor x_21_pad_0 = const()[name = string("x_21_pad_0"), val = tensor([0, 0])]; + tensor x_21_dilations_0 = const()[name = string("x_21_dilations_0"), val = tensor([1])]; + tensor encoder_layers_0_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19445120))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19454400))))[name = string("encoder_layers_0_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_21_cast_fp16 = conv(dilations = x_21_dilations_0, groups = x_21_groups_0, pad = x_21_pad_0, pad_type = x_21_pad_type_0, strides = x_21_strides_0, weight = encoder_layers_0_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_3_cast_fp16)[name = string("x_21_cast_fp16")]; + tensor input_55_perm_0 = const()[name = string("input_55_perm_0"), val = tensor([0, 2, 1])]; + tensor x_23_axes_0 = const()[name = string("x_23_axes_0"), val = tensor([-1])]; + tensor encoder_layers_0_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_0_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19456512)))]; + tensor encoder_layers_0_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_0_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19458624)))]; + tensor input_55_cast_fp16 = transpose(perm = input_55_perm_0, x = x_21_cast_fp16)[name = string("transpose_356")]; + tensor x_23_cast_fp16 = layer_norm(axes = x_23_axes_0, beta = encoder_layers_0_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_0_conv_batch_norm_weight_to_fp16, x = input_55_cast_fp16)[name = string("x_23_cast_fp16")]; + tensor input_57_perm_0 = const()[name = string("input_57_perm_0"), val = tensor([0, 2, 1])]; + tensor input_57_cast_fp16 = transpose(perm = input_57_perm_0, x = x_23_cast_fp16)[name = string("transpose_355")]; + tensor input_59_cast_fp16 = silu(x = input_57_cast_fp16)[name = string("input_59_cast_fp16")]; + string x_25_pad_type_0 = const()[name = string("x_25_pad_type_0"), val = string("valid")]; + tensor x_25_strides_0 = const()[name = string("x_25_strides_0"), val = tensor([1])]; + tensor x_25_pad_0 = const()[name = string("x_25_pad_0"), val = tensor([0, 0])]; + tensor x_25_dilations_0 = const()[name = string("x_25_dilations_0"), val = tensor([1])]; + int32 x_25_groups_0 = const()[name = string("x_25_groups_0"), val = int32(1)]; + tensor encoder_layers_0_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19460736))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20509376))))[name = string("encoder_layers_0_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_25_cast_fp16 = conv(dilations = x_25_dilations_0, groups = x_25_groups_0, pad = x_25_pad_0, pad_type = x_25_pad_type_0, strides = x_25_strides_0, weight = encoder_layers_0_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_59_cast_fp16)[name = string("x_25_cast_fp16")]; + tensor input_61_perm_0 = const()[name = string("input_61_perm_0"), val = tensor([0, 2, 1])]; + tensor input_61_cast_fp16 = transpose(perm = input_61_perm_0, x = x_25_cast_fp16)[name = string("transpose_354")]; + tensor input_63_cast_fp16 = add(x = input_47_cast_fp16, y = input_61_cast_fp16)[name = string("input_63_cast_fp16")]; + tensor input_65_axes_0 = const()[name = string("input_65_axes_0"), val = tensor([-1])]; + tensor encoder_layers_0_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_0_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20511488)))]; + tensor encoder_layers_0_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_0_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20513600)))]; + tensor input_65_cast_fp16 = layer_norm(axes = input_65_axes_0, beta = encoder_layers_0_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_0_norm_feed_forward2_weight_to_fp16, x = input_63_cast_fp16)[name = string("input_65_cast_fp16")]; + tensor encoder_layers_0_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20515712))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24710080))))[name = string("encoder_layers_0_feed_forward2_linear1_weight_to_fp16_quantized")]; + tensor encoder_layers_0_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_0_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24718336)))]; + tensor linear_8_cast_fp16 = linear(bias = encoder_layers_0_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_0_feed_forward2_linear1_weight_to_fp16_quantized, x = input_65_cast_fp16)[name = string("linear_8_cast_fp16")]; + tensor input_69_cast_fp16 = silu(x = linear_8_cast_fp16)[name = string("input_69_cast_fp16")]; + tensor encoder_layers_0_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24726592))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28920960))))[name = string("encoder_layers_0_feed_forward2_linear2_weight_to_fp16_quantized")]; + tensor encoder_layers_0_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_0_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28923072)))]; + tensor linear_9_cast_fp16 = linear(bias = encoder_layers_0_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_0_feed_forward2_linear2_weight_to_fp16_quantized, x = input_69_cast_fp16)[name = string("linear_9_cast_fp16")]; + fp16 var_631_to_fp16 = const()[name = string("op_631_to_fp16"), val = fp16(0x1p-1)]; + tensor var_632_cast_fp16 = mul(x = linear_9_cast_fp16, y = var_631_to_fp16)[name = string("op_632_cast_fp16")]; + tensor input_75_cast_fp16 = add(x = input_63_cast_fp16, y = var_632_cast_fp16)[name = string("input_75_cast_fp16")]; + tensor input_77_axes_0 = const()[name = string("input_77_axes_0"), val = tensor([-1])]; + tensor encoder_layers_0_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_0_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28925184)))]; + tensor encoder_layers_0_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_0_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28927296)))]; + tensor input_77_cast_fp16 = layer_norm(axes = input_77_axes_0, beta = encoder_layers_0_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_0_norm_out_weight_to_fp16, x = input_75_cast_fp16)[name = string("input_77_cast_fp16")]; + tensor cache_5_begin_0 = const()[name = string("cache_5_begin_0"), val = tensor([1, 0, 0, 0])]; + tensor cache_5_end_0 = const()[name = string("cache_5_end_0"), val = tensor([2, 1, 42, 1024])]; + tensor cache_5_end_mask_0 = const()[name = string("cache_5_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_5_squeeze_mask_0 = const()[name = string("cache_5_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_5_cast_fp16 = slice_by_index(begin = cache_5_begin_0, end = cache_5_end_0, end_mask = cache_5_end_mask_0, squeeze_mask = cache_5_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_5_cast_fp16")]; + tensor cache_7_begin_0 = const()[name = string("cache_7_begin_0"), val = tensor([1, 0, 0, 0])]; + tensor cache_7_end_0 = const()[name = string("cache_7_end_0"), val = tensor([2, 1, 1024, 8])]; + tensor cache_7_end_mask_0 = const()[name = string("cache_7_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_7_squeeze_mask_0 = const()[name = string("cache_7_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_7_cast_fp16 = slice_by_index(begin = cache_7_begin_0, end = cache_7_end_0, end_mask = cache_7_end_mask_0, squeeze_mask = cache_7_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_7_cast_fp16")]; + tensor input_79_axes_0 = const()[name = string("input_79_axes_0"), val = tensor([-1])]; + tensor encoder_layers_1_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_1_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28929408)))]; + tensor encoder_layers_1_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_1_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28931520)))]; + tensor input_79_cast_fp16 = layer_norm(axes = input_79_axes_0, beta = encoder_layers_1_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_1_norm_feed_forward1_weight_to_fp16, x = input_77_cast_fp16)[name = string("input_79_cast_fp16")]; + tensor encoder_layers_1_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28933632))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33128000))))[name = string("encoder_layers_1_feed_forward1_linear1_weight_to_fp16_quantized")]; + tensor encoder_layers_1_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_1_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33136256)))]; + tensor linear_10_cast_fp16 = linear(bias = encoder_layers_1_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_1_feed_forward1_linear1_weight_to_fp16_quantized, x = input_79_cast_fp16)[name = string("linear_10_cast_fp16")]; + tensor input_83_cast_fp16 = silu(x = linear_10_cast_fp16)[name = string("input_83_cast_fp16")]; + tensor encoder_layers_1_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33144512))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37338880))))[name = string("encoder_layers_1_feed_forward1_linear2_weight_to_fp16_quantized")]; + tensor encoder_layers_1_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_1_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37340992)))]; + tensor linear_11_cast_fp16 = linear(bias = encoder_layers_1_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_1_feed_forward1_linear2_weight_to_fp16_quantized, x = input_83_cast_fp16)[name = string("linear_11_cast_fp16")]; + fp16 var_668_to_fp16 = const()[name = string("op_668_to_fp16"), val = fp16(0x1p-1)]; + tensor var_669_cast_fp16 = mul(x = linear_11_cast_fp16, y = var_668_to_fp16)[name = string("op_669_cast_fp16")]; + tensor input_89_cast_fp16 = add(x = input_77_cast_fp16, y = var_669_cast_fp16)[name = string("input_89_cast_fp16")]; + tensor key_3_axes_0 = const()[name = string("key_3_axes_0"), val = tensor([-1])]; + tensor encoder_layers_1_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_1_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37343104)))]; + tensor encoder_layers_1_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_1_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37345216)))]; + tensor key_3_cast_fp16 = layer_norm(axes = key_3_axes_0, beta = encoder_layers_1_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_1_norm_self_att_weight_to_fp16, x = input_89_cast_fp16)[name = string("key_3_cast_fp16")]; + bool input_91_interleave_0 = const()[name = string("input_91_interleave_0"), val = bool(false)]; + tensor input_91_cast_fp16 = concat(axis = var_68, interleave = input_91_interleave_0, values = (cache_5_cast_fp16, key_3_cast_fp16))[name = string("input_91_cast_fp16")]; + tensor var_691_begin_0 = const()[name = string("op_691_begin_0"), val = tensor([0, 7, 0])]; + tensor var_691_end_0 = const()[name = string("op_691_end_0"), val = tensor([1, 42, 1024])]; + tensor var_691_end_mask_0 = const()[name = string("op_691_end_mask_0"), val = tensor([true, true, true])]; + tensor var_691_cast_fp16 = slice_by_index(begin = var_691_begin_0, end = var_691_end_0, end_mask = var_691_end_mask_0, x = cache_5_cast_fp16)[name = string("op_691_cast_fp16")]; + bool var_697_interleave_0 = const()[name = string("op_697_interleave_0"), val = bool(false)]; + tensor var_697_cast_fp16 = concat(axis = var_68, interleave = var_697_interleave_0, values = (var_691_cast_fp16, key_3_cast_fp16))[name = string("op_697_cast_fp16")]; + tensor encoder_layers_1_self_attn_linear_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37347328))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38395968))))[name = string("encoder_layers_1_self_attn_linear_q_weight_to_fp16_quantized")]; + tensor encoder_layers_1_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_1_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38398080)))]; + tensor linear_12_cast_fp16 = linear(bias = encoder_layers_1_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_1_self_attn_linear_q_weight_to_fp16_quantized, x = key_3_cast_fp16)[name = string("linear_12_cast_fp16")]; + tensor var_702 = const()[name = string("op_702"), val = tensor([1, -1, 8, 128])]; + tensor q_7_cast_fp16 = reshape(shape = var_702, x = linear_12_cast_fp16)[name = string("q_7_cast_fp16")]; + tensor encoder_layers_1_self_attn_linear_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38400192))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39448832))))[name = string("encoder_layers_1_self_attn_linear_k_weight_to_fp16_quantized")]; + tensor encoder_layers_1_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_1_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39450944)))]; + tensor linear_13_cast_fp16 = linear(bias = encoder_layers_1_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_1_self_attn_linear_k_weight_to_fp16_quantized, x = input_91_cast_fp16)[name = string("linear_13_cast_fp16")]; + tensor var_707 = const()[name = string("op_707"), val = tensor([1, -1, 8, 128])]; + tensor k_5_cast_fp16 = reshape(shape = var_707, x = linear_13_cast_fp16)[name = string("k_5_cast_fp16")]; + tensor encoder_layers_1_self_attn_linear_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39453056))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40501696))))[name = string("encoder_layers_1_self_attn_linear_v_weight_to_fp16_quantized")]; + tensor encoder_layers_1_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_1_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40503808)))]; + tensor linear_14_cast_fp16 = linear(bias = encoder_layers_1_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_1_self_attn_linear_v_weight_to_fp16_quantized, x = input_91_cast_fp16)[name = string("linear_14_cast_fp16")]; + tensor var_712 = const()[name = string("op_712"), val = tensor([1, -1, 8, 128])]; + tensor v_3_cast_fp16 = reshape(shape = var_712, x = linear_14_cast_fp16)[name = string("v_3_cast_fp16")]; + tensor value_11_perm_0 = const()[name = string("value_11_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_1_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_1_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40505920)))]; + tensor var_725_cast_fp16 = add(x = q_7_cast_fp16, y = encoder_layers_1_self_attn_pos_bias_u_to_fp16)[name = string("op_725_cast_fp16")]; + tensor encoder_layers_1_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_1_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40508032)))]; + tensor var_727_cast_fp16 = add(x = q_7_cast_fp16, y = encoder_layers_1_self_attn_pos_bias_v_to_fp16)[name = string("op_727_cast_fp16")]; + tensor q_with_bias_v_3_perm_0 = const()[name = string("q_with_bias_v_3_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_33_transpose_x_0 = const()[name = string("x_33_transpose_x_0"), val = bool(false)]; + bool x_33_transpose_y_0 = const()[name = string("x_33_transpose_y_0"), val = bool(false)]; + tensor op_729_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40510144))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40609536))))[name = string("op_729_to_fp16_quantized")]; + tensor q_with_bias_v_3_cast_fp16 = transpose(perm = q_with_bias_v_3_perm_0, x = var_727_cast_fp16)[name = string("transpose_353")]; + tensor x_33_cast_fp16 = matmul(transpose_x = x_33_transpose_x_0, transpose_y = x_33_transpose_y_0, x = q_with_bias_v_3_cast_fp16, y = op_729_to_fp16_quantized)[name = string("x_33_cast_fp16")]; + tensor x_35_pad_0 = const()[name = string("x_35_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_35_mode_0 = const()[name = string("x_35_mode_0"), val = string("constant")]; + fp16 const_92_to_fp16 = const()[name = string("const_92_to_fp16"), val = fp16(0x0p+0)]; + tensor x_35_cast_fp16 = pad(constant_val = const_92_to_fp16, mode = x_35_mode_0, pad = x_35_pad_0, x = x_33_cast_fp16)[name = string("x_35_cast_fp16")]; + tensor var_737 = const()[name = string("op_737"), val = tensor([1, 8, -1, 7])]; + tensor x_37_cast_fp16 = reshape(shape = var_737, x = x_35_cast_fp16)[name = string("x_37_cast_fp16")]; + tensor var_741_begin_0 = const()[name = string("op_741_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_741_end_0 = const()[name = string("op_741_end_0"), val = tensor([1, 8, 98, 7])]; + tensor var_741_end_mask_0 = const()[name = string("op_741_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_741_cast_fp16 = slice_by_index(begin = var_741_begin_0, end = var_741_end_0, end_mask = var_741_end_mask_0, x = x_37_cast_fp16)[name = string("op_741_cast_fp16")]; + tensor var_742 = const()[name = string("op_742"), val = tensor([1, 8, 7, 97])]; + tensor matrix_bd_5_cast_fp16 = reshape(shape = var_742, x = var_741_cast_fp16)[name = string("matrix_bd_5_cast_fp16")]; + bool matrix_ac_3_transpose_x_0 = const()[name = string("matrix_ac_3_transpose_x_0"), val = bool(false)]; + bool matrix_ac_3_transpose_y_0 = const()[name = string("matrix_ac_3_transpose_y_0"), val = bool(false)]; + tensor transpose_98_perm_0 = const()[name = string("transpose_98_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_99_perm_0 = const()[name = string("transpose_99_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_99 = transpose(perm = transpose_99_perm_0, x = k_5_cast_fp16)[name = string("transpose_351")]; + tensor transpose_98 = transpose(perm = transpose_98_perm_0, x = var_725_cast_fp16)[name = string("transpose_352")]; + tensor matrix_ac_3_cast_fp16 = matmul(transpose_x = matrix_ac_3_transpose_x_0, transpose_y = matrix_ac_3_transpose_y_0, x = transpose_98, y = transpose_99)[name = string("matrix_ac_3_cast_fp16")]; + tensor matrix_bd_7_begin_0 = const()[name = string("matrix_bd_7_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_7_end_0 = const()[name = string("matrix_bd_7_end_0"), val = tensor([1, 8, 7, 49])]; + tensor matrix_bd_7_end_mask_0 = const()[name = string("matrix_bd_7_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_7_cast_fp16 = slice_by_index(begin = matrix_bd_7_begin_0, end = matrix_bd_7_end_0, end_mask = matrix_bd_7_end_mask_0, x = matrix_bd_5_cast_fp16)[name = string("matrix_bd_7_cast_fp16")]; + tensor var_751_cast_fp16 = add(x = matrix_ac_3_cast_fp16, y = matrix_bd_7_cast_fp16)[name = string("op_751_cast_fp16")]; + fp16 _inversed_scores_5_y_0_to_fp16 = const()[name = string("_inversed_scores_5_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_5_cast_fp16 = mul(x = var_751_cast_fp16, y = _inversed_scores_5_y_0_to_fp16)[name = string("_inversed_scores_5_cast_fp16")]; + tensor scores_7_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_5_cast_fp16, cond = mask_11)[name = string("scores_7_cast_fp16")]; + tensor var_757_cast_fp16 = softmax(axis = var_59, x = scores_7_cast_fp16)[name = string("op_757_cast_fp16")]; + tensor input_93_cast_fp16 = select(a = var_44_to_fp16, b = var_757_cast_fp16, cond = mask_11)[name = string("input_93_cast_fp16")]; + bool x_39_transpose_x_0 = const()[name = string("x_39_transpose_x_0"), val = bool(false)]; + bool x_39_transpose_y_0 = const()[name = string("x_39_transpose_y_0"), val = bool(false)]; + tensor value_11_cast_fp16 = transpose(perm = value_11_perm_0, x = v_3_cast_fp16)[name = string("transpose_350")]; + tensor x_39_cast_fp16 = matmul(transpose_x = x_39_transpose_x_0, transpose_y = x_39_transpose_y_0, x = input_93_cast_fp16, y = value_11_cast_fp16)[name = string("x_39_cast_fp16")]; + tensor var_761_perm_0 = const()[name = string("op_761_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_762 = const()[name = string("op_762"), val = tensor([1, -1, 1024])]; + tensor var_761_cast_fp16 = transpose(perm = var_761_perm_0, x = x_39_cast_fp16)[name = string("transpose_349")]; + tensor input_95_cast_fp16 = reshape(shape = var_762, x = var_761_cast_fp16)[name = string("input_95_cast_fp16")]; + tensor encoder_layers_1_self_attn_linear_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40609856))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41658496))))[name = string("encoder_layers_1_self_attn_linear_out_weight_to_fp16_quantized")]; + tensor encoder_layers_1_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_1_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41660608)))]; + tensor linear_16_cast_fp16 = linear(bias = encoder_layers_1_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_1_self_attn_linear_out_weight_to_fp16_quantized, x = input_95_cast_fp16)[name = string("linear_16_cast_fp16")]; + tensor input_99_cast_fp16 = add(x = input_89_cast_fp16, y = linear_16_cast_fp16)[name = string("input_99_cast_fp16")]; + tensor x_43_axes_0 = const()[name = string("x_43_axes_0"), val = tensor([-1])]; + tensor encoder_layers_1_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_1_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41662720)))]; + tensor encoder_layers_1_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_1_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41664832)))]; + tensor x_43_cast_fp16 = layer_norm(axes = x_43_axes_0, beta = encoder_layers_1_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_1_norm_conv_weight_to_fp16, x = input_99_cast_fp16)[name = string("x_43_cast_fp16")]; + tensor input_101_perm_0 = const()[name = string("input_101_perm_0"), val = tensor([0, 2, 1])]; + string input_103_pad_type_0 = const()[name = string("input_103_pad_type_0"), val = string("valid")]; + tensor input_103_strides_0 = const()[name = string("input_103_strides_0"), val = tensor([1])]; + tensor input_103_pad_0 = const()[name = string("input_103_pad_0"), val = tensor([0, 0])]; + tensor input_103_dilations_0 = const()[name = string("input_103_dilations_0"), val = tensor([1])]; + int32 input_103_groups_0 = const()[name = string("input_103_groups_0"), val = int32(1)]; + tensor encoder_layers_1_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41666944))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43764160))))[name = string("encoder_layers_1_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_101_cast_fp16 = transpose(perm = input_101_perm_0, x = x_43_cast_fp16)[name = string("transpose_348")]; + tensor input_103_cast_fp16 = conv(dilations = input_103_dilations_0, groups = input_103_groups_0, pad = input_103_pad_0, pad_type = input_103_pad_type_0, strides = input_103_strides_0, weight = encoder_layers_1_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_101_cast_fp16)[name = string("input_103_cast_fp16")]; + int32 x_45_split_num_splits_0 = const()[name = string("x_45_split_num_splits_0"), val = int32(2)]; + int32 x_45_split_axis_0 = const()[name = string("x_45_split_axis_0"), val = int32(1)]; + tensor x_45_split_cast_fp16_0, tensor x_45_split_cast_fp16_1 = split(axis = x_45_split_axis_0, num_splits = x_45_split_num_splits_0, x = input_103_cast_fp16)[name = string("x_45_split_cast_fp16")]; + tensor x_45_split_1_sigmoid_cast_fp16 = sigmoid(x = x_45_split_cast_fp16_1)[name = string("x_45_split_1_sigmoid_cast_fp16")]; + tensor x_45_cast_fp16 = mul(x = x_45_split_cast_fp16_0, y = x_45_split_1_sigmoid_cast_fp16)[name = string("x_45_cast_fp16")]; + tensor input_105_cast_fp16 = select(a = var_44_to_fp16, b = x_45_cast_fp16, cond = var_575)[name = string("input_105_cast_fp16")]; + bool new_x_7_interleave_0 = const()[name = string("new_x_7_interleave_0"), val = bool(false)]; + tensor new_x_7_cast_fp16 = concat(axis = var_59, interleave = new_x_7_interleave_0, values = (cache_7_cast_fp16, input_105_cast_fp16))[name = string("new_x_7_cast_fp16")]; + tensor var_801_begin_0 = const()[name = string("op_801_begin_0"), val = tensor([0, 0, 7])]; + tensor var_801_end_0 = const()[name = string("op_801_end_0"), val = tensor([1, 1024, 15])]; + tensor var_801_end_mask_0 = const()[name = string("op_801_end_mask_0"), val = tensor([true, true, true])]; + tensor var_801_cast_fp16 = slice_by_index(begin = var_801_begin_0, end = var_801_end_0, end_mask = var_801_end_mask_0, x = new_x_7_cast_fp16)[name = string("op_801_cast_fp16")]; + string x_47_pad_type_0 = const()[name = string("x_47_pad_type_0"), val = string("valid")]; + int32 x_47_groups_0 = const()[name = string("x_47_groups_0"), val = int32(1024)]; + tensor x_47_strides_0 = const()[name = string("x_47_strides_0"), val = tensor([1])]; + tensor x_47_pad_0 = const()[name = string("x_47_pad_0"), val = tensor([0, 0])]; + tensor x_47_dilations_0 = const()[name = string("x_47_dilations_0"), val = tensor([1])]; + tensor encoder_layers_1_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43768320))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43777600))))[name = string("encoder_layers_1_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_47_cast_fp16 = conv(dilations = x_47_dilations_0, groups = x_47_groups_0, pad = x_47_pad_0, pad_type = x_47_pad_type_0, strides = x_47_strides_0, weight = encoder_layers_1_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_7_cast_fp16)[name = string("x_47_cast_fp16")]; + tensor input_107_perm_0 = const()[name = string("input_107_perm_0"), val = tensor([0, 2, 1])]; + tensor x_49_axes_0 = const()[name = string("x_49_axes_0"), val = tensor([-1])]; + tensor encoder_layers_1_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_1_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43779712)))]; + tensor encoder_layers_1_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_1_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43781824)))]; + tensor input_107_cast_fp16 = transpose(perm = input_107_perm_0, x = x_47_cast_fp16)[name = string("transpose_347")]; + tensor x_49_cast_fp16 = layer_norm(axes = x_49_axes_0, beta = encoder_layers_1_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_1_conv_batch_norm_weight_to_fp16, x = input_107_cast_fp16)[name = string("x_49_cast_fp16")]; + tensor input_109_perm_0 = const()[name = string("input_109_perm_0"), val = tensor([0, 2, 1])]; + tensor input_109_cast_fp16 = transpose(perm = input_109_perm_0, x = x_49_cast_fp16)[name = string("transpose_346")]; + tensor input_111_cast_fp16 = silu(x = input_109_cast_fp16)[name = string("input_111_cast_fp16")]; + string x_51_pad_type_0 = const()[name = string("x_51_pad_type_0"), val = string("valid")]; + tensor x_51_strides_0 = const()[name = string("x_51_strides_0"), val = tensor([1])]; + tensor x_51_pad_0 = const()[name = string("x_51_pad_0"), val = tensor([0, 0])]; + tensor x_51_dilations_0 = const()[name = string("x_51_dilations_0"), val = tensor([1])]; + int32 x_51_groups_0 = const()[name = string("x_51_groups_0"), val = int32(1)]; + tensor encoder_layers_1_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43783936))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44832576))))[name = string("encoder_layers_1_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_51_cast_fp16 = conv(dilations = x_51_dilations_0, groups = x_51_groups_0, pad = x_51_pad_0, pad_type = x_51_pad_type_0, strides = x_51_strides_0, weight = encoder_layers_1_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_111_cast_fp16)[name = string("x_51_cast_fp16")]; + tensor input_113_perm_0 = const()[name = string("input_113_perm_0"), val = tensor([0, 2, 1])]; + tensor input_113_cast_fp16 = transpose(perm = input_113_perm_0, x = x_51_cast_fp16)[name = string("transpose_345")]; + tensor input_115_cast_fp16 = add(x = input_99_cast_fp16, y = input_113_cast_fp16)[name = string("input_115_cast_fp16")]; + tensor input_117_axes_0 = const()[name = string("input_117_axes_0"), val = tensor([-1])]; + tensor encoder_layers_1_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_1_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44834688)))]; + tensor encoder_layers_1_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_1_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44836800)))]; + tensor input_117_cast_fp16 = layer_norm(axes = input_117_axes_0, beta = encoder_layers_1_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_1_norm_feed_forward2_weight_to_fp16, x = input_115_cast_fp16)[name = string("input_117_cast_fp16")]; + tensor encoder_layers_1_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44838912))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49033280))))[name = string("encoder_layers_1_feed_forward2_linear1_weight_to_fp16_quantized")]; + tensor encoder_layers_1_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_1_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49041536)))]; + tensor linear_17_cast_fp16 = linear(bias = encoder_layers_1_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_1_feed_forward2_linear1_weight_to_fp16_quantized, x = input_117_cast_fp16)[name = string("linear_17_cast_fp16")]; + tensor input_121_cast_fp16 = silu(x = linear_17_cast_fp16)[name = string("input_121_cast_fp16")]; + tensor encoder_layers_1_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49049792))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53244160))))[name = string("encoder_layers_1_feed_forward2_linear2_weight_to_fp16_quantized")]; + tensor encoder_layers_1_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_1_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53246272)))]; + tensor linear_18_cast_fp16 = linear(bias = encoder_layers_1_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_1_feed_forward2_linear2_weight_to_fp16_quantized, x = input_121_cast_fp16)[name = string("linear_18_cast_fp16")]; + fp16 var_844_to_fp16 = const()[name = string("op_844_to_fp16"), val = fp16(0x1p-1)]; + tensor var_845_cast_fp16 = mul(x = linear_18_cast_fp16, y = var_844_to_fp16)[name = string("op_845_cast_fp16")]; + tensor input_127_cast_fp16 = add(x = input_115_cast_fp16, y = var_845_cast_fp16)[name = string("input_127_cast_fp16")]; + tensor input_129_axes_0 = const()[name = string("input_129_axes_0"), val = tensor([-1])]; + tensor encoder_layers_1_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_1_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53248384)))]; + tensor encoder_layers_1_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_1_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53250496)))]; + tensor input_129_cast_fp16 = layer_norm(axes = input_129_axes_0, beta = encoder_layers_1_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_1_norm_out_weight_to_fp16, x = input_127_cast_fp16)[name = string("input_129_cast_fp16")]; + tensor cache_9_begin_0 = const()[name = string("cache_9_begin_0"), val = tensor([2, 0, 0, 0])]; + tensor cache_9_end_0 = const()[name = string("cache_9_end_0"), val = tensor([3, 1, 42, 1024])]; + tensor cache_9_end_mask_0 = const()[name = string("cache_9_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_9_squeeze_mask_0 = const()[name = string("cache_9_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_9_cast_fp16 = slice_by_index(begin = cache_9_begin_0, end = cache_9_end_0, end_mask = cache_9_end_mask_0, squeeze_mask = cache_9_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_9_cast_fp16")]; + tensor cache_11_begin_0 = const()[name = string("cache_11_begin_0"), val = tensor([2, 0, 0, 0])]; + tensor cache_11_end_0 = const()[name = string("cache_11_end_0"), val = tensor([3, 1, 1024, 8])]; + tensor cache_11_end_mask_0 = const()[name = string("cache_11_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_11_squeeze_mask_0 = const()[name = string("cache_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_11_cast_fp16 = slice_by_index(begin = cache_11_begin_0, end = cache_11_end_0, end_mask = cache_11_end_mask_0, squeeze_mask = cache_11_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_11_cast_fp16")]; + tensor input_131_axes_0 = const()[name = string("input_131_axes_0"), val = tensor([-1])]; + tensor encoder_layers_2_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_2_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53252608)))]; + tensor encoder_layers_2_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_2_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53254720)))]; + tensor input_131_cast_fp16 = layer_norm(axes = input_131_axes_0, beta = encoder_layers_2_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_2_norm_feed_forward1_weight_to_fp16, x = input_129_cast_fp16)[name = string("input_131_cast_fp16")]; + tensor encoder_layers_2_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53256832))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57451200))))[name = string("encoder_layers_2_feed_forward1_linear1_weight_to_fp16_quantized")]; + tensor encoder_layers_2_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_2_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57459456)))]; + tensor linear_19_cast_fp16 = linear(bias = encoder_layers_2_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_2_feed_forward1_linear1_weight_to_fp16_quantized, x = input_131_cast_fp16)[name = string("linear_19_cast_fp16")]; + tensor input_135_cast_fp16 = silu(x = linear_19_cast_fp16)[name = string("input_135_cast_fp16")]; + tensor encoder_layers_2_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57467712))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61662080))))[name = string("encoder_layers_2_feed_forward1_linear2_weight_to_fp16_quantized")]; + tensor encoder_layers_2_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_2_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61664192)))]; + tensor linear_20_cast_fp16 = linear(bias = encoder_layers_2_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_2_feed_forward1_linear2_weight_to_fp16_quantized, x = input_135_cast_fp16)[name = string("linear_20_cast_fp16")]; + fp16 var_881_to_fp16 = const()[name = string("op_881_to_fp16"), val = fp16(0x1p-1)]; + tensor var_882_cast_fp16 = mul(x = linear_20_cast_fp16, y = var_881_to_fp16)[name = string("op_882_cast_fp16")]; + tensor input_141_cast_fp16 = add(x = input_129_cast_fp16, y = var_882_cast_fp16)[name = string("input_141_cast_fp16")]; + tensor key_5_axes_0 = const()[name = string("key_5_axes_0"), val = tensor([-1])]; + tensor encoder_layers_2_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_2_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61666304)))]; + tensor encoder_layers_2_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_2_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61668416)))]; + tensor key_5_cast_fp16 = layer_norm(axes = key_5_axes_0, beta = encoder_layers_2_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_2_norm_self_att_weight_to_fp16, x = input_141_cast_fp16)[name = string("key_5_cast_fp16")]; + bool input_143_interleave_0 = const()[name = string("input_143_interleave_0"), val = bool(false)]; + tensor input_143_cast_fp16 = concat(axis = var_68, interleave = input_143_interleave_0, values = (cache_9_cast_fp16, key_5_cast_fp16))[name = string("input_143_cast_fp16")]; + tensor var_904_begin_0 = const()[name = string("op_904_begin_0"), val = tensor([0, 7, 0])]; + tensor var_904_end_0 = const()[name = string("op_904_end_0"), val = tensor([1, 42, 1024])]; + tensor var_904_end_mask_0 = const()[name = string("op_904_end_mask_0"), val = tensor([true, true, true])]; + tensor var_904_cast_fp16 = slice_by_index(begin = var_904_begin_0, end = var_904_end_0, end_mask = var_904_end_mask_0, x = cache_9_cast_fp16)[name = string("op_904_cast_fp16")]; + bool var_910_interleave_0 = const()[name = string("op_910_interleave_0"), val = bool(false)]; + tensor var_910_cast_fp16 = concat(axis = var_68, interleave = var_910_interleave_0, values = (var_904_cast_fp16, key_5_cast_fp16))[name = string("op_910_cast_fp16")]; + tensor encoder_layers_2_self_attn_linear_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61670528))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62719168))))[name = string("encoder_layers_2_self_attn_linear_q_weight_to_fp16_quantized")]; + tensor encoder_layers_2_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_2_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62721280)))]; + tensor linear_21_cast_fp16 = linear(bias = encoder_layers_2_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_2_self_attn_linear_q_weight_to_fp16_quantized, x = key_5_cast_fp16)[name = string("linear_21_cast_fp16")]; + tensor var_915 = const()[name = string("op_915"), val = tensor([1, -1, 8, 128])]; + tensor q_13_cast_fp16 = reshape(shape = var_915, x = linear_21_cast_fp16)[name = string("q_13_cast_fp16")]; + tensor encoder_layers_2_self_attn_linear_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62723392))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(63772032))))[name = string("encoder_layers_2_self_attn_linear_k_weight_to_fp16_quantized")]; + tensor encoder_layers_2_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_2_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(63774144)))]; + tensor linear_22_cast_fp16 = linear(bias = encoder_layers_2_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_2_self_attn_linear_k_weight_to_fp16_quantized, x = input_143_cast_fp16)[name = string("linear_22_cast_fp16")]; + tensor var_920 = const()[name = string("op_920"), val = tensor([1, -1, 8, 128])]; + tensor k_9_cast_fp16 = reshape(shape = var_920, x = linear_22_cast_fp16)[name = string("k_9_cast_fp16")]; + tensor encoder_layers_2_self_attn_linear_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(63776256))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64824896))))[name = string("encoder_layers_2_self_attn_linear_v_weight_to_fp16_quantized")]; + tensor encoder_layers_2_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_2_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64827008)))]; + tensor linear_23_cast_fp16 = linear(bias = encoder_layers_2_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_2_self_attn_linear_v_weight_to_fp16_quantized, x = input_143_cast_fp16)[name = string("linear_23_cast_fp16")]; + tensor var_925 = const()[name = string("op_925"), val = tensor([1, -1, 8, 128])]; + tensor v_5_cast_fp16 = reshape(shape = var_925, x = linear_23_cast_fp16)[name = string("v_5_cast_fp16")]; + tensor value_13_perm_0 = const()[name = string("value_13_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_2_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_2_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64829120)))]; + tensor var_938_cast_fp16 = add(x = q_13_cast_fp16, y = encoder_layers_2_self_attn_pos_bias_u_to_fp16)[name = string("op_938_cast_fp16")]; + tensor encoder_layers_2_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_2_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64831232)))]; + tensor var_940_cast_fp16 = add(x = q_13_cast_fp16, y = encoder_layers_2_self_attn_pos_bias_v_to_fp16)[name = string("op_940_cast_fp16")]; + tensor q_with_bias_v_5_perm_0 = const()[name = string("q_with_bias_v_5_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_59_transpose_x_0 = const()[name = string("x_59_transpose_x_0"), val = bool(false)]; + bool x_59_transpose_y_0 = const()[name = string("x_59_transpose_y_0"), val = bool(false)]; + tensor op_942_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64833344))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64932736))))[name = string("op_942_to_fp16_quantized")]; + tensor q_with_bias_v_5_cast_fp16 = transpose(perm = q_with_bias_v_5_perm_0, x = var_940_cast_fp16)[name = string("transpose_344")]; + tensor x_59_cast_fp16 = matmul(transpose_x = x_59_transpose_x_0, transpose_y = x_59_transpose_y_0, x = q_with_bias_v_5_cast_fp16, y = op_942_to_fp16_quantized)[name = string("x_59_cast_fp16")]; + tensor x_61_pad_0 = const()[name = string("x_61_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_61_mode_0 = const()[name = string("x_61_mode_0"), val = string("constant")]; + fp16 const_105_to_fp16 = const()[name = string("const_105_to_fp16"), val = fp16(0x0p+0)]; + tensor x_61_cast_fp16 = pad(constant_val = const_105_to_fp16, mode = x_61_mode_0, pad = x_61_pad_0, x = x_59_cast_fp16)[name = string("x_61_cast_fp16")]; + tensor var_950 = const()[name = string("op_950"), val = tensor([1, 8, -1, 7])]; + tensor x_63_cast_fp16 = reshape(shape = var_950, x = x_61_cast_fp16)[name = string("x_63_cast_fp16")]; + tensor var_954_begin_0 = const()[name = string("op_954_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_954_end_0 = const()[name = string("op_954_end_0"), val = tensor([1, 8, 98, 7])]; + tensor var_954_end_mask_0 = const()[name = string("op_954_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_954_cast_fp16 = slice_by_index(begin = var_954_begin_0, end = var_954_end_0, end_mask = var_954_end_mask_0, x = x_63_cast_fp16)[name = string("op_954_cast_fp16")]; + tensor var_955 = const()[name = string("op_955"), val = tensor([1, 8, 7, 97])]; + tensor matrix_bd_9_cast_fp16 = reshape(shape = var_955, x = var_954_cast_fp16)[name = string("matrix_bd_9_cast_fp16")]; + bool matrix_ac_5_transpose_x_0 = const()[name = string("matrix_ac_5_transpose_x_0"), val = bool(false)]; + bool matrix_ac_5_transpose_y_0 = const()[name = string("matrix_ac_5_transpose_y_0"), val = bool(false)]; + tensor transpose_100_perm_0 = const()[name = string("transpose_100_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_101_perm_0 = const()[name = string("transpose_101_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_101 = transpose(perm = transpose_101_perm_0, x = k_9_cast_fp16)[name = string("transpose_342")]; + tensor transpose_100 = transpose(perm = transpose_100_perm_0, x = var_938_cast_fp16)[name = string("transpose_343")]; + tensor matrix_ac_5_cast_fp16 = matmul(transpose_x = matrix_ac_5_transpose_x_0, transpose_y = matrix_ac_5_transpose_y_0, x = transpose_100, y = transpose_101)[name = string("matrix_ac_5_cast_fp16")]; + tensor matrix_bd_11_begin_0 = const()[name = string("matrix_bd_11_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_11_end_0 = const()[name = string("matrix_bd_11_end_0"), val = tensor([1, 8, 7, 49])]; + tensor matrix_bd_11_end_mask_0 = const()[name = string("matrix_bd_11_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_11_cast_fp16 = slice_by_index(begin = matrix_bd_11_begin_0, end = matrix_bd_11_end_0, end_mask = matrix_bd_11_end_mask_0, x = matrix_bd_9_cast_fp16)[name = string("matrix_bd_11_cast_fp16")]; + tensor var_964_cast_fp16 = add(x = matrix_ac_5_cast_fp16, y = matrix_bd_11_cast_fp16)[name = string("op_964_cast_fp16")]; + fp16 _inversed_scores_9_y_0_to_fp16 = const()[name = string("_inversed_scores_9_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_9_cast_fp16 = mul(x = var_964_cast_fp16, y = _inversed_scores_9_y_0_to_fp16)[name = string("_inversed_scores_9_cast_fp16")]; + tensor scores_11_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_9_cast_fp16, cond = mask_11)[name = string("scores_11_cast_fp16")]; + tensor var_970_cast_fp16 = softmax(axis = var_59, x = scores_11_cast_fp16)[name = string("op_970_cast_fp16")]; + tensor input_145_cast_fp16 = select(a = var_44_to_fp16, b = var_970_cast_fp16, cond = mask_11)[name = string("input_145_cast_fp16")]; + bool x_65_transpose_x_0 = const()[name = string("x_65_transpose_x_0"), val = bool(false)]; + bool x_65_transpose_y_0 = const()[name = string("x_65_transpose_y_0"), val = bool(false)]; + tensor value_13_cast_fp16 = transpose(perm = value_13_perm_0, x = v_5_cast_fp16)[name = string("transpose_341")]; + tensor x_65_cast_fp16 = matmul(transpose_x = x_65_transpose_x_0, transpose_y = x_65_transpose_y_0, x = input_145_cast_fp16, y = value_13_cast_fp16)[name = string("x_65_cast_fp16")]; + tensor var_974_perm_0 = const()[name = string("op_974_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_975 = const()[name = string("op_975"), val = tensor([1, -1, 1024])]; + tensor var_974_cast_fp16 = transpose(perm = var_974_perm_0, x = x_65_cast_fp16)[name = string("transpose_340")]; + tensor input_147_cast_fp16 = reshape(shape = var_975, x = var_974_cast_fp16)[name = string("input_147_cast_fp16")]; + tensor encoder_layers_2_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64933056))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65719552))))[name = string("encoder_layers_2_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_2_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_2_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65719744)))]; + tensor linear_25_cast_fp16 = linear(bias = encoder_layers_2_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_2_self_attn_linear_out_weight_to_fp16_palettized, x = input_147_cast_fp16)[name = string("linear_25_cast_fp16")]; + tensor input_151_cast_fp16 = add(x = input_141_cast_fp16, y = linear_25_cast_fp16)[name = string("input_151_cast_fp16")]; + tensor x_69_axes_0 = const()[name = string("x_69_axes_0"), val = tensor([-1])]; + tensor encoder_layers_2_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_2_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65721856)))]; + tensor encoder_layers_2_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_2_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65723968)))]; + tensor x_69_cast_fp16 = layer_norm(axes = x_69_axes_0, beta = encoder_layers_2_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_2_norm_conv_weight_to_fp16, x = input_151_cast_fp16)[name = string("x_69_cast_fp16")]; + tensor input_153_perm_0 = const()[name = string("input_153_perm_0"), val = tensor([0, 2, 1])]; + string input_155_pad_type_0 = const()[name = string("input_155_pad_type_0"), val = string("valid")]; + tensor input_155_strides_0 = const()[name = string("input_155_strides_0"), val = tensor([1])]; + tensor input_155_pad_0 = const()[name = string("input_155_pad_0"), val = tensor([0, 0])]; + tensor input_155_dilations_0 = const()[name = string("input_155_dilations_0"), val = tensor([1])]; + int32 input_155_groups_0 = const()[name = string("input_155_groups_0"), val = int32(1)]; + tensor encoder_layers_2_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65726080))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67823296))))[name = string("encoder_layers_2_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_153_cast_fp16 = transpose(perm = input_153_perm_0, x = x_69_cast_fp16)[name = string("transpose_339")]; + tensor input_155_cast_fp16 = conv(dilations = input_155_dilations_0, groups = input_155_groups_0, pad = input_155_pad_0, pad_type = input_155_pad_type_0, strides = input_155_strides_0, weight = encoder_layers_2_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_153_cast_fp16)[name = string("input_155_cast_fp16")]; + int32 x_71_split_num_splits_0 = const()[name = string("x_71_split_num_splits_0"), val = int32(2)]; + int32 x_71_split_axis_0 = const()[name = string("x_71_split_axis_0"), val = int32(1)]; + tensor x_71_split_cast_fp16_0, tensor x_71_split_cast_fp16_1 = split(axis = x_71_split_axis_0, num_splits = x_71_split_num_splits_0, x = input_155_cast_fp16)[name = string("x_71_split_cast_fp16")]; + tensor x_71_split_1_sigmoid_cast_fp16 = sigmoid(x = x_71_split_cast_fp16_1)[name = string("x_71_split_1_sigmoid_cast_fp16")]; + tensor x_71_cast_fp16 = mul(x = x_71_split_cast_fp16_0, y = x_71_split_1_sigmoid_cast_fp16)[name = string("x_71_cast_fp16")]; + tensor input_157_cast_fp16 = select(a = var_44_to_fp16, b = x_71_cast_fp16, cond = var_575)[name = string("input_157_cast_fp16")]; + bool new_x_11_interleave_0 = const()[name = string("new_x_11_interleave_0"), val = bool(false)]; + tensor new_x_11_cast_fp16 = concat(axis = var_59, interleave = new_x_11_interleave_0, values = (cache_11_cast_fp16, input_157_cast_fp16))[name = string("new_x_11_cast_fp16")]; + tensor var_1014_begin_0 = const()[name = string("op_1014_begin_0"), val = tensor([0, 0, 7])]; + tensor var_1014_end_0 = const()[name = string("op_1014_end_0"), val = tensor([1, 1024, 15])]; + tensor var_1014_end_mask_0 = const()[name = string("op_1014_end_mask_0"), val = tensor([true, true, true])]; + tensor var_1014_cast_fp16 = slice_by_index(begin = var_1014_begin_0, end = var_1014_end_0, end_mask = var_1014_end_mask_0, x = new_x_11_cast_fp16)[name = string("op_1014_cast_fp16")]; + string x_73_pad_type_0 = const()[name = string("x_73_pad_type_0"), val = string("valid")]; + int32 x_73_groups_0 = const()[name = string("x_73_groups_0"), val = int32(1024)]; + tensor x_73_strides_0 = const()[name = string("x_73_strides_0"), val = tensor([1])]; + tensor x_73_pad_0 = const()[name = string("x_73_pad_0"), val = tensor([0, 0])]; + tensor x_73_dilations_0 = const()[name = string("x_73_dilations_0"), val = tensor([1])]; + tensor encoder_layers_2_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67827456))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67836736))))[name = string("encoder_layers_2_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_73_cast_fp16 = conv(dilations = x_73_dilations_0, groups = x_73_groups_0, pad = x_73_pad_0, pad_type = x_73_pad_type_0, strides = x_73_strides_0, weight = encoder_layers_2_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_11_cast_fp16)[name = string("x_73_cast_fp16")]; + tensor input_159_perm_0 = const()[name = string("input_159_perm_0"), val = tensor([0, 2, 1])]; + tensor x_75_axes_0 = const()[name = string("x_75_axes_0"), val = tensor([-1])]; + tensor encoder_layers_2_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_2_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67838848)))]; + tensor encoder_layers_2_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_2_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67840960)))]; + tensor input_159_cast_fp16 = transpose(perm = input_159_perm_0, x = x_73_cast_fp16)[name = string("transpose_338")]; + tensor x_75_cast_fp16 = layer_norm(axes = x_75_axes_0, beta = encoder_layers_2_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_2_conv_batch_norm_weight_to_fp16, x = input_159_cast_fp16)[name = string("x_75_cast_fp16")]; + tensor input_161_perm_0 = const()[name = string("input_161_perm_0"), val = tensor([0, 2, 1])]; + tensor input_161_cast_fp16 = transpose(perm = input_161_perm_0, x = x_75_cast_fp16)[name = string("transpose_337")]; + tensor input_163_cast_fp16 = silu(x = input_161_cast_fp16)[name = string("input_163_cast_fp16")]; + string x_77_pad_type_0 = const()[name = string("x_77_pad_type_0"), val = string("valid")]; + tensor x_77_strides_0 = const()[name = string("x_77_strides_0"), val = tensor([1])]; + tensor x_77_pad_0 = const()[name = string("x_77_pad_0"), val = tensor([0, 0])]; + tensor x_77_dilations_0 = const()[name = string("x_77_dilations_0"), val = tensor([1])]; + int32 x_77_groups_0 = const()[name = string("x_77_groups_0"), val = int32(1)]; + tensor encoder_layers_2_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67843072))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68891712))))[name = string("encoder_layers_2_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_77_cast_fp16 = conv(dilations = x_77_dilations_0, groups = x_77_groups_0, pad = x_77_pad_0, pad_type = x_77_pad_type_0, strides = x_77_strides_0, weight = encoder_layers_2_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_163_cast_fp16)[name = string("x_77_cast_fp16")]; + tensor input_165_perm_0 = const()[name = string("input_165_perm_0"), val = tensor([0, 2, 1])]; + tensor input_165_cast_fp16 = transpose(perm = input_165_perm_0, x = x_77_cast_fp16)[name = string("transpose_336")]; + tensor input_167_cast_fp16 = add(x = input_151_cast_fp16, y = input_165_cast_fp16)[name = string("input_167_cast_fp16")]; + tensor input_169_axes_0 = const()[name = string("input_169_axes_0"), val = tensor([-1])]; + tensor encoder_layers_2_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_2_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68893824)))]; + tensor encoder_layers_2_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_2_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68895936)))]; + tensor input_169_cast_fp16 = layer_norm(axes = input_169_axes_0, beta = encoder_layers_2_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_2_norm_feed_forward2_weight_to_fp16, x = input_167_cast_fp16)[name = string("input_169_cast_fp16")]; + tensor encoder_layers_2_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68898048))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72043840))))[name = string("encoder_layers_2_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_2_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_2_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72044032)))]; + tensor linear_26_cast_fp16 = linear(bias = encoder_layers_2_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_2_feed_forward2_linear1_weight_to_fp16_palettized, x = input_169_cast_fp16)[name = string("linear_26_cast_fp16")]; + tensor input_173_cast_fp16 = silu(x = linear_26_cast_fp16)[name = string("input_173_cast_fp16")]; + tensor encoder_layers_2_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72052288))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75198080))))[name = string("encoder_layers_2_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_2_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_2_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75198272)))]; + tensor linear_27_cast_fp16 = linear(bias = encoder_layers_2_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_2_feed_forward2_linear2_weight_to_fp16_palettized, x = input_173_cast_fp16)[name = string("linear_27_cast_fp16")]; + fp16 var_1057_to_fp16 = const()[name = string("op_1057_to_fp16"), val = fp16(0x1p-1)]; + tensor var_1058_cast_fp16 = mul(x = linear_27_cast_fp16, y = var_1057_to_fp16)[name = string("op_1058_cast_fp16")]; + tensor input_179_cast_fp16 = add(x = input_167_cast_fp16, y = var_1058_cast_fp16)[name = string("input_179_cast_fp16")]; + tensor input_181_axes_0 = const()[name = string("input_181_axes_0"), val = tensor([-1])]; + tensor encoder_layers_2_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_2_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75200384)))]; + tensor encoder_layers_2_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_2_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75202496)))]; + tensor input_181_cast_fp16 = layer_norm(axes = input_181_axes_0, beta = encoder_layers_2_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_2_norm_out_weight_to_fp16, x = input_179_cast_fp16)[name = string("input_181_cast_fp16")]; + tensor cache_13_begin_0 = const()[name = string("cache_13_begin_0"), val = tensor([3, 0, 0, 0])]; + tensor cache_13_end_0 = const()[name = string("cache_13_end_0"), val = tensor([4, 1, 42, 1024])]; + tensor cache_13_end_mask_0 = const()[name = string("cache_13_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_13_squeeze_mask_0 = const()[name = string("cache_13_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_13_cast_fp16 = slice_by_index(begin = cache_13_begin_0, end = cache_13_end_0, end_mask = cache_13_end_mask_0, squeeze_mask = cache_13_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_13_cast_fp16")]; + tensor cache_15_begin_0 = const()[name = string("cache_15_begin_0"), val = tensor([3, 0, 0, 0])]; + tensor cache_15_end_0 = const()[name = string("cache_15_end_0"), val = tensor([4, 1, 1024, 8])]; + tensor cache_15_end_mask_0 = const()[name = string("cache_15_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_15_squeeze_mask_0 = const()[name = string("cache_15_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_15_cast_fp16 = slice_by_index(begin = cache_15_begin_0, end = cache_15_end_0, end_mask = cache_15_end_mask_0, squeeze_mask = cache_15_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_15_cast_fp16")]; + tensor input_183_axes_0 = const()[name = string("input_183_axes_0"), val = tensor([-1])]; + tensor encoder_layers_3_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_3_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75204608)))]; + tensor encoder_layers_3_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_3_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75206720)))]; + tensor input_183_cast_fp16 = layer_norm(axes = input_183_axes_0, beta = encoder_layers_3_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_3_norm_feed_forward1_weight_to_fp16, x = input_181_cast_fp16)[name = string("input_183_cast_fp16")]; + tensor encoder_layers_3_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75208832))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78354624))))[name = string("encoder_layers_3_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_3_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_3_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78354816)))]; + tensor linear_28_cast_fp16 = linear(bias = encoder_layers_3_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_3_feed_forward1_linear1_weight_to_fp16_palettized, x = input_183_cast_fp16)[name = string("linear_28_cast_fp16")]; + tensor input_187_cast_fp16 = silu(x = linear_28_cast_fp16)[name = string("input_187_cast_fp16")]; + tensor encoder_layers_3_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78363072))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81508864))))[name = string("encoder_layers_3_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_3_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_3_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81509056)))]; + tensor linear_29_cast_fp16 = linear(bias = encoder_layers_3_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_3_feed_forward1_linear2_weight_to_fp16_palettized, x = input_187_cast_fp16)[name = string("linear_29_cast_fp16")]; + fp16 var_1094_to_fp16 = const()[name = string("op_1094_to_fp16"), val = fp16(0x1p-1)]; + tensor var_1095_cast_fp16 = mul(x = linear_29_cast_fp16, y = var_1094_to_fp16)[name = string("op_1095_cast_fp16")]; + tensor input_193_cast_fp16 = add(x = input_181_cast_fp16, y = var_1095_cast_fp16)[name = string("input_193_cast_fp16")]; + tensor key_7_axes_0 = const()[name = string("key_7_axes_0"), val = tensor([-1])]; + tensor encoder_layers_3_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_3_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81511168)))]; + tensor encoder_layers_3_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_3_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81513280)))]; + tensor key_7_cast_fp16 = layer_norm(axes = key_7_axes_0, beta = encoder_layers_3_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_3_norm_self_att_weight_to_fp16, x = input_193_cast_fp16)[name = string("key_7_cast_fp16")]; + bool input_195_interleave_0 = const()[name = string("input_195_interleave_0"), val = bool(false)]; + tensor input_195_cast_fp16 = concat(axis = var_68, interleave = input_195_interleave_0, values = (cache_13_cast_fp16, key_7_cast_fp16))[name = string("input_195_cast_fp16")]; + tensor var_1117_begin_0 = const()[name = string("op_1117_begin_0"), val = tensor([0, 7, 0])]; + tensor var_1117_end_0 = const()[name = string("op_1117_end_0"), val = tensor([1, 42, 1024])]; + tensor var_1117_end_mask_0 = const()[name = string("op_1117_end_mask_0"), val = tensor([true, true, true])]; + tensor var_1117_cast_fp16 = slice_by_index(begin = var_1117_begin_0, end = var_1117_end_0, end_mask = var_1117_end_mask_0, x = cache_13_cast_fp16)[name = string("op_1117_cast_fp16")]; + bool var_1123_interleave_0 = const()[name = string("op_1123_interleave_0"), val = bool(false)]; + tensor var_1123_cast_fp16 = concat(axis = var_68, interleave = var_1123_interleave_0, values = (var_1117_cast_fp16, key_7_cast_fp16))[name = string("op_1123_cast_fp16")]; + tensor encoder_layers_3_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81515392))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82301888))))[name = string("encoder_layers_3_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_3_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_3_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82302080)))]; + tensor linear_30_cast_fp16 = linear(bias = encoder_layers_3_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_3_self_attn_linear_q_weight_to_fp16_palettized, x = key_7_cast_fp16)[name = string("linear_30_cast_fp16")]; + tensor var_1128 = const()[name = string("op_1128"), val = tensor([1, -1, 8, 128])]; + tensor q_19_cast_fp16 = reshape(shape = var_1128, x = linear_30_cast_fp16)[name = string("q_19_cast_fp16")]; + tensor encoder_layers_3_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82304192))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83090688))))[name = string("encoder_layers_3_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_3_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_3_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83090880)))]; + tensor linear_31_cast_fp16 = linear(bias = encoder_layers_3_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_3_self_attn_linear_k_weight_to_fp16_palettized, x = input_195_cast_fp16)[name = string("linear_31_cast_fp16")]; + tensor var_1133 = const()[name = string("op_1133"), val = tensor([1, -1, 8, 128])]; + tensor k_13_cast_fp16 = reshape(shape = var_1133, x = linear_31_cast_fp16)[name = string("k_13_cast_fp16")]; + tensor encoder_layers_3_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83092992))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83879488))))[name = string("encoder_layers_3_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_3_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_3_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83879680)))]; + tensor linear_32_cast_fp16 = linear(bias = encoder_layers_3_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_3_self_attn_linear_v_weight_to_fp16_palettized, x = input_195_cast_fp16)[name = string("linear_32_cast_fp16")]; + tensor var_1138 = const()[name = string("op_1138"), val = tensor([1, -1, 8, 128])]; + tensor v_7_cast_fp16 = reshape(shape = var_1138, x = linear_32_cast_fp16)[name = string("v_7_cast_fp16")]; + tensor value_15_perm_0 = const()[name = string("value_15_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_3_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_3_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83881792)))]; + tensor var_1151_cast_fp16 = add(x = q_19_cast_fp16, y = encoder_layers_3_self_attn_pos_bias_u_to_fp16)[name = string("op_1151_cast_fp16")]; + tensor encoder_layers_3_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_3_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83883904)))]; + tensor var_1153_cast_fp16 = add(x = q_19_cast_fp16, y = encoder_layers_3_self_attn_pos_bias_v_to_fp16)[name = string("op_1153_cast_fp16")]; + tensor q_with_bias_v_7_perm_0 = const()[name = string("q_with_bias_v_7_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_85_transpose_x_0 = const()[name = string("x_85_transpose_x_0"), val = bool(false)]; + bool x_85_transpose_y_0 = const()[name = string("x_85_transpose_y_0"), val = bool(false)]; + tensor op_1155_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83886016))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83985408))))[name = string("op_1155_to_fp16_quantized")]; + tensor q_with_bias_v_7_cast_fp16 = transpose(perm = q_with_bias_v_7_perm_0, x = var_1153_cast_fp16)[name = string("transpose_335")]; + tensor x_85_cast_fp16 = matmul(transpose_x = x_85_transpose_x_0, transpose_y = x_85_transpose_y_0, x = q_with_bias_v_7_cast_fp16, y = op_1155_to_fp16_quantized)[name = string("x_85_cast_fp16")]; + tensor x_87_pad_0 = const()[name = string("x_87_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_87_mode_0 = const()[name = string("x_87_mode_0"), val = string("constant")]; + fp16 const_118_to_fp16 = const()[name = string("const_118_to_fp16"), val = fp16(0x0p+0)]; + tensor x_87_cast_fp16 = pad(constant_val = const_118_to_fp16, mode = x_87_mode_0, pad = x_87_pad_0, x = x_85_cast_fp16)[name = string("x_87_cast_fp16")]; + tensor var_1163 = const()[name = string("op_1163"), val = tensor([1, 8, -1, 7])]; + tensor x_89_cast_fp16 = reshape(shape = var_1163, x = x_87_cast_fp16)[name = string("x_89_cast_fp16")]; + tensor var_1167_begin_0 = const()[name = string("op_1167_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_1167_end_0 = const()[name = string("op_1167_end_0"), val = tensor([1, 8, 98, 7])]; + tensor var_1167_end_mask_0 = const()[name = string("op_1167_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_1167_cast_fp16 = slice_by_index(begin = var_1167_begin_0, end = var_1167_end_0, end_mask = var_1167_end_mask_0, x = x_89_cast_fp16)[name = string("op_1167_cast_fp16")]; + tensor var_1168 = const()[name = string("op_1168"), val = tensor([1, 8, 7, 97])]; + tensor matrix_bd_13_cast_fp16 = reshape(shape = var_1168, x = var_1167_cast_fp16)[name = string("matrix_bd_13_cast_fp16")]; + bool matrix_ac_7_transpose_x_0 = const()[name = string("matrix_ac_7_transpose_x_0"), val = bool(false)]; + bool matrix_ac_7_transpose_y_0 = const()[name = string("matrix_ac_7_transpose_y_0"), val = bool(false)]; + tensor transpose_102_perm_0 = const()[name = string("transpose_102_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_103_perm_0 = const()[name = string("transpose_103_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_103 = transpose(perm = transpose_103_perm_0, x = k_13_cast_fp16)[name = string("transpose_333")]; + tensor transpose_102 = transpose(perm = transpose_102_perm_0, x = var_1151_cast_fp16)[name = string("transpose_334")]; + tensor matrix_ac_7_cast_fp16 = matmul(transpose_x = matrix_ac_7_transpose_x_0, transpose_y = matrix_ac_7_transpose_y_0, x = transpose_102, y = transpose_103)[name = string("matrix_ac_7_cast_fp16")]; + tensor matrix_bd_15_begin_0 = const()[name = string("matrix_bd_15_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_15_end_0 = const()[name = string("matrix_bd_15_end_0"), val = tensor([1, 8, 7, 49])]; + tensor matrix_bd_15_end_mask_0 = const()[name = string("matrix_bd_15_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_15_cast_fp16 = slice_by_index(begin = matrix_bd_15_begin_0, end = matrix_bd_15_end_0, end_mask = matrix_bd_15_end_mask_0, x = matrix_bd_13_cast_fp16)[name = string("matrix_bd_15_cast_fp16")]; + tensor var_1177_cast_fp16 = add(x = matrix_ac_7_cast_fp16, y = matrix_bd_15_cast_fp16)[name = string("op_1177_cast_fp16")]; + fp16 _inversed_scores_13_y_0_to_fp16 = const()[name = string("_inversed_scores_13_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_13_cast_fp16 = mul(x = var_1177_cast_fp16, y = _inversed_scores_13_y_0_to_fp16)[name = string("_inversed_scores_13_cast_fp16")]; + tensor scores_15_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_13_cast_fp16, cond = mask_11)[name = string("scores_15_cast_fp16")]; + tensor var_1183_cast_fp16 = softmax(axis = var_59, x = scores_15_cast_fp16)[name = string("op_1183_cast_fp16")]; + tensor input_197_cast_fp16 = select(a = var_44_to_fp16, b = var_1183_cast_fp16, cond = mask_11)[name = string("input_197_cast_fp16")]; + bool x_91_transpose_x_0 = const()[name = string("x_91_transpose_x_0"), val = bool(false)]; + bool x_91_transpose_y_0 = const()[name = string("x_91_transpose_y_0"), val = bool(false)]; + tensor value_15_cast_fp16 = transpose(perm = value_15_perm_0, x = v_7_cast_fp16)[name = string("transpose_332")]; + tensor x_91_cast_fp16 = matmul(transpose_x = x_91_transpose_x_0, transpose_y = x_91_transpose_y_0, x = input_197_cast_fp16, y = value_15_cast_fp16)[name = string("x_91_cast_fp16")]; + tensor var_1187_perm_0 = const()[name = string("op_1187_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1188 = const()[name = string("op_1188"), val = tensor([1, -1, 1024])]; + tensor var_1187_cast_fp16 = transpose(perm = var_1187_perm_0, x = x_91_cast_fp16)[name = string("transpose_331")]; + tensor input_199_cast_fp16 = reshape(shape = var_1188, x = var_1187_cast_fp16)[name = string("input_199_cast_fp16")]; + tensor encoder_layers_3_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83985728))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84772224))))[name = string("encoder_layers_3_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_3_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_3_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84772416)))]; + tensor linear_34_cast_fp16 = linear(bias = encoder_layers_3_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_3_self_attn_linear_out_weight_to_fp16_palettized, x = input_199_cast_fp16)[name = string("linear_34_cast_fp16")]; + tensor input_203_cast_fp16 = add(x = input_193_cast_fp16, y = linear_34_cast_fp16)[name = string("input_203_cast_fp16")]; + tensor x_95_axes_0 = const()[name = string("x_95_axes_0"), val = tensor([-1])]; + tensor encoder_layers_3_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_3_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84774528)))]; + tensor encoder_layers_3_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_3_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84776640)))]; + tensor x_95_cast_fp16 = layer_norm(axes = x_95_axes_0, beta = encoder_layers_3_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_3_norm_conv_weight_to_fp16, x = input_203_cast_fp16)[name = string("x_95_cast_fp16")]; + tensor input_205_perm_0 = const()[name = string("input_205_perm_0"), val = tensor([0, 2, 1])]; + string input_207_pad_type_0 = const()[name = string("input_207_pad_type_0"), val = string("valid")]; + tensor input_207_strides_0 = const()[name = string("input_207_strides_0"), val = tensor([1])]; + tensor input_207_pad_0 = const()[name = string("input_207_pad_0"), val = tensor([0, 0])]; + tensor input_207_dilations_0 = const()[name = string("input_207_dilations_0"), val = tensor([1])]; + int32 input_207_groups_0 = const()[name = string("input_207_groups_0"), val = int32(1)]; + tensor encoder_layers_3_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84778752))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86875968))))[name = string("encoder_layers_3_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_205_cast_fp16 = transpose(perm = input_205_perm_0, x = x_95_cast_fp16)[name = string("transpose_330")]; + tensor input_207_cast_fp16 = conv(dilations = input_207_dilations_0, groups = input_207_groups_0, pad = input_207_pad_0, pad_type = input_207_pad_type_0, strides = input_207_strides_0, weight = encoder_layers_3_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_205_cast_fp16)[name = string("input_207_cast_fp16")]; + int32 x_97_split_num_splits_0 = const()[name = string("x_97_split_num_splits_0"), val = int32(2)]; + int32 x_97_split_axis_0 = const()[name = string("x_97_split_axis_0"), val = int32(1)]; + tensor x_97_split_cast_fp16_0, tensor x_97_split_cast_fp16_1 = split(axis = x_97_split_axis_0, num_splits = x_97_split_num_splits_0, x = input_207_cast_fp16)[name = string("x_97_split_cast_fp16")]; + tensor x_97_split_1_sigmoid_cast_fp16 = sigmoid(x = x_97_split_cast_fp16_1)[name = string("x_97_split_1_sigmoid_cast_fp16")]; + tensor x_97_cast_fp16 = mul(x = x_97_split_cast_fp16_0, y = x_97_split_1_sigmoid_cast_fp16)[name = string("x_97_cast_fp16")]; + tensor input_209_cast_fp16 = select(a = var_44_to_fp16, b = x_97_cast_fp16, cond = var_575)[name = string("input_209_cast_fp16")]; + bool new_x_15_interleave_0 = const()[name = string("new_x_15_interleave_0"), val = bool(false)]; + tensor new_x_15_cast_fp16 = concat(axis = var_59, interleave = new_x_15_interleave_0, values = (cache_15_cast_fp16, input_209_cast_fp16))[name = string("new_x_15_cast_fp16")]; + tensor var_1227_begin_0 = const()[name = string("op_1227_begin_0"), val = tensor([0, 0, 7])]; + tensor var_1227_end_0 = const()[name = string("op_1227_end_0"), val = tensor([1, 1024, 15])]; + tensor var_1227_end_mask_0 = const()[name = string("op_1227_end_mask_0"), val = tensor([true, true, true])]; + tensor var_1227_cast_fp16 = slice_by_index(begin = var_1227_begin_0, end = var_1227_end_0, end_mask = var_1227_end_mask_0, x = new_x_15_cast_fp16)[name = string("op_1227_cast_fp16")]; + string x_99_pad_type_0 = const()[name = string("x_99_pad_type_0"), val = string("valid")]; + int32 x_99_groups_0 = const()[name = string("x_99_groups_0"), val = int32(1024)]; + tensor x_99_strides_0 = const()[name = string("x_99_strides_0"), val = tensor([1])]; + tensor x_99_pad_0 = const()[name = string("x_99_pad_0"), val = tensor([0, 0])]; + tensor x_99_dilations_0 = const()[name = string("x_99_dilations_0"), val = tensor([1])]; + tensor encoder_layers_3_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86880128))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86889408))))[name = string("encoder_layers_3_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_99_cast_fp16 = conv(dilations = x_99_dilations_0, groups = x_99_groups_0, pad = x_99_pad_0, pad_type = x_99_pad_type_0, strides = x_99_strides_0, weight = encoder_layers_3_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_15_cast_fp16)[name = string("x_99_cast_fp16")]; + tensor input_211_perm_0 = const()[name = string("input_211_perm_0"), val = tensor([0, 2, 1])]; + tensor x_101_axes_0 = const()[name = string("x_101_axes_0"), val = tensor([-1])]; + tensor encoder_layers_3_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_3_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86891520)))]; + tensor encoder_layers_3_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_3_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86893632)))]; + tensor input_211_cast_fp16 = transpose(perm = input_211_perm_0, x = x_99_cast_fp16)[name = string("transpose_329")]; + tensor x_101_cast_fp16 = layer_norm(axes = x_101_axes_0, beta = encoder_layers_3_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_3_conv_batch_norm_weight_to_fp16, x = input_211_cast_fp16)[name = string("x_101_cast_fp16")]; + tensor input_213_perm_0 = const()[name = string("input_213_perm_0"), val = tensor([0, 2, 1])]; + tensor input_213_cast_fp16 = transpose(perm = input_213_perm_0, x = x_101_cast_fp16)[name = string("transpose_328")]; + tensor input_215_cast_fp16 = silu(x = input_213_cast_fp16)[name = string("input_215_cast_fp16")]; + string x_103_pad_type_0 = const()[name = string("x_103_pad_type_0"), val = string("valid")]; + tensor x_103_strides_0 = const()[name = string("x_103_strides_0"), val = tensor([1])]; + tensor x_103_pad_0 = const()[name = string("x_103_pad_0"), val = tensor([0, 0])]; + tensor x_103_dilations_0 = const()[name = string("x_103_dilations_0"), val = tensor([1])]; + int32 x_103_groups_0 = const()[name = string("x_103_groups_0"), val = int32(1)]; + tensor encoder_layers_3_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86895744))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87944384))))[name = string("encoder_layers_3_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_103_cast_fp16 = conv(dilations = x_103_dilations_0, groups = x_103_groups_0, pad = x_103_pad_0, pad_type = x_103_pad_type_0, strides = x_103_strides_0, weight = encoder_layers_3_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_215_cast_fp16)[name = string("x_103_cast_fp16")]; + tensor input_217_perm_0 = const()[name = string("input_217_perm_0"), val = tensor([0, 2, 1])]; + tensor input_217_cast_fp16 = transpose(perm = input_217_perm_0, x = x_103_cast_fp16)[name = string("transpose_327")]; + tensor input_219_cast_fp16 = add(x = input_203_cast_fp16, y = input_217_cast_fp16)[name = string("input_219_cast_fp16")]; + tensor input_221_axes_0 = const()[name = string("input_221_axes_0"), val = tensor([-1])]; + tensor encoder_layers_3_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_3_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87946496)))]; + tensor encoder_layers_3_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_3_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87948608)))]; + tensor input_221_cast_fp16 = layer_norm(axes = input_221_axes_0, beta = encoder_layers_3_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_3_norm_feed_forward2_weight_to_fp16, x = input_219_cast_fp16)[name = string("input_221_cast_fp16")]; + tensor encoder_layers_3_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87950720))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(91096512))))[name = string("encoder_layers_3_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_3_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_3_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(91096704)))]; + tensor linear_35_cast_fp16 = linear(bias = encoder_layers_3_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_3_feed_forward2_linear1_weight_to_fp16_palettized, x = input_221_cast_fp16)[name = string("linear_35_cast_fp16")]; + tensor input_225_cast_fp16 = silu(x = linear_35_cast_fp16)[name = string("input_225_cast_fp16")]; + tensor encoder_layers_3_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(91104960))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94250752))))[name = string("encoder_layers_3_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_3_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_3_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94250944)))]; + tensor linear_36_cast_fp16 = linear(bias = encoder_layers_3_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_3_feed_forward2_linear2_weight_to_fp16_palettized, x = input_225_cast_fp16)[name = string("linear_36_cast_fp16")]; + fp16 var_1270_to_fp16 = const()[name = string("op_1270_to_fp16"), val = fp16(0x1p-1)]; + tensor var_1271_cast_fp16 = mul(x = linear_36_cast_fp16, y = var_1270_to_fp16)[name = string("op_1271_cast_fp16")]; + tensor input_231_cast_fp16 = add(x = input_219_cast_fp16, y = var_1271_cast_fp16)[name = string("input_231_cast_fp16")]; + tensor input_233_axes_0 = const()[name = string("input_233_axes_0"), val = tensor([-1])]; + tensor encoder_layers_3_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_3_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94253056)))]; + tensor encoder_layers_3_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_3_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94255168)))]; + tensor input_233_cast_fp16 = layer_norm(axes = input_233_axes_0, beta = encoder_layers_3_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_3_norm_out_weight_to_fp16, x = input_231_cast_fp16)[name = string("input_233_cast_fp16")]; + tensor cache_17_begin_0 = const()[name = string("cache_17_begin_0"), val = tensor([4, 0, 0, 0])]; + tensor cache_17_end_0 = const()[name = string("cache_17_end_0"), val = tensor([5, 1, 42, 1024])]; + tensor cache_17_end_mask_0 = const()[name = string("cache_17_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_17_squeeze_mask_0 = const()[name = string("cache_17_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_17_cast_fp16 = slice_by_index(begin = cache_17_begin_0, end = cache_17_end_0, end_mask = cache_17_end_mask_0, squeeze_mask = cache_17_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_17_cast_fp16")]; + tensor cache_19_begin_0 = const()[name = string("cache_19_begin_0"), val = tensor([4, 0, 0, 0])]; + tensor cache_19_end_0 = const()[name = string("cache_19_end_0"), val = tensor([5, 1, 1024, 8])]; + tensor cache_19_end_mask_0 = const()[name = string("cache_19_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_19_squeeze_mask_0 = const()[name = string("cache_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_19_cast_fp16 = slice_by_index(begin = cache_19_begin_0, end = cache_19_end_0, end_mask = cache_19_end_mask_0, squeeze_mask = cache_19_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_19_cast_fp16")]; + tensor input_235_axes_0 = const()[name = string("input_235_axes_0"), val = tensor([-1])]; + tensor encoder_layers_4_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_4_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94257280)))]; + tensor encoder_layers_4_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_4_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94259392)))]; + tensor input_235_cast_fp16 = layer_norm(axes = input_235_axes_0, beta = encoder_layers_4_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_4_norm_feed_forward1_weight_to_fp16, x = input_233_cast_fp16)[name = string("input_235_cast_fp16")]; + tensor encoder_layers_4_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94261504))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97407296))))[name = string("encoder_layers_4_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_4_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_4_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97407488)))]; + tensor linear_37_cast_fp16 = linear(bias = encoder_layers_4_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_4_feed_forward1_linear1_weight_to_fp16_palettized, x = input_235_cast_fp16)[name = string("linear_37_cast_fp16")]; + tensor input_239_cast_fp16 = silu(x = linear_37_cast_fp16)[name = string("input_239_cast_fp16")]; + tensor encoder_layers_4_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97415744))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100561536))))[name = string("encoder_layers_4_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_4_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_4_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100561728)))]; + tensor linear_38_cast_fp16 = linear(bias = encoder_layers_4_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_4_feed_forward1_linear2_weight_to_fp16_palettized, x = input_239_cast_fp16)[name = string("linear_38_cast_fp16")]; + fp16 var_1307_to_fp16 = const()[name = string("op_1307_to_fp16"), val = fp16(0x1p-1)]; + tensor var_1308_cast_fp16 = mul(x = linear_38_cast_fp16, y = var_1307_to_fp16)[name = string("op_1308_cast_fp16")]; + tensor input_245_cast_fp16 = add(x = input_233_cast_fp16, y = var_1308_cast_fp16)[name = string("input_245_cast_fp16")]; + tensor key_9_axes_0 = const()[name = string("key_9_axes_0"), val = tensor([-1])]; + tensor encoder_layers_4_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_4_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100563840)))]; + tensor encoder_layers_4_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_4_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100565952)))]; + tensor key_9_cast_fp16 = layer_norm(axes = key_9_axes_0, beta = encoder_layers_4_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_4_norm_self_att_weight_to_fp16, x = input_245_cast_fp16)[name = string("key_9_cast_fp16")]; + bool input_247_interleave_0 = const()[name = string("input_247_interleave_0"), val = bool(false)]; + tensor input_247_cast_fp16 = concat(axis = var_68, interleave = input_247_interleave_0, values = (cache_17_cast_fp16, key_9_cast_fp16))[name = string("input_247_cast_fp16")]; + tensor var_1330_begin_0 = const()[name = string("op_1330_begin_0"), val = tensor([0, 7, 0])]; + tensor var_1330_end_0 = const()[name = string("op_1330_end_0"), val = tensor([1, 42, 1024])]; + tensor var_1330_end_mask_0 = const()[name = string("op_1330_end_mask_0"), val = tensor([true, true, true])]; + tensor var_1330_cast_fp16 = slice_by_index(begin = var_1330_begin_0, end = var_1330_end_0, end_mask = var_1330_end_mask_0, x = cache_17_cast_fp16)[name = string("op_1330_cast_fp16")]; + bool var_1336_interleave_0 = const()[name = string("op_1336_interleave_0"), val = bool(false)]; + tensor var_1336_cast_fp16 = concat(axis = var_68, interleave = var_1336_interleave_0, values = (var_1330_cast_fp16, key_9_cast_fp16))[name = string("op_1336_cast_fp16")]; + tensor encoder_layers_4_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100568064))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(101354560))))[name = string("encoder_layers_4_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_4_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_4_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(101354752)))]; + tensor linear_39_cast_fp16 = linear(bias = encoder_layers_4_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_4_self_attn_linear_q_weight_to_fp16_palettized, x = key_9_cast_fp16)[name = string("linear_39_cast_fp16")]; + tensor var_1341 = const()[name = string("op_1341"), val = tensor([1, -1, 8, 128])]; + tensor q_25_cast_fp16 = reshape(shape = var_1341, x = linear_39_cast_fp16)[name = string("q_25_cast_fp16")]; + tensor encoder_layers_4_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(101356864))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102143360))))[name = string("encoder_layers_4_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_4_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_4_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102143552)))]; + tensor linear_40_cast_fp16 = linear(bias = encoder_layers_4_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_4_self_attn_linear_k_weight_to_fp16_palettized, x = input_247_cast_fp16)[name = string("linear_40_cast_fp16")]; + tensor var_1346 = const()[name = string("op_1346"), val = tensor([1, -1, 8, 128])]; + tensor k_17_cast_fp16 = reshape(shape = var_1346, x = linear_40_cast_fp16)[name = string("k_17_cast_fp16")]; + tensor encoder_layers_4_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102145664))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102932160))))[name = string("encoder_layers_4_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_4_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_4_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102932352)))]; + tensor linear_41_cast_fp16 = linear(bias = encoder_layers_4_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_4_self_attn_linear_v_weight_to_fp16_palettized, x = input_247_cast_fp16)[name = string("linear_41_cast_fp16")]; + tensor var_1351 = const()[name = string("op_1351"), val = tensor([1, -1, 8, 128])]; + tensor v_9_cast_fp16 = reshape(shape = var_1351, x = linear_41_cast_fp16)[name = string("v_9_cast_fp16")]; + tensor value_17_perm_0 = const()[name = string("value_17_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_4_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_4_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102934464)))]; + tensor var_1364_cast_fp16 = add(x = q_25_cast_fp16, y = encoder_layers_4_self_attn_pos_bias_u_to_fp16)[name = string("op_1364_cast_fp16")]; + tensor encoder_layers_4_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_4_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102936576)))]; + tensor var_1366_cast_fp16 = add(x = q_25_cast_fp16, y = encoder_layers_4_self_attn_pos_bias_v_to_fp16)[name = string("op_1366_cast_fp16")]; + tensor q_with_bias_v_9_perm_0 = const()[name = string("q_with_bias_v_9_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_111_transpose_x_0 = const()[name = string("x_111_transpose_x_0"), val = bool(false)]; + bool x_111_transpose_y_0 = const()[name = string("x_111_transpose_y_0"), val = bool(false)]; + tensor op_1368_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102938688))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103038080))))[name = string("op_1368_to_fp16_quantized")]; + tensor q_with_bias_v_9_cast_fp16 = transpose(perm = q_with_bias_v_9_perm_0, x = var_1366_cast_fp16)[name = string("transpose_326")]; + tensor x_111_cast_fp16 = matmul(transpose_x = x_111_transpose_x_0, transpose_y = x_111_transpose_y_0, x = q_with_bias_v_9_cast_fp16, y = op_1368_to_fp16_quantized)[name = string("x_111_cast_fp16")]; + tensor x_113_pad_0 = const()[name = string("x_113_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_113_mode_0 = const()[name = string("x_113_mode_0"), val = string("constant")]; + fp16 const_131_to_fp16 = const()[name = string("const_131_to_fp16"), val = fp16(0x0p+0)]; + tensor x_113_cast_fp16 = pad(constant_val = const_131_to_fp16, mode = x_113_mode_0, pad = x_113_pad_0, x = x_111_cast_fp16)[name = string("x_113_cast_fp16")]; + tensor var_1376 = const()[name = string("op_1376"), val = tensor([1, 8, -1, 7])]; + tensor x_115_cast_fp16 = reshape(shape = var_1376, x = x_113_cast_fp16)[name = string("x_115_cast_fp16")]; + tensor var_1380_begin_0 = const()[name = string("op_1380_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_1380_end_0 = const()[name = string("op_1380_end_0"), val = tensor([1, 8, 98, 7])]; + tensor var_1380_end_mask_0 = const()[name = string("op_1380_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_1380_cast_fp16 = slice_by_index(begin = var_1380_begin_0, end = var_1380_end_0, end_mask = var_1380_end_mask_0, x = x_115_cast_fp16)[name = string("op_1380_cast_fp16")]; + tensor var_1381 = const()[name = string("op_1381"), val = tensor([1, 8, 7, 97])]; + tensor matrix_bd_17_cast_fp16 = reshape(shape = var_1381, x = var_1380_cast_fp16)[name = string("matrix_bd_17_cast_fp16")]; + bool matrix_ac_9_transpose_x_0 = const()[name = string("matrix_ac_9_transpose_x_0"), val = bool(false)]; + bool matrix_ac_9_transpose_y_0 = const()[name = string("matrix_ac_9_transpose_y_0"), val = bool(false)]; + tensor transpose_104_perm_0 = const()[name = string("transpose_104_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_105_perm_0 = const()[name = string("transpose_105_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_105 = transpose(perm = transpose_105_perm_0, x = k_17_cast_fp16)[name = string("transpose_324")]; + tensor transpose_104 = transpose(perm = transpose_104_perm_0, x = var_1364_cast_fp16)[name = string("transpose_325")]; + tensor matrix_ac_9_cast_fp16 = matmul(transpose_x = matrix_ac_9_transpose_x_0, transpose_y = matrix_ac_9_transpose_y_0, x = transpose_104, y = transpose_105)[name = string("matrix_ac_9_cast_fp16")]; + tensor matrix_bd_19_begin_0 = const()[name = string("matrix_bd_19_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_19_end_0 = const()[name = string("matrix_bd_19_end_0"), val = tensor([1, 8, 7, 49])]; + tensor matrix_bd_19_end_mask_0 = const()[name = string("matrix_bd_19_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_19_cast_fp16 = slice_by_index(begin = matrix_bd_19_begin_0, end = matrix_bd_19_end_0, end_mask = matrix_bd_19_end_mask_0, x = matrix_bd_17_cast_fp16)[name = string("matrix_bd_19_cast_fp16")]; + tensor var_1390_cast_fp16 = add(x = matrix_ac_9_cast_fp16, y = matrix_bd_19_cast_fp16)[name = string("op_1390_cast_fp16")]; + fp16 _inversed_scores_17_y_0_to_fp16 = const()[name = string("_inversed_scores_17_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_17_cast_fp16 = mul(x = var_1390_cast_fp16, y = _inversed_scores_17_y_0_to_fp16)[name = string("_inversed_scores_17_cast_fp16")]; + tensor scores_19_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_17_cast_fp16, cond = mask_11)[name = string("scores_19_cast_fp16")]; + tensor var_1396_cast_fp16 = softmax(axis = var_59, x = scores_19_cast_fp16)[name = string("op_1396_cast_fp16")]; + tensor input_249_cast_fp16 = select(a = var_44_to_fp16, b = var_1396_cast_fp16, cond = mask_11)[name = string("input_249_cast_fp16")]; + bool x_117_transpose_x_0 = const()[name = string("x_117_transpose_x_0"), val = bool(false)]; + bool x_117_transpose_y_0 = const()[name = string("x_117_transpose_y_0"), val = bool(false)]; + tensor value_17_cast_fp16 = transpose(perm = value_17_perm_0, x = v_9_cast_fp16)[name = string("transpose_323")]; + tensor x_117_cast_fp16 = matmul(transpose_x = x_117_transpose_x_0, transpose_y = x_117_transpose_y_0, x = input_249_cast_fp16, y = value_17_cast_fp16)[name = string("x_117_cast_fp16")]; + tensor var_1400_perm_0 = const()[name = string("op_1400_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1401 = const()[name = string("op_1401"), val = tensor([1, -1, 1024])]; + tensor var_1400_cast_fp16 = transpose(perm = var_1400_perm_0, x = x_117_cast_fp16)[name = string("transpose_322")]; + tensor input_251_cast_fp16 = reshape(shape = var_1401, x = var_1400_cast_fp16)[name = string("input_251_cast_fp16")]; + tensor encoder_layers_4_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103038400))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103824896))))[name = string("encoder_layers_4_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_4_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_4_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103825088)))]; + tensor linear_43_cast_fp16 = linear(bias = encoder_layers_4_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_4_self_attn_linear_out_weight_to_fp16_palettized, x = input_251_cast_fp16)[name = string("linear_43_cast_fp16")]; + tensor input_255_cast_fp16 = add(x = input_245_cast_fp16, y = linear_43_cast_fp16)[name = string("input_255_cast_fp16")]; + tensor x_121_axes_0 = const()[name = string("x_121_axes_0"), val = tensor([-1])]; + tensor encoder_layers_4_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_4_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103827200)))]; + tensor encoder_layers_4_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_4_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103829312)))]; + tensor x_121_cast_fp16 = layer_norm(axes = x_121_axes_0, beta = encoder_layers_4_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_4_norm_conv_weight_to_fp16, x = input_255_cast_fp16)[name = string("x_121_cast_fp16")]; + tensor input_257_perm_0 = const()[name = string("input_257_perm_0"), val = tensor([0, 2, 1])]; + string input_259_pad_type_0 = const()[name = string("input_259_pad_type_0"), val = string("valid")]; + tensor input_259_strides_0 = const()[name = string("input_259_strides_0"), val = tensor([1])]; + tensor input_259_pad_0 = const()[name = string("input_259_pad_0"), val = tensor([0, 0])]; + tensor input_259_dilations_0 = const()[name = string("input_259_dilations_0"), val = tensor([1])]; + int32 input_259_groups_0 = const()[name = string("input_259_groups_0"), val = int32(1)]; + tensor encoder_layers_4_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103831424))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105928640))))[name = string("encoder_layers_4_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_257_cast_fp16 = transpose(perm = input_257_perm_0, x = x_121_cast_fp16)[name = string("transpose_321")]; + tensor input_259_cast_fp16 = conv(dilations = input_259_dilations_0, groups = input_259_groups_0, pad = input_259_pad_0, pad_type = input_259_pad_type_0, strides = input_259_strides_0, weight = encoder_layers_4_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_257_cast_fp16)[name = string("input_259_cast_fp16")]; + int32 x_123_split_num_splits_0 = const()[name = string("x_123_split_num_splits_0"), val = int32(2)]; + int32 x_123_split_axis_0 = const()[name = string("x_123_split_axis_0"), val = int32(1)]; + tensor x_123_split_cast_fp16_0, tensor x_123_split_cast_fp16_1 = split(axis = x_123_split_axis_0, num_splits = x_123_split_num_splits_0, x = input_259_cast_fp16)[name = string("x_123_split_cast_fp16")]; + tensor x_123_split_1_sigmoid_cast_fp16 = sigmoid(x = x_123_split_cast_fp16_1)[name = string("x_123_split_1_sigmoid_cast_fp16")]; + tensor x_123_cast_fp16 = mul(x = x_123_split_cast_fp16_0, y = x_123_split_1_sigmoid_cast_fp16)[name = string("x_123_cast_fp16")]; + tensor input_261_cast_fp16 = select(a = var_44_to_fp16, b = x_123_cast_fp16, cond = var_575)[name = string("input_261_cast_fp16")]; + bool new_x_19_interleave_0 = const()[name = string("new_x_19_interleave_0"), val = bool(false)]; + tensor new_x_19_cast_fp16 = concat(axis = var_59, interleave = new_x_19_interleave_0, values = (cache_19_cast_fp16, input_261_cast_fp16))[name = string("new_x_19_cast_fp16")]; + tensor var_1440_begin_0 = const()[name = string("op_1440_begin_0"), val = tensor([0, 0, 7])]; + tensor var_1440_end_0 = const()[name = string("op_1440_end_0"), val = tensor([1, 1024, 15])]; + tensor var_1440_end_mask_0 = const()[name = string("op_1440_end_mask_0"), val = tensor([true, true, true])]; + tensor var_1440_cast_fp16 = slice_by_index(begin = var_1440_begin_0, end = var_1440_end_0, end_mask = var_1440_end_mask_0, x = new_x_19_cast_fp16)[name = string("op_1440_cast_fp16")]; + string x_125_pad_type_0 = const()[name = string("x_125_pad_type_0"), val = string("valid")]; + int32 x_125_groups_0 = const()[name = string("x_125_groups_0"), val = int32(1024)]; + tensor x_125_strides_0 = const()[name = string("x_125_strides_0"), val = tensor([1])]; + tensor x_125_pad_0 = const()[name = string("x_125_pad_0"), val = tensor([0, 0])]; + tensor x_125_dilations_0 = const()[name = string("x_125_dilations_0"), val = tensor([1])]; + tensor encoder_layers_4_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105932800))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105942080))))[name = string("encoder_layers_4_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_125_cast_fp16 = conv(dilations = x_125_dilations_0, groups = x_125_groups_0, pad = x_125_pad_0, pad_type = x_125_pad_type_0, strides = x_125_strides_0, weight = encoder_layers_4_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_19_cast_fp16)[name = string("x_125_cast_fp16")]; + tensor input_263_perm_0 = const()[name = string("input_263_perm_0"), val = tensor([0, 2, 1])]; + tensor x_127_axes_0 = const()[name = string("x_127_axes_0"), val = tensor([-1])]; + tensor encoder_layers_4_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_4_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105944192)))]; + tensor encoder_layers_4_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_4_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105946304)))]; + tensor input_263_cast_fp16 = transpose(perm = input_263_perm_0, x = x_125_cast_fp16)[name = string("transpose_320")]; + tensor x_127_cast_fp16 = layer_norm(axes = x_127_axes_0, beta = encoder_layers_4_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_4_conv_batch_norm_weight_to_fp16, x = input_263_cast_fp16)[name = string("x_127_cast_fp16")]; + tensor input_265_perm_0 = const()[name = string("input_265_perm_0"), val = tensor([0, 2, 1])]; + tensor input_265_cast_fp16 = transpose(perm = input_265_perm_0, x = x_127_cast_fp16)[name = string("transpose_319")]; + tensor input_267_cast_fp16 = silu(x = input_265_cast_fp16)[name = string("input_267_cast_fp16")]; + string x_129_pad_type_0 = const()[name = string("x_129_pad_type_0"), val = string("valid")]; + tensor x_129_strides_0 = const()[name = string("x_129_strides_0"), val = tensor([1])]; + tensor x_129_pad_0 = const()[name = string("x_129_pad_0"), val = tensor([0, 0])]; + tensor x_129_dilations_0 = const()[name = string("x_129_dilations_0"), val = tensor([1])]; + int32 x_129_groups_0 = const()[name = string("x_129_groups_0"), val = int32(1)]; + tensor encoder_layers_4_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105948416))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106997056))))[name = string("encoder_layers_4_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_129_cast_fp16 = conv(dilations = x_129_dilations_0, groups = x_129_groups_0, pad = x_129_pad_0, pad_type = x_129_pad_type_0, strides = x_129_strides_0, weight = encoder_layers_4_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_267_cast_fp16)[name = string("x_129_cast_fp16")]; + tensor input_269_perm_0 = const()[name = string("input_269_perm_0"), val = tensor([0, 2, 1])]; + tensor input_269_cast_fp16 = transpose(perm = input_269_perm_0, x = x_129_cast_fp16)[name = string("transpose_318")]; + tensor input_271_cast_fp16 = add(x = input_255_cast_fp16, y = input_269_cast_fp16)[name = string("input_271_cast_fp16")]; + tensor input_273_axes_0 = const()[name = string("input_273_axes_0"), val = tensor([-1])]; + tensor encoder_layers_4_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_4_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106999168)))]; + tensor encoder_layers_4_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_4_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107001280)))]; + tensor input_273_cast_fp16 = layer_norm(axes = input_273_axes_0, beta = encoder_layers_4_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_4_norm_feed_forward2_weight_to_fp16, x = input_271_cast_fp16)[name = string("input_273_cast_fp16")]; + tensor encoder_layers_4_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107003392))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110149184))))[name = string("encoder_layers_4_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_4_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_4_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110149376)))]; + tensor linear_44_cast_fp16 = linear(bias = encoder_layers_4_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_4_feed_forward2_linear1_weight_to_fp16_palettized, x = input_273_cast_fp16)[name = string("linear_44_cast_fp16")]; + tensor input_277_cast_fp16 = silu(x = linear_44_cast_fp16)[name = string("input_277_cast_fp16")]; + tensor encoder_layers_4_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110157632))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113303424))))[name = string("encoder_layers_4_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_4_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_4_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113303616)))]; + tensor linear_45_cast_fp16 = linear(bias = encoder_layers_4_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_4_feed_forward2_linear2_weight_to_fp16_palettized, x = input_277_cast_fp16)[name = string("linear_45_cast_fp16")]; + fp16 var_1483_to_fp16 = const()[name = string("op_1483_to_fp16"), val = fp16(0x1p-1)]; + tensor var_1484_cast_fp16 = mul(x = linear_45_cast_fp16, y = var_1483_to_fp16)[name = string("op_1484_cast_fp16")]; + tensor input_283_cast_fp16 = add(x = input_271_cast_fp16, y = var_1484_cast_fp16)[name = string("input_283_cast_fp16")]; + tensor input_285_axes_0 = const()[name = string("input_285_axes_0"), val = tensor([-1])]; + tensor encoder_layers_4_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_4_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113305728)))]; + tensor encoder_layers_4_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_4_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113307840)))]; + tensor input_285_cast_fp16 = layer_norm(axes = input_285_axes_0, beta = encoder_layers_4_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_4_norm_out_weight_to_fp16, x = input_283_cast_fp16)[name = string("input_285_cast_fp16")]; + tensor cache_21_begin_0 = const()[name = string("cache_21_begin_0"), val = tensor([5, 0, 0, 0])]; + tensor cache_21_end_0 = const()[name = string("cache_21_end_0"), val = tensor([6, 1, 42, 1024])]; + tensor cache_21_end_mask_0 = const()[name = string("cache_21_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_21_squeeze_mask_0 = const()[name = string("cache_21_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_21_cast_fp16 = slice_by_index(begin = cache_21_begin_0, end = cache_21_end_0, end_mask = cache_21_end_mask_0, squeeze_mask = cache_21_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_21_cast_fp16")]; + tensor cache_23_begin_0 = const()[name = string("cache_23_begin_0"), val = tensor([5, 0, 0, 0])]; + tensor cache_23_end_0 = const()[name = string("cache_23_end_0"), val = tensor([6, 1, 1024, 8])]; + tensor cache_23_end_mask_0 = const()[name = string("cache_23_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_23_squeeze_mask_0 = const()[name = string("cache_23_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_23_cast_fp16 = slice_by_index(begin = cache_23_begin_0, end = cache_23_end_0, end_mask = cache_23_end_mask_0, squeeze_mask = cache_23_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_23_cast_fp16")]; + tensor input_287_axes_0 = const()[name = string("input_287_axes_0"), val = tensor([-1])]; + tensor encoder_layers_5_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_5_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113309952)))]; + tensor encoder_layers_5_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_5_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113312064)))]; + tensor input_287_cast_fp16 = layer_norm(axes = input_287_axes_0, beta = encoder_layers_5_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_5_norm_feed_forward1_weight_to_fp16, x = input_285_cast_fp16)[name = string("input_287_cast_fp16")]; + tensor encoder_layers_5_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113314176))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116459968))))[name = string("encoder_layers_5_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_5_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_5_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116460160)))]; + tensor linear_46_cast_fp16 = linear(bias = encoder_layers_5_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_5_feed_forward1_linear1_weight_to_fp16_palettized, x = input_287_cast_fp16)[name = string("linear_46_cast_fp16")]; + tensor input_291_cast_fp16 = silu(x = linear_46_cast_fp16)[name = string("input_291_cast_fp16")]; + tensor encoder_layers_5_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116468416))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119614208))))[name = string("encoder_layers_5_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_5_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_5_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119614400)))]; + tensor linear_47_cast_fp16 = linear(bias = encoder_layers_5_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_5_feed_forward1_linear2_weight_to_fp16_palettized, x = input_291_cast_fp16)[name = string("linear_47_cast_fp16")]; + fp16 var_1520_to_fp16 = const()[name = string("op_1520_to_fp16"), val = fp16(0x1p-1)]; + tensor var_1521_cast_fp16 = mul(x = linear_47_cast_fp16, y = var_1520_to_fp16)[name = string("op_1521_cast_fp16")]; + tensor input_297_cast_fp16 = add(x = input_285_cast_fp16, y = var_1521_cast_fp16)[name = string("input_297_cast_fp16")]; + tensor key_11_axes_0 = const()[name = string("key_11_axes_0"), val = tensor([-1])]; + tensor encoder_layers_5_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_5_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119616512)))]; + tensor encoder_layers_5_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_5_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119618624)))]; + tensor key_11_cast_fp16 = layer_norm(axes = key_11_axes_0, beta = encoder_layers_5_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_5_norm_self_att_weight_to_fp16, x = input_297_cast_fp16)[name = string("key_11_cast_fp16")]; + bool input_299_interleave_0 = const()[name = string("input_299_interleave_0"), val = bool(false)]; + tensor input_299_cast_fp16 = concat(axis = var_68, interleave = input_299_interleave_0, values = (cache_21_cast_fp16, key_11_cast_fp16))[name = string("input_299_cast_fp16")]; + tensor var_1543_begin_0 = const()[name = string("op_1543_begin_0"), val = tensor([0, 7, 0])]; + tensor var_1543_end_0 = const()[name = string("op_1543_end_0"), val = tensor([1, 42, 1024])]; + tensor var_1543_end_mask_0 = const()[name = string("op_1543_end_mask_0"), val = tensor([true, true, true])]; + tensor var_1543_cast_fp16 = slice_by_index(begin = var_1543_begin_0, end = var_1543_end_0, end_mask = var_1543_end_mask_0, x = cache_21_cast_fp16)[name = string("op_1543_cast_fp16")]; + bool var_1549_interleave_0 = const()[name = string("op_1549_interleave_0"), val = bool(false)]; + tensor var_1549_cast_fp16 = concat(axis = var_68, interleave = var_1549_interleave_0, values = (var_1543_cast_fp16, key_11_cast_fp16))[name = string("op_1549_cast_fp16")]; + tensor encoder_layers_5_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119620736))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120407232))))[name = string("encoder_layers_5_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_5_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_5_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120407424)))]; + tensor linear_48_cast_fp16 = linear(bias = encoder_layers_5_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_5_self_attn_linear_q_weight_to_fp16_palettized, x = key_11_cast_fp16)[name = string("linear_48_cast_fp16")]; + tensor var_1554 = const()[name = string("op_1554"), val = tensor([1, -1, 8, 128])]; + tensor q_31_cast_fp16 = reshape(shape = var_1554, x = linear_48_cast_fp16)[name = string("q_31_cast_fp16")]; + tensor encoder_layers_5_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120409536))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121196032))))[name = string("encoder_layers_5_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_5_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_5_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121196224)))]; + tensor linear_49_cast_fp16 = linear(bias = encoder_layers_5_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_5_self_attn_linear_k_weight_to_fp16_palettized, x = input_299_cast_fp16)[name = string("linear_49_cast_fp16")]; + tensor var_1559 = const()[name = string("op_1559"), val = tensor([1, -1, 8, 128])]; + tensor k_21_cast_fp16 = reshape(shape = var_1559, x = linear_49_cast_fp16)[name = string("k_21_cast_fp16")]; + tensor encoder_layers_5_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121198336))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121984832))))[name = string("encoder_layers_5_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_5_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_5_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121985024)))]; + tensor linear_50_cast_fp16 = linear(bias = encoder_layers_5_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_5_self_attn_linear_v_weight_to_fp16_palettized, x = input_299_cast_fp16)[name = string("linear_50_cast_fp16")]; + tensor var_1564 = const()[name = string("op_1564"), val = tensor([1, -1, 8, 128])]; + tensor v_11_cast_fp16 = reshape(shape = var_1564, x = linear_50_cast_fp16)[name = string("v_11_cast_fp16")]; + tensor value_19_perm_0 = const()[name = string("value_19_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_5_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_5_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121987136)))]; + tensor var_1577_cast_fp16 = add(x = q_31_cast_fp16, y = encoder_layers_5_self_attn_pos_bias_u_to_fp16)[name = string("op_1577_cast_fp16")]; + tensor encoder_layers_5_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_5_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121989248)))]; + tensor var_1579_cast_fp16 = add(x = q_31_cast_fp16, y = encoder_layers_5_self_attn_pos_bias_v_to_fp16)[name = string("op_1579_cast_fp16")]; + tensor q_with_bias_v_11_perm_0 = const()[name = string("q_with_bias_v_11_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_137_transpose_x_0 = const()[name = string("x_137_transpose_x_0"), val = bool(false)]; + bool x_137_transpose_y_0 = const()[name = string("x_137_transpose_y_0"), val = bool(false)]; + tensor op_1581_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121991360))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122090752))))[name = string("op_1581_to_fp16_quantized")]; + tensor q_with_bias_v_11_cast_fp16 = transpose(perm = q_with_bias_v_11_perm_0, x = var_1579_cast_fp16)[name = string("transpose_317")]; + tensor x_137_cast_fp16 = matmul(transpose_x = x_137_transpose_x_0, transpose_y = x_137_transpose_y_0, x = q_with_bias_v_11_cast_fp16, y = op_1581_to_fp16_quantized)[name = string("x_137_cast_fp16")]; + tensor x_139_pad_0 = const()[name = string("x_139_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_139_mode_0 = const()[name = string("x_139_mode_0"), val = string("constant")]; + fp16 const_144_to_fp16 = const()[name = string("const_144_to_fp16"), val = fp16(0x0p+0)]; + tensor x_139_cast_fp16 = pad(constant_val = const_144_to_fp16, mode = x_139_mode_0, pad = x_139_pad_0, x = x_137_cast_fp16)[name = string("x_139_cast_fp16")]; + tensor var_1589 = const()[name = string("op_1589"), val = tensor([1, 8, -1, 7])]; + tensor x_141_cast_fp16 = reshape(shape = var_1589, x = x_139_cast_fp16)[name = string("x_141_cast_fp16")]; + tensor var_1593_begin_0 = const()[name = string("op_1593_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_1593_end_0 = const()[name = string("op_1593_end_0"), val = tensor([1, 8, 98, 7])]; + tensor var_1593_end_mask_0 = const()[name = string("op_1593_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_1593_cast_fp16 = slice_by_index(begin = var_1593_begin_0, end = var_1593_end_0, end_mask = var_1593_end_mask_0, x = x_141_cast_fp16)[name = string("op_1593_cast_fp16")]; + tensor var_1594 = const()[name = string("op_1594"), val = tensor([1, 8, 7, 97])]; + tensor matrix_bd_21_cast_fp16 = reshape(shape = var_1594, x = var_1593_cast_fp16)[name = string("matrix_bd_21_cast_fp16")]; + bool matrix_ac_11_transpose_x_0 = const()[name = string("matrix_ac_11_transpose_x_0"), val = bool(false)]; + bool matrix_ac_11_transpose_y_0 = const()[name = string("matrix_ac_11_transpose_y_0"), val = bool(false)]; + tensor transpose_106_perm_0 = const()[name = string("transpose_106_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_107_perm_0 = const()[name = string("transpose_107_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_107 = transpose(perm = transpose_107_perm_0, x = k_21_cast_fp16)[name = string("transpose_315")]; + tensor transpose_106 = transpose(perm = transpose_106_perm_0, x = var_1577_cast_fp16)[name = string("transpose_316")]; + tensor matrix_ac_11_cast_fp16 = matmul(transpose_x = matrix_ac_11_transpose_x_0, transpose_y = matrix_ac_11_transpose_y_0, x = transpose_106, y = transpose_107)[name = string("matrix_ac_11_cast_fp16")]; + tensor matrix_bd_23_begin_0 = const()[name = string("matrix_bd_23_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_23_end_0 = const()[name = string("matrix_bd_23_end_0"), val = tensor([1, 8, 7, 49])]; + tensor matrix_bd_23_end_mask_0 = const()[name = string("matrix_bd_23_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_23_cast_fp16 = slice_by_index(begin = matrix_bd_23_begin_0, end = matrix_bd_23_end_0, end_mask = matrix_bd_23_end_mask_0, x = matrix_bd_21_cast_fp16)[name = string("matrix_bd_23_cast_fp16")]; + tensor var_1603_cast_fp16 = add(x = matrix_ac_11_cast_fp16, y = matrix_bd_23_cast_fp16)[name = string("op_1603_cast_fp16")]; + fp16 _inversed_scores_21_y_0_to_fp16 = const()[name = string("_inversed_scores_21_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_21_cast_fp16 = mul(x = var_1603_cast_fp16, y = _inversed_scores_21_y_0_to_fp16)[name = string("_inversed_scores_21_cast_fp16")]; + tensor scores_23_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_21_cast_fp16, cond = mask_11)[name = string("scores_23_cast_fp16")]; + tensor var_1609_cast_fp16 = softmax(axis = var_59, x = scores_23_cast_fp16)[name = string("op_1609_cast_fp16")]; + tensor input_301_cast_fp16 = select(a = var_44_to_fp16, b = var_1609_cast_fp16, cond = mask_11)[name = string("input_301_cast_fp16")]; + bool x_143_transpose_x_0 = const()[name = string("x_143_transpose_x_0"), val = bool(false)]; + bool x_143_transpose_y_0 = const()[name = string("x_143_transpose_y_0"), val = bool(false)]; + tensor value_19_cast_fp16 = transpose(perm = value_19_perm_0, x = v_11_cast_fp16)[name = string("transpose_314")]; + tensor x_143_cast_fp16 = matmul(transpose_x = x_143_transpose_x_0, transpose_y = x_143_transpose_y_0, x = input_301_cast_fp16, y = value_19_cast_fp16)[name = string("x_143_cast_fp16")]; + tensor var_1613_perm_0 = const()[name = string("op_1613_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1614 = const()[name = string("op_1614"), val = tensor([1, -1, 1024])]; + tensor var_1613_cast_fp16 = transpose(perm = var_1613_perm_0, x = x_143_cast_fp16)[name = string("transpose_313")]; + tensor input_303_cast_fp16 = reshape(shape = var_1614, x = var_1613_cast_fp16)[name = string("input_303_cast_fp16")]; + tensor encoder_layers_5_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122091072))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122877568))))[name = string("encoder_layers_5_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_5_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_5_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122877760)))]; + tensor linear_52_cast_fp16 = linear(bias = encoder_layers_5_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_5_self_attn_linear_out_weight_to_fp16_palettized, x = input_303_cast_fp16)[name = string("linear_52_cast_fp16")]; + tensor input_307_cast_fp16 = add(x = input_297_cast_fp16, y = linear_52_cast_fp16)[name = string("input_307_cast_fp16")]; + tensor x_147_axes_0 = const()[name = string("x_147_axes_0"), val = tensor([-1])]; + tensor encoder_layers_5_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_5_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122879872)))]; + tensor encoder_layers_5_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_5_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122881984)))]; + tensor x_147_cast_fp16 = layer_norm(axes = x_147_axes_0, beta = encoder_layers_5_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_5_norm_conv_weight_to_fp16, x = input_307_cast_fp16)[name = string("x_147_cast_fp16")]; + tensor input_309_perm_0 = const()[name = string("input_309_perm_0"), val = tensor([0, 2, 1])]; + string input_311_pad_type_0 = const()[name = string("input_311_pad_type_0"), val = string("valid")]; + tensor input_311_strides_0 = const()[name = string("input_311_strides_0"), val = tensor([1])]; + tensor input_311_pad_0 = const()[name = string("input_311_pad_0"), val = tensor([0, 0])]; + tensor input_311_dilations_0 = const()[name = string("input_311_dilations_0"), val = tensor([1])]; + int32 input_311_groups_0 = const()[name = string("input_311_groups_0"), val = int32(1)]; + tensor encoder_layers_5_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122884096))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(124981312))))[name = string("encoder_layers_5_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_309_cast_fp16 = transpose(perm = input_309_perm_0, x = x_147_cast_fp16)[name = string("transpose_312")]; + tensor input_311_cast_fp16 = conv(dilations = input_311_dilations_0, groups = input_311_groups_0, pad = input_311_pad_0, pad_type = input_311_pad_type_0, strides = input_311_strides_0, weight = encoder_layers_5_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_309_cast_fp16)[name = string("input_311_cast_fp16")]; + int32 x_149_split_num_splits_0 = const()[name = string("x_149_split_num_splits_0"), val = int32(2)]; + int32 x_149_split_axis_0 = const()[name = string("x_149_split_axis_0"), val = int32(1)]; + tensor x_149_split_cast_fp16_0, tensor x_149_split_cast_fp16_1 = split(axis = x_149_split_axis_0, num_splits = x_149_split_num_splits_0, x = input_311_cast_fp16)[name = string("x_149_split_cast_fp16")]; + tensor x_149_split_1_sigmoid_cast_fp16 = sigmoid(x = x_149_split_cast_fp16_1)[name = string("x_149_split_1_sigmoid_cast_fp16")]; + tensor x_149_cast_fp16 = mul(x = x_149_split_cast_fp16_0, y = x_149_split_1_sigmoid_cast_fp16)[name = string("x_149_cast_fp16")]; + tensor input_313_cast_fp16 = select(a = var_44_to_fp16, b = x_149_cast_fp16, cond = var_575)[name = string("input_313_cast_fp16")]; + bool new_x_23_interleave_0 = const()[name = string("new_x_23_interleave_0"), val = bool(false)]; + tensor new_x_23_cast_fp16 = concat(axis = var_59, interleave = new_x_23_interleave_0, values = (cache_23_cast_fp16, input_313_cast_fp16))[name = string("new_x_23_cast_fp16")]; + tensor var_1653_begin_0 = const()[name = string("op_1653_begin_0"), val = tensor([0, 0, 7])]; + tensor var_1653_end_0 = const()[name = string("op_1653_end_0"), val = tensor([1, 1024, 15])]; + tensor var_1653_end_mask_0 = const()[name = string("op_1653_end_mask_0"), val = tensor([true, true, true])]; + tensor var_1653_cast_fp16 = slice_by_index(begin = var_1653_begin_0, end = var_1653_end_0, end_mask = var_1653_end_mask_0, x = new_x_23_cast_fp16)[name = string("op_1653_cast_fp16")]; + string x_151_pad_type_0 = const()[name = string("x_151_pad_type_0"), val = string("valid")]; + int32 x_151_groups_0 = const()[name = string("x_151_groups_0"), val = int32(1024)]; + tensor x_151_strides_0 = const()[name = string("x_151_strides_0"), val = tensor([1])]; + tensor x_151_pad_0 = const()[name = string("x_151_pad_0"), val = tensor([0, 0])]; + tensor x_151_dilations_0 = const()[name = string("x_151_dilations_0"), val = tensor([1])]; + tensor encoder_layers_5_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(124985472))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(124994752))))[name = string("encoder_layers_5_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_151_cast_fp16 = conv(dilations = x_151_dilations_0, groups = x_151_groups_0, pad = x_151_pad_0, pad_type = x_151_pad_type_0, strides = x_151_strides_0, weight = encoder_layers_5_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_23_cast_fp16)[name = string("x_151_cast_fp16")]; + tensor input_315_perm_0 = const()[name = string("input_315_perm_0"), val = tensor([0, 2, 1])]; + tensor x_153_axes_0 = const()[name = string("x_153_axes_0"), val = tensor([-1])]; + tensor encoder_layers_5_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_5_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(124996864)))]; + tensor encoder_layers_5_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_5_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(124998976)))]; + tensor input_315_cast_fp16 = transpose(perm = input_315_perm_0, x = x_151_cast_fp16)[name = string("transpose_311")]; + tensor x_153_cast_fp16 = layer_norm(axes = x_153_axes_0, beta = encoder_layers_5_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_5_conv_batch_norm_weight_to_fp16, x = input_315_cast_fp16)[name = string("x_153_cast_fp16")]; + tensor input_317_perm_0 = const()[name = string("input_317_perm_0"), val = tensor([0, 2, 1])]; + tensor input_317_cast_fp16 = transpose(perm = input_317_perm_0, x = x_153_cast_fp16)[name = string("transpose_310")]; + tensor input_319_cast_fp16 = silu(x = input_317_cast_fp16)[name = string("input_319_cast_fp16")]; + string x_155_pad_type_0 = const()[name = string("x_155_pad_type_0"), val = string("valid")]; + tensor x_155_strides_0 = const()[name = string("x_155_strides_0"), val = tensor([1])]; + tensor x_155_pad_0 = const()[name = string("x_155_pad_0"), val = tensor([0, 0])]; + tensor x_155_dilations_0 = const()[name = string("x_155_dilations_0"), val = tensor([1])]; + int32 x_155_groups_0 = const()[name = string("x_155_groups_0"), val = int32(1)]; + tensor encoder_layers_5_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125001088))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126049728))))[name = string("encoder_layers_5_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_155_cast_fp16 = conv(dilations = x_155_dilations_0, groups = x_155_groups_0, pad = x_155_pad_0, pad_type = x_155_pad_type_0, strides = x_155_strides_0, weight = encoder_layers_5_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_319_cast_fp16)[name = string("x_155_cast_fp16")]; + tensor input_321_perm_0 = const()[name = string("input_321_perm_0"), val = tensor([0, 2, 1])]; + tensor input_321_cast_fp16 = transpose(perm = input_321_perm_0, x = x_155_cast_fp16)[name = string("transpose_309")]; + tensor input_323_cast_fp16 = add(x = input_307_cast_fp16, y = input_321_cast_fp16)[name = string("input_323_cast_fp16")]; + tensor input_325_axes_0 = const()[name = string("input_325_axes_0"), val = tensor([-1])]; + tensor encoder_layers_5_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_5_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126051840)))]; + tensor encoder_layers_5_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_5_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126053952)))]; + tensor input_325_cast_fp16 = layer_norm(axes = input_325_axes_0, beta = encoder_layers_5_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_5_norm_feed_forward2_weight_to_fp16, x = input_323_cast_fp16)[name = string("input_325_cast_fp16")]; + tensor encoder_layers_5_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126056064))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(129201856))))[name = string("encoder_layers_5_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_5_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_5_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(129202048)))]; + tensor linear_53_cast_fp16 = linear(bias = encoder_layers_5_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_5_feed_forward2_linear1_weight_to_fp16_palettized, x = input_325_cast_fp16)[name = string("linear_53_cast_fp16")]; + tensor input_329_cast_fp16 = silu(x = linear_53_cast_fp16)[name = string("input_329_cast_fp16")]; + tensor encoder_layers_5_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(129210304))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132356096))))[name = string("encoder_layers_5_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_5_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_5_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132356288)))]; + tensor linear_54_cast_fp16 = linear(bias = encoder_layers_5_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_5_feed_forward2_linear2_weight_to_fp16_palettized, x = input_329_cast_fp16)[name = string("linear_54_cast_fp16")]; + fp16 var_1696_to_fp16 = const()[name = string("op_1696_to_fp16"), val = fp16(0x1p-1)]; + tensor var_1697_cast_fp16 = mul(x = linear_54_cast_fp16, y = var_1696_to_fp16)[name = string("op_1697_cast_fp16")]; + tensor input_335_cast_fp16 = add(x = input_323_cast_fp16, y = var_1697_cast_fp16)[name = string("input_335_cast_fp16")]; + tensor input_337_axes_0 = const()[name = string("input_337_axes_0"), val = tensor([-1])]; + tensor encoder_layers_5_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_5_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132358400)))]; + tensor encoder_layers_5_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_5_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132360512)))]; + tensor input_337_cast_fp16 = layer_norm(axes = input_337_axes_0, beta = encoder_layers_5_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_5_norm_out_weight_to_fp16, x = input_335_cast_fp16)[name = string("input_337_cast_fp16")]; + tensor cache_25_begin_0 = const()[name = string("cache_25_begin_0"), val = tensor([6, 0, 0, 0])]; + tensor cache_25_end_0 = const()[name = string("cache_25_end_0"), val = tensor([7, 1, 42, 1024])]; + tensor cache_25_end_mask_0 = const()[name = string("cache_25_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_25_squeeze_mask_0 = const()[name = string("cache_25_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_25_cast_fp16 = slice_by_index(begin = cache_25_begin_0, end = cache_25_end_0, end_mask = cache_25_end_mask_0, squeeze_mask = cache_25_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_25_cast_fp16")]; + tensor cache_27_begin_0 = const()[name = string("cache_27_begin_0"), val = tensor([6, 0, 0, 0])]; + tensor cache_27_end_0 = const()[name = string("cache_27_end_0"), val = tensor([7, 1, 1024, 8])]; + tensor cache_27_end_mask_0 = const()[name = string("cache_27_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_27_squeeze_mask_0 = const()[name = string("cache_27_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_27_cast_fp16 = slice_by_index(begin = cache_27_begin_0, end = cache_27_end_0, end_mask = cache_27_end_mask_0, squeeze_mask = cache_27_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_27_cast_fp16")]; + tensor input_339_axes_0 = const()[name = string("input_339_axes_0"), val = tensor([-1])]; + tensor encoder_layers_6_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_6_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132362624)))]; + tensor encoder_layers_6_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_6_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132364736)))]; + tensor input_339_cast_fp16 = layer_norm(axes = input_339_axes_0, beta = encoder_layers_6_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_6_norm_feed_forward1_weight_to_fp16, x = input_337_cast_fp16)[name = string("input_339_cast_fp16")]; + tensor encoder_layers_6_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132366848))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135512640))))[name = string("encoder_layers_6_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_6_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_6_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135512832)))]; + tensor linear_55_cast_fp16 = linear(bias = encoder_layers_6_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_6_feed_forward1_linear1_weight_to_fp16_palettized, x = input_339_cast_fp16)[name = string("linear_55_cast_fp16")]; + tensor input_343_cast_fp16 = silu(x = linear_55_cast_fp16)[name = string("input_343_cast_fp16")]; + tensor encoder_layers_6_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135521088))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138666880))))[name = string("encoder_layers_6_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_6_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_6_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138667072)))]; + tensor linear_56_cast_fp16 = linear(bias = encoder_layers_6_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_6_feed_forward1_linear2_weight_to_fp16_palettized, x = input_343_cast_fp16)[name = string("linear_56_cast_fp16")]; + fp16 var_1733_to_fp16 = const()[name = string("op_1733_to_fp16"), val = fp16(0x1p-1)]; + tensor var_1734_cast_fp16 = mul(x = linear_56_cast_fp16, y = var_1733_to_fp16)[name = string("op_1734_cast_fp16")]; + tensor input_349_cast_fp16 = add(x = input_337_cast_fp16, y = var_1734_cast_fp16)[name = string("input_349_cast_fp16")]; + tensor key_13_axes_0 = const()[name = string("key_13_axes_0"), val = tensor([-1])]; + tensor encoder_layers_6_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_6_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138669184)))]; + tensor encoder_layers_6_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_6_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138671296)))]; + tensor key_13_cast_fp16 = layer_norm(axes = key_13_axes_0, beta = encoder_layers_6_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_6_norm_self_att_weight_to_fp16, x = input_349_cast_fp16)[name = string("key_13_cast_fp16")]; + bool input_351_interleave_0 = const()[name = string("input_351_interleave_0"), val = bool(false)]; + tensor input_351_cast_fp16 = concat(axis = var_68, interleave = input_351_interleave_0, values = (cache_25_cast_fp16, key_13_cast_fp16))[name = string("input_351_cast_fp16")]; + tensor var_1756_begin_0 = const()[name = string("op_1756_begin_0"), val = tensor([0, 7, 0])]; + tensor var_1756_end_0 = const()[name = string("op_1756_end_0"), val = tensor([1, 42, 1024])]; + tensor var_1756_end_mask_0 = const()[name = string("op_1756_end_mask_0"), val = tensor([true, true, true])]; + tensor var_1756_cast_fp16 = slice_by_index(begin = var_1756_begin_0, end = var_1756_end_0, end_mask = var_1756_end_mask_0, x = cache_25_cast_fp16)[name = string("op_1756_cast_fp16")]; + bool var_1762_interleave_0 = const()[name = string("op_1762_interleave_0"), val = bool(false)]; + tensor var_1762_cast_fp16 = concat(axis = var_68, interleave = var_1762_interleave_0, values = (var_1756_cast_fp16, key_13_cast_fp16))[name = string("op_1762_cast_fp16")]; + tensor encoder_layers_6_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138673408))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(139459904))))[name = string("encoder_layers_6_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_6_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_6_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(139460096)))]; + tensor linear_57_cast_fp16 = linear(bias = encoder_layers_6_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_6_self_attn_linear_q_weight_to_fp16_palettized, x = key_13_cast_fp16)[name = string("linear_57_cast_fp16")]; + tensor var_1767 = const()[name = string("op_1767"), val = tensor([1, -1, 8, 128])]; + tensor q_37_cast_fp16 = reshape(shape = var_1767, x = linear_57_cast_fp16)[name = string("q_37_cast_fp16")]; + tensor encoder_layers_6_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(139462208))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140248704))))[name = string("encoder_layers_6_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_6_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_6_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140248896)))]; + tensor linear_58_cast_fp16 = linear(bias = encoder_layers_6_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_6_self_attn_linear_k_weight_to_fp16_palettized, x = input_351_cast_fp16)[name = string("linear_58_cast_fp16")]; + tensor var_1772 = const()[name = string("op_1772"), val = tensor([1, -1, 8, 128])]; + tensor k_25_cast_fp16 = reshape(shape = var_1772, x = linear_58_cast_fp16)[name = string("k_25_cast_fp16")]; + tensor encoder_layers_6_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140251008))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141037504))))[name = string("encoder_layers_6_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_6_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_6_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141037696)))]; + tensor linear_59_cast_fp16 = linear(bias = encoder_layers_6_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_6_self_attn_linear_v_weight_to_fp16_palettized, x = input_351_cast_fp16)[name = string("linear_59_cast_fp16")]; + tensor var_1777 = const()[name = string("op_1777"), val = tensor([1, -1, 8, 128])]; + tensor v_13_cast_fp16 = reshape(shape = var_1777, x = linear_59_cast_fp16)[name = string("v_13_cast_fp16")]; + tensor value_21_perm_0 = const()[name = string("value_21_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_6_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_6_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141039808)))]; + tensor var_1790_cast_fp16 = add(x = q_37_cast_fp16, y = encoder_layers_6_self_attn_pos_bias_u_to_fp16)[name = string("op_1790_cast_fp16")]; + tensor encoder_layers_6_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_6_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141041920)))]; + tensor var_1792_cast_fp16 = add(x = q_37_cast_fp16, y = encoder_layers_6_self_attn_pos_bias_v_to_fp16)[name = string("op_1792_cast_fp16")]; + tensor q_with_bias_v_13_perm_0 = const()[name = string("q_with_bias_v_13_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_163_transpose_x_0 = const()[name = string("x_163_transpose_x_0"), val = bool(false)]; + bool x_163_transpose_y_0 = const()[name = string("x_163_transpose_y_0"), val = bool(false)]; + tensor op_1794_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141044032))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141143424))))[name = string("op_1794_to_fp16_quantized")]; + tensor q_with_bias_v_13_cast_fp16 = transpose(perm = q_with_bias_v_13_perm_0, x = var_1792_cast_fp16)[name = string("transpose_308")]; + tensor x_163_cast_fp16 = matmul(transpose_x = x_163_transpose_x_0, transpose_y = x_163_transpose_y_0, x = q_with_bias_v_13_cast_fp16, y = op_1794_to_fp16_quantized)[name = string("x_163_cast_fp16")]; + tensor x_165_pad_0 = const()[name = string("x_165_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_165_mode_0 = const()[name = string("x_165_mode_0"), val = string("constant")]; + fp16 const_157_to_fp16 = const()[name = string("const_157_to_fp16"), val = fp16(0x0p+0)]; + tensor x_165_cast_fp16 = pad(constant_val = const_157_to_fp16, mode = x_165_mode_0, pad = x_165_pad_0, x = x_163_cast_fp16)[name = string("x_165_cast_fp16")]; + tensor var_1802 = const()[name = string("op_1802"), val = tensor([1, 8, -1, 7])]; + tensor x_167_cast_fp16 = reshape(shape = var_1802, x = x_165_cast_fp16)[name = string("x_167_cast_fp16")]; + tensor var_1806_begin_0 = const()[name = string("op_1806_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_1806_end_0 = const()[name = string("op_1806_end_0"), val = tensor([1, 8, 98, 7])]; + tensor var_1806_end_mask_0 = const()[name = string("op_1806_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_1806_cast_fp16 = slice_by_index(begin = var_1806_begin_0, end = var_1806_end_0, end_mask = var_1806_end_mask_0, x = x_167_cast_fp16)[name = string("op_1806_cast_fp16")]; + tensor var_1807 = const()[name = string("op_1807"), val = tensor([1, 8, 7, 97])]; + tensor matrix_bd_25_cast_fp16 = reshape(shape = var_1807, x = var_1806_cast_fp16)[name = string("matrix_bd_25_cast_fp16")]; + bool matrix_ac_13_transpose_x_0 = const()[name = string("matrix_ac_13_transpose_x_0"), val = bool(false)]; + bool matrix_ac_13_transpose_y_0 = const()[name = string("matrix_ac_13_transpose_y_0"), val = bool(false)]; + tensor transpose_108_perm_0 = const()[name = string("transpose_108_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_109_perm_0 = const()[name = string("transpose_109_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_109 = transpose(perm = transpose_109_perm_0, x = k_25_cast_fp16)[name = string("transpose_306")]; + tensor transpose_108 = transpose(perm = transpose_108_perm_0, x = var_1790_cast_fp16)[name = string("transpose_307")]; + tensor matrix_ac_13_cast_fp16 = matmul(transpose_x = matrix_ac_13_transpose_x_0, transpose_y = matrix_ac_13_transpose_y_0, x = transpose_108, y = transpose_109)[name = string("matrix_ac_13_cast_fp16")]; + tensor matrix_bd_27_begin_0 = const()[name = string("matrix_bd_27_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_27_end_0 = const()[name = string("matrix_bd_27_end_0"), val = tensor([1, 8, 7, 49])]; + tensor matrix_bd_27_end_mask_0 = const()[name = string("matrix_bd_27_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_27_cast_fp16 = slice_by_index(begin = matrix_bd_27_begin_0, end = matrix_bd_27_end_0, end_mask = matrix_bd_27_end_mask_0, x = matrix_bd_25_cast_fp16)[name = string("matrix_bd_27_cast_fp16")]; + tensor var_1816_cast_fp16 = add(x = matrix_ac_13_cast_fp16, y = matrix_bd_27_cast_fp16)[name = string("op_1816_cast_fp16")]; + fp16 _inversed_scores_25_y_0_to_fp16 = const()[name = string("_inversed_scores_25_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_25_cast_fp16 = mul(x = var_1816_cast_fp16, y = _inversed_scores_25_y_0_to_fp16)[name = string("_inversed_scores_25_cast_fp16")]; + tensor scores_27_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_25_cast_fp16, cond = mask_11)[name = string("scores_27_cast_fp16")]; + tensor var_1822_cast_fp16 = softmax(axis = var_59, x = scores_27_cast_fp16)[name = string("op_1822_cast_fp16")]; + tensor input_353_cast_fp16 = select(a = var_44_to_fp16, b = var_1822_cast_fp16, cond = mask_11)[name = string("input_353_cast_fp16")]; + bool x_169_transpose_x_0 = const()[name = string("x_169_transpose_x_0"), val = bool(false)]; + bool x_169_transpose_y_0 = const()[name = string("x_169_transpose_y_0"), val = bool(false)]; + tensor value_21_cast_fp16 = transpose(perm = value_21_perm_0, x = v_13_cast_fp16)[name = string("transpose_305")]; + tensor x_169_cast_fp16 = matmul(transpose_x = x_169_transpose_x_0, transpose_y = x_169_transpose_y_0, x = input_353_cast_fp16, y = value_21_cast_fp16)[name = string("x_169_cast_fp16")]; + tensor var_1826_perm_0 = const()[name = string("op_1826_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1827 = const()[name = string("op_1827"), val = tensor([1, -1, 1024])]; + tensor var_1826_cast_fp16 = transpose(perm = var_1826_perm_0, x = x_169_cast_fp16)[name = string("transpose_304")]; + tensor input_355_cast_fp16 = reshape(shape = var_1827, x = var_1826_cast_fp16)[name = string("input_355_cast_fp16")]; + tensor encoder_layers_6_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141143744))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141930240))))[name = string("encoder_layers_6_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_6_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_6_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141930432)))]; + tensor linear_61_cast_fp16 = linear(bias = encoder_layers_6_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_6_self_attn_linear_out_weight_to_fp16_palettized, x = input_355_cast_fp16)[name = string("linear_61_cast_fp16")]; + tensor input_359_cast_fp16 = add(x = input_349_cast_fp16, y = linear_61_cast_fp16)[name = string("input_359_cast_fp16")]; + tensor x_173_axes_0 = const()[name = string("x_173_axes_0"), val = tensor([-1])]; + tensor encoder_layers_6_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_6_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141932544)))]; + tensor encoder_layers_6_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_6_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141934656)))]; + tensor x_173_cast_fp16 = layer_norm(axes = x_173_axes_0, beta = encoder_layers_6_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_6_norm_conv_weight_to_fp16, x = input_359_cast_fp16)[name = string("x_173_cast_fp16")]; + tensor input_361_perm_0 = const()[name = string("input_361_perm_0"), val = tensor([0, 2, 1])]; + string input_363_pad_type_0 = const()[name = string("input_363_pad_type_0"), val = string("valid")]; + tensor input_363_strides_0 = const()[name = string("input_363_strides_0"), val = tensor([1])]; + tensor input_363_pad_0 = const()[name = string("input_363_pad_0"), val = tensor([0, 0])]; + tensor input_363_dilations_0 = const()[name = string("input_363_dilations_0"), val = tensor([1])]; + int32 input_363_groups_0 = const()[name = string("input_363_groups_0"), val = int32(1)]; + tensor encoder_layers_6_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141936768))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144033984))))[name = string("encoder_layers_6_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_361_cast_fp16 = transpose(perm = input_361_perm_0, x = x_173_cast_fp16)[name = string("transpose_303")]; + tensor input_363_cast_fp16 = conv(dilations = input_363_dilations_0, groups = input_363_groups_0, pad = input_363_pad_0, pad_type = input_363_pad_type_0, strides = input_363_strides_0, weight = encoder_layers_6_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_361_cast_fp16)[name = string("input_363_cast_fp16")]; + int32 x_175_split_num_splits_0 = const()[name = string("x_175_split_num_splits_0"), val = int32(2)]; + int32 x_175_split_axis_0 = const()[name = string("x_175_split_axis_0"), val = int32(1)]; + tensor x_175_split_cast_fp16_0, tensor x_175_split_cast_fp16_1 = split(axis = x_175_split_axis_0, num_splits = x_175_split_num_splits_0, x = input_363_cast_fp16)[name = string("x_175_split_cast_fp16")]; + tensor x_175_split_1_sigmoid_cast_fp16 = sigmoid(x = x_175_split_cast_fp16_1)[name = string("x_175_split_1_sigmoid_cast_fp16")]; + tensor x_175_cast_fp16 = mul(x = x_175_split_cast_fp16_0, y = x_175_split_1_sigmoid_cast_fp16)[name = string("x_175_cast_fp16")]; + tensor input_365_cast_fp16 = select(a = var_44_to_fp16, b = x_175_cast_fp16, cond = var_575)[name = string("input_365_cast_fp16")]; + bool new_x_27_interleave_0 = const()[name = string("new_x_27_interleave_0"), val = bool(false)]; + tensor new_x_27_cast_fp16 = concat(axis = var_59, interleave = new_x_27_interleave_0, values = (cache_27_cast_fp16, input_365_cast_fp16))[name = string("new_x_27_cast_fp16")]; + tensor var_1866_begin_0 = const()[name = string("op_1866_begin_0"), val = tensor([0, 0, 7])]; + tensor var_1866_end_0 = const()[name = string("op_1866_end_0"), val = tensor([1, 1024, 15])]; + tensor var_1866_end_mask_0 = const()[name = string("op_1866_end_mask_0"), val = tensor([true, true, true])]; + tensor var_1866_cast_fp16 = slice_by_index(begin = var_1866_begin_0, end = var_1866_end_0, end_mask = var_1866_end_mask_0, x = new_x_27_cast_fp16)[name = string("op_1866_cast_fp16")]; + string x_177_pad_type_0 = const()[name = string("x_177_pad_type_0"), val = string("valid")]; + int32 x_177_groups_0 = const()[name = string("x_177_groups_0"), val = int32(1024)]; + tensor x_177_strides_0 = const()[name = string("x_177_strides_0"), val = tensor([1])]; + tensor x_177_pad_0 = const()[name = string("x_177_pad_0"), val = tensor([0, 0])]; + tensor x_177_dilations_0 = const()[name = string("x_177_dilations_0"), val = tensor([1])]; + tensor encoder_layers_6_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144038144))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144047424))))[name = string("encoder_layers_6_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_177_cast_fp16 = conv(dilations = x_177_dilations_0, groups = x_177_groups_0, pad = x_177_pad_0, pad_type = x_177_pad_type_0, strides = x_177_strides_0, weight = encoder_layers_6_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_27_cast_fp16)[name = string("x_177_cast_fp16")]; + tensor input_367_perm_0 = const()[name = string("input_367_perm_0"), val = tensor([0, 2, 1])]; + tensor x_179_axes_0 = const()[name = string("x_179_axes_0"), val = tensor([-1])]; + tensor encoder_layers_6_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_6_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144049536)))]; + tensor encoder_layers_6_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_6_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144051648)))]; + tensor input_367_cast_fp16 = transpose(perm = input_367_perm_0, x = x_177_cast_fp16)[name = string("transpose_302")]; + tensor x_179_cast_fp16 = layer_norm(axes = x_179_axes_0, beta = encoder_layers_6_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_6_conv_batch_norm_weight_to_fp16, x = input_367_cast_fp16)[name = string("x_179_cast_fp16")]; + tensor input_369_perm_0 = const()[name = string("input_369_perm_0"), val = tensor([0, 2, 1])]; + tensor input_369_cast_fp16 = transpose(perm = input_369_perm_0, x = x_179_cast_fp16)[name = string("transpose_301")]; + tensor input_371_cast_fp16 = silu(x = input_369_cast_fp16)[name = string("input_371_cast_fp16")]; + string x_181_pad_type_0 = const()[name = string("x_181_pad_type_0"), val = string("valid")]; + tensor x_181_strides_0 = const()[name = string("x_181_strides_0"), val = tensor([1])]; + tensor x_181_pad_0 = const()[name = string("x_181_pad_0"), val = tensor([0, 0])]; + tensor x_181_dilations_0 = const()[name = string("x_181_dilations_0"), val = tensor([1])]; + int32 x_181_groups_0 = const()[name = string("x_181_groups_0"), val = int32(1)]; + tensor encoder_layers_6_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144053760))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(145102400))))[name = string("encoder_layers_6_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_181_cast_fp16 = conv(dilations = x_181_dilations_0, groups = x_181_groups_0, pad = x_181_pad_0, pad_type = x_181_pad_type_0, strides = x_181_strides_0, weight = encoder_layers_6_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_371_cast_fp16)[name = string("x_181_cast_fp16")]; + tensor input_373_perm_0 = const()[name = string("input_373_perm_0"), val = tensor([0, 2, 1])]; + tensor input_373_cast_fp16 = transpose(perm = input_373_perm_0, x = x_181_cast_fp16)[name = string("transpose_300")]; + tensor input_375_cast_fp16 = add(x = input_359_cast_fp16, y = input_373_cast_fp16)[name = string("input_375_cast_fp16")]; + tensor input_377_axes_0 = const()[name = string("input_377_axes_0"), val = tensor([-1])]; + tensor encoder_layers_6_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_6_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(145104512)))]; + tensor encoder_layers_6_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_6_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(145106624)))]; + tensor input_377_cast_fp16 = layer_norm(axes = input_377_axes_0, beta = encoder_layers_6_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_6_norm_feed_forward2_weight_to_fp16, x = input_375_cast_fp16)[name = string("input_377_cast_fp16")]; + tensor encoder_layers_6_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(145108736))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(148254528))))[name = string("encoder_layers_6_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_6_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_6_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(148254720)))]; + tensor linear_62_cast_fp16 = linear(bias = encoder_layers_6_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_6_feed_forward2_linear1_weight_to_fp16_palettized, x = input_377_cast_fp16)[name = string("linear_62_cast_fp16")]; + tensor input_381_cast_fp16 = silu(x = linear_62_cast_fp16)[name = string("input_381_cast_fp16")]; + tensor encoder_layers_6_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(148262976))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151408768))))[name = string("encoder_layers_6_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_6_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_6_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151408960)))]; + tensor linear_63_cast_fp16 = linear(bias = encoder_layers_6_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_6_feed_forward2_linear2_weight_to_fp16_palettized, x = input_381_cast_fp16)[name = string("linear_63_cast_fp16")]; + fp16 var_1909_to_fp16 = const()[name = string("op_1909_to_fp16"), val = fp16(0x1p-1)]; + tensor var_1910_cast_fp16 = mul(x = linear_63_cast_fp16, y = var_1909_to_fp16)[name = string("op_1910_cast_fp16")]; + tensor input_387_cast_fp16 = add(x = input_375_cast_fp16, y = var_1910_cast_fp16)[name = string("input_387_cast_fp16")]; + tensor input_389_axes_0 = const()[name = string("input_389_axes_0"), val = tensor([-1])]; + tensor encoder_layers_6_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_6_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151411072)))]; + tensor encoder_layers_6_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_6_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151413184)))]; + tensor input_389_cast_fp16 = layer_norm(axes = input_389_axes_0, beta = encoder_layers_6_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_6_norm_out_weight_to_fp16, x = input_387_cast_fp16)[name = string("input_389_cast_fp16")]; + tensor cache_29_begin_0 = const()[name = string("cache_29_begin_0"), val = tensor([7, 0, 0, 0])]; + tensor cache_29_end_0 = const()[name = string("cache_29_end_0"), val = tensor([8, 1, 42, 1024])]; + tensor cache_29_end_mask_0 = const()[name = string("cache_29_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_29_squeeze_mask_0 = const()[name = string("cache_29_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_29_cast_fp16 = slice_by_index(begin = cache_29_begin_0, end = cache_29_end_0, end_mask = cache_29_end_mask_0, squeeze_mask = cache_29_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_29_cast_fp16")]; + tensor cache_31_begin_0 = const()[name = string("cache_31_begin_0"), val = tensor([7, 0, 0, 0])]; + tensor cache_31_end_0 = const()[name = string("cache_31_end_0"), val = tensor([8, 1, 1024, 8])]; + tensor cache_31_end_mask_0 = const()[name = string("cache_31_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_31_squeeze_mask_0 = const()[name = string("cache_31_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_31_cast_fp16 = slice_by_index(begin = cache_31_begin_0, end = cache_31_end_0, end_mask = cache_31_end_mask_0, squeeze_mask = cache_31_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_31_cast_fp16")]; + tensor input_391_axes_0 = const()[name = string("input_391_axes_0"), val = tensor([-1])]; + tensor encoder_layers_7_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_7_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151415296)))]; + tensor encoder_layers_7_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_7_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151417408)))]; + tensor input_391_cast_fp16 = layer_norm(axes = input_391_axes_0, beta = encoder_layers_7_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_7_norm_feed_forward1_weight_to_fp16, x = input_389_cast_fp16)[name = string("input_391_cast_fp16")]; + tensor encoder_layers_7_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151419520))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(154565312))))[name = string("encoder_layers_7_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_7_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_7_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(154565504)))]; + tensor linear_64_cast_fp16 = linear(bias = encoder_layers_7_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_7_feed_forward1_linear1_weight_to_fp16_palettized, x = input_391_cast_fp16)[name = string("linear_64_cast_fp16")]; + tensor input_395_cast_fp16 = silu(x = linear_64_cast_fp16)[name = string("input_395_cast_fp16")]; + tensor encoder_layers_7_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(154573760))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(157719552))))[name = string("encoder_layers_7_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_7_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_7_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(157719744)))]; + tensor linear_65_cast_fp16 = linear(bias = encoder_layers_7_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_7_feed_forward1_linear2_weight_to_fp16_palettized, x = input_395_cast_fp16)[name = string("linear_65_cast_fp16")]; + fp16 var_1946_to_fp16 = const()[name = string("op_1946_to_fp16"), val = fp16(0x1p-1)]; + tensor var_1947_cast_fp16 = mul(x = linear_65_cast_fp16, y = var_1946_to_fp16)[name = string("op_1947_cast_fp16")]; + tensor input_401_cast_fp16 = add(x = input_389_cast_fp16, y = var_1947_cast_fp16)[name = string("input_401_cast_fp16")]; + tensor key_15_axes_0 = const()[name = string("key_15_axes_0"), val = tensor([-1])]; + tensor encoder_layers_7_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_7_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(157721856)))]; + tensor encoder_layers_7_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_7_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(157723968)))]; + tensor key_15_cast_fp16 = layer_norm(axes = key_15_axes_0, beta = encoder_layers_7_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_7_norm_self_att_weight_to_fp16, x = input_401_cast_fp16)[name = string("key_15_cast_fp16")]; + bool input_403_interleave_0 = const()[name = string("input_403_interleave_0"), val = bool(false)]; + tensor input_403_cast_fp16 = concat(axis = var_68, interleave = input_403_interleave_0, values = (cache_29_cast_fp16, key_15_cast_fp16))[name = string("input_403_cast_fp16")]; + tensor var_1969_begin_0 = const()[name = string("op_1969_begin_0"), val = tensor([0, 7, 0])]; + tensor var_1969_end_0 = const()[name = string("op_1969_end_0"), val = tensor([1, 42, 1024])]; + tensor var_1969_end_mask_0 = const()[name = string("op_1969_end_mask_0"), val = tensor([true, true, true])]; + tensor var_1969_cast_fp16 = slice_by_index(begin = var_1969_begin_0, end = var_1969_end_0, end_mask = var_1969_end_mask_0, x = cache_29_cast_fp16)[name = string("op_1969_cast_fp16")]; + bool var_1975_interleave_0 = const()[name = string("op_1975_interleave_0"), val = bool(false)]; + tensor var_1975_cast_fp16 = concat(axis = var_68, interleave = var_1975_interleave_0, values = (var_1969_cast_fp16, key_15_cast_fp16))[name = string("op_1975_cast_fp16")]; + tensor encoder_layers_7_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(157726080))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158512576))))[name = string("encoder_layers_7_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_7_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_7_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158512768)))]; + tensor linear_66_cast_fp16 = linear(bias = encoder_layers_7_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_7_self_attn_linear_q_weight_to_fp16_palettized, x = key_15_cast_fp16)[name = string("linear_66_cast_fp16")]; + tensor var_1980 = const()[name = string("op_1980"), val = tensor([1, -1, 8, 128])]; + tensor q_43_cast_fp16 = reshape(shape = var_1980, x = linear_66_cast_fp16)[name = string("q_43_cast_fp16")]; + tensor encoder_layers_7_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158514880))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(159301376))))[name = string("encoder_layers_7_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_7_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_7_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(159301568)))]; + tensor linear_67_cast_fp16 = linear(bias = encoder_layers_7_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_7_self_attn_linear_k_weight_to_fp16_palettized, x = input_403_cast_fp16)[name = string("linear_67_cast_fp16")]; + tensor var_1985 = const()[name = string("op_1985"), val = tensor([1, -1, 8, 128])]; + tensor k_29_cast_fp16 = reshape(shape = var_1985, x = linear_67_cast_fp16)[name = string("k_29_cast_fp16")]; + tensor encoder_layers_7_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(159303680))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160090176))))[name = string("encoder_layers_7_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_7_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_7_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160090368)))]; + tensor linear_68_cast_fp16 = linear(bias = encoder_layers_7_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_7_self_attn_linear_v_weight_to_fp16_palettized, x = input_403_cast_fp16)[name = string("linear_68_cast_fp16")]; + tensor var_1990 = const()[name = string("op_1990"), val = tensor([1, -1, 8, 128])]; + tensor v_15_cast_fp16 = reshape(shape = var_1990, x = linear_68_cast_fp16)[name = string("v_15_cast_fp16")]; + tensor value_23_perm_0 = const()[name = string("value_23_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_7_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_7_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160092480)))]; + tensor var_2003_cast_fp16 = add(x = q_43_cast_fp16, y = encoder_layers_7_self_attn_pos_bias_u_to_fp16)[name = string("op_2003_cast_fp16")]; + tensor encoder_layers_7_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_7_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160094592)))]; + tensor var_2005_cast_fp16 = add(x = q_43_cast_fp16, y = encoder_layers_7_self_attn_pos_bias_v_to_fp16)[name = string("op_2005_cast_fp16")]; + tensor q_with_bias_v_15_perm_0 = const()[name = string("q_with_bias_v_15_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_189_transpose_x_0 = const()[name = string("x_189_transpose_x_0"), val = bool(false)]; + bool x_189_transpose_y_0 = const()[name = string("x_189_transpose_y_0"), val = bool(false)]; + tensor op_2007_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160096704))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160196096))))[name = string("op_2007_to_fp16_quantized")]; + tensor q_with_bias_v_15_cast_fp16 = transpose(perm = q_with_bias_v_15_perm_0, x = var_2005_cast_fp16)[name = string("transpose_299")]; + tensor x_189_cast_fp16 = matmul(transpose_x = x_189_transpose_x_0, transpose_y = x_189_transpose_y_0, x = q_with_bias_v_15_cast_fp16, y = op_2007_to_fp16_quantized)[name = string("x_189_cast_fp16")]; + tensor x_191_pad_0 = const()[name = string("x_191_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_191_mode_0 = const()[name = string("x_191_mode_0"), val = string("constant")]; + fp16 const_170_to_fp16 = const()[name = string("const_170_to_fp16"), val = fp16(0x0p+0)]; + tensor x_191_cast_fp16 = pad(constant_val = const_170_to_fp16, mode = x_191_mode_0, pad = x_191_pad_0, x = x_189_cast_fp16)[name = string("x_191_cast_fp16")]; + tensor var_2015 = const()[name = string("op_2015"), val = tensor([1, 8, -1, 7])]; + tensor x_193_cast_fp16 = reshape(shape = var_2015, x = x_191_cast_fp16)[name = string("x_193_cast_fp16")]; + tensor var_2019_begin_0 = const()[name = string("op_2019_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_2019_end_0 = const()[name = string("op_2019_end_0"), val = tensor([1, 8, 98, 7])]; + tensor var_2019_end_mask_0 = const()[name = string("op_2019_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_2019_cast_fp16 = slice_by_index(begin = var_2019_begin_0, end = var_2019_end_0, end_mask = var_2019_end_mask_0, x = x_193_cast_fp16)[name = string("op_2019_cast_fp16")]; + tensor var_2020 = const()[name = string("op_2020"), val = tensor([1, 8, 7, 97])]; + tensor matrix_bd_29_cast_fp16 = reshape(shape = var_2020, x = var_2019_cast_fp16)[name = string("matrix_bd_29_cast_fp16")]; + bool matrix_ac_15_transpose_x_0 = const()[name = string("matrix_ac_15_transpose_x_0"), val = bool(false)]; + bool matrix_ac_15_transpose_y_0 = const()[name = string("matrix_ac_15_transpose_y_0"), val = bool(false)]; + tensor transpose_110_perm_0 = const()[name = string("transpose_110_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_111_perm_0 = const()[name = string("transpose_111_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_111 = transpose(perm = transpose_111_perm_0, x = k_29_cast_fp16)[name = string("transpose_297")]; + tensor transpose_110 = transpose(perm = transpose_110_perm_0, x = var_2003_cast_fp16)[name = string("transpose_298")]; + tensor matrix_ac_15_cast_fp16 = matmul(transpose_x = matrix_ac_15_transpose_x_0, transpose_y = matrix_ac_15_transpose_y_0, x = transpose_110, y = transpose_111)[name = string("matrix_ac_15_cast_fp16")]; + tensor matrix_bd_31_begin_0 = const()[name = string("matrix_bd_31_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_31_end_0 = const()[name = string("matrix_bd_31_end_0"), val = tensor([1, 8, 7, 49])]; + tensor matrix_bd_31_end_mask_0 = const()[name = string("matrix_bd_31_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_31_cast_fp16 = slice_by_index(begin = matrix_bd_31_begin_0, end = matrix_bd_31_end_0, end_mask = matrix_bd_31_end_mask_0, x = matrix_bd_29_cast_fp16)[name = string("matrix_bd_31_cast_fp16")]; + tensor var_2029_cast_fp16 = add(x = matrix_ac_15_cast_fp16, y = matrix_bd_31_cast_fp16)[name = string("op_2029_cast_fp16")]; + fp16 _inversed_scores_29_y_0_to_fp16 = const()[name = string("_inversed_scores_29_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_29_cast_fp16 = mul(x = var_2029_cast_fp16, y = _inversed_scores_29_y_0_to_fp16)[name = string("_inversed_scores_29_cast_fp16")]; + tensor scores_31_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_29_cast_fp16, cond = mask_11)[name = string("scores_31_cast_fp16")]; + tensor var_2035_cast_fp16 = softmax(axis = var_59, x = scores_31_cast_fp16)[name = string("op_2035_cast_fp16")]; + tensor input_405_cast_fp16 = select(a = var_44_to_fp16, b = var_2035_cast_fp16, cond = mask_11)[name = string("input_405_cast_fp16")]; + bool x_195_transpose_x_0 = const()[name = string("x_195_transpose_x_0"), val = bool(false)]; + bool x_195_transpose_y_0 = const()[name = string("x_195_transpose_y_0"), val = bool(false)]; + tensor value_23_cast_fp16 = transpose(perm = value_23_perm_0, x = v_15_cast_fp16)[name = string("transpose_296")]; + tensor x_195_cast_fp16 = matmul(transpose_x = x_195_transpose_x_0, transpose_y = x_195_transpose_y_0, x = input_405_cast_fp16, y = value_23_cast_fp16)[name = string("x_195_cast_fp16")]; + tensor var_2039_perm_0 = const()[name = string("op_2039_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_2040 = const()[name = string("op_2040"), val = tensor([1, -1, 1024])]; + tensor var_2039_cast_fp16 = transpose(perm = var_2039_perm_0, x = x_195_cast_fp16)[name = string("transpose_295")]; + tensor input_407_cast_fp16 = reshape(shape = var_2040, x = var_2039_cast_fp16)[name = string("input_407_cast_fp16")]; + tensor encoder_layers_7_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160196416))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160982912))))[name = string("encoder_layers_7_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_7_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_7_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160983104)))]; + tensor linear_70_cast_fp16 = linear(bias = encoder_layers_7_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_7_self_attn_linear_out_weight_to_fp16_palettized, x = input_407_cast_fp16)[name = string("linear_70_cast_fp16")]; + tensor input_411_cast_fp16 = add(x = input_401_cast_fp16, y = linear_70_cast_fp16)[name = string("input_411_cast_fp16")]; + tensor x_199_axes_0 = const()[name = string("x_199_axes_0"), val = tensor([-1])]; + tensor encoder_layers_7_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_7_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160985216)))]; + tensor encoder_layers_7_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_7_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160987328)))]; + tensor x_199_cast_fp16 = layer_norm(axes = x_199_axes_0, beta = encoder_layers_7_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_7_norm_conv_weight_to_fp16, x = input_411_cast_fp16)[name = string("x_199_cast_fp16")]; + tensor input_413_perm_0 = const()[name = string("input_413_perm_0"), val = tensor([0, 2, 1])]; + string input_415_pad_type_0 = const()[name = string("input_415_pad_type_0"), val = string("valid")]; + tensor input_415_strides_0 = const()[name = string("input_415_strides_0"), val = tensor([1])]; + tensor input_415_pad_0 = const()[name = string("input_415_pad_0"), val = tensor([0, 0])]; + tensor input_415_dilations_0 = const()[name = string("input_415_dilations_0"), val = tensor([1])]; + int32 input_415_groups_0 = const()[name = string("input_415_groups_0"), val = int32(1)]; + tensor encoder_layers_7_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160989440))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163086656))))[name = string("encoder_layers_7_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_413_cast_fp16 = transpose(perm = input_413_perm_0, x = x_199_cast_fp16)[name = string("transpose_294")]; + tensor input_415_cast_fp16 = conv(dilations = input_415_dilations_0, groups = input_415_groups_0, pad = input_415_pad_0, pad_type = input_415_pad_type_0, strides = input_415_strides_0, weight = encoder_layers_7_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_413_cast_fp16)[name = string("input_415_cast_fp16")]; + int32 x_201_split_num_splits_0 = const()[name = string("x_201_split_num_splits_0"), val = int32(2)]; + int32 x_201_split_axis_0 = const()[name = string("x_201_split_axis_0"), val = int32(1)]; + tensor x_201_split_cast_fp16_0, tensor x_201_split_cast_fp16_1 = split(axis = x_201_split_axis_0, num_splits = x_201_split_num_splits_0, x = input_415_cast_fp16)[name = string("x_201_split_cast_fp16")]; + tensor x_201_split_1_sigmoid_cast_fp16 = sigmoid(x = x_201_split_cast_fp16_1)[name = string("x_201_split_1_sigmoid_cast_fp16")]; + tensor x_201_cast_fp16 = mul(x = x_201_split_cast_fp16_0, y = x_201_split_1_sigmoid_cast_fp16)[name = string("x_201_cast_fp16")]; + tensor input_417_cast_fp16 = select(a = var_44_to_fp16, b = x_201_cast_fp16, cond = var_575)[name = string("input_417_cast_fp16")]; + bool new_x_31_interleave_0 = const()[name = string("new_x_31_interleave_0"), val = bool(false)]; + tensor new_x_31_cast_fp16 = concat(axis = var_59, interleave = new_x_31_interleave_0, values = (cache_31_cast_fp16, input_417_cast_fp16))[name = string("new_x_31_cast_fp16")]; + tensor var_2079_begin_0 = const()[name = string("op_2079_begin_0"), val = tensor([0, 0, 7])]; + tensor var_2079_end_0 = const()[name = string("op_2079_end_0"), val = tensor([1, 1024, 15])]; + tensor var_2079_end_mask_0 = const()[name = string("op_2079_end_mask_0"), val = tensor([true, true, true])]; + tensor var_2079_cast_fp16 = slice_by_index(begin = var_2079_begin_0, end = var_2079_end_0, end_mask = var_2079_end_mask_0, x = new_x_31_cast_fp16)[name = string("op_2079_cast_fp16")]; + string x_203_pad_type_0 = const()[name = string("x_203_pad_type_0"), val = string("valid")]; + int32 x_203_groups_0 = const()[name = string("x_203_groups_0"), val = int32(1024)]; + tensor x_203_strides_0 = const()[name = string("x_203_strides_0"), val = tensor([1])]; + tensor x_203_pad_0 = const()[name = string("x_203_pad_0"), val = tensor([0, 0])]; + tensor x_203_dilations_0 = const()[name = string("x_203_dilations_0"), val = tensor([1])]; + tensor encoder_layers_7_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163090816))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163100096))))[name = string("encoder_layers_7_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_203_cast_fp16 = conv(dilations = x_203_dilations_0, groups = x_203_groups_0, pad = x_203_pad_0, pad_type = x_203_pad_type_0, strides = x_203_strides_0, weight = encoder_layers_7_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_31_cast_fp16)[name = string("x_203_cast_fp16")]; + tensor input_419_perm_0 = const()[name = string("input_419_perm_0"), val = tensor([0, 2, 1])]; + tensor x_205_axes_0 = const()[name = string("x_205_axes_0"), val = tensor([-1])]; + tensor encoder_layers_7_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_7_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163102208)))]; + tensor encoder_layers_7_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_7_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163104320)))]; + tensor input_419_cast_fp16 = transpose(perm = input_419_perm_0, x = x_203_cast_fp16)[name = string("transpose_293")]; + tensor x_205_cast_fp16 = layer_norm(axes = x_205_axes_0, beta = encoder_layers_7_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_7_conv_batch_norm_weight_to_fp16, x = input_419_cast_fp16)[name = string("x_205_cast_fp16")]; + tensor input_421_perm_0 = const()[name = string("input_421_perm_0"), val = tensor([0, 2, 1])]; + tensor input_421_cast_fp16 = transpose(perm = input_421_perm_0, x = x_205_cast_fp16)[name = string("transpose_292")]; + tensor input_423_cast_fp16 = silu(x = input_421_cast_fp16)[name = string("input_423_cast_fp16")]; + string x_207_pad_type_0 = const()[name = string("x_207_pad_type_0"), val = string("valid")]; + tensor x_207_strides_0 = const()[name = string("x_207_strides_0"), val = tensor([1])]; + tensor x_207_pad_0 = const()[name = string("x_207_pad_0"), val = tensor([0, 0])]; + tensor x_207_dilations_0 = const()[name = string("x_207_dilations_0"), val = tensor([1])]; + int32 x_207_groups_0 = const()[name = string("x_207_groups_0"), val = int32(1)]; + tensor encoder_layers_7_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163106432))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(164155072))))[name = string("encoder_layers_7_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_207_cast_fp16 = conv(dilations = x_207_dilations_0, groups = x_207_groups_0, pad = x_207_pad_0, pad_type = x_207_pad_type_0, strides = x_207_strides_0, weight = encoder_layers_7_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_423_cast_fp16)[name = string("x_207_cast_fp16")]; + tensor input_425_perm_0 = const()[name = string("input_425_perm_0"), val = tensor([0, 2, 1])]; + tensor input_425_cast_fp16 = transpose(perm = input_425_perm_0, x = x_207_cast_fp16)[name = string("transpose_291")]; + tensor input_427_cast_fp16 = add(x = input_411_cast_fp16, y = input_425_cast_fp16)[name = string("input_427_cast_fp16")]; + tensor input_429_axes_0 = const()[name = string("input_429_axes_0"), val = tensor([-1])]; + tensor encoder_layers_7_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_7_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(164157184)))]; + tensor encoder_layers_7_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_7_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(164159296)))]; + tensor input_429_cast_fp16 = layer_norm(axes = input_429_axes_0, beta = encoder_layers_7_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_7_norm_feed_forward2_weight_to_fp16, x = input_427_cast_fp16)[name = string("input_429_cast_fp16")]; + tensor encoder_layers_7_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(164161408))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167307200))))[name = string("encoder_layers_7_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_7_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_7_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167307392)))]; + tensor linear_71_cast_fp16 = linear(bias = encoder_layers_7_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_7_feed_forward2_linear1_weight_to_fp16_palettized, x = input_429_cast_fp16)[name = string("linear_71_cast_fp16")]; + tensor input_433_cast_fp16 = silu(x = linear_71_cast_fp16)[name = string("input_433_cast_fp16")]; + tensor encoder_layers_7_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167315648))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170461440))))[name = string("encoder_layers_7_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_7_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_7_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170461632)))]; + tensor linear_72_cast_fp16 = linear(bias = encoder_layers_7_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_7_feed_forward2_linear2_weight_to_fp16_palettized, x = input_433_cast_fp16)[name = string("linear_72_cast_fp16")]; + fp16 var_2122_to_fp16 = const()[name = string("op_2122_to_fp16"), val = fp16(0x1p-1)]; + tensor var_2123_cast_fp16 = mul(x = linear_72_cast_fp16, y = var_2122_to_fp16)[name = string("op_2123_cast_fp16")]; + tensor input_439_cast_fp16 = add(x = input_427_cast_fp16, y = var_2123_cast_fp16)[name = string("input_439_cast_fp16")]; + tensor input_441_axes_0 = const()[name = string("input_441_axes_0"), val = tensor([-1])]; + tensor encoder_layers_7_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_7_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170463744)))]; + tensor encoder_layers_7_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_7_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170465856)))]; + tensor input_441_cast_fp16 = layer_norm(axes = input_441_axes_0, beta = encoder_layers_7_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_7_norm_out_weight_to_fp16, x = input_439_cast_fp16)[name = string("input_441_cast_fp16")]; + tensor cache_33_begin_0 = const()[name = string("cache_33_begin_0"), val = tensor([8, 0, 0, 0])]; + tensor cache_33_end_0 = const()[name = string("cache_33_end_0"), val = tensor([9, 1, 42, 1024])]; + tensor cache_33_end_mask_0 = const()[name = string("cache_33_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_33_squeeze_mask_0 = const()[name = string("cache_33_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_33_cast_fp16 = slice_by_index(begin = cache_33_begin_0, end = cache_33_end_0, end_mask = cache_33_end_mask_0, squeeze_mask = cache_33_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_33_cast_fp16")]; + tensor cache_35_begin_0 = const()[name = string("cache_35_begin_0"), val = tensor([8, 0, 0, 0])]; + tensor cache_35_end_0 = const()[name = string("cache_35_end_0"), val = tensor([9, 1, 1024, 8])]; + tensor cache_35_end_mask_0 = const()[name = string("cache_35_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_35_squeeze_mask_0 = const()[name = string("cache_35_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_35_cast_fp16 = slice_by_index(begin = cache_35_begin_0, end = cache_35_end_0, end_mask = cache_35_end_mask_0, squeeze_mask = cache_35_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_35_cast_fp16")]; + tensor input_443_axes_0 = const()[name = string("input_443_axes_0"), val = tensor([-1])]; + tensor encoder_layers_8_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_8_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170467968)))]; + tensor encoder_layers_8_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_8_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170470080)))]; + tensor input_443_cast_fp16 = layer_norm(axes = input_443_axes_0, beta = encoder_layers_8_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_8_norm_feed_forward1_weight_to_fp16, x = input_441_cast_fp16)[name = string("input_443_cast_fp16")]; + tensor encoder_layers_8_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170472192))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(173617984))))[name = string("encoder_layers_8_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_8_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_8_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(173618176)))]; + tensor linear_73_cast_fp16 = linear(bias = encoder_layers_8_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_8_feed_forward1_linear1_weight_to_fp16_palettized, x = input_443_cast_fp16)[name = string("linear_73_cast_fp16")]; + tensor input_447_cast_fp16 = silu(x = linear_73_cast_fp16)[name = string("input_447_cast_fp16")]; + tensor encoder_layers_8_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(173626432))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(176772224))))[name = string("encoder_layers_8_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_8_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_8_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(176772416)))]; + tensor linear_74_cast_fp16 = linear(bias = encoder_layers_8_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_8_feed_forward1_linear2_weight_to_fp16_palettized, x = input_447_cast_fp16)[name = string("linear_74_cast_fp16")]; + fp16 var_2159_to_fp16 = const()[name = string("op_2159_to_fp16"), val = fp16(0x1p-1)]; + tensor var_2160_cast_fp16 = mul(x = linear_74_cast_fp16, y = var_2159_to_fp16)[name = string("op_2160_cast_fp16")]; + tensor input_453_cast_fp16 = add(x = input_441_cast_fp16, y = var_2160_cast_fp16)[name = string("input_453_cast_fp16")]; + tensor key_17_axes_0 = const()[name = string("key_17_axes_0"), val = tensor([-1])]; + tensor encoder_layers_8_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_8_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(176774528)))]; + tensor encoder_layers_8_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_8_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(176776640)))]; + tensor key_17_cast_fp16 = layer_norm(axes = key_17_axes_0, beta = encoder_layers_8_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_8_norm_self_att_weight_to_fp16, x = input_453_cast_fp16)[name = string("key_17_cast_fp16")]; + bool input_455_interleave_0 = const()[name = string("input_455_interleave_0"), val = bool(false)]; + tensor input_455_cast_fp16 = concat(axis = var_68, interleave = input_455_interleave_0, values = (cache_33_cast_fp16, key_17_cast_fp16))[name = string("input_455_cast_fp16")]; + tensor var_2182_begin_0 = const()[name = string("op_2182_begin_0"), val = tensor([0, 7, 0])]; + tensor var_2182_end_0 = const()[name = string("op_2182_end_0"), val = tensor([1, 42, 1024])]; + tensor var_2182_end_mask_0 = const()[name = string("op_2182_end_mask_0"), val = tensor([true, true, true])]; + tensor var_2182_cast_fp16 = slice_by_index(begin = var_2182_begin_0, end = var_2182_end_0, end_mask = var_2182_end_mask_0, x = cache_33_cast_fp16)[name = string("op_2182_cast_fp16")]; + bool var_2188_interleave_0 = const()[name = string("op_2188_interleave_0"), val = bool(false)]; + tensor var_2188_cast_fp16 = concat(axis = var_68, interleave = var_2188_interleave_0, values = (var_2182_cast_fp16, key_17_cast_fp16))[name = string("op_2188_cast_fp16")]; + tensor encoder_layers_8_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(176778752))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(177565248))))[name = string("encoder_layers_8_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_8_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_8_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(177565440)))]; + tensor linear_75_cast_fp16 = linear(bias = encoder_layers_8_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_8_self_attn_linear_q_weight_to_fp16_palettized, x = key_17_cast_fp16)[name = string("linear_75_cast_fp16")]; + tensor var_2193 = const()[name = string("op_2193"), val = tensor([1, -1, 8, 128])]; + tensor q_49_cast_fp16 = reshape(shape = var_2193, x = linear_75_cast_fp16)[name = string("q_49_cast_fp16")]; + tensor encoder_layers_8_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(177567552))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(178354048))))[name = string("encoder_layers_8_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_8_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_8_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(178354240)))]; + tensor linear_76_cast_fp16 = linear(bias = encoder_layers_8_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_8_self_attn_linear_k_weight_to_fp16_palettized, x = input_455_cast_fp16)[name = string("linear_76_cast_fp16")]; + tensor var_2198 = const()[name = string("op_2198"), val = tensor([1, -1, 8, 128])]; + tensor k_33_cast_fp16 = reshape(shape = var_2198, x = linear_76_cast_fp16)[name = string("k_33_cast_fp16")]; + tensor encoder_layers_8_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(178356352))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179142848))))[name = string("encoder_layers_8_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_8_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_8_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179143040)))]; + tensor linear_77_cast_fp16 = linear(bias = encoder_layers_8_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_8_self_attn_linear_v_weight_to_fp16_palettized, x = input_455_cast_fp16)[name = string("linear_77_cast_fp16")]; + tensor var_2203 = const()[name = string("op_2203"), val = tensor([1, -1, 8, 128])]; + tensor v_17_cast_fp16 = reshape(shape = var_2203, x = linear_77_cast_fp16)[name = string("v_17_cast_fp16")]; + tensor value_25_perm_0 = const()[name = string("value_25_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_8_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_8_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179145152)))]; + tensor var_2216_cast_fp16 = add(x = q_49_cast_fp16, y = encoder_layers_8_self_attn_pos_bias_u_to_fp16)[name = string("op_2216_cast_fp16")]; + tensor encoder_layers_8_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_8_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179147264)))]; + tensor var_2218_cast_fp16 = add(x = q_49_cast_fp16, y = encoder_layers_8_self_attn_pos_bias_v_to_fp16)[name = string("op_2218_cast_fp16")]; + tensor q_with_bias_v_17_perm_0 = const()[name = string("q_with_bias_v_17_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_215_transpose_x_0 = const()[name = string("x_215_transpose_x_0"), val = bool(false)]; + bool x_215_transpose_y_0 = const()[name = string("x_215_transpose_y_0"), val = bool(false)]; + tensor op_2220_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179149376))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179248768))))[name = string("op_2220_to_fp16_quantized")]; + tensor q_with_bias_v_17_cast_fp16 = transpose(perm = q_with_bias_v_17_perm_0, x = var_2218_cast_fp16)[name = string("transpose_290")]; + tensor x_215_cast_fp16 = matmul(transpose_x = x_215_transpose_x_0, transpose_y = x_215_transpose_y_0, x = q_with_bias_v_17_cast_fp16, y = op_2220_to_fp16_quantized)[name = string("x_215_cast_fp16")]; + tensor x_217_pad_0 = const()[name = string("x_217_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_217_mode_0 = const()[name = string("x_217_mode_0"), val = string("constant")]; + fp16 const_183_to_fp16 = const()[name = string("const_183_to_fp16"), val = fp16(0x0p+0)]; + tensor x_217_cast_fp16 = pad(constant_val = const_183_to_fp16, mode = x_217_mode_0, pad = x_217_pad_0, x = x_215_cast_fp16)[name = string("x_217_cast_fp16")]; + tensor var_2228 = const()[name = string("op_2228"), val = tensor([1, 8, -1, 7])]; + tensor x_219_cast_fp16 = reshape(shape = var_2228, x = x_217_cast_fp16)[name = string("x_219_cast_fp16")]; + tensor var_2232_begin_0 = const()[name = string("op_2232_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_2232_end_0 = const()[name = string("op_2232_end_0"), val = tensor([1, 8, 98, 7])]; + tensor var_2232_end_mask_0 = const()[name = string("op_2232_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_2232_cast_fp16 = slice_by_index(begin = var_2232_begin_0, end = var_2232_end_0, end_mask = var_2232_end_mask_0, x = x_219_cast_fp16)[name = string("op_2232_cast_fp16")]; + tensor var_2233 = const()[name = string("op_2233"), val = tensor([1, 8, 7, 97])]; + tensor matrix_bd_33_cast_fp16 = reshape(shape = var_2233, x = var_2232_cast_fp16)[name = string("matrix_bd_33_cast_fp16")]; + bool matrix_ac_17_transpose_x_0 = const()[name = string("matrix_ac_17_transpose_x_0"), val = bool(false)]; + bool matrix_ac_17_transpose_y_0 = const()[name = string("matrix_ac_17_transpose_y_0"), val = bool(false)]; + tensor transpose_112_perm_0 = const()[name = string("transpose_112_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_113_perm_0 = const()[name = string("transpose_113_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_113 = transpose(perm = transpose_113_perm_0, x = k_33_cast_fp16)[name = string("transpose_288")]; + tensor transpose_112 = transpose(perm = transpose_112_perm_0, x = var_2216_cast_fp16)[name = string("transpose_289")]; + tensor matrix_ac_17_cast_fp16 = matmul(transpose_x = matrix_ac_17_transpose_x_0, transpose_y = matrix_ac_17_transpose_y_0, x = transpose_112, y = transpose_113)[name = string("matrix_ac_17_cast_fp16")]; + tensor matrix_bd_35_begin_0 = const()[name = string("matrix_bd_35_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_35_end_0 = const()[name = string("matrix_bd_35_end_0"), val = tensor([1, 8, 7, 49])]; + tensor matrix_bd_35_end_mask_0 = const()[name = string("matrix_bd_35_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_35_cast_fp16 = slice_by_index(begin = matrix_bd_35_begin_0, end = matrix_bd_35_end_0, end_mask = matrix_bd_35_end_mask_0, x = matrix_bd_33_cast_fp16)[name = string("matrix_bd_35_cast_fp16")]; + tensor var_2242_cast_fp16 = add(x = matrix_ac_17_cast_fp16, y = matrix_bd_35_cast_fp16)[name = string("op_2242_cast_fp16")]; + fp16 _inversed_scores_33_y_0_to_fp16 = const()[name = string("_inversed_scores_33_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_33_cast_fp16 = mul(x = var_2242_cast_fp16, y = _inversed_scores_33_y_0_to_fp16)[name = string("_inversed_scores_33_cast_fp16")]; + tensor scores_35_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_33_cast_fp16, cond = mask_11)[name = string("scores_35_cast_fp16")]; + tensor var_2248_cast_fp16 = softmax(axis = var_59, x = scores_35_cast_fp16)[name = string("op_2248_cast_fp16")]; + tensor input_457_cast_fp16 = select(a = var_44_to_fp16, b = var_2248_cast_fp16, cond = mask_11)[name = string("input_457_cast_fp16")]; + bool x_221_transpose_x_0 = const()[name = string("x_221_transpose_x_0"), val = bool(false)]; + bool x_221_transpose_y_0 = const()[name = string("x_221_transpose_y_0"), val = bool(false)]; + tensor value_25_cast_fp16 = transpose(perm = value_25_perm_0, x = v_17_cast_fp16)[name = string("transpose_287")]; + tensor x_221_cast_fp16 = matmul(transpose_x = x_221_transpose_x_0, transpose_y = x_221_transpose_y_0, x = input_457_cast_fp16, y = value_25_cast_fp16)[name = string("x_221_cast_fp16")]; + tensor var_2252_perm_0 = const()[name = string("op_2252_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_2253 = const()[name = string("op_2253"), val = tensor([1, -1, 1024])]; + tensor var_2252_cast_fp16 = transpose(perm = var_2252_perm_0, x = x_221_cast_fp16)[name = string("transpose_286")]; + tensor input_459_cast_fp16 = reshape(shape = var_2253, x = var_2252_cast_fp16)[name = string("input_459_cast_fp16")]; + tensor encoder_layers_8_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179249088))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(180035584))))[name = string("encoder_layers_8_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_8_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_8_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(180035776)))]; + tensor linear_79_cast_fp16 = linear(bias = encoder_layers_8_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_8_self_attn_linear_out_weight_to_fp16_palettized, x = input_459_cast_fp16)[name = string("linear_79_cast_fp16")]; + tensor input_463_cast_fp16 = add(x = input_453_cast_fp16, y = linear_79_cast_fp16)[name = string("input_463_cast_fp16")]; + tensor x_225_axes_0 = const()[name = string("x_225_axes_0"), val = tensor([-1])]; + tensor encoder_layers_8_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_8_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(180037888)))]; + tensor encoder_layers_8_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_8_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(180040000)))]; + tensor x_225_cast_fp16 = layer_norm(axes = x_225_axes_0, beta = encoder_layers_8_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_8_norm_conv_weight_to_fp16, x = input_463_cast_fp16)[name = string("x_225_cast_fp16")]; + tensor input_465_perm_0 = const()[name = string("input_465_perm_0"), val = tensor([0, 2, 1])]; + string input_467_pad_type_0 = const()[name = string("input_467_pad_type_0"), val = string("valid")]; + tensor input_467_strides_0 = const()[name = string("input_467_strides_0"), val = tensor([1])]; + tensor input_467_pad_0 = const()[name = string("input_467_pad_0"), val = tensor([0, 0])]; + tensor input_467_dilations_0 = const()[name = string("input_467_dilations_0"), val = tensor([1])]; + int32 input_467_groups_0 = const()[name = string("input_467_groups_0"), val = int32(1)]; + tensor encoder_layers_8_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(180042112))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182139328))))[name = string("encoder_layers_8_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_465_cast_fp16 = transpose(perm = input_465_perm_0, x = x_225_cast_fp16)[name = string("transpose_285")]; + tensor input_467_cast_fp16 = conv(dilations = input_467_dilations_0, groups = input_467_groups_0, pad = input_467_pad_0, pad_type = input_467_pad_type_0, strides = input_467_strides_0, weight = encoder_layers_8_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_465_cast_fp16)[name = string("input_467_cast_fp16")]; + int32 x_227_split_num_splits_0 = const()[name = string("x_227_split_num_splits_0"), val = int32(2)]; + int32 x_227_split_axis_0 = const()[name = string("x_227_split_axis_0"), val = int32(1)]; + tensor x_227_split_cast_fp16_0, tensor x_227_split_cast_fp16_1 = split(axis = x_227_split_axis_0, num_splits = x_227_split_num_splits_0, x = input_467_cast_fp16)[name = string("x_227_split_cast_fp16")]; + tensor x_227_split_1_sigmoid_cast_fp16 = sigmoid(x = x_227_split_cast_fp16_1)[name = string("x_227_split_1_sigmoid_cast_fp16")]; + tensor x_227_cast_fp16 = mul(x = x_227_split_cast_fp16_0, y = x_227_split_1_sigmoid_cast_fp16)[name = string("x_227_cast_fp16")]; + tensor input_469_cast_fp16 = select(a = var_44_to_fp16, b = x_227_cast_fp16, cond = var_575)[name = string("input_469_cast_fp16")]; + bool new_x_35_interleave_0 = const()[name = string("new_x_35_interleave_0"), val = bool(false)]; + tensor new_x_35_cast_fp16 = concat(axis = var_59, interleave = new_x_35_interleave_0, values = (cache_35_cast_fp16, input_469_cast_fp16))[name = string("new_x_35_cast_fp16")]; + tensor var_2292_begin_0 = const()[name = string("op_2292_begin_0"), val = tensor([0, 0, 7])]; + tensor var_2292_end_0 = const()[name = string("op_2292_end_0"), val = tensor([1, 1024, 15])]; + tensor var_2292_end_mask_0 = const()[name = string("op_2292_end_mask_0"), val = tensor([true, true, true])]; + tensor var_2292_cast_fp16 = slice_by_index(begin = var_2292_begin_0, end = var_2292_end_0, end_mask = var_2292_end_mask_0, x = new_x_35_cast_fp16)[name = string("op_2292_cast_fp16")]; + string x_229_pad_type_0 = const()[name = string("x_229_pad_type_0"), val = string("valid")]; + int32 x_229_groups_0 = const()[name = string("x_229_groups_0"), val = int32(1024)]; + tensor x_229_strides_0 = const()[name = string("x_229_strides_0"), val = tensor([1])]; + tensor x_229_pad_0 = const()[name = string("x_229_pad_0"), val = tensor([0, 0])]; + tensor x_229_dilations_0 = const()[name = string("x_229_dilations_0"), val = tensor([1])]; + tensor encoder_layers_8_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182143488))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182152768))))[name = string("encoder_layers_8_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_229_cast_fp16 = conv(dilations = x_229_dilations_0, groups = x_229_groups_0, pad = x_229_pad_0, pad_type = x_229_pad_type_0, strides = x_229_strides_0, weight = encoder_layers_8_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_35_cast_fp16)[name = string("x_229_cast_fp16")]; + tensor input_471_perm_0 = const()[name = string("input_471_perm_0"), val = tensor([0, 2, 1])]; + tensor x_231_axes_0 = const()[name = string("x_231_axes_0"), val = tensor([-1])]; + tensor encoder_layers_8_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_8_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182154880)))]; + tensor encoder_layers_8_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_8_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182156992)))]; + tensor input_471_cast_fp16 = transpose(perm = input_471_perm_0, x = x_229_cast_fp16)[name = string("transpose_284")]; + tensor x_231_cast_fp16 = layer_norm(axes = x_231_axes_0, beta = encoder_layers_8_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_8_conv_batch_norm_weight_to_fp16, x = input_471_cast_fp16)[name = string("x_231_cast_fp16")]; + tensor input_473_perm_0 = const()[name = string("input_473_perm_0"), val = tensor([0, 2, 1])]; + tensor input_473_cast_fp16 = transpose(perm = input_473_perm_0, x = x_231_cast_fp16)[name = string("transpose_283")]; + tensor input_475_cast_fp16 = silu(x = input_473_cast_fp16)[name = string("input_475_cast_fp16")]; + string x_233_pad_type_0 = const()[name = string("x_233_pad_type_0"), val = string("valid")]; + tensor x_233_strides_0 = const()[name = string("x_233_strides_0"), val = tensor([1])]; + tensor x_233_pad_0 = const()[name = string("x_233_pad_0"), val = tensor([0, 0])]; + tensor x_233_dilations_0 = const()[name = string("x_233_dilations_0"), val = tensor([1])]; + int32 x_233_groups_0 = const()[name = string("x_233_groups_0"), val = int32(1)]; + tensor encoder_layers_8_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182159104))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183207744))))[name = string("encoder_layers_8_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_233_cast_fp16 = conv(dilations = x_233_dilations_0, groups = x_233_groups_0, pad = x_233_pad_0, pad_type = x_233_pad_type_0, strides = x_233_strides_0, weight = encoder_layers_8_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_475_cast_fp16)[name = string("x_233_cast_fp16")]; + tensor input_477_perm_0 = const()[name = string("input_477_perm_0"), val = tensor([0, 2, 1])]; + tensor input_477_cast_fp16 = transpose(perm = input_477_perm_0, x = x_233_cast_fp16)[name = string("transpose_282")]; + tensor input_479_cast_fp16 = add(x = input_463_cast_fp16, y = input_477_cast_fp16)[name = string("input_479_cast_fp16")]; + tensor input_481_axes_0 = const()[name = string("input_481_axes_0"), val = tensor([-1])]; + tensor encoder_layers_8_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_8_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183209856)))]; + tensor encoder_layers_8_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_8_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183211968)))]; + tensor input_481_cast_fp16 = layer_norm(axes = input_481_axes_0, beta = encoder_layers_8_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_8_norm_feed_forward2_weight_to_fp16, x = input_479_cast_fp16)[name = string("input_481_cast_fp16")]; + tensor encoder_layers_8_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183214080))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186359872))))[name = string("encoder_layers_8_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_8_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_8_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186360064)))]; + tensor linear_80_cast_fp16 = linear(bias = encoder_layers_8_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_8_feed_forward2_linear1_weight_to_fp16_palettized, x = input_481_cast_fp16)[name = string("linear_80_cast_fp16")]; + tensor input_485_cast_fp16 = silu(x = linear_80_cast_fp16)[name = string("input_485_cast_fp16")]; + tensor encoder_layers_8_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186368320))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(189514112))))[name = string("encoder_layers_8_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_8_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_8_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(189514304)))]; + tensor linear_81_cast_fp16 = linear(bias = encoder_layers_8_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_8_feed_forward2_linear2_weight_to_fp16_palettized, x = input_485_cast_fp16)[name = string("linear_81_cast_fp16")]; + fp16 var_2335_to_fp16 = const()[name = string("op_2335_to_fp16"), val = fp16(0x1p-1)]; + tensor var_2336_cast_fp16 = mul(x = linear_81_cast_fp16, y = var_2335_to_fp16)[name = string("op_2336_cast_fp16")]; + tensor input_491_cast_fp16 = add(x = input_479_cast_fp16, y = var_2336_cast_fp16)[name = string("input_491_cast_fp16")]; + tensor input_493_axes_0 = const()[name = string("input_493_axes_0"), val = tensor([-1])]; + tensor encoder_layers_8_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_8_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(189516416)))]; + tensor encoder_layers_8_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_8_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(189518528)))]; + tensor input_493_cast_fp16 = layer_norm(axes = input_493_axes_0, beta = encoder_layers_8_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_8_norm_out_weight_to_fp16, x = input_491_cast_fp16)[name = string("input_493_cast_fp16")]; + tensor cache_37_begin_0 = const()[name = string("cache_37_begin_0"), val = tensor([9, 0, 0, 0])]; + tensor cache_37_end_0 = const()[name = string("cache_37_end_0"), val = tensor([10, 1, 42, 1024])]; + tensor cache_37_end_mask_0 = const()[name = string("cache_37_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_37_squeeze_mask_0 = const()[name = string("cache_37_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_37_cast_fp16 = slice_by_index(begin = cache_37_begin_0, end = cache_37_end_0, end_mask = cache_37_end_mask_0, squeeze_mask = cache_37_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_37_cast_fp16")]; + tensor cache_39_begin_0 = const()[name = string("cache_39_begin_0"), val = tensor([9, 0, 0, 0])]; + tensor cache_39_end_0 = const()[name = string("cache_39_end_0"), val = tensor([10, 1, 1024, 8])]; + tensor cache_39_end_mask_0 = const()[name = string("cache_39_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_39_squeeze_mask_0 = const()[name = string("cache_39_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_39_cast_fp16 = slice_by_index(begin = cache_39_begin_0, end = cache_39_end_0, end_mask = cache_39_end_mask_0, squeeze_mask = cache_39_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_39_cast_fp16")]; + tensor input_495_axes_0 = const()[name = string("input_495_axes_0"), val = tensor([-1])]; + tensor encoder_layers_9_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_9_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(189520640)))]; + tensor encoder_layers_9_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_9_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(189522752)))]; + tensor input_495_cast_fp16 = layer_norm(axes = input_495_axes_0, beta = encoder_layers_9_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_9_norm_feed_forward1_weight_to_fp16, x = input_493_cast_fp16)[name = string("input_495_cast_fp16")]; + tensor encoder_layers_9_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(189524864))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(192670656))))[name = string("encoder_layers_9_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_9_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_9_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(192670848)))]; + tensor linear_82_cast_fp16 = linear(bias = encoder_layers_9_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_9_feed_forward1_linear1_weight_to_fp16_palettized, x = input_495_cast_fp16)[name = string("linear_82_cast_fp16")]; + tensor input_499_cast_fp16 = silu(x = linear_82_cast_fp16)[name = string("input_499_cast_fp16")]; + tensor encoder_layers_9_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(192679104))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(195824896))))[name = string("encoder_layers_9_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_9_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_9_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(195825088)))]; + tensor linear_83_cast_fp16 = linear(bias = encoder_layers_9_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_9_feed_forward1_linear2_weight_to_fp16_palettized, x = input_499_cast_fp16)[name = string("linear_83_cast_fp16")]; + fp16 var_2372_to_fp16 = const()[name = string("op_2372_to_fp16"), val = fp16(0x1p-1)]; + tensor var_2373_cast_fp16 = mul(x = linear_83_cast_fp16, y = var_2372_to_fp16)[name = string("op_2373_cast_fp16")]; + tensor input_505_cast_fp16 = add(x = input_493_cast_fp16, y = var_2373_cast_fp16)[name = string("input_505_cast_fp16")]; + tensor key_19_axes_0 = const()[name = string("key_19_axes_0"), val = tensor([-1])]; + tensor encoder_layers_9_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_9_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(195827200)))]; + tensor encoder_layers_9_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_9_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(195829312)))]; + tensor key_19_cast_fp16 = layer_norm(axes = key_19_axes_0, beta = encoder_layers_9_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_9_norm_self_att_weight_to_fp16, x = input_505_cast_fp16)[name = string("key_19_cast_fp16")]; + bool input_507_interleave_0 = const()[name = string("input_507_interleave_0"), val = bool(false)]; + tensor input_507_cast_fp16 = concat(axis = var_68, interleave = input_507_interleave_0, values = (cache_37_cast_fp16, key_19_cast_fp16))[name = string("input_507_cast_fp16")]; + tensor var_2395_begin_0 = const()[name = string("op_2395_begin_0"), val = tensor([0, 7, 0])]; + tensor var_2395_end_0 = const()[name = string("op_2395_end_0"), val = tensor([1, 42, 1024])]; + tensor var_2395_end_mask_0 = const()[name = string("op_2395_end_mask_0"), val = tensor([true, true, true])]; + tensor var_2395_cast_fp16 = slice_by_index(begin = var_2395_begin_0, end = var_2395_end_0, end_mask = var_2395_end_mask_0, x = cache_37_cast_fp16)[name = string("op_2395_cast_fp16")]; + bool var_2401_interleave_0 = const()[name = string("op_2401_interleave_0"), val = bool(false)]; + tensor var_2401_cast_fp16 = concat(axis = var_68, interleave = var_2401_interleave_0, values = (var_2395_cast_fp16, key_19_cast_fp16))[name = string("op_2401_cast_fp16")]; + tensor encoder_layers_9_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(195831424))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196617920))))[name = string("encoder_layers_9_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_9_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_9_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196618112)))]; + tensor linear_84_cast_fp16 = linear(bias = encoder_layers_9_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_9_self_attn_linear_q_weight_to_fp16_palettized, x = key_19_cast_fp16)[name = string("linear_84_cast_fp16")]; + tensor var_2406 = const()[name = string("op_2406"), val = tensor([1, -1, 8, 128])]; + tensor q_55_cast_fp16 = reshape(shape = var_2406, x = linear_84_cast_fp16)[name = string("q_55_cast_fp16")]; + tensor encoder_layers_9_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196620224))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197406720))))[name = string("encoder_layers_9_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_9_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_9_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197406912)))]; + tensor linear_85_cast_fp16 = linear(bias = encoder_layers_9_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_9_self_attn_linear_k_weight_to_fp16_palettized, x = input_507_cast_fp16)[name = string("linear_85_cast_fp16")]; + tensor var_2411 = const()[name = string("op_2411"), val = tensor([1, -1, 8, 128])]; + tensor k_37_cast_fp16 = reshape(shape = var_2411, x = linear_85_cast_fp16)[name = string("k_37_cast_fp16")]; + tensor encoder_layers_9_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197409024))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198195520))))[name = string("encoder_layers_9_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_9_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_9_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198195712)))]; + tensor linear_86_cast_fp16 = linear(bias = encoder_layers_9_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_9_self_attn_linear_v_weight_to_fp16_palettized, x = input_507_cast_fp16)[name = string("linear_86_cast_fp16")]; + tensor var_2416 = const()[name = string("op_2416"), val = tensor([1, -1, 8, 128])]; + tensor v_19_cast_fp16 = reshape(shape = var_2416, x = linear_86_cast_fp16)[name = string("v_19_cast_fp16")]; + tensor value_27_perm_0 = const()[name = string("value_27_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_9_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_9_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198197824)))]; + tensor var_2429_cast_fp16 = add(x = q_55_cast_fp16, y = encoder_layers_9_self_attn_pos_bias_u_to_fp16)[name = string("op_2429_cast_fp16")]; + tensor encoder_layers_9_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_9_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198199936)))]; + tensor var_2431_cast_fp16 = add(x = q_55_cast_fp16, y = encoder_layers_9_self_attn_pos_bias_v_to_fp16)[name = string("op_2431_cast_fp16")]; + tensor q_with_bias_v_19_perm_0 = const()[name = string("q_with_bias_v_19_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_241_transpose_x_0 = const()[name = string("x_241_transpose_x_0"), val = bool(false)]; + bool x_241_transpose_y_0 = const()[name = string("x_241_transpose_y_0"), val = bool(false)]; + tensor op_2433_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198202048))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198301440))))[name = string("op_2433_to_fp16_quantized")]; + tensor q_with_bias_v_19_cast_fp16 = transpose(perm = q_with_bias_v_19_perm_0, x = var_2431_cast_fp16)[name = string("transpose_281")]; + tensor x_241_cast_fp16 = matmul(transpose_x = x_241_transpose_x_0, transpose_y = x_241_transpose_y_0, x = q_with_bias_v_19_cast_fp16, y = op_2433_to_fp16_quantized)[name = string("x_241_cast_fp16")]; + tensor x_243_pad_0 = const()[name = string("x_243_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_243_mode_0 = const()[name = string("x_243_mode_0"), val = string("constant")]; + fp16 const_196_to_fp16 = const()[name = string("const_196_to_fp16"), val = fp16(0x0p+0)]; + tensor x_243_cast_fp16 = pad(constant_val = const_196_to_fp16, mode = x_243_mode_0, pad = x_243_pad_0, x = x_241_cast_fp16)[name = string("x_243_cast_fp16")]; + tensor var_2441 = const()[name = string("op_2441"), val = tensor([1, 8, -1, 7])]; + tensor x_245_cast_fp16 = reshape(shape = var_2441, x = x_243_cast_fp16)[name = string("x_245_cast_fp16")]; + tensor var_2445_begin_0 = const()[name = string("op_2445_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_2445_end_0 = const()[name = string("op_2445_end_0"), val = tensor([1, 8, 98, 7])]; + tensor var_2445_end_mask_0 = const()[name = string("op_2445_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_2445_cast_fp16 = slice_by_index(begin = var_2445_begin_0, end = var_2445_end_0, end_mask = var_2445_end_mask_0, x = x_245_cast_fp16)[name = string("op_2445_cast_fp16")]; + tensor var_2446 = const()[name = string("op_2446"), val = tensor([1, 8, 7, 97])]; + tensor matrix_bd_37_cast_fp16 = reshape(shape = var_2446, x = var_2445_cast_fp16)[name = string("matrix_bd_37_cast_fp16")]; + bool matrix_ac_19_transpose_x_0 = const()[name = string("matrix_ac_19_transpose_x_0"), val = bool(false)]; + bool matrix_ac_19_transpose_y_0 = const()[name = string("matrix_ac_19_transpose_y_0"), val = bool(false)]; + tensor transpose_114_perm_0 = const()[name = string("transpose_114_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_115_perm_0 = const()[name = string("transpose_115_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_115 = transpose(perm = transpose_115_perm_0, x = k_37_cast_fp16)[name = string("transpose_279")]; + tensor transpose_114 = transpose(perm = transpose_114_perm_0, x = var_2429_cast_fp16)[name = string("transpose_280")]; + tensor matrix_ac_19_cast_fp16 = matmul(transpose_x = matrix_ac_19_transpose_x_0, transpose_y = matrix_ac_19_transpose_y_0, x = transpose_114, y = transpose_115)[name = string("matrix_ac_19_cast_fp16")]; + tensor matrix_bd_39_begin_0 = const()[name = string("matrix_bd_39_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_39_end_0 = const()[name = string("matrix_bd_39_end_0"), val = tensor([1, 8, 7, 49])]; + tensor matrix_bd_39_end_mask_0 = const()[name = string("matrix_bd_39_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_39_cast_fp16 = slice_by_index(begin = matrix_bd_39_begin_0, end = matrix_bd_39_end_0, end_mask = matrix_bd_39_end_mask_0, x = matrix_bd_37_cast_fp16)[name = string("matrix_bd_39_cast_fp16")]; + tensor var_2455_cast_fp16 = add(x = matrix_ac_19_cast_fp16, y = matrix_bd_39_cast_fp16)[name = string("op_2455_cast_fp16")]; + fp16 _inversed_scores_37_y_0_to_fp16 = const()[name = string("_inversed_scores_37_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_37_cast_fp16 = mul(x = var_2455_cast_fp16, y = _inversed_scores_37_y_0_to_fp16)[name = string("_inversed_scores_37_cast_fp16")]; + tensor scores_39_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_37_cast_fp16, cond = mask_11)[name = string("scores_39_cast_fp16")]; + tensor var_2461_cast_fp16 = softmax(axis = var_59, x = scores_39_cast_fp16)[name = string("op_2461_cast_fp16")]; + tensor input_509_cast_fp16 = select(a = var_44_to_fp16, b = var_2461_cast_fp16, cond = mask_11)[name = string("input_509_cast_fp16")]; + bool x_247_transpose_x_0 = const()[name = string("x_247_transpose_x_0"), val = bool(false)]; + bool x_247_transpose_y_0 = const()[name = string("x_247_transpose_y_0"), val = bool(false)]; + tensor value_27_cast_fp16 = transpose(perm = value_27_perm_0, x = v_19_cast_fp16)[name = string("transpose_278")]; + tensor x_247_cast_fp16 = matmul(transpose_x = x_247_transpose_x_0, transpose_y = x_247_transpose_y_0, x = input_509_cast_fp16, y = value_27_cast_fp16)[name = string("x_247_cast_fp16")]; + tensor var_2465_perm_0 = const()[name = string("op_2465_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_2466 = const()[name = string("op_2466"), val = tensor([1, -1, 1024])]; + tensor var_2465_cast_fp16 = transpose(perm = var_2465_perm_0, x = x_247_cast_fp16)[name = string("transpose_277")]; + tensor input_511_cast_fp16 = reshape(shape = var_2466, x = var_2465_cast_fp16)[name = string("input_511_cast_fp16")]; + tensor encoder_layers_9_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198301760))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199088256))))[name = string("encoder_layers_9_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_9_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_9_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199088448)))]; + tensor linear_88_cast_fp16 = linear(bias = encoder_layers_9_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_9_self_attn_linear_out_weight_to_fp16_palettized, x = input_511_cast_fp16)[name = string("linear_88_cast_fp16")]; + tensor input_515_cast_fp16 = add(x = input_505_cast_fp16, y = linear_88_cast_fp16)[name = string("input_515_cast_fp16")]; + tensor x_251_axes_0 = const()[name = string("x_251_axes_0"), val = tensor([-1])]; + tensor encoder_layers_9_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_9_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199090560)))]; + tensor encoder_layers_9_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_9_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199092672)))]; + tensor x_251_cast_fp16 = layer_norm(axes = x_251_axes_0, beta = encoder_layers_9_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_9_norm_conv_weight_to_fp16, x = input_515_cast_fp16)[name = string("x_251_cast_fp16")]; + tensor input_517_perm_0 = const()[name = string("input_517_perm_0"), val = tensor([0, 2, 1])]; + string input_519_pad_type_0 = const()[name = string("input_519_pad_type_0"), val = string("valid")]; + tensor input_519_strides_0 = const()[name = string("input_519_strides_0"), val = tensor([1])]; + tensor input_519_pad_0 = const()[name = string("input_519_pad_0"), val = tensor([0, 0])]; + tensor input_519_dilations_0 = const()[name = string("input_519_dilations_0"), val = tensor([1])]; + int32 input_519_groups_0 = const()[name = string("input_519_groups_0"), val = int32(1)]; + tensor encoder_layers_9_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199094784))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(201192000))))[name = string("encoder_layers_9_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_517_cast_fp16 = transpose(perm = input_517_perm_0, x = x_251_cast_fp16)[name = string("transpose_276")]; + tensor input_519_cast_fp16 = conv(dilations = input_519_dilations_0, groups = input_519_groups_0, pad = input_519_pad_0, pad_type = input_519_pad_type_0, strides = input_519_strides_0, weight = encoder_layers_9_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_517_cast_fp16)[name = string("input_519_cast_fp16")]; + int32 x_253_split_num_splits_0 = const()[name = string("x_253_split_num_splits_0"), val = int32(2)]; + int32 x_253_split_axis_0 = const()[name = string("x_253_split_axis_0"), val = int32(1)]; + tensor x_253_split_cast_fp16_0, tensor x_253_split_cast_fp16_1 = split(axis = x_253_split_axis_0, num_splits = x_253_split_num_splits_0, x = input_519_cast_fp16)[name = string("x_253_split_cast_fp16")]; + tensor x_253_split_1_sigmoid_cast_fp16 = sigmoid(x = x_253_split_cast_fp16_1)[name = string("x_253_split_1_sigmoid_cast_fp16")]; + tensor x_253_cast_fp16 = mul(x = x_253_split_cast_fp16_0, y = x_253_split_1_sigmoid_cast_fp16)[name = string("x_253_cast_fp16")]; + tensor input_521_cast_fp16 = select(a = var_44_to_fp16, b = x_253_cast_fp16, cond = var_575)[name = string("input_521_cast_fp16")]; + bool new_x_39_interleave_0 = const()[name = string("new_x_39_interleave_0"), val = bool(false)]; + tensor new_x_39_cast_fp16 = concat(axis = var_59, interleave = new_x_39_interleave_0, values = (cache_39_cast_fp16, input_521_cast_fp16))[name = string("new_x_39_cast_fp16")]; + tensor var_2505_begin_0 = const()[name = string("op_2505_begin_0"), val = tensor([0, 0, 7])]; + tensor var_2505_end_0 = const()[name = string("op_2505_end_0"), val = tensor([1, 1024, 15])]; + tensor var_2505_end_mask_0 = const()[name = string("op_2505_end_mask_0"), val = tensor([true, true, true])]; + tensor var_2505_cast_fp16 = slice_by_index(begin = var_2505_begin_0, end = var_2505_end_0, end_mask = var_2505_end_mask_0, x = new_x_39_cast_fp16)[name = string("op_2505_cast_fp16")]; + string x_255_pad_type_0 = const()[name = string("x_255_pad_type_0"), val = string("valid")]; + int32 x_255_groups_0 = const()[name = string("x_255_groups_0"), val = int32(1024)]; + tensor x_255_strides_0 = const()[name = string("x_255_strides_0"), val = tensor([1])]; + tensor x_255_pad_0 = const()[name = string("x_255_pad_0"), val = tensor([0, 0])]; + tensor x_255_dilations_0 = const()[name = string("x_255_dilations_0"), val = tensor([1])]; + tensor encoder_layers_9_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(201196160))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(201205440))))[name = string("encoder_layers_9_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_255_cast_fp16 = conv(dilations = x_255_dilations_0, groups = x_255_groups_0, pad = x_255_pad_0, pad_type = x_255_pad_type_0, strides = x_255_strides_0, weight = encoder_layers_9_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_39_cast_fp16)[name = string("x_255_cast_fp16")]; + tensor input_523_perm_0 = const()[name = string("input_523_perm_0"), val = tensor([0, 2, 1])]; + tensor x_257_axes_0 = const()[name = string("x_257_axes_0"), val = tensor([-1])]; + tensor encoder_layers_9_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_9_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(201207552)))]; + tensor encoder_layers_9_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_9_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(201209664)))]; + tensor input_523_cast_fp16 = transpose(perm = input_523_perm_0, x = x_255_cast_fp16)[name = string("transpose_275")]; + tensor x_257_cast_fp16 = layer_norm(axes = x_257_axes_0, beta = encoder_layers_9_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_9_conv_batch_norm_weight_to_fp16, x = input_523_cast_fp16)[name = string("x_257_cast_fp16")]; + tensor input_525_perm_0 = const()[name = string("input_525_perm_0"), val = tensor([0, 2, 1])]; + tensor input_525_cast_fp16 = transpose(perm = input_525_perm_0, x = x_257_cast_fp16)[name = string("transpose_274")]; + tensor input_527_cast_fp16 = silu(x = input_525_cast_fp16)[name = string("input_527_cast_fp16")]; + string x_259_pad_type_0 = const()[name = string("x_259_pad_type_0"), val = string("valid")]; + tensor x_259_strides_0 = const()[name = string("x_259_strides_0"), val = tensor([1])]; + tensor x_259_pad_0 = const()[name = string("x_259_pad_0"), val = tensor([0, 0])]; + tensor x_259_dilations_0 = const()[name = string("x_259_dilations_0"), val = tensor([1])]; + int32 x_259_groups_0 = const()[name = string("x_259_groups_0"), val = int32(1)]; + tensor encoder_layers_9_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(201211776))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202260416))))[name = string("encoder_layers_9_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_259_cast_fp16 = conv(dilations = x_259_dilations_0, groups = x_259_groups_0, pad = x_259_pad_0, pad_type = x_259_pad_type_0, strides = x_259_strides_0, weight = encoder_layers_9_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_527_cast_fp16)[name = string("x_259_cast_fp16")]; + tensor input_529_perm_0 = const()[name = string("input_529_perm_0"), val = tensor([0, 2, 1])]; + tensor input_529_cast_fp16 = transpose(perm = input_529_perm_0, x = x_259_cast_fp16)[name = string("transpose_273")]; + tensor input_531_cast_fp16 = add(x = input_515_cast_fp16, y = input_529_cast_fp16)[name = string("input_531_cast_fp16")]; + tensor input_533_axes_0 = const()[name = string("input_533_axes_0"), val = tensor([-1])]; + tensor encoder_layers_9_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_9_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202262528)))]; + tensor encoder_layers_9_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_9_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202264640)))]; + tensor input_533_cast_fp16 = layer_norm(axes = input_533_axes_0, beta = encoder_layers_9_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_9_norm_feed_forward2_weight_to_fp16, x = input_531_cast_fp16)[name = string("input_533_cast_fp16")]; + tensor encoder_layers_9_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202266752))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(205412544))))[name = string("encoder_layers_9_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_9_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_9_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(205412736)))]; + tensor linear_89_cast_fp16 = linear(bias = encoder_layers_9_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_9_feed_forward2_linear1_weight_to_fp16_palettized, x = input_533_cast_fp16)[name = string("linear_89_cast_fp16")]; + tensor input_537_cast_fp16 = silu(x = linear_89_cast_fp16)[name = string("input_537_cast_fp16")]; + tensor encoder_layers_9_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(205420992))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(208566784))))[name = string("encoder_layers_9_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_9_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_9_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(208566976)))]; + tensor linear_90_cast_fp16 = linear(bias = encoder_layers_9_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_9_feed_forward2_linear2_weight_to_fp16_palettized, x = input_537_cast_fp16)[name = string("linear_90_cast_fp16")]; + fp16 var_2548_to_fp16 = const()[name = string("op_2548_to_fp16"), val = fp16(0x1p-1)]; + tensor var_2549_cast_fp16 = mul(x = linear_90_cast_fp16, y = var_2548_to_fp16)[name = string("op_2549_cast_fp16")]; + tensor input_543_cast_fp16 = add(x = input_531_cast_fp16, y = var_2549_cast_fp16)[name = string("input_543_cast_fp16")]; + tensor input_545_axes_0 = const()[name = string("input_545_axes_0"), val = tensor([-1])]; + tensor encoder_layers_9_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_9_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(208569088)))]; + tensor encoder_layers_9_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_9_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(208571200)))]; + tensor input_545_cast_fp16 = layer_norm(axes = input_545_axes_0, beta = encoder_layers_9_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_9_norm_out_weight_to_fp16, x = input_543_cast_fp16)[name = string("input_545_cast_fp16")]; + tensor cache_41_begin_0 = const()[name = string("cache_41_begin_0"), val = tensor([10, 0, 0, 0])]; + tensor cache_41_end_0 = const()[name = string("cache_41_end_0"), val = tensor([11, 1, 42, 1024])]; + tensor cache_41_end_mask_0 = const()[name = string("cache_41_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_41_squeeze_mask_0 = const()[name = string("cache_41_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_41_cast_fp16 = slice_by_index(begin = cache_41_begin_0, end = cache_41_end_0, end_mask = cache_41_end_mask_0, squeeze_mask = cache_41_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_41_cast_fp16")]; + tensor cache_43_begin_0 = const()[name = string("cache_43_begin_0"), val = tensor([10, 0, 0, 0])]; + tensor cache_43_end_0 = const()[name = string("cache_43_end_0"), val = tensor([11, 1, 1024, 8])]; + tensor cache_43_end_mask_0 = const()[name = string("cache_43_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_43_squeeze_mask_0 = const()[name = string("cache_43_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_43_cast_fp16 = slice_by_index(begin = cache_43_begin_0, end = cache_43_end_0, end_mask = cache_43_end_mask_0, squeeze_mask = cache_43_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_43_cast_fp16")]; + tensor input_547_axes_0 = const()[name = string("input_547_axes_0"), val = tensor([-1])]; + tensor encoder_layers_10_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_10_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(208573312)))]; + tensor encoder_layers_10_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_10_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(208575424)))]; + tensor input_547_cast_fp16 = layer_norm(axes = input_547_axes_0, beta = encoder_layers_10_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_10_norm_feed_forward1_weight_to_fp16, x = input_545_cast_fp16)[name = string("input_547_cast_fp16")]; + tensor encoder_layers_10_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(208577536))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211723328))))[name = string("encoder_layers_10_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_10_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_10_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211723520)))]; + tensor linear_91_cast_fp16 = linear(bias = encoder_layers_10_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_10_feed_forward1_linear1_weight_to_fp16_palettized, x = input_547_cast_fp16)[name = string("linear_91_cast_fp16")]; + tensor input_551_cast_fp16 = silu(x = linear_91_cast_fp16)[name = string("input_551_cast_fp16")]; + tensor encoder_layers_10_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211731776))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(214877568))))[name = string("encoder_layers_10_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_10_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_10_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(214877760)))]; + tensor linear_92_cast_fp16 = linear(bias = encoder_layers_10_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_10_feed_forward1_linear2_weight_to_fp16_palettized, x = input_551_cast_fp16)[name = string("linear_92_cast_fp16")]; + fp16 var_2585_to_fp16 = const()[name = string("op_2585_to_fp16"), val = fp16(0x1p-1)]; + tensor var_2586_cast_fp16 = mul(x = linear_92_cast_fp16, y = var_2585_to_fp16)[name = string("op_2586_cast_fp16")]; + tensor input_557_cast_fp16 = add(x = input_545_cast_fp16, y = var_2586_cast_fp16)[name = string("input_557_cast_fp16")]; + tensor key_21_axes_0 = const()[name = string("key_21_axes_0"), val = tensor([-1])]; + tensor encoder_layers_10_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_10_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(214879872)))]; + tensor encoder_layers_10_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_10_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(214881984)))]; + tensor key_21_cast_fp16 = layer_norm(axes = key_21_axes_0, beta = encoder_layers_10_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_10_norm_self_att_weight_to_fp16, x = input_557_cast_fp16)[name = string("key_21_cast_fp16")]; + bool input_559_interleave_0 = const()[name = string("input_559_interleave_0"), val = bool(false)]; + tensor input_559_cast_fp16 = concat(axis = var_68, interleave = input_559_interleave_0, values = (cache_41_cast_fp16, key_21_cast_fp16))[name = string("input_559_cast_fp16")]; + tensor var_2608_begin_0 = const()[name = string("op_2608_begin_0"), val = tensor([0, 7, 0])]; + tensor var_2608_end_0 = const()[name = string("op_2608_end_0"), val = tensor([1, 42, 1024])]; + tensor var_2608_end_mask_0 = const()[name = string("op_2608_end_mask_0"), val = tensor([true, true, true])]; + tensor var_2608_cast_fp16 = slice_by_index(begin = var_2608_begin_0, end = var_2608_end_0, end_mask = var_2608_end_mask_0, x = cache_41_cast_fp16)[name = string("op_2608_cast_fp16")]; + bool var_2614_interleave_0 = const()[name = string("op_2614_interleave_0"), val = bool(false)]; + tensor var_2614_cast_fp16 = concat(axis = var_68, interleave = var_2614_interleave_0, values = (var_2608_cast_fp16, key_21_cast_fp16))[name = string("op_2614_cast_fp16")]; + tensor encoder_layers_10_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(214884096))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215670592))))[name = string("encoder_layers_10_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_10_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_10_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215670784)))]; + tensor linear_93_cast_fp16 = linear(bias = encoder_layers_10_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_10_self_attn_linear_q_weight_to_fp16_palettized, x = key_21_cast_fp16)[name = string("linear_93_cast_fp16")]; + tensor var_2619 = const()[name = string("op_2619"), val = tensor([1, -1, 8, 128])]; + tensor q_61_cast_fp16 = reshape(shape = var_2619, x = linear_93_cast_fp16)[name = string("q_61_cast_fp16")]; + tensor encoder_layers_10_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215672896))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216459392))))[name = string("encoder_layers_10_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_10_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_10_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216459584)))]; + tensor linear_94_cast_fp16 = linear(bias = encoder_layers_10_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_10_self_attn_linear_k_weight_to_fp16_palettized, x = input_559_cast_fp16)[name = string("linear_94_cast_fp16")]; + tensor var_2624 = const()[name = string("op_2624"), val = tensor([1, -1, 8, 128])]; + tensor k_41_cast_fp16 = reshape(shape = var_2624, x = linear_94_cast_fp16)[name = string("k_41_cast_fp16")]; + tensor encoder_layers_10_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216461696))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217248192))))[name = string("encoder_layers_10_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_10_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_10_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217248384)))]; + tensor linear_95_cast_fp16 = linear(bias = encoder_layers_10_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_10_self_attn_linear_v_weight_to_fp16_palettized, x = input_559_cast_fp16)[name = string("linear_95_cast_fp16")]; + tensor var_2629 = const()[name = string("op_2629"), val = tensor([1, -1, 8, 128])]; + tensor v_21_cast_fp16 = reshape(shape = var_2629, x = linear_95_cast_fp16)[name = string("v_21_cast_fp16")]; + tensor value_29_perm_0 = const()[name = string("value_29_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_10_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_10_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217250496)))]; + tensor var_2642_cast_fp16 = add(x = q_61_cast_fp16, y = encoder_layers_10_self_attn_pos_bias_u_to_fp16)[name = string("op_2642_cast_fp16")]; + tensor encoder_layers_10_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_10_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217252608)))]; + tensor var_2644_cast_fp16 = add(x = q_61_cast_fp16, y = encoder_layers_10_self_attn_pos_bias_v_to_fp16)[name = string("op_2644_cast_fp16")]; + tensor q_with_bias_v_21_perm_0 = const()[name = string("q_with_bias_v_21_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_267_transpose_x_0 = const()[name = string("x_267_transpose_x_0"), val = bool(false)]; + bool x_267_transpose_y_0 = const()[name = string("x_267_transpose_y_0"), val = bool(false)]; + tensor op_2646_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217254720))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217354112))))[name = string("op_2646_to_fp16_quantized")]; + tensor q_with_bias_v_21_cast_fp16 = transpose(perm = q_with_bias_v_21_perm_0, x = var_2644_cast_fp16)[name = string("transpose_272")]; + tensor x_267_cast_fp16 = matmul(transpose_x = x_267_transpose_x_0, transpose_y = x_267_transpose_y_0, x = q_with_bias_v_21_cast_fp16, y = op_2646_to_fp16_quantized)[name = string("x_267_cast_fp16")]; + tensor x_269_pad_0 = const()[name = string("x_269_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_269_mode_0 = const()[name = string("x_269_mode_0"), val = string("constant")]; + fp16 const_209_to_fp16 = const()[name = string("const_209_to_fp16"), val = fp16(0x0p+0)]; + tensor x_269_cast_fp16 = pad(constant_val = const_209_to_fp16, mode = x_269_mode_0, pad = x_269_pad_0, x = x_267_cast_fp16)[name = string("x_269_cast_fp16")]; + tensor var_2654 = const()[name = string("op_2654"), val = tensor([1, 8, -1, 7])]; + tensor x_271_cast_fp16 = reshape(shape = var_2654, x = x_269_cast_fp16)[name = string("x_271_cast_fp16")]; + tensor var_2658_begin_0 = const()[name = string("op_2658_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_2658_end_0 = const()[name = string("op_2658_end_0"), val = tensor([1, 8, 98, 7])]; + tensor var_2658_end_mask_0 = const()[name = string("op_2658_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_2658_cast_fp16 = slice_by_index(begin = var_2658_begin_0, end = var_2658_end_0, end_mask = var_2658_end_mask_0, x = x_271_cast_fp16)[name = string("op_2658_cast_fp16")]; + tensor var_2659 = const()[name = string("op_2659"), val = tensor([1, 8, 7, 97])]; + tensor matrix_bd_41_cast_fp16 = reshape(shape = var_2659, x = var_2658_cast_fp16)[name = string("matrix_bd_41_cast_fp16")]; + bool matrix_ac_21_transpose_x_0 = const()[name = string("matrix_ac_21_transpose_x_0"), val = bool(false)]; + bool matrix_ac_21_transpose_y_0 = const()[name = string("matrix_ac_21_transpose_y_0"), val = bool(false)]; + tensor transpose_116_perm_0 = const()[name = string("transpose_116_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_117_perm_0 = const()[name = string("transpose_117_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_117 = transpose(perm = transpose_117_perm_0, x = k_41_cast_fp16)[name = string("transpose_270")]; + tensor transpose_116 = transpose(perm = transpose_116_perm_0, x = var_2642_cast_fp16)[name = string("transpose_271")]; + tensor matrix_ac_21_cast_fp16 = matmul(transpose_x = matrix_ac_21_transpose_x_0, transpose_y = matrix_ac_21_transpose_y_0, x = transpose_116, y = transpose_117)[name = string("matrix_ac_21_cast_fp16")]; + tensor matrix_bd_43_begin_0 = const()[name = string("matrix_bd_43_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_43_end_0 = const()[name = string("matrix_bd_43_end_0"), val = tensor([1, 8, 7, 49])]; + tensor matrix_bd_43_end_mask_0 = const()[name = string("matrix_bd_43_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_43_cast_fp16 = slice_by_index(begin = matrix_bd_43_begin_0, end = matrix_bd_43_end_0, end_mask = matrix_bd_43_end_mask_0, x = matrix_bd_41_cast_fp16)[name = string("matrix_bd_43_cast_fp16")]; + tensor var_2668_cast_fp16 = add(x = matrix_ac_21_cast_fp16, y = matrix_bd_43_cast_fp16)[name = string("op_2668_cast_fp16")]; + fp16 _inversed_scores_41_y_0_to_fp16 = const()[name = string("_inversed_scores_41_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_41_cast_fp16 = mul(x = var_2668_cast_fp16, y = _inversed_scores_41_y_0_to_fp16)[name = string("_inversed_scores_41_cast_fp16")]; + tensor scores_43_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_41_cast_fp16, cond = mask_11)[name = string("scores_43_cast_fp16")]; + tensor var_2674_cast_fp16 = softmax(axis = var_59, x = scores_43_cast_fp16)[name = string("op_2674_cast_fp16")]; + tensor input_561_cast_fp16 = select(a = var_44_to_fp16, b = var_2674_cast_fp16, cond = mask_11)[name = string("input_561_cast_fp16")]; + bool x_273_transpose_x_0 = const()[name = string("x_273_transpose_x_0"), val = bool(false)]; + bool x_273_transpose_y_0 = const()[name = string("x_273_transpose_y_0"), val = bool(false)]; + tensor value_29_cast_fp16 = transpose(perm = value_29_perm_0, x = v_21_cast_fp16)[name = string("transpose_269")]; + tensor x_273_cast_fp16 = matmul(transpose_x = x_273_transpose_x_0, transpose_y = x_273_transpose_y_0, x = input_561_cast_fp16, y = value_29_cast_fp16)[name = string("x_273_cast_fp16")]; + tensor var_2678_perm_0 = const()[name = string("op_2678_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_2679 = const()[name = string("op_2679"), val = tensor([1, -1, 1024])]; + tensor var_2678_cast_fp16 = transpose(perm = var_2678_perm_0, x = x_273_cast_fp16)[name = string("transpose_268")]; + tensor input_563_cast_fp16 = reshape(shape = var_2679, x = var_2678_cast_fp16)[name = string("input_563_cast_fp16")]; + tensor encoder_layers_10_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217354432))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218140928))))[name = string("encoder_layers_10_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_10_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_10_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218141120)))]; + tensor linear_97_cast_fp16 = linear(bias = encoder_layers_10_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_10_self_attn_linear_out_weight_to_fp16_palettized, x = input_563_cast_fp16)[name = string("linear_97_cast_fp16")]; + tensor input_567_cast_fp16 = add(x = input_557_cast_fp16, y = linear_97_cast_fp16)[name = string("input_567_cast_fp16")]; + tensor x_277_axes_0 = const()[name = string("x_277_axes_0"), val = tensor([-1])]; + tensor encoder_layers_10_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_10_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218143232)))]; + tensor encoder_layers_10_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_10_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218145344)))]; + tensor x_277_cast_fp16 = layer_norm(axes = x_277_axes_0, beta = encoder_layers_10_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_10_norm_conv_weight_to_fp16, x = input_567_cast_fp16)[name = string("x_277_cast_fp16")]; + tensor input_569_perm_0 = const()[name = string("input_569_perm_0"), val = tensor([0, 2, 1])]; + string input_571_pad_type_0 = const()[name = string("input_571_pad_type_0"), val = string("valid")]; + tensor input_571_strides_0 = const()[name = string("input_571_strides_0"), val = tensor([1])]; + tensor input_571_pad_0 = const()[name = string("input_571_pad_0"), val = tensor([0, 0])]; + tensor input_571_dilations_0 = const()[name = string("input_571_dilations_0"), val = tensor([1])]; + int32 input_571_groups_0 = const()[name = string("input_571_groups_0"), val = int32(1)]; + tensor encoder_layers_10_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218147456))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220244672))))[name = string("encoder_layers_10_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_569_cast_fp16 = transpose(perm = input_569_perm_0, x = x_277_cast_fp16)[name = string("transpose_267")]; + tensor input_571_cast_fp16 = conv(dilations = input_571_dilations_0, groups = input_571_groups_0, pad = input_571_pad_0, pad_type = input_571_pad_type_0, strides = input_571_strides_0, weight = encoder_layers_10_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_569_cast_fp16)[name = string("input_571_cast_fp16")]; + int32 x_279_split_num_splits_0 = const()[name = string("x_279_split_num_splits_0"), val = int32(2)]; + int32 x_279_split_axis_0 = const()[name = string("x_279_split_axis_0"), val = int32(1)]; + tensor x_279_split_cast_fp16_0, tensor x_279_split_cast_fp16_1 = split(axis = x_279_split_axis_0, num_splits = x_279_split_num_splits_0, x = input_571_cast_fp16)[name = string("x_279_split_cast_fp16")]; + tensor x_279_split_1_sigmoid_cast_fp16 = sigmoid(x = x_279_split_cast_fp16_1)[name = string("x_279_split_1_sigmoid_cast_fp16")]; + tensor x_279_cast_fp16 = mul(x = x_279_split_cast_fp16_0, y = x_279_split_1_sigmoid_cast_fp16)[name = string("x_279_cast_fp16")]; + tensor input_573_cast_fp16 = select(a = var_44_to_fp16, b = x_279_cast_fp16, cond = var_575)[name = string("input_573_cast_fp16")]; + bool new_x_43_interleave_0 = const()[name = string("new_x_43_interleave_0"), val = bool(false)]; + tensor new_x_43_cast_fp16 = concat(axis = var_59, interleave = new_x_43_interleave_0, values = (cache_43_cast_fp16, input_573_cast_fp16))[name = string("new_x_43_cast_fp16")]; + tensor var_2718_begin_0 = const()[name = string("op_2718_begin_0"), val = tensor([0, 0, 7])]; + tensor var_2718_end_0 = const()[name = string("op_2718_end_0"), val = tensor([1, 1024, 15])]; + tensor var_2718_end_mask_0 = const()[name = string("op_2718_end_mask_0"), val = tensor([true, true, true])]; + tensor var_2718_cast_fp16 = slice_by_index(begin = var_2718_begin_0, end = var_2718_end_0, end_mask = var_2718_end_mask_0, x = new_x_43_cast_fp16)[name = string("op_2718_cast_fp16")]; + string x_281_pad_type_0 = const()[name = string("x_281_pad_type_0"), val = string("valid")]; + int32 x_281_groups_0 = const()[name = string("x_281_groups_0"), val = int32(1024)]; + tensor x_281_strides_0 = const()[name = string("x_281_strides_0"), val = tensor([1])]; + tensor x_281_pad_0 = const()[name = string("x_281_pad_0"), val = tensor([0, 0])]; + tensor x_281_dilations_0 = const()[name = string("x_281_dilations_0"), val = tensor([1])]; + tensor encoder_layers_10_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220248832))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220258112))))[name = string("encoder_layers_10_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_281_cast_fp16 = conv(dilations = x_281_dilations_0, groups = x_281_groups_0, pad = x_281_pad_0, pad_type = x_281_pad_type_0, strides = x_281_strides_0, weight = encoder_layers_10_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_43_cast_fp16)[name = string("x_281_cast_fp16")]; + tensor input_575_perm_0 = const()[name = string("input_575_perm_0"), val = tensor([0, 2, 1])]; + tensor x_283_axes_0 = const()[name = string("x_283_axes_0"), val = tensor([-1])]; + tensor encoder_layers_10_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_10_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220260224)))]; + tensor encoder_layers_10_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_10_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220262336)))]; + tensor input_575_cast_fp16 = transpose(perm = input_575_perm_0, x = x_281_cast_fp16)[name = string("transpose_266")]; + tensor x_283_cast_fp16 = layer_norm(axes = x_283_axes_0, beta = encoder_layers_10_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_10_conv_batch_norm_weight_to_fp16, x = input_575_cast_fp16)[name = string("x_283_cast_fp16")]; + tensor input_577_perm_0 = const()[name = string("input_577_perm_0"), val = tensor([0, 2, 1])]; + tensor input_577_cast_fp16 = transpose(perm = input_577_perm_0, x = x_283_cast_fp16)[name = string("transpose_265")]; + tensor input_579_cast_fp16 = silu(x = input_577_cast_fp16)[name = string("input_579_cast_fp16")]; + string x_285_pad_type_0 = const()[name = string("x_285_pad_type_0"), val = string("valid")]; + tensor x_285_strides_0 = const()[name = string("x_285_strides_0"), val = tensor([1])]; + tensor x_285_pad_0 = const()[name = string("x_285_pad_0"), val = tensor([0, 0])]; + tensor x_285_dilations_0 = const()[name = string("x_285_dilations_0"), val = tensor([1])]; + int32 x_285_groups_0 = const()[name = string("x_285_groups_0"), val = int32(1)]; + tensor encoder_layers_10_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220264448))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221313088))))[name = string("encoder_layers_10_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_285_cast_fp16 = conv(dilations = x_285_dilations_0, groups = x_285_groups_0, pad = x_285_pad_0, pad_type = x_285_pad_type_0, strides = x_285_strides_0, weight = encoder_layers_10_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_579_cast_fp16)[name = string("x_285_cast_fp16")]; + tensor input_581_perm_0 = const()[name = string("input_581_perm_0"), val = tensor([0, 2, 1])]; + tensor input_581_cast_fp16 = transpose(perm = input_581_perm_0, x = x_285_cast_fp16)[name = string("transpose_264")]; + tensor input_583_cast_fp16 = add(x = input_567_cast_fp16, y = input_581_cast_fp16)[name = string("input_583_cast_fp16")]; + tensor input_585_axes_0 = const()[name = string("input_585_axes_0"), val = tensor([-1])]; + tensor encoder_layers_10_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_10_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221315200)))]; + tensor encoder_layers_10_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_10_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221317312)))]; + tensor input_585_cast_fp16 = layer_norm(axes = input_585_axes_0, beta = encoder_layers_10_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_10_norm_feed_forward2_weight_to_fp16, x = input_583_cast_fp16)[name = string("input_585_cast_fp16")]; + tensor encoder_layers_10_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221319424))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(224465216))))[name = string("encoder_layers_10_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_10_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_10_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(224465408)))]; + tensor linear_98_cast_fp16 = linear(bias = encoder_layers_10_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_10_feed_forward2_linear1_weight_to_fp16_palettized, x = input_585_cast_fp16)[name = string("linear_98_cast_fp16")]; + tensor input_589_cast_fp16 = silu(x = linear_98_cast_fp16)[name = string("input_589_cast_fp16")]; + tensor encoder_layers_10_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(224473664))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227619456))))[name = string("encoder_layers_10_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_10_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_10_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227619648)))]; + tensor linear_99_cast_fp16 = linear(bias = encoder_layers_10_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_10_feed_forward2_linear2_weight_to_fp16_palettized, x = input_589_cast_fp16)[name = string("linear_99_cast_fp16")]; + fp16 var_2761_to_fp16 = const()[name = string("op_2761_to_fp16"), val = fp16(0x1p-1)]; + tensor var_2762_cast_fp16 = mul(x = linear_99_cast_fp16, y = var_2761_to_fp16)[name = string("op_2762_cast_fp16")]; + tensor input_595_cast_fp16 = add(x = input_583_cast_fp16, y = var_2762_cast_fp16)[name = string("input_595_cast_fp16")]; + tensor input_597_axes_0 = const()[name = string("input_597_axes_0"), val = tensor([-1])]; + tensor encoder_layers_10_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_10_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227621760)))]; + tensor encoder_layers_10_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_10_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227623872)))]; + tensor input_597_cast_fp16 = layer_norm(axes = input_597_axes_0, beta = encoder_layers_10_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_10_norm_out_weight_to_fp16, x = input_595_cast_fp16)[name = string("input_597_cast_fp16")]; + tensor cache_45_begin_0 = const()[name = string("cache_45_begin_0"), val = tensor([11, 0, 0, 0])]; + tensor cache_45_end_0 = const()[name = string("cache_45_end_0"), val = tensor([12, 1, 42, 1024])]; + tensor cache_45_end_mask_0 = const()[name = string("cache_45_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_45_squeeze_mask_0 = const()[name = string("cache_45_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_45_cast_fp16 = slice_by_index(begin = cache_45_begin_0, end = cache_45_end_0, end_mask = cache_45_end_mask_0, squeeze_mask = cache_45_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_45_cast_fp16")]; + tensor cache_47_begin_0 = const()[name = string("cache_47_begin_0"), val = tensor([11, 0, 0, 0])]; + tensor cache_47_end_0 = const()[name = string("cache_47_end_0"), val = tensor([12, 1, 1024, 8])]; + tensor cache_47_end_mask_0 = const()[name = string("cache_47_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_47_squeeze_mask_0 = const()[name = string("cache_47_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_47_cast_fp16 = slice_by_index(begin = cache_47_begin_0, end = cache_47_end_0, end_mask = cache_47_end_mask_0, squeeze_mask = cache_47_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_47_cast_fp16")]; + tensor input_599_axes_0 = const()[name = string("input_599_axes_0"), val = tensor([-1])]; + tensor encoder_layers_11_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_11_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227625984)))]; + tensor encoder_layers_11_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_11_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227628096)))]; + tensor input_599_cast_fp16 = layer_norm(axes = input_599_axes_0, beta = encoder_layers_11_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_11_norm_feed_forward1_weight_to_fp16, x = input_597_cast_fp16)[name = string("input_599_cast_fp16")]; + tensor encoder_layers_11_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227630208))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(230776000))))[name = string("encoder_layers_11_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_11_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_11_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(230776192)))]; + tensor linear_100_cast_fp16 = linear(bias = encoder_layers_11_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_11_feed_forward1_linear1_weight_to_fp16_palettized, x = input_599_cast_fp16)[name = string("linear_100_cast_fp16")]; + tensor input_603_cast_fp16 = silu(x = linear_100_cast_fp16)[name = string("input_603_cast_fp16")]; + tensor encoder_layers_11_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(230784448))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(233930240))))[name = string("encoder_layers_11_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_11_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_11_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(233930432)))]; + tensor linear_101_cast_fp16 = linear(bias = encoder_layers_11_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_11_feed_forward1_linear2_weight_to_fp16_palettized, x = input_603_cast_fp16)[name = string("linear_101_cast_fp16")]; + fp16 var_2798_to_fp16 = const()[name = string("op_2798_to_fp16"), val = fp16(0x1p-1)]; + tensor var_2799_cast_fp16 = mul(x = linear_101_cast_fp16, y = var_2798_to_fp16)[name = string("op_2799_cast_fp16")]; + tensor input_609_cast_fp16 = add(x = input_597_cast_fp16, y = var_2799_cast_fp16)[name = string("input_609_cast_fp16")]; + tensor key_23_axes_0 = const()[name = string("key_23_axes_0"), val = tensor([-1])]; + tensor encoder_layers_11_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_11_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(233932544)))]; + tensor encoder_layers_11_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_11_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(233934656)))]; + tensor key_23_cast_fp16 = layer_norm(axes = key_23_axes_0, beta = encoder_layers_11_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_11_norm_self_att_weight_to_fp16, x = input_609_cast_fp16)[name = string("key_23_cast_fp16")]; + bool input_611_interleave_0 = const()[name = string("input_611_interleave_0"), val = bool(false)]; + tensor input_611_cast_fp16 = concat(axis = var_68, interleave = input_611_interleave_0, values = (cache_45_cast_fp16, key_23_cast_fp16))[name = string("input_611_cast_fp16")]; + tensor var_2821_begin_0 = const()[name = string("op_2821_begin_0"), val = tensor([0, 7, 0])]; + tensor var_2821_end_0 = const()[name = string("op_2821_end_0"), val = tensor([1, 42, 1024])]; + tensor var_2821_end_mask_0 = const()[name = string("op_2821_end_mask_0"), val = tensor([true, true, true])]; + tensor var_2821_cast_fp16 = slice_by_index(begin = var_2821_begin_0, end = var_2821_end_0, end_mask = var_2821_end_mask_0, x = cache_45_cast_fp16)[name = string("op_2821_cast_fp16")]; + bool var_2827_interleave_0 = const()[name = string("op_2827_interleave_0"), val = bool(false)]; + tensor var_2827_cast_fp16 = concat(axis = var_68, interleave = var_2827_interleave_0, values = (var_2821_cast_fp16, key_23_cast_fp16))[name = string("op_2827_cast_fp16")]; + tensor encoder_layers_11_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(233936768))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234723264))))[name = string("encoder_layers_11_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_11_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_11_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234723456)))]; + tensor linear_102_cast_fp16 = linear(bias = encoder_layers_11_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_11_self_attn_linear_q_weight_to_fp16_palettized, x = key_23_cast_fp16)[name = string("linear_102_cast_fp16")]; + tensor var_2832 = const()[name = string("op_2832"), val = tensor([1, -1, 8, 128])]; + tensor q_67_cast_fp16 = reshape(shape = var_2832, x = linear_102_cast_fp16)[name = string("q_67_cast_fp16")]; + tensor encoder_layers_11_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234725568))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(235512064))))[name = string("encoder_layers_11_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_11_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_11_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(235512256)))]; + tensor linear_103_cast_fp16 = linear(bias = encoder_layers_11_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_11_self_attn_linear_k_weight_to_fp16_palettized, x = input_611_cast_fp16)[name = string("linear_103_cast_fp16")]; + tensor var_2837 = const()[name = string("op_2837"), val = tensor([1, -1, 8, 128])]; + tensor k_45_cast_fp16 = reshape(shape = var_2837, x = linear_103_cast_fp16)[name = string("k_45_cast_fp16")]; + tensor encoder_layers_11_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(235514368))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236300864))))[name = string("encoder_layers_11_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_11_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_11_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236301056)))]; + tensor linear_104_cast_fp16 = linear(bias = encoder_layers_11_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_11_self_attn_linear_v_weight_to_fp16_palettized, x = input_611_cast_fp16)[name = string("linear_104_cast_fp16")]; + tensor var_2842 = const()[name = string("op_2842"), val = tensor([1, -1, 8, 128])]; + tensor v_23_cast_fp16 = reshape(shape = var_2842, x = linear_104_cast_fp16)[name = string("v_23_cast_fp16")]; + tensor value_31_perm_0 = const()[name = string("value_31_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_11_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_11_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236303168)))]; + tensor var_2855_cast_fp16 = add(x = q_67_cast_fp16, y = encoder_layers_11_self_attn_pos_bias_u_to_fp16)[name = string("op_2855_cast_fp16")]; + tensor encoder_layers_11_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_11_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236305280)))]; + tensor var_2857_cast_fp16 = add(x = q_67_cast_fp16, y = encoder_layers_11_self_attn_pos_bias_v_to_fp16)[name = string("op_2857_cast_fp16")]; + tensor q_with_bias_v_23_perm_0 = const()[name = string("q_with_bias_v_23_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_293_transpose_x_0 = const()[name = string("x_293_transpose_x_0"), val = bool(false)]; + bool x_293_transpose_y_0 = const()[name = string("x_293_transpose_y_0"), val = bool(false)]; + tensor op_2859_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236307392))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236406784))))[name = string("op_2859_to_fp16_quantized")]; + tensor q_with_bias_v_23_cast_fp16 = transpose(perm = q_with_bias_v_23_perm_0, x = var_2857_cast_fp16)[name = string("transpose_263")]; + tensor x_293_cast_fp16 = matmul(transpose_x = x_293_transpose_x_0, transpose_y = x_293_transpose_y_0, x = q_with_bias_v_23_cast_fp16, y = op_2859_to_fp16_quantized)[name = string("x_293_cast_fp16")]; + tensor x_295_pad_0 = const()[name = string("x_295_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_295_mode_0 = const()[name = string("x_295_mode_0"), val = string("constant")]; + fp16 const_222_to_fp16 = const()[name = string("const_222_to_fp16"), val = fp16(0x0p+0)]; + tensor x_295_cast_fp16 = pad(constant_val = const_222_to_fp16, mode = x_295_mode_0, pad = x_295_pad_0, x = x_293_cast_fp16)[name = string("x_295_cast_fp16")]; + tensor var_2867 = const()[name = string("op_2867"), val = tensor([1, 8, -1, 7])]; + tensor x_297_cast_fp16 = reshape(shape = var_2867, x = x_295_cast_fp16)[name = string("x_297_cast_fp16")]; + tensor var_2871_begin_0 = const()[name = string("op_2871_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_2871_end_0 = const()[name = string("op_2871_end_0"), val = tensor([1, 8, 98, 7])]; + tensor var_2871_end_mask_0 = const()[name = string("op_2871_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_2871_cast_fp16 = slice_by_index(begin = var_2871_begin_0, end = var_2871_end_0, end_mask = var_2871_end_mask_0, x = x_297_cast_fp16)[name = string("op_2871_cast_fp16")]; + tensor var_2872 = const()[name = string("op_2872"), val = tensor([1, 8, 7, 97])]; + tensor matrix_bd_45_cast_fp16 = reshape(shape = var_2872, x = var_2871_cast_fp16)[name = string("matrix_bd_45_cast_fp16")]; + bool matrix_ac_23_transpose_x_0 = const()[name = string("matrix_ac_23_transpose_x_0"), val = bool(false)]; + bool matrix_ac_23_transpose_y_0 = const()[name = string("matrix_ac_23_transpose_y_0"), val = bool(false)]; + tensor transpose_118_perm_0 = const()[name = string("transpose_118_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_119_perm_0 = const()[name = string("transpose_119_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_119 = transpose(perm = transpose_119_perm_0, x = k_45_cast_fp16)[name = string("transpose_261")]; + tensor transpose_118 = transpose(perm = transpose_118_perm_0, x = var_2855_cast_fp16)[name = string("transpose_262")]; + tensor matrix_ac_23_cast_fp16 = matmul(transpose_x = matrix_ac_23_transpose_x_0, transpose_y = matrix_ac_23_transpose_y_0, x = transpose_118, y = transpose_119)[name = string("matrix_ac_23_cast_fp16")]; + tensor matrix_bd_47_begin_0 = const()[name = string("matrix_bd_47_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_47_end_0 = const()[name = string("matrix_bd_47_end_0"), val = tensor([1, 8, 7, 49])]; + tensor matrix_bd_47_end_mask_0 = const()[name = string("matrix_bd_47_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_47_cast_fp16 = slice_by_index(begin = matrix_bd_47_begin_0, end = matrix_bd_47_end_0, end_mask = matrix_bd_47_end_mask_0, x = matrix_bd_45_cast_fp16)[name = string("matrix_bd_47_cast_fp16")]; + tensor var_2881_cast_fp16 = add(x = matrix_ac_23_cast_fp16, y = matrix_bd_47_cast_fp16)[name = string("op_2881_cast_fp16")]; + fp16 _inversed_scores_45_y_0_to_fp16 = const()[name = string("_inversed_scores_45_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_45_cast_fp16 = mul(x = var_2881_cast_fp16, y = _inversed_scores_45_y_0_to_fp16)[name = string("_inversed_scores_45_cast_fp16")]; + tensor scores_47_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_45_cast_fp16, cond = mask_11)[name = string("scores_47_cast_fp16")]; + tensor var_2887_cast_fp16 = softmax(axis = var_59, x = scores_47_cast_fp16)[name = string("op_2887_cast_fp16")]; + tensor input_613_cast_fp16 = select(a = var_44_to_fp16, b = var_2887_cast_fp16, cond = mask_11)[name = string("input_613_cast_fp16")]; + bool x_299_transpose_x_0 = const()[name = string("x_299_transpose_x_0"), val = bool(false)]; + bool x_299_transpose_y_0 = const()[name = string("x_299_transpose_y_0"), val = bool(false)]; + tensor value_31_cast_fp16 = transpose(perm = value_31_perm_0, x = v_23_cast_fp16)[name = string("transpose_260")]; + tensor x_299_cast_fp16 = matmul(transpose_x = x_299_transpose_x_0, transpose_y = x_299_transpose_y_0, x = input_613_cast_fp16, y = value_31_cast_fp16)[name = string("x_299_cast_fp16")]; + tensor var_2891_perm_0 = const()[name = string("op_2891_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_2892 = const()[name = string("op_2892"), val = tensor([1, -1, 1024])]; + tensor var_2891_cast_fp16 = transpose(perm = var_2891_perm_0, x = x_299_cast_fp16)[name = string("transpose_259")]; + tensor input_615_cast_fp16 = reshape(shape = var_2892, x = var_2891_cast_fp16)[name = string("input_615_cast_fp16")]; + tensor encoder_layers_11_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236407104))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(237193600))))[name = string("encoder_layers_11_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_11_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_11_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(237193792)))]; + tensor linear_106_cast_fp16 = linear(bias = encoder_layers_11_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_11_self_attn_linear_out_weight_to_fp16_palettized, x = input_615_cast_fp16)[name = string("linear_106_cast_fp16")]; + tensor input_619_cast_fp16 = add(x = input_609_cast_fp16, y = linear_106_cast_fp16)[name = string("input_619_cast_fp16")]; + tensor x_303_axes_0 = const()[name = string("x_303_axes_0"), val = tensor([-1])]; + tensor encoder_layers_11_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_11_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(237195904)))]; + tensor encoder_layers_11_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_11_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(237198016)))]; + tensor x_303_cast_fp16 = layer_norm(axes = x_303_axes_0, beta = encoder_layers_11_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_11_norm_conv_weight_to_fp16, x = input_619_cast_fp16)[name = string("x_303_cast_fp16")]; + tensor input_621_perm_0 = const()[name = string("input_621_perm_0"), val = tensor([0, 2, 1])]; + string input_623_pad_type_0 = const()[name = string("input_623_pad_type_0"), val = string("valid")]; + tensor input_623_strides_0 = const()[name = string("input_623_strides_0"), val = tensor([1])]; + tensor input_623_pad_0 = const()[name = string("input_623_pad_0"), val = tensor([0, 0])]; + tensor input_623_dilations_0 = const()[name = string("input_623_dilations_0"), val = tensor([1])]; + int32 input_623_groups_0 = const()[name = string("input_623_groups_0"), val = int32(1)]; + tensor encoder_layers_11_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(237200128))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(239297344))))[name = string("encoder_layers_11_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_621_cast_fp16 = transpose(perm = input_621_perm_0, x = x_303_cast_fp16)[name = string("transpose_258")]; + tensor input_623_cast_fp16 = conv(dilations = input_623_dilations_0, groups = input_623_groups_0, pad = input_623_pad_0, pad_type = input_623_pad_type_0, strides = input_623_strides_0, weight = encoder_layers_11_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_621_cast_fp16)[name = string("input_623_cast_fp16")]; + int32 x_305_split_num_splits_0 = const()[name = string("x_305_split_num_splits_0"), val = int32(2)]; + int32 x_305_split_axis_0 = const()[name = string("x_305_split_axis_0"), val = int32(1)]; + tensor x_305_split_cast_fp16_0, tensor x_305_split_cast_fp16_1 = split(axis = x_305_split_axis_0, num_splits = x_305_split_num_splits_0, x = input_623_cast_fp16)[name = string("x_305_split_cast_fp16")]; + tensor x_305_split_1_sigmoid_cast_fp16 = sigmoid(x = x_305_split_cast_fp16_1)[name = string("x_305_split_1_sigmoid_cast_fp16")]; + tensor x_305_cast_fp16 = mul(x = x_305_split_cast_fp16_0, y = x_305_split_1_sigmoid_cast_fp16)[name = string("x_305_cast_fp16")]; + tensor input_625_cast_fp16 = select(a = var_44_to_fp16, b = x_305_cast_fp16, cond = var_575)[name = string("input_625_cast_fp16")]; + bool new_x_47_interleave_0 = const()[name = string("new_x_47_interleave_0"), val = bool(false)]; + tensor new_x_47_cast_fp16 = concat(axis = var_59, interleave = new_x_47_interleave_0, values = (cache_47_cast_fp16, input_625_cast_fp16))[name = string("new_x_47_cast_fp16")]; + tensor var_2931_begin_0 = const()[name = string("op_2931_begin_0"), val = tensor([0, 0, 7])]; + tensor var_2931_end_0 = const()[name = string("op_2931_end_0"), val = tensor([1, 1024, 15])]; + tensor var_2931_end_mask_0 = const()[name = string("op_2931_end_mask_0"), val = tensor([true, true, true])]; + tensor var_2931_cast_fp16 = slice_by_index(begin = var_2931_begin_0, end = var_2931_end_0, end_mask = var_2931_end_mask_0, x = new_x_47_cast_fp16)[name = string("op_2931_cast_fp16")]; + string x_307_pad_type_0 = const()[name = string("x_307_pad_type_0"), val = string("valid")]; + int32 x_307_groups_0 = const()[name = string("x_307_groups_0"), val = int32(1024)]; + tensor x_307_strides_0 = const()[name = string("x_307_strides_0"), val = tensor([1])]; + tensor x_307_pad_0 = const()[name = string("x_307_pad_0"), val = tensor([0, 0])]; + tensor x_307_dilations_0 = const()[name = string("x_307_dilations_0"), val = tensor([1])]; + tensor encoder_layers_11_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(239301504))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(239310784))))[name = string("encoder_layers_11_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_307_cast_fp16 = conv(dilations = x_307_dilations_0, groups = x_307_groups_0, pad = x_307_pad_0, pad_type = x_307_pad_type_0, strides = x_307_strides_0, weight = encoder_layers_11_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_47_cast_fp16)[name = string("x_307_cast_fp16")]; + tensor input_627_perm_0 = const()[name = string("input_627_perm_0"), val = tensor([0, 2, 1])]; + tensor x_309_axes_0 = const()[name = string("x_309_axes_0"), val = tensor([-1])]; + tensor encoder_layers_11_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_11_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(239312896)))]; + tensor encoder_layers_11_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_11_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(239315008)))]; + tensor input_627_cast_fp16 = transpose(perm = input_627_perm_0, x = x_307_cast_fp16)[name = string("transpose_257")]; + tensor x_309_cast_fp16 = layer_norm(axes = x_309_axes_0, beta = encoder_layers_11_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_11_conv_batch_norm_weight_to_fp16, x = input_627_cast_fp16)[name = string("x_309_cast_fp16")]; + tensor input_629_perm_0 = const()[name = string("input_629_perm_0"), val = tensor([0, 2, 1])]; + tensor input_629_cast_fp16 = transpose(perm = input_629_perm_0, x = x_309_cast_fp16)[name = string("transpose_256")]; + tensor input_631_cast_fp16 = silu(x = input_629_cast_fp16)[name = string("input_631_cast_fp16")]; + string x_311_pad_type_0 = const()[name = string("x_311_pad_type_0"), val = string("valid")]; + tensor x_311_strides_0 = const()[name = string("x_311_strides_0"), val = tensor([1])]; + tensor x_311_pad_0 = const()[name = string("x_311_pad_0"), val = tensor([0, 0])]; + tensor x_311_dilations_0 = const()[name = string("x_311_dilations_0"), val = tensor([1])]; + int32 x_311_groups_0 = const()[name = string("x_311_groups_0"), val = int32(1)]; + tensor encoder_layers_11_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(239317120))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240365760))))[name = string("encoder_layers_11_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_311_cast_fp16 = conv(dilations = x_311_dilations_0, groups = x_311_groups_0, pad = x_311_pad_0, pad_type = x_311_pad_type_0, strides = x_311_strides_0, weight = encoder_layers_11_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_631_cast_fp16)[name = string("x_311_cast_fp16")]; + tensor input_633_perm_0 = const()[name = string("input_633_perm_0"), val = tensor([0, 2, 1])]; + tensor input_633_cast_fp16 = transpose(perm = input_633_perm_0, x = x_311_cast_fp16)[name = string("transpose_255")]; + tensor input_635_cast_fp16 = add(x = input_619_cast_fp16, y = input_633_cast_fp16)[name = string("input_635_cast_fp16")]; + tensor input_637_axes_0 = const()[name = string("input_637_axes_0"), val = tensor([-1])]; + tensor encoder_layers_11_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_11_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240367872)))]; + tensor encoder_layers_11_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_11_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240369984)))]; + tensor input_637_cast_fp16 = layer_norm(axes = input_637_axes_0, beta = encoder_layers_11_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_11_norm_feed_forward2_weight_to_fp16, x = input_635_cast_fp16)[name = string("input_637_cast_fp16")]; + tensor encoder_layers_11_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240372096))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(243517888))))[name = string("encoder_layers_11_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_11_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_11_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(243518080)))]; + tensor linear_107_cast_fp16 = linear(bias = encoder_layers_11_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_11_feed_forward2_linear1_weight_to_fp16_palettized, x = input_637_cast_fp16)[name = string("linear_107_cast_fp16")]; + tensor input_641_cast_fp16 = silu(x = linear_107_cast_fp16)[name = string("input_641_cast_fp16")]; + tensor encoder_layers_11_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(243526336))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(246672128))))[name = string("encoder_layers_11_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_11_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_11_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(246672320)))]; + tensor linear_108_cast_fp16 = linear(bias = encoder_layers_11_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_11_feed_forward2_linear2_weight_to_fp16_palettized, x = input_641_cast_fp16)[name = string("linear_108_cast_fp16")]; + fp16 var_2974_to_fp16 = const()[name = string("op_2974_to_fp16"), val = fp16(0x1p-1)]; + tensor var_2975_cast_fp16 = mul(x = linear_108_cast_fp16, y = var_2974_to_fp16)[name = string("op_2975_cast_fp16")]; + tensor input_647_cast_fp16 = add(x = input_635_cast_fp16, y = var_2975_cast_fp16)[name = string("input_647_cast_fp16")]; + tensor input_649_axes_0 = const()[name = string("input_649_axes_0"), val = tensor([-1])]; + tensor encoder_layers_11_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_11_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(246674432)))]; + tensor encoder_layers_11_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_11_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(246676544)))]; + tensor input_649_cast_fp16 = layer_norm(axes = input_649_axes_0, beta = encoder_layers_11_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_11_norm_out_weight_to_fp16, x = input_647_cast_fp16)[name = string("input_649_cast_fp16")]; + tensor cache_49_begin_0 = const()[name = string("cache_49_begin_0"), val = tensor([12, 0, 0, 0])]; + tensor cache_49_end_0 = const()[name = string("cache_49_end_0"), val = tensor([13, 1, 42, 1024])]; + tensor cache_49_end_mask_0 = const()[name = string("cache_49_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_49_squeeze_mask_0 = const()[name = string("cache_49_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_49_cast_fp16 = slice_by_index(begin = cache_49_begin_0, end = cache_49_end_0, end_mask = cache_49_end_mask_0, squeeze_mask = cache_49_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_49_cast_fp16")]; + tensor cache_51_begin_0 = const()[name = string("cache_51_begin_0"), val = tensor([12, 0, 0, 0])]; + tensor cache_51_end_0 = const()[name = string("cache_51_end_0"), val = tensor([13, 1, 1024, 8])]; + tensor cache_51_end_mask_0 = const()[name = string("cache_51_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_51_squeeze_mask_0 = const()[name = string("cache_51_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_51_cast_fp16 = slice_by_index(begin = cache_51_begin_0, end = cache_51_end_0, end_mask = cache_51_end_mask_0, squeeze_mask = cache_51_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_51_cast_fp16")]; + tensor input_651_axes_0 = const()[name = string("input_651_axes_0"), val = tensor([-1])]; + tensor encoder_layers_12_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_12_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(246678656)))]; + tensor encoder_layers_12_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_12_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(246680768)))]; + tensor input_651_cast_fp16 = layer_norm(axes = input_651_axes_0, beta = encoder_layers_12_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_12_norm_feed_forward1_weight_to_fp16, x = input_649_cast_fp16)[name = string("input_651_cast_fp16")]; + tensor encoder_layers_12_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(246682880))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(249828672))))[name = string("encoder_layers_12_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_12_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_12_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(249828864)))]; + tensor linear_109_cast_fp16 = linear(bias = encoder_layers_12_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_12_feed_forward1_linear1_weight_to_fp16_palettized, x = input_651_cast_fp16)[name = string("linear_109_cast_fp16")]; + tensor input_655_cast_fp16 = silu(x = linear_109_cast_fp16)[name = string("input_655_cast_fp16")]; + tensor encoder_layers_12_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(249837120))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(252982912))))[name = string("encoder_layers_12_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_12_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_12_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(252983104)))]; + tensor linear_110_cast_fp16 = linear(bias = encoder_layers_12_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_12_feed_forward1_linear2_weight_to_fp16_palettized, x = input_655_cast_fp16)[name = string("linear_110_cast_fp16")]; + fp16 var_3011_to_fp16 = const()[name = string("op_3011_to_fp16"), val = fp16(0x1p-1)]; + tensor var_3012_cast_fp16 = mul(x = linear_110_cast_fp16, y = var_3011_to_fp16)[name = string("op_3012_cast_fp16")]; + tensor input_661_cast_fp16 = add(x = input_649_cast_fp16, y = var_3012_cast_fp16)[name = string("input_661_cast_fp16")]; + tensor key_25_axes_0 = const()[name = string("key_25_axes_0"), val = tensor([-1])]; + tensor encoder_layers_12_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_12_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(252985216)))]; + tensor encoder_layers_12_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_12_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(252987328)))]; + tensor key_25_cast_fp16 = layer_norm(axes = key_25_axes_0, beta = encoder_layers_12_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_12_norm_self_att_weight_to_fp16, x = input_661_cast_fp16)[name = string("key_25_cast_fp16")]; + bool input_663_interleave_0 = const()[name = string("input_663_interleave_0"), val = bool(false)]; + tensor input_663_cast_fp16 = concat(axis = var_68, interleave = input_663_interleave_0, values = (cache_49_cast_fp16, key_25_cast_fp16))[name = string("input_663_cast_fp16")]; + tensor var_3034_begin_0 = const()[name = string("op_3034_begin_0"), val = tensor([0, 7, 0])]; + tensor var_3034_end_0 = const()[name = string("op_3034_end_0"), val = tensor([1, 42, 1024])]; + tensor var_3034_end_mask_0 = const()[name = string("op_3034_end_mask_0"), val = tensor([true, true, true])]; + tensor var_3034_cast_fp16 = slice_by_index(begin = var_3034_begin_0, end = var_3034_end_0, end_mask = var_3034_end_mask_0, x = cache_49_cast_fp16)[name = string("op_3034_cast_fp16")]; + bool var_3040_interleave_0 = const()[name = string("op_3040_interleave_0"), val = bool(false)]; + tensor var_3040_cast_fp16 = concat(axis = var_68, interleave = var_3040_interleave_0, values = (var_3034_cast_fp16, key_25_cast_fp16))[name = string("op_3040_cast_fp16")]; + tensor encoder_layers_12_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(252989440))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253775936))))[name = string("encoder_layers_12_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_12_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_12_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253776128)))]; + tensor linear_111_cast_fp16 = linear(bias = encoder_layers_12_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_12_self_attn_linear_q_weight_to_fp16_palettized, x = key_25_cast_fp16)[name = string("linear_111_cast_fp16")]; + tensor var_3045 = const()[name = string("op_3045"), val = tensor([1, -1, 8, 128])]; + tensor q_73_cast_fp16 = reshape(shape = var_3045, x = linear_111_cast_fp16)[name = string("q_73_cast_fp16")]; + tensor encoder_layers_12_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253778240))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(254564736))))[name = string("encoder_layers_12_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_12_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_12_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(254564928)))]; + tensor linear_112_cast_fp16 = linear(bias = encoder_layers_12_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_12_self_attn_linear_k_weight_to_fp16_palettized, x = input_663_cast_fp16)[name = string("linear_112_cast_fp16")]; + tensor var_3050 = const()[name = string("op_3050"), val = tensor([1, -1, 8, 128])]; + tensor k_49_cast_fp16 = reshape(shape = var_3050, x = linear_112_cast_fp16)[name = string("k_49_cast_fp16")]; + tensor encoder_layers_12_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(254567040))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255353536))))[name = string("encoder_layers_12_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_12_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_12_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255353728)))]; + tensor linear_113_cast_fp16 = linear(bias = encoder_layers_12_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_12_self_attn_linear_v_weight_to_fp16_palettized, x = input_663_cast_fp16)[name = string("linear_113_cast_fp16")]; + tensor var_3055 = const()[name = string("op_3055"), val = tensor([1, -1, 8, 128])]; + tensor v_25_cast_fp16 = reshape(shape = var_3055, x = linear_113_cast_fp16)[name = string("v_25_cast_fp16")]; + tensor value_33_perm_0 = const()[name = string("value_33_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_12_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_12_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255355840)))]; + tensor var_3068_cast_fp16 = add(x = q_73_cast_fp16, y = encoder_layers_12_self_attn_pos_bias_u_to_fp16)[name = string("op_3068_cast_fp16")]; + tensor encoder_layers_12_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_12_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255357952)))]; + tensor var_3070_cast_fp16 = add(x = q_73_cast_fp16, y = encoder_layers_12_self_attn_pos_bias_v_to_fp16)[name = string("op_3070_cast_fp16")]; + tensor q_with_bias_v_25_perm_0 = const()[name = string("q_with_bias_v_25_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_319_transpose_x_0 = const()[name = string("x_319_transpose_x_0"), val = bool(false)]; + bool x_319_transpose_y_0 = const()[name = string("x_319_transpose_y_0"), val = bool(false)]; + tensor op_3072_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255360064))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255459456))))[name = string("op_3072_to_fp16_quantized")]; + tensor q_with_bias_v_25_cast_fp16 = transpose(perm = q_with_bias_v_25_perm_0, x = var_3070_cast_fp16)[name = string("transpose_254")]; + tensor x_319_cast_fp16 = matmul(transpose_x = x_319_transpose_x_0, transpose_y = x_319_transpose_y_0, x = q_with_bias_v_25_cast_fp16, y = op_3072_to_fp16_quantized)[name = string("x_319_cast_fp16")]; + tensor x_321_pad_0 = const()[name = string("x_321_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_321_mode_0 = const()[name = string("x_321_mode_0"), val = string("constant")]; + fp16 const_235_to_fp16 = const()[name = string("const_235_to_fp16"), val = fp16(0x0p+0)]; + tensor x_321_cast_fp16 = pad(constant_val = const_235_to_fp16, mode = x_321_mode_0, pad = x_321_pad_0, x = x_319_cast_fp16)[name = string("x_321_cast_fp16")]; + tensor var_3080 = const()[name = string("op_3080"), val = tensor([1, 8, -1, 7])]; + tensor x_323_cast_fp16 = reshape(shape = var_3080, x = x_321_cast_fp16)[name = string("x_323_cast_fp16")]; + tensor var_3084_begin_0 = const()[name = string("op_3084_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_3084_end_0 = const()[name = string("op_3084_end_0"), val = tensor([1, 8, 98, 7])]; + tensor var_3084_end_mask_0 = const()[name = string("op_3084_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_3084_cast_fp16 = slice_by_index(begin = var_3084_begin_0, end = var_3084_end_0, end_mask = var_3084_end_mask_0, x = x_323_cast_fp16)[name = string("op_3084_cast_fp16")]; + tensor var_3085 = const()[name = string("op_3085"), val = tensor([1, 8, 7, 97])]; + tensor matrix_bd_49_cast_fp16 = reshape(shape = var_3085, x = var_3084_cast_fp16)[name = string("matrix_bd_49_cast_fp16")]; + bool matrix_ac_25_transpose_x_0 = const()[name = string("matrix_ac_25_transpose_x_0"), val = bool(false)]; + bool matrix_ac_25_transpose_y_0 = const()[name = string("matrix_ac_25_transpose_y_0"), val = bool(false)]; + tensor transpose_120_perm_0 = const()[name = string("transpose_120_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_121_perm_0 = const()[name = string("transpose_121_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_121 = transpose(perm = transpose_121_perm_0, x = k_49_cast_fp16)[name = string("transpose_252")]; + tensor transpose_120 = transpose(perm = transpose_120_perm_0, x = var_3068_cast_fp16)[name = string("transpose_253")]; + tensor matrix_ac_25_cast_fp16 = matmul(transpose_x = matrix_ac_25_transpose_x_0, transpose_y = matrix_ac_25_transpose_y_0, x = transpose_120, y = transpose_121)[name = string("matrix_ac_25_cast_fp16")]; + tensor matrix_bd_51_begin_0 = const()[name = string("matrix_bd_51_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_51_end_0 = const()[name = string("matrix_bd_51_end_0"), val = tensor([1, 8, 7, 49])]; + tensor matrix_bd_51_end_mask_0 = const()[name = string("matrix_bd_51_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_51_cast_fp16 = slice_by_index(begin = matrix_bd_51_begin_0, end = matrix_bd_51_end_0, end_mask = matrix_bd_51_end_mask_0, x = matrix_bd_49_cast_fp16)[name = string("matrix_bd_51_cast_fp16")]; + tensor var_3094_cast_fp16 = add(x = matrix_ac_25_cast_fp16, y = matrix_bd_51_cast_fp16)[name = string("op_3094_cast_fp16")]; + fp16 _inversed_scores_49_y_0_to_fp16 = const()[name = string("_inversed_scores_49_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_49_cast_fp16 = mul(x = var_3094_cast_fp16, y = _inversed_scores_49_y_0_to_fp16)[name = string("_inversed_scores_49_cast_fp16")]; + tensor scores_51_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_49_cast_fp16, cond = mask_11)[name = string("scores_51_cast_fp16")]; + tensor var_3100_cast_fp16 = softmax(axis = var_59, x = scores_51_cast_fp16)[name = string("op_3100_cast_fp16")]; + tensor input_665_cast_fp16 = select(a = var_44_to_fp16, b = var_3100_cast_fp16, cond = mask_11)[name = string("input_665_cast_fp16")]; + bool x_325_transpose_x_0 = const()[name = string("x_325_transpose_x_0"), val = bool(false)]; + bool x_325_transpose_y_0 = const()[name = string("x_325_transpose_y_0"), val = bool(false)]; + tensor value_33_cast_fp16 = transpose(perm = value_33_perm_0, x = v_25_cast_fp16)[name = string("transpose_251")]; + tensor x_325_cast_fp16 = matmul(transpose_x = x_325_transpose_x_0, transpose_y = x_325_transpose_y_0, x = input_665_cast_fp16, y = value_33_cast_fp16)[name = string("x_325_cast_fp16")]; + tensor var_3104_perm_0 = const()[name = string("op_3104_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_3105 = const()[name = string("op_3105"), val = tensor([1, -1, 1024])]; + tensor var_3104_cast_fp16 = transpose(perm = var_3104_perm_0, x = x_325_cast_fp16)[name = string("transpose_250")]; + tensor input_667_cast_fp16 = reshape(shape = var_3105, x = var_3104_cast_fp16)[name = string("input_667_cast_fp16")]; + tensor encoder_layers_12_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255459776))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256246272))))[name = string("encoder_layers_12_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_12_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_12_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256246464)))]; + tensor linear_115_cast_fp16 = linear(bias = encoder_layers_12_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_12_self_attn_linear_out_weight_to_fp16_palettized, x = input_667_cast_fp16)[name = string("linear_115_cast_fp16")]; + tensor input_671_cast_fp16 = add(x = input_661_cast_fp16, y = linear_115_cast_fp16)[name = string("input_671_cast_fp16")]; + tensor x_329_axes_0 = const()[name = string("x_329_axes_0"), val = tensor([-1])]; + tensor encoder_layers_12_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_12_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256248576)))]; + tensor encoder_layers_12_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_12_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256250688)))]; + tensor x_329_cast_fp16 = layer_norm(axes = x_329_axes_0, beta = encoder_layers_12_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_12_norm_conv_weight_to_fp16, x = input_671_cast_fp16)[name = string("x_329_cast_fp16")]; + tensor input_673_perm_0 = const()[name = string("input_673_perm_0"), val = tensor([0, 2, 1])]; + string input_675_pad_type_0 = const()[name = string("input_675_pad_type_0"), val = string("valid")]; + tensor input_675_strides_0 = const()[name = string("input_675_strides_0"), val = tensor([1])]; + tensor input_675_pad_0 = const()[name = string("input_675_pad_0"), val = tensor([0, 0])]; + tensor input_675_dilations_0 = const()[name = string("input_675_dilations_0"), val = tensor([1])]; + int32 input_675_groups_0 = const()[name = string("input_675_groups_0"), val = int32(1)]; + tensor encoder_layers_12_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256252800))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(258350016))))[name = string("encoder_layers_12_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_673_cast_fp16 = transpose(perm = input_673_perm_0, x = x_329_cast_fp16)[name = string("transpose_249")]; + tensor input_675_cast_fp16 = conv(dilations = input_675_dilations_0, groups = input_675_groups_0, pad = input_675_pad_0, pad_type = input_675_pad_type_0, strides = input_675_strides_0, weight = encoder_layers_12_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_673_cast_fp16)[name = string("input_675_cast_fp16")]; + int32 x_331_split_num_splits_0 = const()[name = string("x_331_split_num_splits_0"), val = int32(2)]; + int32 x_331_split_axis_0 = const()[name = string("x_331_split_axis_0"), val = int32(1)]; + tensor x_331_split_cast_fp16_0, tensor x_331_split_cast_fp16_1 = split(axis = x_331_split_axis_0, num_splits = x_331_split_num_splits_0, x = input_675_cast_fp16)[name = string("x_331_split_cast_fp16")]; + tensor x_331_split_1_sigmoid_cast_fp16 = sigmoid(x = x_331_split_cast_fp16_1)[name = string("x_331_split_1_sigmoid_cast_fp16")]; + tensor x_331_cast_fp16 = mul(x = x_331_split_cast_fp16_0, y = x_331_split_1_sigmoid_cast_fp16)[name = string("x_331_cast_fp16")]; + tensor input_677_cast_fp16 = select(a = var_44_to_fp16, b = x_331_cast_fp16, cond = var_575)[name = string("input_677_cast_fp16")]; + bool new_x_51_interleave_0 = const()[name = string("new_x_51_interleave_0"), val = bool(false)]; + tensor new_x_51_cast_fp16 = concat(axis = var_59, interleave = new_x_51_interleave_0, values = (cache_51_cast_fp16, input_677_cast_fp16))[name = string("new_x_51_cast_fp16")]; + tensor var_3144_begin_0 = const()[name = string("op_3144_begin_0"), val = tensor([0, 0, 7])]; + tensor var_3144_end_0 = const()[name = string("op_3144_end_0"), val = tensor([1, 1024, 15])]; + tensor var_3144_end_mask_0 = const()[name = string("op_3144_end_mask_0"), val = tensor([true, true, true])]; + tensor var_3144_cast_fp16 = slice_by_index(begin = var_3144_begin_0, end = var_3144_end_0, end_mask = var_3144_end_mask_0, x = new_x_51_cast_fp16)[name = string("op_3144_cast_fp16")]; + string x_333_pad_type_0 = const()[name = string("x_333_pad_type_0"), val = string("valid")]; + int32 x_333_groups_0 = const()[name = string("x_333_groups_0"), val = int32(1024)]; + tensor x_333_strides_0 = const()[name = string("x_333_strides_0"), val = tensor([1])]; + tensor x_333_pad_0 = const()[name = string("x_333_pad_0"), val = tensor([0, 0])]; + tensor x_333_dilations_0 = const()[name = string("x_333_dilations_0"), val = tensor([1])]; + tensor encoder_layers_12_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(258354176))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(258363456))))[name = string("encoder_layers_12_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_333_cast_fp16 = conv(dilations = x_333_dilations_0, groups = x_333_groups_0, pad = x_333_pad_0, pad_type = x_333_pad_type_0, strides = x_333_strides_0, weight = encoder_layers_12_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_51_cast_fp16)[name = string("x_333_cast_fp16")]; + tensor input_679_perm_0 = const()[name = string("input_679_perm_0"), val = tensor([0, 2, 1])]; + tensor x_335_axes_0 = const()[name = string("x_335_axes_0"), val = tensor([-1])]; + tensor encoder_layers_12_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_12_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(258365568)))]; + tensor encoder_layers_12_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_12_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(258367680)))]; + tensor input_679_cast_fp16 = transpose(perm = input_679_perm_0, x = x_333_cast_fp16)[name = string("transpose_248")]; + tensor x_335_cast_fp16 = layer_norm(axes = x_335_axes_0, beta = encoder_layers_12_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_12_conv_batch_norm_weight_to_fp16, x = input_679_cast_fp16)[name = string("x_335_cast_fp16")]; + tensor input_681_perm_0 = const()[name = string("input_681_perm_0"), val = tensor([0, 2, 1])]; + tensor input_681_cast_fp16 = transpose(perm = input_681_perm_0, x = x_335_cast_fp16)[name = string("transpose_247")]; + tensor input_683_cast_fp16 = silu(x = input_681_cast_fp16)[name = string("input_683_cast_fp16")]; + string x_337_pad_type_0 = const()[name = string("x_337_pad_type_0"), val = string("valid")]; + tensor x_337_strides_0 = const()[name = string("x_337_strides_0"), val = tensor([1])]; + tensor x_337_pad_0 = const()[name = string("x_337_pad_0"), val = tensor([0, 0])]; + tensor x_337_dilations_0 = const()[name = string("x_337_dilations_0"), val = tensor([1])]; + int32 x_337_groups_0 = const()[name = string("x_337_groups_0"), val = int32(1)]; + tensor encoder_layers_12_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(258369792))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259418432))))[name = string("encoder_layers_12_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_337_cast_fp16 = conv(dilations = x_337_dilations_0, groups = x_337_groups_0, pad = x_337_pad_0, pad_type = x_337_pad_type_0, strides = x_337_strides_0, weight = encoder_layers_12_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_683_cast_fp16)[name = string("x_337_cast_fp16")]; + tensor input_685_perm_0 = const()[name = string("input_685_perm_0"), val = tensor([0, 2, 1])]; + tensor input_685_cast_fp16 = transpose(perm = input_685_perm_0, x = x_337_cast_fp16)[name = string("transpose_246")]; + tensor input_687_cast_fp16 = add(x = input_671_cast_fp16, y = input_685_cast_fp16)[name = string("input_687_cast_fp16")]; + tensor input_689_axes_0 = const()[name = string("input_689_axes_0"), val = tensor([-1])]; + tensor encoder_layers_12_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_12_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259420544)))]; + tensor encoder_layers_12_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_12_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259422656)))]; + tensor input_689_cast_fp16 = layer_norm(axes = input_689_axes_0, beta = encoder_layers_12_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_12_norm_feed_forward2_weight_to_fp16, x = input_687_cast_fp16)[name = string("input_689_cast_fp16")]; + tensor encoder_layers_12_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259424768))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(262570560))))[name = string("encoder_layers_12_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_12_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_12_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(262570752)))]; + tensor linear_116_cast_fp16 = linear(bias = encoder_layers_12_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_12_feed_forward2_linear1_weight_to_fp16_palettized, x = input_689_cast_fp16)[name = string("linear_116_cast_fp16")]; + tensor input_693_cast_fp16 = silu(x = linear_116_cast_fp16)[name = string("input_693_cast_fp16")]; + tensor encoder_layers_12_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(262579008))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(265724800))))[name = string("encoder_layers_12_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_12_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_12_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(265724992)))]; + tensor linear_117_cast_fp16 = linear(bias = encoder_layers_12_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_12_feed_forward2_linear2_weight_to_fp16_palettized, x = input_693_cast_fp16)[name = string("linear_117_cast_fp16")]; + fp16 var_3187_to_fp16 = const()[name = string("op_3187_to_fp16"), val = fp16(0x1p-1)]; + tensor var_3188_cast_fp16 = mul(x = linear_117_cast_fp16, y = var_3187_to_fp16)[name = string("op_3188_cast_fp16")]; + tensor input_699_cast_fp16 = add(x = input_687_cast_fp16, y = var_3188_cast_fp16)[name = string("input_699_cast_fp16")]; + tensor input_701_axes_0 = const()[name = string("input_701_axes_0"), val = tensor([-1])]; + tensor encoder_layers_12_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_12_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(265727104)))]; + tensor encoder_layers_12_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_12_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(265729216)))]; + tensor input_701_cast_fp16 = layer_norm(axes = input_701_axes_0, beta = encoder_layers_12_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_12_norm_out_weight_to_fp16, x = input_699_cast_fp16)[name = string("input_701_cast_fp16")]; + tensor cache_53_begin_0 = const()[name = string("cache_53_begin_0"), val = tensor([13, 0, 0, 0])]; + tensor cache_53_end_0 = const()[name = string("cache_53_end_0"), val = tensor([14, 1, 42, 1024])]; + tensor cache_53_end_mask_0 = const()[name = string("cache_53_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_53_squeeze_mask_0 = const()[name = string("cache_53_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_53_cast_fp16 = slice_by_index(begin = cache_53_begin_0, end = cache_53_end_0, end_mask = cache_53_end_mask_0, squeeze_mask = cache_53_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_53_cast_fp16")]; + tensor cache_55_begin_0 = const()[name = string("cache_55_begin_0"), val = tensor([13, 0, 0, 0])]; + tensor cache_55_end_0 = const()[name = string("cache_55_end_0"), val = tensor([14, 1, 1024, 8])]; + tensor cache_55_end_mask_0 = const()[name = string("cache_55_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_55_squeeze_mask_0 = const()[name = string("cache_55_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_55_cast_fp16 = slice_by_index(begin = cache_55_begin_0, end = cache_55_end_0, end_mask = cache_55_end_mask_0, squeeze_mask = cache_55_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_55_cast_fp16")]; + tensor input_703_axes_0 = const()[name = string("input_703_axes_0"), val = tensor([-1])]; + tensor encoder_layers_13_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_13_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(265731328)))]; + tensor encoder_layers_13_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_13_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(265733440)))]; + tensor input_703_cast_fp16 = layer_norm(axes = input_703_axes_0, beta = encoder_layers_13_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_13_norm_feed_forward1_weight_to_fp16, x = input_701_cast_fp16)[name = string("input_703_cast_fp16")]; + tensor encoder_layers_13_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(265735552))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(268881344))))[name = string("encoder_layers_13_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_13_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_13_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(268881536)))]; + tensor linear_118_cast_fp16 = linear(bias = encoder_layers_13_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_13_feed_forward1_linear1_weight_to_fp16_palettized, x = input_703_cast_fp16)[name = string("linear_118_cast_fp16")]; + tensor input_707_cast_fp16 = silu(x = linear_118_cast_fp16)[name = string("input_707_cast_fp16")]; + tensor encoder_layers_13_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(268889792))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272035584))))[name = string("encoder_layers_13_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_13_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_13_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272035776)))]; + tensor linear_119_cast_fp16 = linear(bias = encoder_layers_13_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_13_feed_forward1_linear2_weight_to_fp16_palettized, x = input_707_cast_fp16)[name = string("linear_119_cast_fp16")]; + fp16 var_3224_to_fp16 = const()[name = string("op_3224_to_fp16"), val = fp16(0x1p-1)]; + tensor var_3225_cast_fp16 = mul(x = linear_119_cast_fp16, y = var_3224_to_fp16)[name = string("op_3225_cast_fp16")]; + tensor input_713_cast_fp16 = add(x = input_701_cast_fp16, y = var_3225_cast_fp16)[name = string("input_713_cast_fp16")]; + tensor key_27_axes_0 = const()[name = string("key_27_axes_0"), val = tensor([-1])]; + tensor encoder_layers_13_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_13_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272037888)))]; + tensor encoder_layers_13_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_13_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272040000)))]; + tensor key_27_cast_fp16 = layer_norm(axes = key_27_axes_0, beta = encoder_layers_13_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_13_norm_self_att_weight_to_fp16, x = input_713_cast_fp16)[name = string("key_27_cast_fp16")]; + bool input_715_interleave_0 = const()[name = string("input_715_interleave_0"), val = bool(false)]; + tensor input_715_cast_fp16 = concat(axis = var_68, interleave = input_715_interleave_0, values = (cache_53_cast_fp16, key_27_cast_fp16))[name = string("input_715_cast_fp16")]; + tensor var_3247_begin_0 = const()[name = string("op_3247_begin_0"), val = tensor([0, 7, 0])]; + tensor var_3247_end_0 = const()[name = string("op_3247_end_0"), val = tensor([1, 42, 1024])]; + tensor var_3247_end_mask_0 = const()[name = string("op_3247_end_mask_0"), val = tensor([true, true, true])]; + tensor var_3247_cast_fp16 = slice_by_index(begin = var_3247_begin_0, end = var_3247_end_0, end_mask = var_3247_end_mask_0, x = cache_53_cast_fp16)[name = string("op_3247_cast_fp16")]; + bool var_3253_interleave_0 = const()[name = string("op_3253_interleave_0"), val = bool(false)]; + tensor var_3253_cast_fp16 = concat(axis = var_68, interleave = var_3253_interleave_0, values = (var_3247_cast_fp16, key_27_cast_fp16))[name = string("op_3253_cast_fp16")]; + tensor encoder_layers_13_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272042112))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272828608))))[name = string("encoder_layers_13_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_13_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_13_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272828800)))]; + tensor linear_120_cast_fp16 = linear(bias = encoder_layers_13_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_13_self_attn_linear_q_weight_to_fp16_palettized, x = key_27_cast_fp16)[name = string("linear_120_cast_fp16")]; + tensor var_3258 = const()[name = string("op_3258"), val = tensor([1, -1, 8, 128])]; + tensor q_79_cast_fp16 = reshape(shape = var_3258, x = linear_120_cast_fp16)[name = string("q_79_cast_fp16")]; + tensor encoder_layers_13_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272830912))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(273617408))))[name = string("encoder_layers_13_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_13_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_13_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(273617600)))]; + tensor linear_121_cast_fp16 = linear(bias = encoder_layers_13_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_13_self_attn_linear_k_weight_to_fp16_palettized, x = input_715_cast_fp16)[name = string("linear_121_cast_fp16")]; + tensor var_3263 = const()[name = string("op_3263"), val = tensor([1, -1, 8, 128])]; + tensor k_53_cast_fp16 = reshape(shape = var_3263, x = linear_121_cast_fp16)[name = string("k_53_cast_fp16")]; + tensor encoder_layers_13_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(273619712))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274406208))))[name = string("encoder_layers_13_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_13_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_13_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274406400)))]; + tensor linear_122_cast_fp16 = linear(bias = encoder_layers_13_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_13_self_attn_linear_v_weight_to_fp16_palettized, x = input_715_cast_fp16)[name = string("linear_122_cast_fp16")]; + tensor var_3268 = const()[name = string("op_3268"), val = tensor([1, -1, 8, 128])]; + tensor v_27_cast_fp16 = reshape(shape = var_3268, x = linear_122_cast_fp16)[name = string("v_27_cast_fp16")]; + tensor value_35_perm_0 = const()[name = string("value_35_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_13_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_13_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274408512)))]; + tensor var_3281_cast_fp16 = add(x = q_79_cast_fp16, y = encoder_layers_13_self_attn_pos_bias_u_to_fp16)[name = string("op_3281_cast_fp16")]; + tensor encoder_layers_13_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_13_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274410624)))]; + tensor var_3283_cast_fp16 = add(x = q_79_cast_fp16, y = encoder_layers_13_self_attn_pos_bias_v_to_fp16)[name = string("op_3283_cast_fp16")]; + tensor q_with_bias_v_27_perm_0 = const()[name = string("q_with_bias_v_27_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_345_transpose_x_0 = const()[name = string("x_345_transpose_x_0"), val = bool(false)]; + bool x_345_transpose_y_0 = const()[name = string("x_345_transpose_y_0"), val = bool(false)]; + tensor op_3285_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274412736))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274512128))))[name = string("op_3285_to_fp16_quantized")]; + tensor q_with_bias_v_27_cast_fp16 = transpose(perm = q_with_bias_v_27_perm_0, x = var_3283_cast_fp16)[name = string("transpose_245")]; + tensor x_345_cast_fp16 = matmul(transpose_x = x_345_transpose_x_0, transpose_y = x_345_transpose_y_0, x = q_with_bias_v_27_cast_fp16, y = op_3285_to_fp16_quantized)[name = string("x_345_cast_fp16")]; + tensor x_347_pad_0 = const()[name = string("x_347_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_347_mode_0 = const()[name = string("x_347_mode_0"), val = string("constant")]; + fp16 const_248_to_fp16 = const()[name = string("const_248_to_fp16"), val = fp16(0x0p+0)]; + tensor x_347_cast_fp16 = pad(constant_val = const_248_to_fp16, mode = x_347_mode_0, pad = x_347_pad_0, x = x_345_cast_fp16)[name = string("x_347_cast_fp16")]; + tensor var_3293 = const()[name = string("op_3293"), val = tensor([1, 8, -1, 7])]; + tensor x_349_cast_fp16 = reshape(shape = var_3293, x = x_347_cast_fp16)[name = string("x_349_cast_fp16")]; + tensor var_3297_begin_0 = const()[name = string("op_3297_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_3297_end_0 = const()[name = string("op_3297_end_0"), val = tensor([1, 8, 98, 7])]; + tensor var_3297_end_mask_0 = const()[name = string("op_3297_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_3297_cast_fp16 = slice_by_index(begin = var_3297_begin_0, end = var_3297_end_0, end_mask = var_3297_end_mask_0, x = x_349_cast_fp16)[name = string("op_3297_cast_fp16")]; + tensor var_3298 = const()[name = string("op_3298"), val = tensor([1, 8, 7, 97])]; + tensor matrix_bd_53_cast_fp16 = reshape(shape = var_3298, x = var_3297_cast_fp16)[name = string("matrix_bd_53_cast_fp16")]; + bool matrix_ac_27_transpose_x_0 = const()[name = string("matrix_ac_27_transpose_x_0"), val = bool(false)]; + bool matrix_ac_27_transpose_y_0 = const()[name = string("matrix_ac_27_transpose_y_0"), val = bool(false)]; + tensor transpose_122_perm_0 = const()[name = string("transpose_122_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_123_perm_0 = const()[name = string("transpose_123_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_123 = transpose(perm = transpose_123_perm_0, x = k_53_cast_fp16)[name = string("transpose_243")]; + tensor transpose_122 = transpose(perm = transpose_122_perm_0, x = var_3281_cast_fp16)[name = string("transpose_244")]; + tensor matrix_ac_27_cast_fp16 = matmul(transpose_x = matrix_ac_27_transpose_x_0, transpose_y = matrix_ac_27_transpose_y_0, x = transpose_122, y = transpose_123)[name = string("matrix_ac_27_cast_fp16")]; + tensor matrix_bd_55_begin_0 = const()[name = string("matrix_bd_55_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_55_end_0 = const()[name = string("matrix_bd_55_end_0"), val = tensor([1, 8, 7, 49])]; + tensor matrix_bd_55_end_mask_0 = const()[name = string("matrix_bd_55_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_55_cast_fp16 = slice_by_index(begin = matrix_bd_55_begin_0, end = matrix_bd_55_end_0, end_mask = matrix_bd_55_end_mask_0, x = matrix_bd_53_cast_fp16)[name = string("matrix_bd_55_cast_fp16")]; + tensor var_3307_cast_fp16 = add(x = matrix_ac_27_cast_fp16, y = matrix_bd_55_cast_fp16)[name = string("op_3307_cast_fp16")]; + fp16 _inversed_scores_53_y_0_to_fp16 = const()[name = string("_inversed_scores_53_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_53_cast_fp16 = mul(x = var_3307_cast_fp16, y = _inversed_scores_53_y_0_to_fp16)[name = string("_inversed_scores_53_cast_fp16")]; + tensor scores_55_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_53_cast_fp16, cond = mask_11)[name = string("scores_55_cast_fp16")]; + tensor var_3313_cast_fp16 = softmax(axis = var_59, x = scores_55_cast_fp16)[name = string("op_3313_cast_fp16")]; + tensor input_717_cast_fp16 = select(a = var_44_to_fp16, b = var_3313_cast_fp16, cond = mask_11)[name = string("input_717_cast_fp16")]; + bool x_351_transpose_x_0 = const()[name = string("x_351_transpose_x_0"), val = bool(false)]; + bool x_351_transpose_y_0 = const()[name = string("x_351_transpose_y_0"), val = bool(false)]; + tensor value_35_cast_fp16 = transpose(perm = value_35_perm_0, x = v_27_cast_fp16)[name = string("transpose_242")]; + tensor x_351_cast_fp16 = matmul(transpose_x = x_351_transpose_x_0, transpose_y = x_351_transpose_y_0, x = input_717_cast_fp16, y = value_35_cast_fp16)[name = string("x_351_cast_fp16")]; + tensor var_3317_perm_0 = const()[name = string("op_3317_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_3318 = const()[name = string("op_3318"), val = tensor([1, -1, 1024])]; + tensor var_3317_cast_fp16 = transpose(perm = var_3317_perm_0, x = x_351_cast_fp16)[name = string("transpose_241")]; + tensor input_719_cast_fp16 = reshape(shape = var_3318, x = var_3317_cast_fp16)[name = string("input_719_cast_fp16")]; + tensor encoder_layers_13_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274512448))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275298944))))[name = string("encoder_layers_13_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_13_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_13_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275299136)))]; + tensor linear_124_cast_fp16 = linear(bias = encoder_layers_13_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_13_self_attn_linear_out_weight_to_fp16_palettized, x = input_719_cast_fp16)[name = string("linear_124_cast_fp16")]; + tensor input_723_cast_fp16 = add(x = input_713_cast_fp16, y = linear_124_cast_fp16)[name = string("input_723_cast_fp16")]; + tensor x_355_axes_0 = const()[name = string("x_355_axes_0"), val = tensor([-1])]; + tensor encoder_layers_13_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_13_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275301248)))]; + tensor encoder_layers_13_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_13_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275303360)))]; + tensor x_355_cast_fp16 = layer_norm(axes = x_355_axes_0, beta = encoder_layers_13_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_13_norm_conv_weight_to_fp16, x = input_723_cast_fp16)[name = string("x_355_cast_fp16")]; + tensor input_725_perm_0 = const()[name = string("input_725_perm_0"), val = tensor([0, 2, 1])]; + string input_727_pad_type_0 = const()[name = string("input_727_pad_type_0"), val = string("valid")]; + tensor input_727_strides_0 = const()[name = string("input_727_strides_0"), val = tensor([1])]; + tensor input_727_pad_0 = const()[name = string("input_727_pad_0"), val = tensor([0, 0])]; + tensor input_727_dilations_0 = const()[name = string("input_727_dilations_0"), val = tensor([1])]; + int32 input_727_groups_0 = const()[name = string("input_727_groups_0"), val = int32(1)]; + tensor encoder_layers_13_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275305472))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(277402688))))[name = string("encoder_layers_13_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_725_cast_fp16 = transpose(perm = input_725_perm_0, x = x_355_cast_fp16)[name = string("transpose_240")]; + tensor input_727_cast_fp16 = conv(dilations = input_727_dilations_0, groups = input_727_groups_0, pad = input_727_pad_0, pad_type = input_727_pad_type_0, strides = input_727_strides_0, weight = encoder_layers_13_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_725_cast_fp16)[name = string("input_727_cast_fp16")]; + int32 x_357_split_num_splits_0 = const()[name = string("x_357_split_num_splits_0"), val = int32(2)]; + int32 x_357_split_axis_0 = const()[name = string("x_357_split_axis_0"), val = int32(1)]; + tensor x_357_split_cast_fp16_0, tensor x_357_split_cast_fp16_1 = split(axis = x_357_split_axis_0, num_splits = x_357_split_num_splits_0, x = input_727_cast_fp16)[name = string("x_357_split_cast_fp16")]; + tensor x_357_split_1_sigmoid_cast_fp16 = sigmoid(x = x_357_split_cast_fp16_1)[name = string("x_357_split_1_sigmoid_cast_fp16")]; + tensor x_357_cast_fp16 = mul(x = x_357_split_cast_fp16_0, y = x_357_split_1_sigmoid_cast_fp16)[name = string("x_357_cast_fp16")]; + tensor input_729_cast_fp16 = select(a = var_44_to_fp16, b = x_357_cast_fp16, cond = var_575)[name = string("input_729_cast_fp16")]; + bool new_x_55_interleave_0 = const()[name = string("new_x_55_interleave_0"), val = bool(false)]; + tensor new_x_55_cast_fp16 = concat(axis = var_59, interleave = new_x_55_interleave_0, values = (cache_55_cast_fp16, input_729_cast_fp16))[name = string("new_x_55_cast_fp16")]; + tensor var_3357_begin_0 = const()[name = string("op_3357_begin_0"), val = tensor([0, 0, 7])]; + tensor var_3357_end_0 = const()[name = string("op_3357_end_0"), val = tensor([1, 1024, 15])]; + tensor var_3357_end_mask_0 = const()[name = string("op_3357_end_mask_0"), val = tensor([true, true, true])]; + tensor var_3357_cast_fp16 = slice_by_index(begin = var_3357_begin_0, end = var_3357_end_0, end_mask = var_3357_end_mask_0, x = new_x_55_cast_fp16)[name = string("op_3357_cast_fp16")]; + string x_359_pad_type_0 = const()[name = string("x_359_pad_type_0"), val = string("valid")]; + int32 x_359_groups_0 = const()[name = string("x_359_groups_0"), val = int32(1024)]; + tensor x_359_strides_0 = const()[name = string("x_359_strides_0"), val = tensor([1])]; + tensor x_359_pad_0 = const()[name = string("x_359_pad_0"), val = tensor([0, 0])]; + tensor x_359_dilations_0 = const()[name = string("x_359_dilations_0"), val = tensor([1])]; + tensor encoder_layers_13_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(277406848))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(277416128))))[name = string("encoder_layers_13_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_359_cast_fp16 = conv(dilations = x_359_dilations_0, groups = x_359_groups_0, pad = x_359_pad_0, pad_type = x_359_pad_type_0, strides = x_359_strides_0, weight = encoder_layers_13_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_55_cast_fp16)[name = string("x_359_cast_fp16")]; + tensor input_731_perm_0 = const()[name = string("input_731_perm_0"), val = tensor([0, 2, 1])]; + tensor x_361_axes_0 = const()[name = string("x_361_axes_0"), val = tensor([-1])]; + tensor encoder_layers_13_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_13_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(277418240)))]; + tensor encoder_layers_13_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_13_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(277420352)))]; + tensor input_731_cast_fp16 = transpose(perm = input_731_perm_0, x = x_359_cast_fp16)[name = string("transpose_239")]; + tensor x_361_cast_fp16 = layer_norm(axes = x_361_axes_0, beta = encoder_layers_13_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_13_conv_batch_norm_weight_to_fp16, x = input_731_cast_fp16)[name = string("x_361_cast_fp16")]; + tensor input_733_perm_0 = const()[name = string("input_733_perm_0"), val = tensor([0, 2, 1])]; + tensor input_733_cast_fp16 = transpose(perm = input_733_perm_0, x = x_361_cast_fp16)[name = string("transpose_238")]; + tensor input_735_cast_fp16 = silu(x = input_733_cast_fp16)[name = string("input_735_cast_fp16")]; + string x_363_pad_type_0 = const()[name = string("x_363_pad_type_0"), val = string("valid")]; + tensor x_363_strides_0 = const()[name = string("x_363_strides_0"), val = tensor([1])]; + tensor x_363_pad_0 = const()[name = string("x_363_pad_0"), val = tensor([0, 0])]; + tensor x_363_dilations_0 = const()[name = string("x_363_dilations_0"), val = tensor([1])]; + int32 x_363_groups_0 = const()[name = string("x_363_groups_0"), val = int32(1)]; + tensor encoder_layers_13_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(277422464))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278471104))))[name = string("encoder_layers_13_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_363_cast_fp16 = conv(dilations = x_363_dilations_0, groups = x_363_groups_0, pad = x_363_pad_0, pad_type = x_363_pad_type_0, strides = x_363_strides_0, weight = encoder_layers_13_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_735_cast_fp16)[name = string("x_363_cast_fp16")]; + tensor input_737_perm_0 = const()[name = string("input_737_perm_0"), val = tensor([0, 2, 1])]; + tensor input_737_cast_fp16 = transpose(perm = input_737_perm_0, x = x_363_cast_fp16)[name = string("transpose_237")]; + tensor input_739_cast_fp16 = add(x = input_723_cast_fp16, y = input_737_cast_fp16)[name = string("input_739_cast_fp16")]; + tensor input_741_axes_0 = const()[name = string("input_741_axes_0"), val = tensor([-1])]; + tensor encoder_layers_13_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_13_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278473216)))]; + tensor encoder_layers_13_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_13_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278475328)))]; + tensor input_741_cast_fp16 = layer_norm(axes = input_741_axes_0, beta = encoder_layers_13_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_13_norm_feed_forward2_weight_to_fp16, x = input_739_cast_fp16)[name = string("input_741_cast_fp16")]; + tensor encoder_layers_13_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278477440))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(281623232))))[name = string("encoder_layers_13_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_13_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_13_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(281623424)))]; + tensor linear_125_cast_fp16 = linear(bias = encoder_layers_13_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_13_feed_forward2_linear1_weight_to_fp16_palettized, x = input_741_cast_fp16)[name = string("linear_125_cast_fp16")]; + tensor input_745_cast_fp16 = silu(x = linear_125_cast_fp16)[name = string("input_745_cast_fp16")]; + tensor encoder_layers_13_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(281631680))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(284777472))))[name = string("encoder_layers_13_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_13_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_13_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(284777664)))]; + tensor linear_126_cast_fp16 = linear(bias = encoder_layers_13_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_13_feed_forward2_linear2_weight_to_fp16_palettized, x = input_745_cast_fp16)[name = string("linear_126_cast_fp16")]; + fp16 var_3400_to_fp16 = const()[name = string("op_3400_to_fp16"), val = fp16(0x1p-1)]; + tensor var_3401_cast_fp16 = mul(x = linear_126_cast_fp16, y = var_3400_to_fp16)[name = string("op_3401_cast_fp16")]; + tensor input_751_cast_fp16 = add(x = input_739_cast_fp16, y = var_3401_cast_fp16)[name = string("input_751_cast_fp16")]; + tensor input_753_axes_0 = const()[name = string("input_753_axes_0"), val = tensor([-1])]; + tensor encoder_layers_13_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_13_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(284779776)))]; + tensor encoder_layers_13_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_13_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(284781888)))]; + tensor input_753_cast_fp16 = layer_norm(axes = input_753_axes_0, beta = encoder_layers_13_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_13_norm_out_weight_to_fp16, x = input_751_cast_fp16)[name = string("input_753_cast_fp16")]; + tensor cache_57_begin_0 = const()[name = string("cache_57_begin_0"), val = tensor([14, 0, 0, 0])]; + tensor cache_57_end_0 = const()[name = string("cache_57_end_0"), val = tensor([15, 1, 42, 1024])]; + tensor cache_57_end_mask_0 = const()[name = string("cache_57_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_57_squeeze_mask_0 = const()[name = string("cache_57_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_57_cast_fp16 = slice_by_index(begin = cache_57_begin_0, end = cache_57_end_0, end_mask = cache_57_end_mask_0, squeeze_mask = cache_57_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_57_cast_fp16")]; + tensor cache_59_begin_0 = const()[name = string("cache_59_begin_0"), val = tensor([14, 0, 0, 0])]; + tensor cache_59_end_0 = const()[name = string("cache_59_end_0"), val = tensor([15, 1, 1024, 8])]; + tensor cache_59_end_mask_0 = const()[name = string("cache_59_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_59_squeeze_mask_0 = const()[name = string("cache_59_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_59_cast_fp16 = slice_by_index(begin = cache_59_begin_0, end = cache_59_end_0, end_mask = cache_59_end_mask_0, squeeze_mask = cache_59_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_59_cast_fp16")]; + tensor input_755_axes_0 = const()[name = string("input_755_axes_0"), val = tensor([-1])]; + tensor encoder_layers_14_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_14_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(284784000)))]; + tensor encoder_layers_14_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_14_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(284786112)))]; + tensor input_755_cast_fp16 = layer_norm(axes = input_755_axes_0, beta = encoder_layers_14_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_14_norm_feed_forward1_weight_to_fp16, x = input_753_cast_fp16)[name = string("input_755_cast_fp16")]; + tensor encoder_layers_14_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(284788224))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(287934016))))[name = string("encoder_layers_14_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_14_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_14_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(287934208)))]; + tensor linear_127_cast_fp16 = linear(bias = encoder_layers_14_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_14_feed_forward1_linear1_weight_to_fp16_palettized, x = input_755_cast_fp16)[name = string("linear_127_cast_fp16")]; + tensor input_759_cast_fp16 = silu(x = linear_127_cast_fp16)[name = string("input_759_cast_fp16")]; + tensor encoder_layers_14_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(287942464))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291088256))))[name = string("encoder_layers_14_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_14_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_14_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291088448)))]; + tensor linear_128_cast_fp16 = linear(bias = encoder_layers_14_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_14_feed_forward1_linear2_weight_to_fp16_palettized, x = input_759_cast_fp16)[name = string("linear_128_cast_fp16")]; + fp16 var_3437_to_fp16 = const()[name = string("op_3437_to_fp16"), val = fp16(0x1p-1)]; + tensor var_3438_cast_fp16 = mul(x = linear_128_cast_fp16, y = var_3437_to_fp16)[name = string("op_3438_cast_fp16")]; + tensor input_765_cast_fp16 = add(x = input_753_cast_fp16, y = var_3438_cast_fp16)[name = string("input_765_cast_fp16")]; + tensor key_29_axes_0 = const()[name = string("key_29_axes_0"), val = tensor([-1])]; + tensor encoder_layers_14_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_14_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291090560)))]; + tensor encoder_layers_14_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_14_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291092672)))]; + tensor key_29_cast_fp16 = layer_norm(axes = key_29_axes_0, beta = encoder_layers_14_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_14_norm_self_att_weight_to_fp16, x = input_765_cast_fp16)[name = string("key_29_cast_fp16")]; + bool input_767_interleave_0 = const()[name = string("input_767_interleave_0"), val = bool(false)]; + tensor input_767_cast_fp16 = concat(axis = var_68, interleave = input_767_interleave_0, values = (cache_57_cast_fp16, key_29_cast_fp16))[name = string("input_767_cast_fp16")]; + tensor var_3460_begin_0 = const()[name = string("op_3460_begin_0"), val = tensor([0, 7, 0])]; + tensor var_3460_end_0 = const()[name = string("op_3460_end_0"), val = tensor([1, 42, 1024])]; + tensor var_3460_end_mask_0 = const()[name = string("op_3460_end_mask_0"), val = tensor([true, true, true])]; + tensor var_3460_cast_fp16 = slice_by_index(begin = var_3460_begin_0, end = var_3460_end_0, end_mask = var_3460_end_mask_0, x = cache_57_cast_fp16)[name = string("op_3460_cast_fp16")]; + bool var_3466_interleave_0 = const()[name = string("op_3466_interleave_0"), val = bool(false)]; + tensor var_3466_cast_fp16 = concat(axis = var_68, interleave = var_3466_interleave_0, values = (var_3460_cast_fp16, key_29_cast_fp16))[name = string("op_3466_cast_fp16")]; + tensor encoder_layers_14_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291094784))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291881280))))[name = string("encoder_layers_14_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_14_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_14_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291881472)))]; + tensor linear_129_cast_fp16 = linear(bias = encoder_layers_14_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_14_self_attn_linear_q_weight_to_fp16_palettized, x = key_29_cast_fp16)[name = string("linear_129_cast_fp16")]; + tensor var_3471 = const()[name = string("op_3471"), val = tensor([1, -1, 8, 128])]; + tensor q_85_cast_fp16 = reshape(shape = var_3471, x = linear_129_cast_fp16)[name = string("q_85_cast_fp16")]; + tensor encoder_layers_14_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291883584))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(292670080))))[name = string("encoder_layers_14_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_14_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_14_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(292670272)))]; + tensor linear_130_cast_fp16 = linear(bias = encoder_layers_14_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_14_self_attn_linear_k_weight_to_fp16_palettized, x = input_767_cast_fp16)[name = string("linear_130_cast_fp16")]; + tensor var_3476 = const()[name = string("op_3476"), val = tensor([1, -1, 8, 128])]; + tensor k_57_cast_fp16 = reshape(shape = var_3476, x = linear_130_cast_fp16)[name = string("k_57_cast_fp16")]; + tensor encoder_layers_14_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(292672384))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293458880))))[name = string("encoder_layers_14_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_14_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_14_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293459072)))]; + tensor linear_131_cast_fp16 = linear(bias = encoder_layers_14_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_14_self_attn_linear_v_weight_to_fp16_palettized, x = input_767_cast_fp16)[name = string("linear_131_cast_fp16")]; + tensor var_3481 = const()[name = string("op_3481"), val = tensor([1, -1, 8, 128])]; + tensor v_29_cast_fp16 = reshape(shape = var_3481, x = linear_131_cast_fp16)[name = string("v_29_cast_fp16")]; + tensor value_37_perm_0 = const()[name = string("value_37_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_14_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_14_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293461184)))]; + tensor var_3494_cast_fp16 = add(x = q_85_cast_fp16, y = encoder_layers_14_self_attn_pos_bias_u_to_fp16)[name = string("op_3494_cast_fp16")]; + tensor encoder_layers_14_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_14_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293463296)))]; + tensor var_3496_cast_fp16 = add(x = q_85_cast_fp16, y = encoder_layers_14_self_attn_pos_bias_v_to_fp16)[name = string("op_3496_cast_fp16")]; + tensor q_with_bias_v_29_perm_0 = const()[name = string("q_with_bias_v_29_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_371_transpose_x_0 = const()[name = string("x_371_transpose_x_0"), val = bool(false)]; + bool x_371_transpose_y_0 = const()[name = string("x_371_transpose_y_0"), val = bool(false)]; + tensor op_3498_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293465408))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293564800))))[name = string("op_3498_to_fp16_quantized")]; + tensor q_with_bias_v_29_cast_fp16 = transpose(perm = q_with_bias_v_29_perm_0, x = var_3496_cast_fp16)[name = string("transpose_236")]; + tensor x_371_cast_fp16 = matmul(transpose_x = x_371_transpose_x_0, transpose_y = x_371_transpose_y_0, x = q_with_bias_v_29_cast_fp16, y = op_3498_to_fp16_quantized)[name = string("x_371_cast_fp16")]; + tensor x_373_pad_0 = const()[name = string("x_373_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_373_mode_0 = const()[name = string("x_373_mode_0"), val = string("constant")]; + fp16 const_261_to_fp16 = const()[name = string("const_261_to_fp16"), val = fp16(0x0p+0)]; + tensor x_373_cast_fp16 = pad(constant_val = const_261_to_fp16, mode = x_373_mode_0, pad = x_373_pad_0, x = x_371_cast_fp16)[name = string("x_373_cast_fp16")]; + tensor var_3506 = const()[name = string("op_3506"), val = tensor([1, 8, -1, 7])]; + tensor x_375_cast_fp16 = reshape(shape = var_3506, x = x_373_cast_fp16)[name = string("x_375_cast_fp16")]; + tensor var_3510_begin_0 = const()[name = string("op_3510_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_3510_end_0 = const()[name = string("op_3510_end_0"), val = tensor([1, 8, 98, 7])]; + tensor var_3510_end_mask_0 = const()[name = string("op_3510_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_3510_cast_fp16 = slice_by_index(begin = var_3510_begin_0, end = var_3510_end_0, end_mask = var_3510_end_mask_0, x = x_375_cast_fp16)[name = string("op_3510_cast_fp16")]; + tensor var_3511 = const()[name = string("op_3511"), val = tensor([1, 8, 7, 97])]; + tensor matrix_bd_57_cast_fp16 = reshape(shape = var_3511, x = var_3510_cast_fp16)[name = string("matrix_bd_57_cast_fp16")]; + bool matrix_ac_29_transpose_x_0 = const()[name = string("matrix_ac_29_transpose_x_0"), val = bool(false)]; + bool matrix_ac_29_transpose_y_0 = const()[name = string("matrix_ac_29_transpose_y_0"), val = bool(false)]; + tensor transpose_124_perm_0 = const()[name = string("transpose_124_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_125_perm_0 = const()[name = string("transpose_125_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_125 = transpose(perm = transpose_125_perm_0, x = k_57_cast_fp16)[name = string("transpose_234")]; + tensor transpose_124 = transpose(perm = transpose_124_perm_0, x = var_3494_cast_fp16)[name = string("transpose_235")]; + tensor matrix_ac_29_cast_fp16 = matmul(transpose_x = matrix_ac_29_transpose_x_0, transpose_y = matrix_ac_29_transpose_y_0, x = transpose_124, y = transpose_125)[name = string("matrix_ac_29_cast_fp16")]; + tensor matrix_bd_59_begin_0 = const()[name = string("matrix_bd_59_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_59_end_0 = const()[name = string("matrix_bd_59_end_0"), val = tensor([1, 8, 7, 49])]; + tensor matrix_bd_59_end_mask_0 = const()[name = string("matrix_bd_59_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_59_cast_fp16 = slice_by_index(begin = matrix_bd_59_begin_0, end = matrix_bd_59_end_0, end_mask = matrix_bd_59_end_mask_0, x = matrix_bd_57_cast_fp16)[name = string("matrix_bd_59_cast_fp16")]; + tensor var_3520_cast_fp16 = add(x = matrix_ac_29_cast_fp16, y = matrix_bd_59_cast_fp16)[name = string("op_3520_cast_fp16")]; + fp16 _inversed_scores_57_y_0_to_fp16 = const()[name = string("_inversed_scores_57_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_57_cast_fp16 = mul(x = var_3520_cast_fp16, y = _inversed_scores_57_y_0_to_fp16)[name = string("_inversed_scores_57_cast_fp16")]; + tensor scores_59_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_57_cast_fp16, cond = mask_11)[name = string("scores_59_cast_fp16")]; + tensor var_3526_cast_fp16 = softmax(axis = var_59, x = scores_59_cast_fp16)[name = string("op_3526_cast_fp16")]; + tensor input_769_cast_fp16 = select(a = var_44_to_fp16, b = var_3526_cast_fp16, cond = mask_11)[name = string("input_769_cast_fp16")]; + bool x_377_transpose_x_0 = const()[name = string("x_377_transpose_x_0"), val = bool(false)]; + bool x_377_transpose_y_0 = const()[name = string("x_377_transpose_y_0"), val = bool(false)]; + tensor value_37_cast_fp16 = transpose(perm = value_37_perm_0, x = v_29_cast_fp16)[name = string("transpose_233")]; + tensor x_377_cast_fp16 = matmul(transpose_x = x_377_transpose_x_0, transpose_y = x_377_transpose_y_0, x = input_769_cast_fp16, y = value_37_cast_fp16)[name = string("x_377_cast_fp16")]; + tensor var_3530_perm_0 = const()[name = string("op_3530_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_3531 = const()[name = string("op_3531"), val = tensor([1, -1, 1024])]; + tensor var_3530_cast_fp16 = transpose(perm = var_3530_perm_0, x = x_377_cast_fp16)[name = string("transpose_232")]; + tensor input_771_cast_fp16 = reshape(shape = var_3531, x = var_3530_cast_fp16)[name = string("input_771_cast_fp16")]; + tensor encoder_layers_14_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293565120))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294351616))))[name = string("encoder_layers_14_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_14_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_14_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294351808)))]; + tensor linear_133_cast_fp16 = linear(bias = encoder_layers_14_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_14_self_attn_linear_out_weight_to_fp16_palettized, x = input_771_cast_fp16)[name = string("linear_133_cast_fp16")]; + tensor input_775_cast_fp16 = add(x = input_765_cast_fp16, y = linear_133_cast_fp16)[name = string("input_775_cast_fp16")]; + tensor x_381_axes_0 = const()[name = string("x_381_axes_0"), val = tensor([-1])]; + tensor encoder_layers_14_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_14_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294353920)))]; + tensor encoder_layers_14_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_14_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294356032)))]; + tensor x_381_cast_fp16 = layer_norm(axes = x_381_axes_0, beta = encoder_layers_14_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_14_norm_conv_weight_to_fp16, x = input_775_cast_fp16)[name = string("x_381_cast_fp16")]; + tensor input_777_perm_0 = const()[name = string("input_777_perm_0"), val = tensor([0, 2, 1])]; + string input_779_pad_type_0 = const()[name = string("input_779_pad_type_0"), val = string("valid")]; + tensor input_779_strides_0 = const()[name = string("input_779_strides_0"), val = tensor([1])]; + tensor input_779_pad_0 = const()[name = string("input_779_pad_0"), val = tensor([0, 0])]; + tensor input_779_dilations_0 = const()[name = string("input_779_dilations_0"), val = tensor([1])]; + int32 input_779_groups_0 = const()[name = string("input_779_groups_0"), val = int32(1)]; + tensor encoder_layers_14_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294358144))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296455360))))[name = string("encoder_layers_14_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_777_cast_fp16 = transpose(perm = input_777_perm_0, x = x_381_cast_fp16)[name = string("transpose_231")]; + tensor input_779_cast_fp16 = conv(dilations = input_779_dilations_0, groups = input_779_groups_0, pad = input_779_pad_0, pad_type = input_779_pad_type_0, strides = input_779_strides_0, weight = encoder_layers_14_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_777_cast_fp16)[name = string("input_779_cast_fp16")]; + int32 x_383_split_num_splits_0 = const()[name = string("x_383_split_num_splits_0"), val = int32(2)]; + int32 x_383_split_axis_0 = const()[name = string("x_383_split_axis_0"), val = int32(1)]; + tensor x_383_split_cast_fp16_0, tensor x_383_split_cast_fp16_1 = split(axis = x_383_split_axis_0, num_splits = x_383_split_num_splits_0, x = input_779_cast_fp16)[name = string("x_383_split_cast_fp16")]; + tensor x_383_split_1_sigmoid_cast_fp16 = sigmoid(x = x_383_split_cast_fp16_1)[name = string("x_383_split_1_sigmoid_cast_fp16")]; + tensor x_383_cast_fp16 = mul(x = x_383_split_cast_fp16_0, y = x_383_split_1_sigmoid_cast_fp16)[name = string("x_383_cast_fp16")]; + tensor input_781_cast_fp16 = select(a = var_44_to_fp16, b = x_383_cast_fp16, cond = var_575)[name = string("input_781_cast_fp16")]; + bool new_x_59_interleave_0 = const()[name = string("new_x_59_interleave_0"), val = bool(false)]; + tensor new_x_59_cast_fp16 = concat(axis = var_59, interleave = new_x_59_interleave_0, values = (cache_59_cast_fp16, input_781_cast_fp16))[name = string("new_x_59_cast_fp16")]; + tensor var_3570_begin_0 = const()[name = string("op_3570_begin_0"), val = tensor([0, 0, 7])]; + tensor var_3570_end_0 = const()[name = string("op_3570_end_0"), val = tensor([1, 1024, 15])]; + tensor var_3570_end_mask_0 = const()[name = string("op_3570_end_mask_0"), val = tensor([true, true, true])]; + tensor var_3570_cast_fp16 = slice_by_index(begin = var_3570_begin_0, end = var_3570_end_0, end_mask = var_3570_end_mask_0, x = new_x_59_cast_fp16)[name = string("op_3570_cast_fp16")]; + string x_385_pad_type_0 = const()[name = string("x_385_pad_type_0"), val = string("valid")]; + int32 x_385_groups_0 = const()[name = string("x_385_groups_0"), val = int32(1024)]; + tensor x_385_strides_0 = const()[name = string("x_385_strides_0"), val = tensor([1])]; + tensor x_385_pad_0 = const()[name = string("x_385_pad_0"), val = tensor([0, 0])]; + tensor x_385_dilations_0 = const()[name = string("x_385_dilations_0"), val = tensor([1])]; + tensor encoder_layers_14_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296459520))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296468800))))[name = string("encoder_layers_14_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_385_cast_fp16 = conv(dilations = x_385_dilations_0, groups = x_385_groups_0, pad = x_385_pad_0, pad_type = x_385_pad_type_0, strides = x_385_strides_0, weight = encoder_layers_14_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_59_cast_fp16)[name = string("x_385_cast_fp16")]; + tensor input_783_perm_0 = const()[name = string("input_783_perm_0"), val = tensor([0, 2, 1])]; + tensor x_387_axes_0 = const()[name = string("x_387_axes_0"), val = tensor([-1])]; + tensor encoder_layers_14_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_14_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296470912)))]; + tensor encoder_layers_14_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_14_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296473024)))]; + tensor input_783_cast_fp16 = transpose(perm = input_783_perm_0, x = x_385_cast_fp16)[name = string("transpose_230")]; + tensor x_387_cast_fp16 = layer_norm(axes = x_387_axes_0, beta = encoder_layers_14_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_14_conv_batch_norm_weight_to_fp16, x = input_783_cast_fp16)[name = string("x_387_cast_fp16")]; + tensor input_785_perm_0 = const()[name = string("input_785_perm_0"), val = tensor([0, 2, 1])]; + tensor input_785_cast_fp16 = transpose(perm = input_785_perm_0, x = x_387_cast_fp16)[name = string("transpose_229")]; + tensor input_787_cast_fp16 = silu(x = input_785_cast_fp16)[name = string("input_787_cast_fp16")]; + string x_389_pad_type_0 = const()[name = string("x_389_pad_type_0"), val = string("valid")]; + tensor x_389_strides_0 = const()[name = string("x_389_strides_0"), val = tensor([1])]; + tensor x_389_pad_0 = const()[name = string("x_389_pad_0"), val = tensor([0, 0])]; + tensor x_389_dilations_0 = const()[name = string("x_389_dilations_0"), val = tensor([1])]; + int32 x_389_groups_0 = const()[name = string("x_389_groups_0"), val = int32(1)]; + tensor encoder_layers_14_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296475136))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297523776))))[name = string("encoder_layers_14_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_389_cast_fp16 = conv(dilations = x_389_dilations_0, groups = x_389_groups_0, pad = x_389_pad_0, pad_type = x_389_pad_type_0, strides = x_389_strides_0, weight = encoder_layers_14_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_787_cast_fp16)[name = string("x_389_cast_fp16")]; + tensor input_789_perm_0 = const()[name = string("input_789_perm_0"), val = tensor([0, 2, 1])]; + tensor input_789_cast_fp16 = transpose(perm = input_789_perm_0, x = x_389_cast_fp16)[name = string("transpose_228")]; + tensor input_791_cast_fp16 = add(x = input_775_cast_fp16, y = input_789_cast_fp16)[name = string("input_791_cast_fp16")]; + tensor input_793_axes_0 = const()[name = string("input_793_axes_0"), val = tensor([-1])]; + tensor encoder_layers_14_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_14_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297525888)))]; + tensor encoder_layers_14_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_14_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297528000)))]; + tensor input_793_cast_fp16 = layer_norm(axes = input_793_axes_0, beta = encoder_layers_14_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_14_norm_feed_forward2_weight_to_fp16, x = input_791_cast_fp16)[name = string("input_793_cast_fp16")]; + tensor encoder_layers_14_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297530112))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(300675904))))[name = string("encoder_layers_14_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_14_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_14_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(300676096)))]; + tensor linear_134_cast_fp16 = linear(bias = encoder_layers_14_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_14_feed_forward2_linear1_weight_to_fp16_palettized, x = input_793_cast_fp16)[name = string("linear_134_cast_fp16")]; + tensor input_797_cast_fp16 = silu(x = linear_134_cast_fp16)[name = string("input_797_cast_fp16")]; + tensor encoder_layers_14_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(300684352))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(303830144))))[name = string("encoder_layers_14_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_14_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_14_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(303830336)))]; + tensor linear_135_cast_fp16 = linear(bias = encoder_layers_14_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_14_feed_forward2_linear2_weight_to_fp16_palettized, x = input_797_cast_fp16)[name = string("linear_135_cast_fp16")]; + fp16 var_3613_to_fp16 = const()[name = string("op_3613_to_fp16"), val = fp16(0x1p-1)]; + tensor var_3614_cast_fp16 = mul(x = linear_135_cast_fp16, y = var_3613_to_fp16)[name = string("op_3614_cast_fp16")]; + tensor input_803_cast_fp16 = add(x = input_791_cast_fp16, y = var_3614_cast_fp16)[name = string("input_803_cast_fp16")]; + tensor input_805_axes_0 = const()[name = string("input_805_axes_0"), val = tensor([-1])]; + tensor encoder_layers_14_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_14_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(303832448)))]; + tensor encoder_layers_14_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_14_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(303834560)))]; + tensor input_805_cast_fp16 = layer_norm(axes = input_805_axes_0, beta = encoder_layers_14_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_14_norm_out_weight_to_fp16, x = input_803_cast_fp16)[name = string("input_805_cast_fp16")]; + tensor cache_61_begin_0 = const()[name = string("cache_61_begin_0"), val = tensor([15, 0, 0, 0])]; + tensor cache_61_end_0 = const()[name = string("cache_61_end_0"), val = tensor([16, 1, 42, 1024])]; + tensor cache_61_end_mask_0 = const()[name = string("cache_61_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_61_squeeze_mask_0 = const()[name = string("cache_61_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_61_cast_fp16 = slice_by_index(begin = cache_61_begin_0, end = cache_61_end_0, end_mask = cache_61_end_mask_0, squeeze_mask = cache_61_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_61_cast_fp16")]; + tensor cache_63_begin_0 = const()[name = string("cache_63_begin_0"), val = tensor([15, 0, 0, 0])]; + tensor cache_63_end_0 = const()[name = string("cache_63_end_0"), val = tensor([16, 1, 1024, 8])]; + tensor cache_63_end_mask_0 = const()[name = string("cache_63_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_63_squeeze_mask_0 = const()[name = string("cache_63_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_63_cast_fp16 = slice_by_index(begin = cache_63_begin_0, end = cache_63_end_0, end_mask = cache_63_end_mask_0, squeeze_mask = cache_63_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_63_cast_fp16")]; + tensor input_807_axes_0 = const()[name = string("input_807_axes_0"), val = tensor([-1])]; + tensor encoder_layers_15_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_15_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(303836672)))]; + tensor encoder_layers_15_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_15_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(303838784)))]; + tensor input_807_cast_fp16 = layer_norm(axes = input_807_axes_0, beta = encoder_layers_15_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_15_norm_feed_forward1_weight_to_fp16, x = input_805_cast_fp16)[name = string("input_807_cast_fp16")]; + tensor encoder_layers_15_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(303840896))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(306986688))))[name = string("encoder_layers_15_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_15_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_15_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(306986880)))]; + tensor linear_136_cast_fp16 = linear(bias = encoder_layers_15_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_15_feed_forward1_linear1_weight_to_fp16_palettized, x = input_807_cast_fp16)[name = string("linear_136_cast_fp16")]; + tensor input_811_cast_fp16 = silu(x = linear_136_cast_fp16)[name = string("input_811_cast_fp16")]; + tensor encoder_layers_15_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(306995136))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310140928))))[name = string("encoder_layers_15_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_15_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_15_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310141120)))]; + tensor linear_137_cast_fp16 = linear(bias = encoder_layers_15_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_15_feed_forward1_linear2_weight_to_fp16_palettized, x = input_811_cast_fp16)[name = string("linear_137_cast_fp16")]; + fp16 var_3650_to_fp16 = const()[name = string("op_3650_to_fp16"), val = fp16(0x1p-1)]; + tensor var_3651_cast_fp16 = mul(x = linear_137_cast_fp16, y = var_3650_to_fp16)[name = string("op_3651_cast_fp16")]; + tensor input_817_cast_fp16 = add(x = input_805_cast_fp16, y = var_3651_cast_fp16)[name = string("input_817_cast_fp16")]; + tensor key_31_axes_0 = const()[name = string("key_31_axes_0"), val = tensor([-1])]; + tensor encoder_layers_15_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_15_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310143232)))]; + tensor encoder_layers_15_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_15_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310145344)))]; + tensor key_31_cast_fp16 = layer_norm(axes = key_31_axes_0, beta = encoder_layers_15_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_15_norm_self_att_weight_to_fp16, x = input_817_cast_fp16)[name = string("key_31_cast_fp16")]; + bool input_819_interleave_0 = const()[name = string("input_819_interleave_0"), val = bool(false)]; + tensor input_819_cast_fp16 = concat(axis = var_68, interleave = input_819_interleave_0, values = (cache_61_cast_fp16, key_31_cast_fp16))[name = string("input_819_cast_fp16")]; + tensor var_3673_begin_0 = const()[name = string("op_3673_begin_0"), val = tensor([0, 7, 0])]; + tensor var_3673_end_0 = const()[name = string("op_3673_end_0"), val = tensor([1, 42, 1024])]; + tensor var_3673_end_mask_0 = const()[name = string("op_3673_end_mask_0"), val = tensor([true, true, true])]; + tensor var_3673_cast_fp16 = slice_by_index(begin = var_3673_begin_0, end = var_3673_end_0, end_mask = var_3673_end_mask_0, x = cache_61_cast_fp16)[name = string("op_3673_cast_fp16")]; + bool var_3679_interleave_0 = const()[name = string("op_3679_interleave_0"), val = bool(false)]; + tensor var_3679_cast_fp16 = concat(axis = var_68, interleave = var_3679_interleave_0, values = (var_3673_cast_fp16, key_31_cast_fp16))[name = string("op_3679_cast_fp16")]; + tensor encoder_layers_15_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310147456))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310933952))))[name = string("encoder_layers_15_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_15_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_15_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310934144)))]; + tensor linear_138_cast_fp16 = linear(bias = encoder_layers_15_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_15_self_attn_linear_q_weight_to_fp16_palettized, x = key_31_cast_fp16)[name = string("linear_138_cast_fp16")]; + tensor var_3684 = const()[name = string("op_3684"), val = tensor([1, -1, 8, 128])]; + tensor q_91_cast_fp16 = reshape(shape = var_3684, x = linear_138_cast_fp16)[name = string("q_91_cast_fp16")]; + tensor encoder_layers_15_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310936256))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311722752))))[name = string("encoder_layers_15_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_15_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_15_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311722944)))]; + tensor linear_139_cast_fp16 = linear(bias = encoder_layers_15_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_15_self_attn_linear_k_weight_to_fp16_palettized, x = input_819_cast_fp16)[name = string("linear_139_cast_fp16")]; + tensor var_3689 = const()[name = string("op_3689"), val = tensor([1, -1, 8, 128])]; + tensor k_61_cast_fp16 = reshape(shape = var_3689, x = linear_139_cast_fp16)[name = string("k_61_cast_fp16")]; + tensor encoder_layers_15_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311725056))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(312511552))))[name = string("encoder_layers_15_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_15_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_15_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(312511744)))]; + tensor linear_140_cast_fp16 = linear(bias = encoder_layers_15_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_15_self_attn_linear_v_weight_to_fp16_palettized, x = input_819_cast_fp16)[name = string("linear_140_cast_fp16")]; + tensor var_3694 = const()[name = string("op_3694"), val = tensor([1, -1, 8, 128])]; + tensor v_31_cast_fp16 = reshape(shape = var_3694, x = linear_140_cast_fp16)[name = string("v_31_cast_fp16")]; + tensor value_39_perm_0 = const()[name = string("value_39_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_15_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_15_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(312513856)))]; + tensor var_3707_cast_fp16 = add(x = q_91_cast_fp16, y = encoder_layers_15_self_attn_pos_bias_u_to_fp16)[name = string("op_3707_cast_fp16")]; + tensor encoder_layers_15_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_15_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(312515968)))]; + tensor var_3709_cast_fp16 = add(x = q_91_cast_fp16, y = encoder_layers_15_self_attn_pos_bias_v_to_fp16)[name = string("op_3709_cast_fp16")]; + tensor q_with_bias_v_31_perm_0 = const()[name = string("q_with_bias_v_31_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_397_transpose_x_0 = const()[name = string("x_397_transpose_x_0"), val = bool(false)]; + bool x_397_transpose_y_0 = const()[name = string("x_397_transpose_y_0"), val = bool(false)]; + tensor op_3711_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(312518080))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(312617472))))[name = string("op_3711_to_fp16_quantized")]; + tensor q_with_bias_v_31_cast_fp16 = transpose(perm = q_with_bias_v_31_perm_0, x = var_3709_cast_fp16)[name = string("transpose_227")]; + tensor x_397_cast_fp16 = matmul(transpose_x = x_397_transpose_x_0, transpose_y = x_397_transpose_y_0, x = q_with_bias_v_31_cast_fp16, y = op_3711_to_fp16_quantized)[name = string("x_397_cast_fp16")]; + tensor x_399_pad_0 = const()[name = string("x_399_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_399_mode_0 = const()[name = string("x_399_mode_0"), val = string("constant")]; + fp16 const_274_to_fp16 = const()[name = string("const_274_to_fp16"), val = fp16(0x0p+0)]; + tensor x_399_cast_fp16 = pad(constant_val = const_274_to_fp16, mode = x_399_mode_0, pad = x_399_pad_0, x = x_397_cast_fp16)[name = string("x_399_cast_fp16")]; + tensor var_3719 = const()[name = string("op_3719"), val = tensor([1, 8, -1, 7])]; + tensor x_401_cast_fp16 = reshape(shape = var_3719, x = x_399_cast_fp16)[name = string("x_401_cast_fp16")]; + tensor var_3723_begin_0 = const()[name = string("op_3723_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_3723_end_0 = const()[name = string("op_3723_end_0"), val = tensor([1, 8, 98, 7])]; + tensor var_3723_end_mask_0 = const()[name = string("op_3723_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_3723_cast_fp16 = slice_by_index(begin = var_3723_begin_0, end = var_3723_end_0, end_mask = var_3723_end_mask_0, x = x_401_cast_fp16)[name = string("op_3723_cast_fp16")]; + tensor var_3724 = const()[name = string("op_3724"), val = tensor([1, 8, 7, 97])]; + tensor matrix_bd_61_cast_fp16 = reshape(shape = var_3724, x = var_3723_cast_fp16)[name = string("matrix_bd_61_cast_fp16")]; + bool matrix_ac_31_transpose_x_0 = const()[name = string("matrix_ac_31_transpose_x_0"), val = bool(false)]; + bool matrix_ac_31_transpose_y_0 = const()[name = string("matrix_ac_31_transpose_y_0"), val = bool(false)]; + tensor transpose_126_perm_0 = const()[name = string("transpose_126_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_127_perm_0 = const()[name = string("transpose_127_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_127 = transpose(perm = transpose_127_perm_0, x = k_61_cast_fp16)[name = string("transpose_225")]; + tensor transpose_126 = transpose(perm = transpose_126_perm_0, x = var_3707_cast_fp16)[name = string("transpose_226")]; + tensor matrix_ac_31_cast_fp16 = matmul(transpose_x = matrix_ac_31_transpose_x_0, transpose_y = matrix_ac_31_transpose_y_0, x = transpose_126, y = transpose_127)[name = string("matrix_ac_31_cast_fp16")]; + tensor matrix_bd_63_begin_0 = const()[name = string("matrix_bd_63_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_63_end_0 = const()[name = string("matrix_bd_63_end_0"), val = tensor([1, 8, 7, 49])]; + tensor matrix_bd_63_end_mask_0 = const()[name = string("matrix_bd_63_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_63_cast_fp16 = slice_by_index(begin = matrix_bd_63_begin_0, end = matrix_bd_63_end_0, end_mask = matrix_bd_63_end_mask_0, x = matrix_bd_61_cast_fp16)[name = string("matrix_bd_63_cast_fp16")]; + tensor var_3733_cast_fp16 = add(x = matrix_ac_31_cast_fp16, y = matrix_bd_63_cast_fp16)[name = string("op_3733_cast_fp16")]; + fp16 _inversed_scores_61_y_0_to_fp16 = const()[name = string("_inversed_scores_61_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_61_cast_fp16 = mul(x = var_3733_cast_fp16, y = _inversed_scores_61_y_0_to_fp16)[name = string("_inversed_scores_61_cast_fp16")]; + tensor scores_63_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_61_cast_fp16, cond = mask_11)[name = string("scores_63_cast_fp16")]; + tensor var_3739_cast_fp16 = softmax(axis = var_59, x = scores_63_cast_fp16)[name = string("op_3739_cast_fp16")]; + tensor input_821_cast_fp16 = select(a = var_44_to_fp16, b = var_3739_cast_fp16, cond = mask_11)[name = string("input_821_cast_fp16")]; + bool x_403_transpose_x_0 = const()[name = string("x_403_transpose_x_0"), val = bool(false)]; + bool x_403_transpose_y_0 = const()[name = string("x_403_transpose_y_0"), val = bool(false)]; + tensor value_39_cast_fp16 = transpose(perm = value_39_perm_0, x = v_31_cast_fp16)[name = string("transpose_224")]; + tensor x_403_cast_fp16 = matmul(transpose_x = x_403_transpose_x_0, transpose_y = x_403_transpose_y_0, x = input_821_cast_fp16, y = value_39_cast_fp16)[name = string("x_403_cast_fp16")]; + tensor var_3743_perm_0 = const()[name = string("op_3743_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_3744 = const()[name = string("op_3744"), val = tensor([1, -1, 1024])]; + tensor var_3743_cast_fp16 = transpose(perm = var_3743_perm_0, x = x_403_cast_fp16)[name = string("transpose_223")]; + tensor input_823_cast_fp16 = reshape(shape = var_3744, x = var_3743_cast_fp16)[name = string("input_823_cast_fp16")]; + tensor encoder_layers_15_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(312617792))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(313404288))))[name = string("encoder_layers_15_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_15_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_15_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(313404480)))]; + tensor linear_142_cast_fp16 = linear(bias = encoder_layers_15_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_15_self_attn_linear_out_weight_to_fp16_palettized, x = input_823_cast_fp16)[name = string("linear_142_cast_fp16")]; + tensor input_827_cast_fp16 = add(x = input_817_cast_fp16, y = linear_142_cast_fp16)[name = string("input_827_cast_fp16")]; + tensor x_407_axes_0 = const()[name = string("x_407_axes_0"), val = tensor([-1])]; + tensor encoder_layers_15_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_15_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(313406592)))]; + tensor encoder_layers_15_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_15_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(313408704)))]; + tensor x_407_cast_fp16 = layer_norm(axes = x_407_axes_0, beta = encoder_layers_15_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_15_norm_conv_weight_to_fp16, x = input_827_cast_fp16)[name = string("x_407_cast_fp16")]; + tensor input_829_perm_0 = const()[name = string("input_829_perm_0"), val = tensor([0, 2, 1])]; + string input_831_pad_type_0 = const()[name = string("input_831_pad_type_0"), val = string("valid")]; + tensor input_831_strides_0 = const()[name = string("input_831_strides_0"), val = tensor([1])]; + tensor input_831_pad_0 = const()[name = string("input_831_pad_0"), val = tensor([0, 0])]; + tensor input_831_dilations_0 = const()[name = string("input_831_dilations_0"), val = tensor([1])]; + int32 input_831_groups_0 = const()[name = string("input_831_groups_0"), val = int32(1)]; + tensor encoder_layers_15_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(313410816))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315508032))))[name = string("encoder_layers_15_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_829_cast_fp16 = transpose(perm = input_829_perm_0, x = x_407_cast_fp16)[name = string("transpose_222")]; + tensor input_831_cast_fp16 = conv(dilations = input_831_dilations_0, groups = input_831_groups_0, pad = input_831_pad_0, pad_type = input_831_pad_type_0, strides = input_831_strides_0, weight = encoder_layers_15_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_829_cast_fp16)[name = string("input_831_cast_fp16")]; + int32 x_409_split_num_splits_0 = const()[name = string("x_409_split_num_splits_0"), val = int32(2)]; + int32 x_409_split_axis_0 = const()[name = string("x_409_split_axis_0"), val = int32(1)]; + tensor x_409_split_cast_fp16_0, tensor x_409_split_cast_fp16_1 = split(axis = x_409_split_axis_0, num_splits = x_409_split_num_splits_0, x = input_831_cast_fp16)[name = string("x_409_split_cast_fp16")]; + tensor x_409_split_1_sigmoid_cast_fp16 = sigmoid(x = x_409_split_cast_fp16_1)[name = string("x_409_split_1_sigmoid_cast_fp16")]; + tensor x_409_cast_fp16 = mul(x = x_409_split_cast_fp16_0, y = x_409_split_1_sigmoid_cast_fp16)[name = string("x_409_cast_fp16")]; + tensor input_833_cast_fp16 = select(a = var_44_to_fp16, b = x_409_cast_fp16, cond = var_575)[name = string("input_833_cast_fp16")]; + bool new_x_63_interleave_0 = const()[name = string("new_x_63_interleave_0"), val = bool(false)]; + tensor new_x_63_cast_fp16 = concat(axis = var_59, interleave = new_x_63_interleave_0, values = (cache_63_cast_fp16, input_833_cast_fp16))[name = string("new_x_63_cast_fp16")]; + tensor var_3783_begin_0 = const()[name = string("op_3783_begin_0"), val = tensor([0, 0, 7])]; + tensor var_3783_end_0 = const()[name = string("op_3783_end_0"), val = tensor([1, 1024, 15])]; + tensor var_3783_end_mask_0 = const()[name = string("op_3783_end_mask_0"), val = tensor([true, true, true])]; + tensor var_3783_cast_fp16 = slice_by_index(begin = var_3783_begin_0, end = var_3783_end_0, end_mask = var_3783_end_mask_0, x = new_x_63_cast_fp16)[name = string("op_3783_cast_fp16")]; + string x_411_pad_type_0 = const()[name = string("x_411_pad_type_0"), val = string("valid")]; + int32 x_411_groups_0 = const()[name = string("x_411_groups_0"), val = int32(1024)]; + tensor x_411_strides_0 = const()[name = string("x_411_strides_0"), val = tensor([1])]; + tensor x_411_pad_0 = const()[name = string("x_411_pad_0"), val = tensor([0, 0])]; + tensor x_411_dilations_0 = const()[name = string("x_411_dilations_0"), val = tensor([1])]; + tensor encoder_layers_15_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315512192))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315521472))))[name = string("encoder_layers_15_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_411_cast_fp16 = conv(dilations = x_411_dilations_0, groups = x_411_groups_0, pad = x_411_pad_0, pad_type = x_411_pad_type_0, strides = x_411_strides_0, weight = encoder_layers_15_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_63_cast_fp16)[name = string("x_411_cast_fp16")]; + tensor input_835_perm_0 = const()[name = string("input_835_perm_0"), val = tensor([0, 2, 1])]; + tensor x_413_axes_0 = const()[name = string("x_413_axes_0"), val = tensor([-1])]; + tensor encoder_layers_15_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_15_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315523584)))]; + tensor encoder_layers_15_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_15_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315525696)))]; + tensor input_835_cast_fp16 = transpose(perm = input_835_perm_0, x = x_411_cast_fp16)[name = string("transpose_221")]; + tensor x_413_cast_fp16 = layer_norm(axes = x_413_axes_0, beta = encoder_layers_15_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_15_conv_batch_norm_weight_to_fp16, x = input_835_cast_fp16)[name = string("x_413_cast_fp16")]; + tensor input_837_perm_0 = const()[name = string("input_837_perm_0"), val = tensor([0, 2, 1])]; + tensor input_837_cast_fp16 = transpose(perm = input_837_perm_0, x = x_413_cast_fp16)[name = string("transpose_220")]; + tensor input_839_cast_fp16 = silu(x = input_837_cast_fp16)[name = string("input_839_cast_fp16")]; + string x_415_pad_type_0 = const()[name = string("x_415_pad_type_0"), val = string("valid")]; + tensor x_415_strides_0 = const()[name = string("x_415_strides_0"), val = tensor([1])]; + tensor x_415_pad_0 = const()[name = string("x_415_pad_0"), val = tensor([0, 0])]; + tensor x_415_dilations_0 = const()[name = string("x_415_dilations_0"), val = tensor([1])]; + int32 x_415_groups_0 = const()[name = string("x_415_groups_0"), val = int32(1)]; + tensor encoder_layers_15_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315527808))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316576448))))[name = string("encoder_layers_15_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_415_cast_fp16 = conv(dilations = x_415_dilations_0, groups = x_415_groups_0, pad = x_415_pad_0, pad_type = x_415_pad_type_0, strides = x_415_strides_0, weight = encoder_layers_15_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_839_cast_fp16)[name = string("x_415_cast_fp16")]; + tensor input_841_perm_0 = const()[name = string("input_841_perm_0"), val = tensor([0, 2, 1])]; + tensor input_841_cast_fp16 = transpose(perm = input_841_perm_0, x = x_415_cast_fp16)[name = string("transpose_219")]; + tensor input_843_cast_fp16 = add(x = input_827_cast_fp16, y = input_841_cast_fp16)[name = string("input_843_cast_fp16")]; + tensor input_845_axes_0 = const()[name = string("input_845_axes_0"), val = tensor([-1])]; + tensor encoder_layers_15_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_15_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316578560)))]; + tensor encoder_layers_15_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_15_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316580672)))]; + tensor input_845_cast_fp16 = layer_norm(axes = input_845_axes_0, beta = encoder_layers_15_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_15_norm_feed_forward2_weight_to_fp16, x = input_843_cast_fp16)[name = string("input_845_cast_fp16")]; + tensor encoder_layers_15_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316582784))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(319728576))))[name = string("encoder_layers_15_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_15_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_15_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(319728768)))]; + tensor linear_143_cast_fp16 = linear(bias = encoder_layers_15_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_15_feed_forward2_linear1_weight_to_fp16_palettized, x = input_845_cast_fp16)[name = string("linear_143_cast_fp16")]; + tensor input_849_cast_fp16 = silu(x = linear_143_cast_fp16)[name = string("input_849_cast_fp16")]; + tensor encoder_layers_15_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(319737024))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(322882816))))[name = string("encoder_layers_15_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_15_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_15_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(322883008)))]; + tensor linear_144_cast_fp16 = linear(bias = encoder_layers_15_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_15_feed_forward2_linear2_weight_to_fp16_palettized, x = input_849_cast_fp16)[name = string("linear_144_cast_fp16")]; + fp16 var_3826_to_fp16 = const()[name = string("op_3826_to_fp16"), val = fp16(0x1p-1)]; + tensor var_3827_cast_fp16 = mul(x = linear_144_cast_fp16, y = var_3826_to_fp16)[name = string("op_3827_cast_fp16")]; + tensor input_855_cast_fp16 = add(x = input_843_cast_fp16, y = var_3827_cast_fp16)[name = string("input_855_cast_fp16")]; + tensor input_857_axes_0 = const()[name = string("input_857_axes_0"), val = tensor([-1])]; + tensor encoder_layers_15_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_15_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(322885120)))]; + tensor encoder_layers_15_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_15_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(322887232)))]; + tensor input_857_cast_fp16 = layer_norm(axes = input_857_axes_0, beta = encoder_layers_15_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_15_norm_out_weight_to_fp16, x = input_855_cast_fp16)[name = string("input_857_cast_fp16")]; + tensor cache_65_begin_0 = const()[name = string("cache_65_begin_0"), val = tensor([16, 0, 0, 0])]; + tensor cache_65_end_0 = const()[name = string("cache_65_end_0"), val = tensor([17, 1, 42, 1024])]; + tensor cache_65_end_mask_0 = const()[name = string("cache_65_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_65_squeeze_mask_0 = const()[name = string("cache_65_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_65_cast_fp16 = slice_by_index(begin = cache_65_begin_0, end = cache_65_end_0, end_mask = cache_65_end_mask_0, squeeze_mask = cache_65_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_65_cast_fp16")]; + tensor cache_67_begin_0 = const()[name = string("cache_67_begin_0"), val = tensor([16, 0, 0, 0])]; + tensor cache_67_end_0 = const()[name = string("cache_67_end_0"), val = tensor([17, 1, 1024, 8])]; + tensor cache_67_end_mask_0 = const()[name = string("cache_67_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_67_squeeze_mask_0 = const()[name = string("cache_67_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_67_cast_fp16 = slice_by_index(begin = cache_67_begin_0, end = cache_67_end_0, end_mask = cache_67_end_mask_0, squeeze_mask = cache_67_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_67_cast_fp16")]; + tensor input_859_axes_0 = const()[name = string("input_859_axes_0"), val = tensor([-1])]; + tensor encoder_layers_16_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_16_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(322889344)))]; + tensor encoder_layers_16_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_16_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(322891456)))]; + tensor input_859_cast_fp16 = layer_norm(axes = input_859_axes_0, beta = encoder_layers_16_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_16_norm_feed_forward1_weight_to_fp16, x = input_857_cast_fp16)[name = string("input_859_cast_fp16")]; + tensor encoder_layers_16_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(322893568))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(326039360))))[name = string("encoder_layers_16_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_16_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_16_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(326039552)))]; + tensor linear_145_cast_fp16 = linear(bias = encoder_layers_16_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_16_feed_forward1_linear1_weight_to_fp16_palettized, x = input_859_cast_fp16)[name = string("linear_145_cast_fp16")]; + tensor input_863_cast_fp16 = silu(x = linear_145_cast_fp16)[name = string("input_863_cast_fp16")]; + tensor encoder_layers_16_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(326047808))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(329193600))))[name = string("encoder_layers_16_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_16_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_16_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(329193792)))]; + tensor linear_146_cast_fp16 = linear(bias = encoder_layers_16_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_16_feed_forward1_linear2_weight_to_fp16_palettized, x = input_863_cast_fp16)[name = string("linear_146_cast_fp16")]; + fp16 var_3863_to_fp16 = const()[name = string("op_3863_to_fp16"), val = fp16(0x1p-1)]; + tensor var_3864_cast_fp16 = mul(x = linear_146_cast_fp16, y = var_3863_to_fp16)[name = string("op_3864_cast_fp16")]; + tensor input_869_cast_fp16 = add(x = input_857_cast_fp16, y = var_3864_cast_fp16)[name = string("input_869_cast_fp16")]; + tensor key_33_axes_0 = const()[name = string("key_33_axes_0"), val = tensor([-1])]; + tensor encoder_layers_16_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_16_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(329195904)))]; + tensor encoder_layers_16_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_16_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(329198016)))]; + tensor key_33_cast_fp16 = layer_norm(axes = key_33_axes_0, beta = encoder_layers_16_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_16_norm_self_att_weight_to_fp16, x = input_869_cast_fp16)[name = string("key_33_cast_fp16")]; + bool input_871_interleave_0 = const()[name = string("input_871_interleave_0"), val = bool(false)]; + tensor input_871_cast_fp16 = concat(axis = var_68, interleave = input_871_interleave_0, values = (cache_65_cast_fp16, key_33_cast_fp16))[name = string("input_871_cast_fp16")]; + tensor var_3886_begin_0 = const()[name = string("op_3886_begin_0"), val = tensor([0, 7, 0])]; + tensor var_3886_end_0 = const()[name = string("op_3886_end_0"), val = tensor([1, 42, 1024])]; + tensor var_3886_end_mask_0 = const()[name = string("op_3886_end_mask_0"), val = tensor([true, true, true])]; + tensor var_3886_cast_fp16 = slice_by_index(begin = var_3886_begin_0, end = var_3886_end_0, end_mask = var_3886_end_mask_0, x = cache_65_cast_fp16)[name = string("op_3886_cast_fp16")]; + bool var_3892_interleave_0 = const()[name = string("op_3892_interleave_0"), val = bool(false)]; + tensor var_3892_cast_fp16 = concat(axis = var_68, interleave = var_3892_interleave_0, values = (var_3886_cast_fp16, key_33_cast_fp16))[name = string("op_3892_cast_fp16")]; + tensor encoder_layers_16_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(329200128))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(329986624))))[name = string("encoder_layers_16_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_16_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_16_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(329986816)))]; + tensor linear_147_cast_fp16 = linear(bias = encoder_layers_16_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_16_self_attn_linear_q_weight_to_fp16_palettized, x = key_33_cast_fp16)[name = string("linear_147_cast_fp16")]; + tensor var_3897 = const()[name = string("op_3897"), val = tensor([1, -1, 8, 128])]; + tensor q_97_cast_fp16 = reshape(shape = var_3897, x = linear_147_cast_fp16)[name = string("q_97_cast_fp16")]; + tensor encoder_layers_16_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(329988928))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(330775424))))[name = string("encoder_layers_16_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_16_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_16_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(330775616)))]; + tensor linear_148_cast_fp16 = linear(bias = encoder_layers_16_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_16_self_attn_linear_k_weight_to_fp16_palettized, x = input_871_cast_fp16)[name = string("linear_148_cast_fp16")]; + tensor var_3902 = const()[name = string("op_3902"), val = tensor([1, -1, 8, 128])]; + tensor k_65_cast_fp16 = reshape(shape = var_3902, x = linear_148_cast_fp16)[name = string("k_65_cast_fp16")]; + tensor encoder_layers_16_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(330777728))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331564224))))[name = string("encoder_layers_16_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_16_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_16_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331564416)))]; + tensor linear_149_cast_fp16 = linear(bias = encoder_layers_16_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_16_self_attn_linear_v_weight_to_fp16_palettized, x = input_871_cast_fp16)[name = string("linear_149_cast_fp16")]; + tensor var_3907 = const()[name = string("op_3907"), val = tensor([1, -1, 8, 128])]; + tensor v_33_cast_fp16 = reshape(shape = var_3907, x = linear_149_cast_fp16)[name = string("v_33_cast_fp16")]; + tensor value_41_perm_0 = const()[name = string("value_41_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_16_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_16_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331566528)))]; + tensor var_3920_cast_fp16 = add(x = q_97_cast_fp16, y = encoder_layers_16_self_attn_pos_bias_u_to_fp16)[name = string("op_3920_cast_fp16")]; + tensor encoder_layers_16_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_16_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331568640)))]; + tensor var_3922_cast_fp16 = add(x = q_97_cast_fp16, y = encoder_layers_16_self_attn_pos_bias_v_to_fp16)[name = string("op_3922_cast_fp16")]; + tensor q_with_bias_v_33_perm_0 = const()[name = string("q_with_bias_v_33_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_423_transpose_x_0 = const()[name = string("x_423_transpose_x_0"), val = bool(false)]; + bool x_423_transpose_y_0 = const()[name = string("x_423_transpose_y_0"), val = bool(false)]; + tensor op_3924_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331570752))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331670144))))[name = string("op_3924_to_fp16_quantized")]; + tensor q_with_bias_v_33_cast_fp16 = transpose(perm = q_with_bias_v_33_perm_0, x = var_3922_cast_fp16)[name = string("transpose_218")]; + tensor x_423_cast_fp16 = matmul(transpose_x = x_423_transpose_x_0, transpose_y = x_423_transpose_y_0, x = q_with_bias_v_33_cast_fp16, y = op_3924_to_fp16_quantized)[name = string("x_423_cast_fp16")]; + tensor x_425_pad_0 = const()[name = string("x_425_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_425_mode_0 = const()[name = string("x_425_mode_0"), val = string("constant")]; + fp16 const_287_to_fp16 = const()[name = string("const_287_to_fp16"), val = fp16(0x0p+0)]; + tensor x_425_cast_fp16 = pad(constant_val = const_287_to_fp16, mode = x_425_mode_0, pad = x_425_pad_0, x = x_423_cast_fp16)[name = string("x_425_cast_fp16")]; + tensor var_3932 = const()[name = string("op_3932"), val = tensor([1, 8, -1, 7])]; + tensor x_427_cast_fp16 = reshape(shape = var_3932, x = x_425_cast_fp16)[name = string("x_427_cast_fp16")]; + tensor var_3936_begin_0 = const()[name = string("op_3936_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_3936_end_0 = const()[name = string("op_3936_end_0"), val = tensor([1, 8, 98, 7])]; + tensor var_3936_end_mask_0 = const()[name = string("op_3936_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_3936_cast_fp16 = slice_by_index(begin = var_3936_begin_0, end = var_3936_end_0, end_mask = var_3936_end_mask_0, x = x_427_cast_fp16)[name = string("op_3936_cast_fp16")]; + tensor var_3937 = const()[name = string("op_3937"), val = tensor([1, 8, 7, 97])]; + tensor matrix_bd_65_cast_fp16 = reshape(shape = var_3937, x = var_3936_cast_fp16)[name = string("matrix_bd_65_cast_fp16")]; + bool matrix_ac_33_transpose_x_0 = const()[name = string("matrix_ac_33_transpose_x_0"), val = bool(false)]; + bool matrix_ac_33_transpose_y_0 = const()[name = string("matrix_ac_33_transpose_y_0"), val = bool(false)]; + tensor transpose_128_perm_0 = const()[name = string("transpose_128_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_129_perm_0 = const()[name = string("transpose_129_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_129 = transpose(perm = transpose_129_perm_0, x = k_65_cast_fp16)[name = string("transpose_216")]; + tensor transpose_128 = transpose(perm = transpose_128_perm_0, x = var_3920_cast_fp16)[name = string("transpose_217")]; + tensor matrix_ac_33_cast_fp16 = matmul(transpose_x = matrix_ac_33_transpose_x_0, transpose_y = matrix_ac_33_transpose_y_0, x = transpose_128, y = transpose_129)[name = string("matrix_ac_33_cast_fp16")]; + tensor matrix_bd_67_begin_0 = const()[name = string("matrix_bd_67_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_67_end_0 = const()[name = string("matrix_bd_67_end_0"), val = tensor([1, 8, 7, 49])]; + tensor matrix_bd_67_end_mask_0 = const()[name = string("matrix_bd_67_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_67_cast_fp16 = slice_by_index(begin = matrix_bd_67_begin_0, end = matrix_bd_67_end_0, end_mask = matrix_bd_67_end_mask_0, x = matrix_bd_65_cast_fp16)[name = string("matrix_bd_67_cast_fp16")]; + tensor var_3946_cast_fp16 = add(x = matrix_ac_33_cast_fp16, y = matrix_bd_67_cast_fp16)[name = string("op_3946_cast_fp16")]; + fp16 _inversed_scores_65_y_0_to_fp16 = const()[name = string("_inversed_scores_65_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_65_cast_fp16 = mul(x = var_3946_cast_fp16, y = _inversed_scores_65_y_0_to_fp16)[name = string("_inversed_scores_65_cast_fp16")]; + tensor scores_67_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_65_cast_fp16, cond = mask_11)[name = string("scores_67_cast_fp16")]; + tensor var_3952_cast_fp16 = softmax(axis = var_59, x = scores_67_cast_fp16)[name = string("op_3952_cast_fp16")]; + tensor input_873_cast_fp16 = select(a = var_44_to_fp16, b = var_3952_cast_fp16, cond = mask_11)[name = string("input_873_cast_fp16")]; + bool x_429_transpose_x_0 = const()[name = string("x_429_transpose_x_0"), val = bool(false)]; + bool x_429_transpose_y_0 = const()[name = string("x_429_transpose_y_0"), val = bool(false)]; + tensor value_41_cast_fp16 = transpose(perm = value_41_perm_0, x = v_33_cast_fp16)[name = string("transpose_215")]; + tensor x_429_cast_fp16 = matmul(transpose_x = x_429_transpose_x_0, transpose_y = x_429_transpose_y_0, x = input_873_cast_fp16, y = value_41_cast_fp16)[name = string("x_429_cast_fp16")]; + tensor var_3956_perm_0 = const()[name = string("op_3956_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_3957 = const()[name = string("op_3957"), val = tensor([1, -1, 1024])]; + tensor var_3956_cast_fp16 = transpose(perm = var_3956_perm_0, x = x_429_cast_fp16)[name = string("transpose_214")]; + tensor input_875_cast_fp16 = reshape(shape = var_3957, x = var_3956_cast_fp16)[name = string("input_875_cast_fp16")]; + tensor encoder_layers_16_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331670464))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332456960))))[name = string("encoder_layers_16_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_16_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_16_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332457152)))]; + tensor linear_151_cast_fp16 = linear(bias = encoder_layers_16_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_16_self_attn_linear_out_weight_to_fp16_palettized, x = input_875_cast_fp16)[name = string("linear_151_cast_fp16")]; + tensor input_879_cast_fp16 = add(x = input_869_cast_fp16, y = linear_151_cast_fp16)[name = string("input_879_cast_fp16")]; + tensor x_433_axes_0 = const()[name = string("x_433_axes_0"), val = tensor([-1])]; + tensor encoder_layers_16_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_16_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332459264)))]; + tensor encoder_layers_16_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_16_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332461376)))]; + tensor x_433_cast_fp16 = layer_norm(axes = x_433_axes_0, beta = encoder_layers_16_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_16_norm_conv_weight_to_fp16, x = input_879_cast_fp16)[name = string("x_433_cast_fp16")]; + tensor input_881_perm_0 = const()[name = string("input_881_perm_0"), val = tensor([0, 2, 1])]; + string input_883_pad_type_0 = const()[name = string("input_883_pad_type_0"), val = string("valid")]; + tensor input_883_strides_0 = const()[name = string("input_883_strides_0"), val = tensor([1])]; + tensor input_883_pad_0 = const()[name = string("input_883_pad_0"), val = tensor([0, 0])]; + tensor input_883_dilations_0 = const()[name = string("input_883_dilations_0"), val = tensor([1])]; + int32 input_883_groups_0 = const()[name = string("input_883_groups_0"), val = int32(1)]; + tensor encoder_layers_16_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332463488))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(334560704))))[name = string("encoder_layers_16_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_881_cast_fp16 = transpose(perm = input_881_perm_0, x = x_433_cast_fp16)[name = string("transpose_213")]; + tensor input_883_cast_fp16 = conv(dilations = input_883_dilations_0, groups = input_883_groups_0, pad = input_883_pad_0, pad_type = input_883_pad_type_0, strides = input_883_strides_0, weight = encoder_layers_16_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_881_cast_fp16)[name = string("input_883_cast_fp16")]; + int32 x_435_split_num_splits_0 = const()[name = string("x_435_split_num_splits_0"), val = int32(2)]; + int32 x_435_split_axis_0 = const()[name = string("x_435_split_axis_0"), val = int32(1)]; + tensor x_435_split_cast_fp16_0, tensor x_435_split_cast_fp16_1 = split(axis = x_435_split_axis_0, num_splits = x_435_split_num_splits_0, x = input_883_cast_fp16)[name = string("x_435_split_cast_fp16")]; + tensor x_435_split_1_sigmoid_cast_fp16 = sigmoid(x = x_435_split_cast_fp16_1)[name = string("x_435_split_1_sigmoid_cast_fp16")]; + tensor x_435_cast_fp16 = mul(x = x_435_split_cast_fp16_0, y = x_435_split_1_sigmoid_cast_fp16)[name = string("x_435_cast_fp16")]; + tensor input_885_cast_fp16 = select(a = var_44_to_fp16, b = x_435_cast_fp16, cond = var_575)[name = string("input_885_cast_fp16")]; + bool new_x_67_interleave_0 = const()[name = string("new_x_67_interleave_0"), val = bool(false)]; + tensor new_x_67_cast_fp16 = concat(axis = var_59, interleave = new_x_67_interleave_0, values = (cache_67_cast_fp16, input_885_cast_fp16))[name = string("new_x_67_cast_fp16")]; + tensor var_3996_begin_0 = const()[name = string("op_3996_begin_0"), val = tensor([0, 0, 7])]; + tensor var_3996_end_0 = const()[name = string("op_3996_end_0"), val = tensor([1, 1024, 15])]; + tensor var_3996_end_mask_0 = const()[name = string("op_3996_end_mask_0"), val = tensor([true, true, true])]; + tensor var_3996_cast_fp16 = slice_by_index(begin = var_3996_begin_0, end = var_3996_end_0, end_mask = var_3996_end_mask_0, x = new_x_67_cast_fp16)[name = string("op_3996_cast_fp16")]; + string x_437_pad_type_0 = const()[name = string("x_437_pad_type_0"), val = string("valid")]; + int32 x_437_groups_0 = const()[name = string("x_437_groups_0"), val = int32(1024)]; + tensor x_437_strides_0 = const()[name = string("x_437_strides_0"), val = tensor([1])]; + tensor x_437_pad_0 = const()[name = string("x_437_pad_0"), val = tensor([0, 0])]; + tensor x_437_dilations_0 = const()[name = string("x_437_dilations_0"), val = tensor([1])]; + tensor encoder_layers_16_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(334564864))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(334574144))))[name = string("encoder_layers_16_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_437_cast_fp16 = conv(dilations = x_437_dilations_0, groups = x_437_groups_0, pad = x_437_pad_0, pad_type = x_437_pad_type_0, strides = x_437_strides_0, weight = encoder_layers_16_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_67_cast_fp16)[name = string("x_437_cast_fp16")]; + tensor input_887_perm_0 = const()[name = string("input_887_perm_0"), val = tensor([0, 2, 1])]; + tensor x_439_axes_0 = const()[name = string("x_439_axes_0"), val = tensor([-1])]; + tensor encoder_layers_16_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_16_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(334576256)))]; + tensor encoder_layers_16_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_16_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(334578368)))]; + tensor input_887_cast_fp16 = transpose(perm = input_887_perm_0, x = x_437_cast_fp16)[name = string("transpose_212")]; + tensor x_439_cast_fp16 = layer_norm(axes = x_439_axes_0, beta = encoder_layers_16_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_16_conv_batch_norm_weight_to_fp16, x = input_887_cast_fp16)[name = string("x_439_cast_fp16")]; + tensor input_889_perm_0 = const()[name = string("input_889_perm_0"), val = tensor([0, 2, 1])]; + tensor input_889_cast_fp16 = transpose(perm = input_889_perm_0, x = x_439_cast_fp16)[name = string("transpose_211")]; + tensor input_891_cast_fp16 = silu(x = input_889_cast_fp16)[name = string("input_891_cast_fp16")]; + string x_441_pad_type_0 = const()[name = string("x_441_pad_type_0"), val = string("valid")]; + tensor x_441_strides_0 = const()[name = string("x_441_strides_0"), val = tensor([1])]; + tensor x_441_pad_0 = const()[name = string("x_441_pad_0"), val = tensor([0, 0])]; + tensor x_441_dilations_0 = const()[name = string("x_441_dilations_0"), val = tensor([1])]; + int32 x_441_groups_0 = const()[name = string("x_441_groups_0"), val = int32(1)]; + tensor encoder_layers_16_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(334580480))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(335629120))))[name = string("encoder_layers_16_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_441_cast_fp16 = conv(dilations = x_441_dilations_0, groups = x_441_groups_0, pad = x_441_pad_0, pad_type = x_441_pad_type_0, strides = x_441_strides_0, weight = encoder_layers_16_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_891_cast_fp16)[name = string("x_441_cast_fp16")]; + tensor input_893_perm_0 = const()[name = string("input_893_perm_0"), val = tensor([0, 2, 1])]; + tensor input_893_cast_fp16 = transpose(perm = input_893_perm_0, x = x_441_cast_fp16)[name = string("transpose_210")]; + tensor input_895_cast_fp16 = add(x = input_879_cast_fp16, y = input_893_cast_fp16)[name = string("input_895_cast_fp16")]; + tensor input_897_axes_0 = const()[name = string("input_897_axes_0"), val = tensor([-1])]; + tensor encoder_layers_16_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_16_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(335631232)))]; + tensor encoder_layers_16_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_16_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(335633344)))]; + tensor input_897_cast_fp16 = layer_norm(axes = input_897_axes_0, beta = encoder_layers_16_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_16_norm_feed_forward2_weight_to_fp16, x = input_895_cast_fp16)[name = string("input_897_cast_fp16")]; + tensor encoder_layers_16_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(335635456))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(338781248))))[name = string("encoder_layers_16_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_16_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_16_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(338781440)))]; + tensor linear_152_cast_fp16 = linear(bias = encoder_layers_16_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_16_feed_forward2_linear1_weight_to_fp16_palettized, x = input_897_cast_fp16)[name = string("linear_152_cast_fp16")]; + tensor input_901_cast_fp16 = silu(x = linear_152_cast_fp16)[name = string("input_901_cast_fp16")]; + tensor encoder_layers_16_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(338789696))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(341935488))))[name = string("encoder_layers_16_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_16_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_16_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(341935680)))]; + tensor linear_153_cast_fp16 = linear(bias = encoder_layers_16_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_16_feed_forward2_linear2_weight_to_fp16_palettized, x = input_901_cast_fp16)[name = string("linear_153_cast_fp16")]; + fp16 var_4039_to_fp16 = const()[name = string("op_4039_to_fp16"), val = fp16(0x1p-1)]; + tensor var_4040_cast_fp16 = mul(x = linear_153_cast_fp16, y = var_4039_to_fp16)[name = string("op_4040_cast_fp16")]; + tensor input_907_cast_fp16 = add(x = input_895_cast_fp16, y = var_4040_cast_fp16)[name = string("input_907_cast_fp16")]; + tensor input_909_axes_0 = const()[name = string("input_909_axes_0"), val = tensor([-1])]; + tensor encoder_layers_16_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_16_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(341937792)))]; + tensor encoder_layers_16_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_16_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(341939904)))]; + tensor input_909_cast_fp16 = layer_norm(axes = input_909_axes_0, beta = encoder_layers_16_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_16_norm_out_weight_to_fp16, x = input_907_cast_fp16)[name = string("input_909_cast_fp16")]; + tensor cache_69_begin_0 = const()[name = string("cache_69_begin_0"), val = tensor([17, 0, 0, 0])]; + tensor cache_69_end_0 = const()[name = string("cache_69_end_0"), val = tensor([18, 1, 42, 1024])]; + tensor cache_69_end_mask_0 = const()[name = string("cache_69_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_69_squeeze_mask_0 = const()[name = string("cache_69_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_69_cast_fp16 = slice_by_index(begin = cache_69_begin_0, end = cache_69_end_0, end_mask = cache_69_end_mask_0, squeeze_mask = cache_69_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_69_cast_fp16")]; + tensor cache_71_begin_0 = const()[name = string("cache_71_begin_0"), val = tensor([17, 0, 0, 0])]; + tensor cache_71_end_0 = const()[name = string("cache_71_end_0"), val = tensor([18, 1, 1024, 8])]; + tensor cache_71_end_mask_0 = const()[name = string("cache_71_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_71_squeeze_mask_0 = const()[name = string("cache_71_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_71_cast_fp16 = slice_by_index(begin = cache_71_begin_0, end = cache_71_end_0, end_mask = cache_71_end_mask_0, squeeze_mask = cache_71_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_71_cast_fp16")]; + tensor input_911_axes_0 = const()[name = string("input_911_axes_0"), val = tensor([-1])]; + tensor encoder_layers_17_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_17_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(341942016)))]; + tensor encoder_layers_17_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_17_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(341944128)))]; + tensor input_911_cast_fp16 = layer_norm(axes = input_911_axes_0, beta = encoder_layers_17_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_17_norm_feed_forward1_weight_to_fp16, x = input_909_cast_fp16)[name = string("input_911_cast_fp16")]; + tensor encoder_layers_17_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(341946240))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(345092032))))[name = string("encoder_layers_17_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_17_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_17_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(345092224)))]; + tensor linear_154_cast_fp16 = linear(bias = encoder_layers_17_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_17_feed_forward1_linear1_weight_to_fp16_palettized, x = input_911_cast_fp16)[name = string("linear_154_cast_fp16")]; + tensor input_915_cast_fp16 = silu(x = linear_154_cast_fp16)[name = string("input_915_cast_fp16")]; + tensor encoder_layers_17_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(345100480))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(348246272))))[name = string("encoder_layers_17_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_17_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_17_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(348246464)))]; + tensor linear_155_cast_fp16 = linear(bias = encoder_layers_17_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_17_feed_forward1_linear2_weight_to_fp16_palettized, x = input_915_cast_fp16)[name = string("linear_155_cast_fp16")]; + fp16 var_4076_to_fp16 = const()[name = string("op_4076_to_fp16"), val = fp16(0x1p-1)]; + tensor var_4077_cast_fp16 = mul(x = linear_155_cast_fp16, y = var_4076_to_fp16)[name = string("op_4077_cast_fp16")]; + tensor input_921_cast_fp16 = add(x = input_909_cast_fp16, y = var_4077_cast_fp16)[name = string("input_921_cast_fp16")]; + tensor key_35_axes_0 = const()[name = string("key_35_axes_0"), val = tensor([-1])]; + tensor encoder_layers_17_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_17_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(348248576)))]; + tensor encoder_layers_17_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_17_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(348250688)))]; + tensor key_35_cast_fp16 = layer_norm(axes = key_35_axes_0, beta = encoder_layers_17_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_17_norm_self_att_weight_to_fp16, x = input_921_cast_fp16)[name = string("key_35_cast_fp16")]; + bool input_923_interleave_0 = const()[name = string("input_923_interleave_0"), val = bool(false)]; + tensor input_923_cast_fp16 = concat(axis = var_68, interleave = input_923_interleave_0, values = (cache_69_cast_fp16, key_35_cast_fp16))[name = string("input_923_cast_fp16")]; + tensor var_4099_begin_0 = const()[name = string("op_4099_begin_0"), val = tensor([0, 7, 0])]; + tensor var_4099_end_0 = const()[name = string("op_4099_end_0"), val = tensor([1, 42, 1024])]; + tensor var_4099_end_mask_0 = const()[name = string("op_4099_end_mask_0"), val = tensor([true, true, true])]; + tensor var_4099_cast_fp16 = slice_by_index(begin = var_4099_begin_0, end = var_4099_end_0, end_mask = var_4099_end_mask_0, x = cache_69_cast_fp16)[name = string("op_4099_cast_fp16")]; + bool var_4105_interleave_0 = const()[name = string("op_4105_interleave_0"), val = bool(false)]; + tensor var_4105_cast_fp16 = concat(axis = var_68, interleave = var_4105_interleave_0, values = (var_4099_cast_fp16, key_35_cast_fp16))[name = string("op_4105_cast_fp16")]; + tensor encoder_layers_17_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(348252800))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(349039296))))[name = string("encoder_layers_17_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_17_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_17_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(349039488)))]; + tensor linear_156_cast_fp16 = linear(bias = encoder_layers_17_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_17_self_attn_linear_q_weight_to_fp16_palettized, x = key_35_cast_fp16)[name = string("linear_156_cast_fp16")]; + tensor var_4110 = const()[name = string("op_4110"), val = tensor([1, -1, 8, 128])]; + tensor q_103_cast_fp16 = reshape(shape = var_4110, x = linear_156_cast_fp16)[name = string("q_103_cast_fp16")]; + tensor encoder_layers_17_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(349041600))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(349828096))))[name = string("encoder_layers_17_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_17_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_17_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(349828288)))]; + tensor linear_157_cast_fp16 = linear(bias = encoder_layers_17_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_17_self_attn_linear_k_weight_to_fp16_palettized, x = input_923_cast_fp16)[name = string("linear_157_cast_fp16")]; + tensor var_4115 = const()[name = string("op_4115"), val = tensor([1, -1, 8, 128])]; + tensor k_69_cast_fp16 = reshape(shape = var_4115, x = linear_157_cast_fp16)[name = string("k_69_cast_fp16")]; + tensor encoder_layers_17_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(349830400))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350616896))))[name = string("encoder_layers_17_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_17_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_17_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350617088)))]; + tensor linear_158_cast_fp16 = linear(bias = encoder_layers_17_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_17_self_attn_linear_v_weight_to_fp16_palettized, x = input_923_cast_fp16)[name = string("linear_158_cast_fp16")]; + tensor var_4120 = const()[name = string("op_4120"), val = tensor([1, -1, 8, 128])]; + tensor v_35_cast_fp16 = reshape(shape = var_4120, x = linear_158_cast_fp16)[name = string("v_35_cast_fp16")]; + tensor value_43_perm_0 = const()[name = string("value_43_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_17_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_17_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350619200)))]; + tensor var_4133_cast_fp16 = add(x = q_103_cast_fp16, y = encoder_layers_17_self_attn_pos_bias_u_to_fp16)[name = string("op_4133_cast_fp16")]; + tensor encoder_layers_17_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_17_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350621312)))]; + tensor var_4135_cast_fp16 = add(x = q_103_cast_fp16, y = encoder_layers_17_self_attn_pos_bias_v_to_fp16)[name = string("op_4135_cast_fp16")]; + tensor q_with_bias_v_35_perm_0 = const()[name = string("q_with_bias_v_35_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_449_transpose_x_0 = const()[name = string("x_449_transpose_x_0"), val = bool(false)]; + bool x_449_transpose_y_0 = const()[name = string("x_449_transpose_y_0"), val = bool(false)]; + tensor op_4137_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350623424))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350722816))))[name = string("op_4137_to_fp16_quantized")]; + tensor q_with_bias_v_35_cast_fp16 = transpose(perm = q_with_bias_v_35_perm_0, x = var_4135_cast_fp16)[name = string("transpose_209")]; + tensor x_449_cast_fp16 = matmul(transpose_x = x_449_transpose_x_0, transpose_y = x_449_transpose_y_0, x = q_with_bias_v_35_cast_fp16, y = op_4137_to_fp16_quantized)[name = string("x_449_cast_fp16")]; + tensor x_451_pad_0 = const()[name = string("x_451_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_451_mode_0 = const()[name = string("x_451_mode_0"), val = string("constant")]; + fp16 const_300_to_fp16 = const()[name = string("const_300_to_fp16"), val = fp16(0x0p+0)]; + tensor x_451_cast_fp16 = pad(constant_val = const_300_to_fp16, mode = x_451_mode_0, pad = x_451_pad_0, x = x_449_cast_fp16)[name = string("x_451_cast_fp16")]; + tensor var_4145 = const()[name = string("op_4145"), val = tensor([1, 8, -1, 7])]; + tensor x_453_cast_fp16 = reshape(shape = var_4145, x = x_451_cast_fp16)[name = string("x_453_cast_fp16")]; + tensor var_4149_begin_0 = const()[name = string("op_4149_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_4149_end_0 = const()[name = string("op_4149_end_0"), val = tensor([1, 8, 98, 7])]; + tensor var_4149_end_mask_0 = const()[name = string("op_4149_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_4149_cast_fp16 = slice_by_index(begin = var_4149_begin_0, end = var_4149_end_0, end_mask = var_4149_end_mask_0, x = x_453_cast_fp16)[name = string("op_4149_cast_fp16")]; + tensor var_4150 = const()[name = string("op_4150"), val = tensor([1, 8, 7, 97])]; + tensor matrix_bd_69_cast_fp16 = reshape(shape = var_4150, x = var_4149_cast_fp16)[name = string("matrix_bd_69_cast_fp16")]; + bool matrix_ac_35_transpose_x_0 = const()[name = string("matrix_ac_35_transpose_x_0"), val = bool(false)]; + bool matrix_ac_35_transpose_y_0 = const()[name = string("matrix_ac_35_transpose_y_0"), val = bool(false)]; + tensor transpose_130_perm_0 = const()[name = string("transpose_130_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_131_perm_0 = const()[name = string("transpose_131_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_131 = transpose(perm = transpose_131_perm_0, x = k_69_cast_fp16)[name = string("transpose_207")]; + tensor transpose_130 = transpose(perm = transpose_130_perm_0, x = var_4133_cast_fp16)[name = string("transpose_208")]; + tensor matrix_ac_35_cast_fp16 = matmul(transpose_x = matrix_ac_35_transpose_x_0, transpose_y = matrix_ac_35_transpose_y_0, x = transpose_130, y = transpose_131)[name = string("matrix_ac_35_cast_fp16")]; + tensor matrix_bd_71_begin_0 = const()[name = string("matrix_bd_71_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_71_end_0 = const()[name = string("matrix_bd_71_end_0"), val = tensor([1, 8, 7, 49])]; + tensor matrix_bd_71_end_mask_0 = const()[name = string("matrix_bd_71_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_71_cast_fp16 = slice_by_index(begin = matrix_bd_71_begin_0, end = matrix_bd_71_end_0, end_mask = matrix_bd_71_end_mask_0, x = matrix_bd_69_cast_fp16)[name = string("matrix_bd_71_cast_fp16")]; + tensor var_4159_cast_fp16 = add(x = matrix_ac_35_cast_fp16, y = matrix_bd_71_cast_fp16)[name = string("op_4159_cast_fp16")]; + fp16 _inversed_scores_69_y_0_to_fp16 = const()[name = string("_inversed_scores_69_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_69_cast_fp16 = mul(x = var_4159_cast_fp16, y = _inversed_scores_69_y_0_to_fp16)[name = string("_inversed_scores_69_cast_fp16")]; + tensor scores_71_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_69_cast_fp16, cond = mask_11)[name = string("scores_71_cast_fp16")]; + tensor var_4165_cast_fp16 = softmax(axis = var_59, x = scores_71_cast_fp16)[name = string("op_4165_cast_fp16")]; + tensor input_925_cast_fp16 = select(a = var_44_to_fp16, b = var_4165_cast_fp16, cond = mask_11)[name = string("input_925_cast_fp16")]; + bool x_455_transpose_x_0 = const()[name = string("x_455_transpose_x_0"), val = bool(false)]; + bool x_455_transpose_y_0 = const()[name = string("x_455_transpose_y_0"), val = bool(false)]; + tensor value_43_cast_fp16 = transpose(perm = value_43_perm_0, x = v_35_cast_fp16)[name = string("transpose_206")]; + tensor x_455_cast_fp16 = matmul(transpose_x = x_455_transpose_x_0, transpose_y = x_455_transpose_y_0, x = input_925_cast_fp16, y = value_43_cast_fp16)[name = string("x_455_cast_fp16")]; + tensor var_4169_perm_0 = const()[name = string("op_4169_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_4170 = const()[name = string("op_4170"), val = tensor([1, -1, 1024])]; + tensor var_4169_cast_fp16 = transpose(perm = var_4169_perm_0, x = x_455_cast_fp16)[name = string("transpose_205")]; + tensor input_927_cast_fp16 = reshape(shape = var_4170, x = var_4169_cast_fp16)[name = string("input_927_cast_fp16")]; + tensor encoder_layers_17_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350723136))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351509632))))[name = string("encoder_layers_17_self_attn_linear_out_weight_to_fp16_palettized")]; + tensor encoder_layers_17_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_17_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351509824)))]; + tensor linear_160_cast_fp16 = linear(bias = encoder_layers_17_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_17_self_attn_linear_out_weight_to_fp16_palettized, x = input_927_cast_fp16)[name = string("linear_160_cast_fp16")]; + tensor input_931_cast_fp16 = add(x = input_921_cast_fp16, y = linear_160_cast_fp16)[name = string("input_931_cast_fp16")]; + tensor x_459_axes_0 = const()[name = string("x_459_axes_0"), val = tensor([-1])]; + tensor encoder_layers_17_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_17_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351511936)))]; + tensor encoder_layers_17_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_17_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351514048)))]; + tensor x_459_cast_fp16 = layer_norm(axes = x_459_axes_0, beta = encoder_layers_17_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_17_norm_conv_weight_to_fp16, x = input_931_cast_fp16)[name = string("x_459_cast_fp16")]; + tensor input_933_perm_0 = const()[name = string("input_933_perm_0"), val = tensor([0, 2, 1])]; + string input_935_pad_type_0 = const()[name = string("input_935_pad_type_0"), val = string("valid")]; + tensor input_935_strides_0 = const()[name = string("input_935_strides_0"), val = tensor([1])]; + tensor input_935_pad_0 = const()[name = string("input_935_pad_0"), val = tensor([0, 0])]; + tensor input_935_dilations_0 = const()[name = string("input_935_dilations_0"), val = tensor([1])]; + int32 input_935_groups_0 = const()[name = string("input_935_groups_0"), val = int32(1)]; + tensor encoder_layers_17_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351516160))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(353613376))))[name = string("encoder_layers_17_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_933_cast_fp16 = transpose(perm = input_933_perm_0, x = x_459_cast_fp16)[name = string("transpose_204")]; + tensor input_935_cast_fp16 = conv(dilations = input_935_dilations_0, groups = input_935_groups_0, pad = input_935_pad_0, pad_type = input_935_pad_type_0, strides = input_935_strides_0, weight = encoder_layers_17_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_933_cast_fp16)[name = string("input_935_cast_fp16")]; + int32 x_461_split_num_splits_0 = const()[name = string("x_461_split_num_splits_0"), val = int32(2)]; + int32 x_461_split_axis_0 = const()[name = string("x_461_split_axis_0"), val = int32(1)]; + tensor x_461_split_cast_fp16_0, tensor x_461_split_cast_fp16_1 = split(axis = x_461_split_axis_0, num_splits = x_461_split_num_splits_0, x = input_935_cast_fp16)[name = string("x_461_split_cast_fp16")]; + tensor x_461_split_1_sigmoid_cast_fp16 = sigmoid(x = x_461_split_cast_fp16_1)[name = string("x_461_split_1_sigmoid_cast_fp16")]; + tensor x_461_cast_fp16 = mul(x = x_461_split_cast_fp16_0, y = x_461_split_1_sigmoid_cast_fp16)[name = string("x_461_cast_fp16")]; + tensor input_937_cast_fp16 = select(a = var_44_to_fp16, b = x_461_cast_fp16, cond = var_575)[name = string("input_937_cast_fp16")]; + bool new_x_71_interleave_0 = const()[name = string("new_x_71_interleave_0"), val = bool(false)]; + tensor new_x_71_cast_fp16 = concat(axis = var_59, interleave = new_x_71_interleave_0, values = (cache_71_cast_fp16, input_937_cast_fp16))[name = string("new_x_71_cast_fp16")]; + tensor var_4209_begin_0 = const()[name = string("op_4209_begin_0"), val = tensor([0, 0, 7])]; + tensor var_4209_end_0 = const()[name = string("op_4209_end_0"), val = tensor([1, 1024, 15])]; + tensor var_4209_end_mask_0 = const()[name = string("op_4209_end_mask_0"), val = tensor([true, true, true])]; + tensor var_4209_cast_fp16 = slice_by_index(begin = var_4209_begin_0, end = var_4209_end_0, end_mask = var_4209_end_mask_0, x = new_x_71_cast_fp16)[name = string("op_4209_cast_fp16")]; + string x_463_pad_type_0 = const()[name = string("x_463_pad_type_0"), val = string("valid")]; + int32 x_463_groups_0 = const()[name = string("x_463_groups_0"), val = int32(1024)]; + tensor x_463_strides_0 = const()[name = string("x_463_strides_0"), val = tensor([1])]; + tensor x_463_pad_0 = const()[name = string("x_463_pad_0"), val = tensor([0, 0])]; + tensor x_463_dilations_0 = const()[name = string("x_463_dilations_0"), val = tensor([1])]; + tensor encoder_layers_17_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(353617536))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(353626816))))[name = string("encoder_layers_17_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_463_cast_fp16 = conv(dilations = x_463_dilations_0, groups = x_463_groups_0, pad = x_463_pad_0, pad_type = x_463_pad_type_0, strides = x_463_strides_0, weight = encoder_layers_17_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_71_cast_fp16)[name = string("x_463_cast_fp16")]; + tensor input_939_perm_0 = const()[name = string("input_939_perm_0"), val = tensor([0, 2, 1])]; + tensor x_465_axes_0 = const()[name = string("x_465_axes_0"), val = tensor([-1])]; + tensor encoder_layers_17_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_17_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(353628928)))]; + tensor encoder_layers_17_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_17_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(353631040)))]; + tensor input_939_cast_fp16 = transpose(perm = input_939_perm_0, x = x_463_cast_fp16)[name = string("transpose_203")]; + tensor x_465_cast_fp16 = layer_norm(axes = x_465_axes_0, beta = encoder_layers_17_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_17_conv_batch_norm_weight_to_fp16, x = input_939_cast_fp16)[name = string("x_465_cast_fp16")]; + tensor input_941_perm_0 = const()[name = string("input_941_perm_0"), val = tensor([0, 2, 1])]; + tensor input_941_cast_fp16 = transpose(perm = input_941_perm_0, x = x_465_cast_fp16)[name = string("transpose_202")]; + tensor input_943_cast_fp16 = silu(x = input_941_cast_fp16)[name = string("input_943_cast_fp16")]; + string x_467_pad_type_0 = const()[name = string("x_467_pad_type_0"), val = string("valid")]; + tensor x_467_strides_0 = const()[name = string("x_467_strides_0"), val = tensor([1])]; + tensor x_467_pad_0 = const()[name = string("x_467_pad_0"), val = tensor([0, 0])]; + tensor x_467_dilations_0 = const()[name = string("x_467_dilations_0"), val = tensor([1])]; + int32 x_467_groups_0 = const()[name = string("x_467_groups_0"), val = int32(1)]; + tensor encoder_layers_17_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(353633152))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(354681792))))[name = string("encoder_layers_17_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_467_cast_fp16 = conv(dilations = x_467_dilations_0, groups = x_467_groups_0, pad = x_467_pad_0, pad_type = x_467_pad_type_0, strides = x_467_strides_0, weight = encoder_layers_17_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_943_cast_fp16)[name = string("x_467_cast_fp16")]; + tensor input_945_perm_0 = const()[name = string("input_945_perm_0"), val = tensor([0, 2, 1])]; + tensor input_945_cast_fp16 = transpose(perm = input_945_perm_0, x = x_467_cast_fp16)[name = string("transpose_201")]; + tensor input_947_cast_fp16 = add(x = input_931_cast_fp16, y = input_945_cast_fp16)[name = string("input_947_cast_fp16")]; + tensor input_949_axes_0 = const()[name = string("input_949_axes_0"), val = tensor([-1])]; + tensor encoder_layers_17_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_17_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(354683904)))]; + tensor encoder_layers_17_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_17_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(354686016)))]; + tensor input_949_cast_fp16 = layer_norm(axes = input_949_axes_0, beta = encoder_layers_17_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_17_norm_feed_forward2_weight_to_fp16, x = input_947_cast_fp16)[name = string("input_949_cast_fp16")]; + tensor encoder_layers_17_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(354688128))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(357833920))))[name = string("encoder_layers_17_feed_forward2_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_17_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_17_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(357834112)))]; + tensor linear_161_cast_fp16 = linear(bias = encoder_layers_17_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_17_feed_forward2_linear1_weight_to_fp16_palettized, x = input_949_cast_fp16)[name = string("linear_161_cast_fp16")]; + tensor input_953_cast_fp16 = silu(x = linear_161_cast_fp16)[name = string("input_953_cast_fp16")]; + tensor encoder_layers_17_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(357842368))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(360988160))))[name = string("encoder_layers_17_feed_forward2_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_17_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_17_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(360988352)))]; + tensor linear_162_cast_fp16 = linear(bias = encoder_layers_17_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_17_feed_forward2_linear2_weight_to_fp16_palettized, x = input_953_cast_fp16)[name = string("linear_162_cast_fp16")]; + fp16 var_4252_to_fp16 = const()[name = string("op_4252_to_fp16"), val = fp16(0x1p-1)]; + tensor var_4253_cast_fp16 = mul(x = linear_162_cast_fp16, y = var_4252_to_fp16)[name = string("op_4253_cast_fp16")]; + tensor input_959_cast_fp16 = add(x = input_947_cast_fp16, y = var_4253_cast_fp16)[name = string("input_959_cast_fp16")]; + tensor input_961_axes_0 = const()[name = string("input_961_axes_0"), val = tensor([-1])]; + tensor encoder_layers_17_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_17_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(360990464)))]; + tensor encoder_layers_17_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_17_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(360992576)))]; + tensor input_961_cast_fp16 = layer_norm(axes = input_961_axes_0, beta = encoder_layers_17_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_17_norm_out_weight_to_fp16, x = input_959_cast_fp16)[name = string("input_961_cast_fp16")]; + tensor cache_73_begin_0 = const()[name = string("cache_73_begin_0"), val = tensor([18, 0, 0, 0])]; + tensor cache_73_end_0 = const()[name = string("cache_73_end_0"), val = tensor([19, 1, 42, 1024])]; + tensor cache_73_end_mask_0 = const()[name = string("cache_73_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_73_squeeze_mask_0 = const()[name = string("cache_73_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_73_cast_fp16 = slice_by_index(begin = cache_73_begin_0, end = cache_73_end_0, end_mask = cache_73_end_mask_0, squeeze_mask = cache_73_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_73_cast_fp16")]; + tensor cache_75_begin_0 = const()[name = string("cache_75_begin_0"), val = tensor([18, 0, 0, 0])]; + tensor cache_75_end_0 = const()[name = string("cache_75_end_0"), val = tensor([19, 1, 1024, 8])]; + tensor cache_75_end_mask_0 = const()[name = string("cache_75_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_75_squeeze_mask_0 = const()[name = string("cache_75_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_75_cast_fp16 = slice_by_index(begin = cache_75_begin_0, end = cache_75_end_0, end_mask = cache_75_end_mask_0, squeeze_mask = cache_75_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_75_cast_fp16")]; + tensor input_963_axes_0 = const()[name = string("input_963_axes_0"), val = tensor([-1])]; + tensor encoder_layers_18_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_18_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(360994688)))]; + tensor encoder_layers_18_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_18_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(360996800)))]; + tensor input_963_cast_fp16 = layer_norm(axes = input_963_axes_0, beta = encoder_layers_18_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_18_norm_feed_forward1_weight_to_fp16, x = input_961_cast_fp16)[name = string("input_963_cast_fp16")]; + tensor encoder_layers_18_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(360998912))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(364144704))))[name = string("encoder_layers_18_feed_forward1_linear1_weight_to_fp16_palettized")]; + tensor encoder_layers_18_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_18_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(364144896)))]; + tensor linear_163_cast_fp16 = linear(bias = encoder_layers_18_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_18_feed_forward1_linear1_weight_to_fp16_palettized, x = input_963_cast_fp16)[name = string("linear_163_cast_fp16")]; + tensor input_967_cast_fp16 = silu(x = linear_163_cast_fp16)[name = string("input_967_cast_fp16")]; + tensor encoder_layers_18_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(364153152))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(367298944))))[name = string("encoder_layers_18_feed_forward1_linear2_weight_to_fp16_palettized")]; + tensor encoder_layers_18_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_18_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(367299136)))]; + tensor linear_164_cast_fp16 = linear(bias = encoder_layers_18_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_18_feed_forward1_linear2_weight_to_fp16_palettized, x = input_967_cast_fp16)[name = string("linear_164_cast_fp16")]; + fp16 var_4289_to_fp16 = const()[name = string("op_4289_to_fp16"), val = fp16(0x1p-1)]; + tensor var_4290_cast_fp16 = mul(x = linear_164_cast_fp16, y = var_4289_to_fp16)[name = string("op_4290_cast_fp16")]; + tensor input_973_cast_fp16 = add(x = input_961_cast_fp16, y = var_4290_cast_fp16)[name = string("input_973_cast_fp16")]; + tensor key_37_axes_0 = const()[name = string("key_37_axes_0"), val = tensor([-1])]; + tensor encoder_layers_18_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_18_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(367301248)))]; + tensor encoder_layers_18_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_18_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(367303360)))]; + tensor key_37_cast_fp16 = layer_norm(axes = key_37_axes_0, beta = encoder_layers_18_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_18_norm_self_att_weight_to_fp16, x = input_973_cast_fp16)[name = string("key_37_cast_fp16")]; + bool input_975_interleave_0 = const()[name = string("input_975_interleave_0"), val = bool(false)]; + tensor input_975_cast_fp16 = concat(axis = var_68, interleave = input_975_interleave_0, values = (cache_73_cast_fp16, key_37_cast_fp16))[name = string("input_975_cast_fp16")]; + tensor var_4312_begin_0 = const()[name = string("op_4312_begin_0"), val = tensor([0, 7, 0])]; + tensor var_4312_end_0 = const()[name = string("op_4312_end_0"), val = tensor([1, 42, 1024])]; + tensor var_4312_end_mask_0 = const()[name = string("op_4312_end_mask_0"), val = tensor([true, true, true])]; + tensor var_4312_cast_fp16 = slice_by_index(begin = var_4312_begin_0, end = var_4312_end_0, end_mask = var_4312_end_mask_0, x = cache_73_cast_fp16)[name = string("op_4312_cast_fp16")]; + bool var_4318_interleave_0 = const()[name = string("op_4318_interleave_0"), val = bool(false)]; + tensor var_4318_cast_fp16 = concat(axis = var_68, interleave = var_4318_interleave_0, values = (var_4312_cast_fp16, key_37_cast_fp16))[name = string("op_4318_cast_fp16")]; + tensor encoder_layers_18_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(367305472))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368091968))))[name = string("encoder_layers_18_self_attn_linear_q_weight_to_fp16_palettized")]; + tensor encoder_layers_18_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_18_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368092160)))]; + tensor linear_165_cast_fp16 = linear(bias = encoder_layers_18_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_18_self_attn_linear_q_weight_to_fp16_palettized, x = key_37_cast_fp16)[name = string("linear_165_cast_fp16")]; + tensor var_4323 = const()[name = string("op_4323"), val = tensor([1, -1, 8, 128])]; + tensor q_109_cast_fp16 = reshape(shape = var_4323, x = linear_165_cast_fp16)[name = string("q_109_cast_fp16")]; + tensor encoder_layers_18_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368094272))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368880768))))[name = string("encoder_layers_18_self_attn_linear_k_weight_to_fp16_palettized")]; + tensor encoder_layers_18_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_18_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368880960)))]; + tensor linear_166_cast_fp16 = linear(bias = encoder_layers_18_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_18_self_attn_linear_k_weight_to_fp16_palettized, x = input_975_cast_fp16)[name = string("linear_166_cast_fp16")]; + tensor var_4328 = const()[name = string("op_4328"), val = tensor([1, -1, 8, 128])]; + tensor k_73_cast_fp16 = reshape(shape = var_4328, x = linear_166_cast_fp16)[name = string("k_73_cast_fp16")]; + tensor encoder_layers_18_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368883072))), lut = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369669568))))[name = string("encoder_layers_18_self_attn_linear_v_weight_to_fp16_palettized")]; + tensor encoder_layers_18_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_18_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369669760)))]; + tensor linear_167_cast_fp16 = linear(bias = encoder_layers_18_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_18_self_attn_linear_v_weight_to_fp16_palettized, x = input_975_cast_fp16)[name = string("linear_167_cast_fp16")]; + tensor var_4333 = const()[name = string("op_4333"), val = tensor([1, -1, 8, 128])]; + tensor v_37_cast_fp16 = reshape(shape = var_4333, x = linear_167_cast_fp16)[name = string("v_37_cast_fp16")]; + tensor value_45_perm_0 = const()[name = string("value_45_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_18_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_18_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369671872)))]; + tensor var_4346_cast_fp16 = add(x = q_109_cast_fp16, y = encoder_layers_18_self_attn_pos_bias_u_to_fp16)[name = string("op_4346_cast_fp16")]; + tensor encoder_layers_18_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_18_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369673984)))]; + tensor var_4348_cast_fp16 = add(x = q_109_cast_fp16, y = encoder_layers_18_self_attn_pos_bias_v_to_fp16)[name = string("op_4348_cast_fp16")]; + tensor q_with_bias_v_37_perm_0 = const()[name = string("q_with_bias_v_37_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_475_transpose_x_0 = const()[name = string("x_475_transpose_x_0"), val = bool(false)]; + bool x_475_transpose_y_0 = const()[name = string("x_475_transpose_y_0"), val = bool(false)]; + tensor op_4350_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369676096))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369775488))))[name = string("op_4350_to_fp16_quantized")]; + tensor q_with_bias_v_37_cast_fp16 = transpose(perm = q_with_bias_v_37_perm_0, x = var_4348_cast_fp16)[name = string("transpose_200")]; + tensor x_475_cast_fp16 = matmul(transpose_x = x_475_transpose_x_0, transpose_y = x_475_transpose_y_0, x = q_with_bias_v_37_cast_fp16, y = op_4350_to_fp16_quantized)[name = string("x_475_cast_fp16")]; + tensor x_477_pad_0 = const()[name = string("x_477_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_477_mode_0 = const()[name = string("x_477_mode_0"), val = string("constant")]; + fp16 const_313_to_fp16 = const()[name = string("const_313_to_fp16"), val = fp16(0x0p+0)]; + tensor x_477_cast_fp16 = pad(constant_val = const_313_to_fp16, mode = x_477_mode_0, pad = x_477_pad_0, x = x_475_cast_fp16)[name = string("x_477_cast_fp16")]; + tensor var_4358 = const()[name = string("op_4358"), val = tensor([1, 8, -1, 7])]; + tensor x_479_cast_fp16 = reshape(shape = var_4358, x = x_477_cast_fp16)[name = string("x_479_cast_fp16")]; + tensor var_4362_begin_0 = const()[name = string("op_4362_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_4362_end_0 = const()[name = string("op_4362_end_0"), val = tensor([1, 8, 98, 7])]; + tensor var_4362_end_mask_0 = const()[name = string("op_4362_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_4362_cast_fp16 = slice_by_index(begin = var_4362_begin_0, end = var_4362_end_0, end_mask = var_4362_end_mask_0, x = x_479_cast_fp16)[name = string("op_4362_cast_fp16")]; + tensor var_4363 = const()[name = string("op_4363"), val = tensor([1, 8, 7, 97])]; + tensor matrix_bd_73_cast_fp16 = reshape(shape = var_4363, x = var_4362_cast_fp16)[name = string("matrix_bd_73_cast_fp16")]; + bool matrix_ac_37_transpose_x_0 = const()[name = string("matrix_ac_37_transpose_x_0"), val = bool(false)]; + bool matrix_ac_37_transpose_y_0 = const()[name = string("matrix_ac_37_transpose_y_0"), val = bool(false)]; + tensor transpose_132_perm_0 = const()[name = string("transpose_132_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_133_perm_0 = const()[name = string("transpose_133_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_133 = transpose(perm = transpose_133_perm_0, x = k_73_cast_fp16)[name = string("transpose_198")]; + tensor transpose_132 = transpose(perm = transpose_132_perm_0, x = var_4346_cast_fp16)[name = string("transpose_199")]; + tensor matrix_ac_37_cast_fp16 = matmul(transpose_x = matrix_ac_37_transpose_x_0, transpose_y = matrix_ac_37_transpose_y_0, x = transpose_132, y = transpose_133)[name = string("matrix_ac_37_cast_fp16")]; + tensor matrix_bd_75_begin_0 = const()[name = string("matrix_bd_75_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_75_end_0 = const()[name = string("matrix_bd_75_end_0"), val = tensor([1, 8, 7, 49])]; + tensor matrix_bd_75_end_mask_0 = const()[name = string("matrix_bd_75_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_75_cast_fp16 = slice_by_index(begin = matrix_bd_75_begin_0, end = matrix_bd_75_end_0, end_mask = matrix_bd_75_end_mask_0, x = matrix_bd_73_cast_fp16)[name = string("matrix_bd_75_cast_fp16")]; + tensor var_4372_cast_fp16 = add(x = matrix_ac_37_cast_fp16, y = matrix_bd_75_cast_fp16)[name = string("op_4372_cast_fp16")]; + fp16 _inversed_scores_73_y_0_to_fp16 = const()[name = string("_inversed_scores_73_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_73_cast_fp16 = mul(x = var_4372_cast_fp16, y = _inversed_scores_73_y_0_to_fp16)[name = string("_inversed_scores_73_cast_fp16")]; + tensor scores_75_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_73_cast_fp16, cond = mask_11)[name = string("scores_75_cast_fp16")]; + tensor var_4378_cast_fp16 = softmax(axis = var_59, x = scores_75_cast_fp16)[name = string("op_4378_cast_fp16")]; + tensor input_977_cast_fp16 = select(a = var_44_to_fp16, b = var_4378_cast_fp16, cond = mask_11)[name = string("input_977_cast_fp16")]; + bool x_481_transpose_x_0 = const()[name = string("x_481_transpose_x_0"), val = bool(false)]; + bool x_481_transpose_y_0 = const()[name = string("x_481_transpose_y_0"), val = bool(false)]; + tensor value_45_cast_fp16 = transpose(perm = value_45_perm_0, x = v_37_cast_fp16)[name = string("transpose_197")]; + tensor x_481_cast_fp16 = matmul(transpose_x = x_481_transpose_x_0, transpose_y = x_481_transpose_y_0, x = input_977_cast_fp16, y = value_45_cast_fp16)[name = string("x_481_cast_fp16")]; + tensor var_4382_perm_0 = const()[name = string("op_4382_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_4383 = const()[name = string("op_4383"), val = tensor([1, -1, 1024])]; + tensor var_4382_cast_fp16 = transpose(perm = var_4382_perm_0, x = x_481_cast_fp16)[name = string("transpose_196")]; + tensor input_979_cast_fp16 = reshape(shape = var_4383, x = var_4382_cast_fp16)[name = string("input_979_cast_fp16")]; + tensor encoder_layers_18_self_attn_linear_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369775808))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(370824448))))[name = string("encoder_layers_18_self_attn_linear_out_weight_to_fp16_quantized")]; + tensor encoder_layers_18_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_18_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(370826560)))]; + tensor linear_169_cast_fp16 = linear(bias = encoder_layers_18_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_18_self_attn_linear_out_weight_to_fp16_quantized, x = input_979_cast_fp16)[name = string("linear_169_cast_fp16")]; + tensor input_983_cast_fp16 = add(x = input_973_cast_fp16, y = linear_169_cast_fp16)[name = string("input_983_cast_fp16")]; + tensor x_485_axes_0 = const()[name = string("x_485_axes_0"), val = tensor([-1])]; + tensor encoder_layers_18_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_18_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(370828672)))]; + tensor encoder_layers_18_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_18_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(370830784)))]; + tensor x_485_cast_fp16 = layer_norm(axes = x_485_axes_0, beta = encoder_layers_18_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_18_norm_conv_weight_to_fp16, x = input_983_cast_fp16)[name = string("x_485_cast_fp16")]; + tensor input_985_perm_0 = const()[name = string("input_985_perm_0"), val = tensor([0, 2, 1])]; + string input_987_pad_type_0 = const()[name = string("input_987_pad_type_0"), val = string("valid")]; + tensor input_987_strides_0 = const()[name = string("input_987_strides_0"), val = tensor([1])]; + tensor input_987_pad_0 = const()[name = string("input_987_pad_0"), val = tensor([0, 0])]; + tensor input_987_dilations_0 = const()[name = string("input_987_dilations_0"), val = tensor([1])]; + int32 input_987_groups_0 = const()[name = string("input_987_groups_0"), val = int32(1)]; + tensor encoder_layers_18_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(370832896))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(372930112))))[name = string("encoder_layers_18_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_985_cast_fp16 = transpose(perm = input_985_perm_0, x = x_485_cast_fp16)[name = string("transpose_195")]; + tensor input_987_cast_fp16 = conv(dilations = input_987_dilations_0, groups = input_987_groups_0, pad = input_987_pad_0, pad_type = input_987_pad_type_0, strides = input_987_strides_0, weight = encoder_layers_18_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_985_cast_fp16)[name = string("input_987_cast_fp16")]; + int32 x_487_split_num_splits_0 = const()[name = string("x_487_split_num_splits_0"), val = int32(2)]; + int32 x_487_split_axis_0 = const()[name = string("x_487_split_axis_0"), val = int32(1)]; + tensor x_487_split_cast_fp16_0, tensor x_487_split_cast_fp16_1 = split(axis = x_487_split_axis_0, num_splits = x_487_split_num_splits_0, x = input_987_cast_fp16)[name = string("x_487_split_cast_fp16")]; + tensor x_487_split_1_sigmoid_cast_fp16 = sigmoid(x = x_487_split_cast_fp16_1)[name = string("x_487_split_1_sigmoid_cast_fp16")]; + tensor x_487_cast_fp16 = mul(x = x_487_split_cast_fp16_0, y = x_487_split_1_sigmoid_cast_fp16)[name = string("x_487_cast_fp16")]; + tensor input_989_cast_fp16 = select(a = var_44_to_fp16, b = x_487_cast_fp16, cond = var_575)[name = string("input_989_cast_fp16")]; + bool new_x_75_interleave_0 = const()[name = string("new_x_75_interleave_0"), val = bool(false)]; + tensor new_x_75_cast_fp16 = concat(axis = var_59, interleave = new_x_75_interleave_0, values = (cache_75_cast_fp16, input_989_cast_fp16))[name = string("new_x_75_cast_fp16")]; + tensor var_4422_begin_0 = const()[name = string("op_4422_begin_0"), val = tensor([0, 0, 7])]; + tensor var_4422_end_0 = const()[name = string("op_4422_end_0"), val = tensor([1, 1024, 15])]; + tensor var_4422_end_mask_0 = const()[name = string("op_4422_end_mask_0"), val = tensor([true, true, true])]; + tensor var_4422_cast_fp16 = slice_by_index(begin = var_4422_begin_0, end = var_4422_end_0, end_mask = var_4422_end_mask_0, x = new_x_75_cast_fp16)[name = string("op_4422_cast_fp16")]; + string x_489_pad_type_0 = const()[name = string("x_489_pad_type_0"), val = string("valid")]; + int32 x_489_groups_0 = const()[name = string("x_489_groups_0"), val = int32(1024)]; + tensor x_489_strides_0 = const()[name = string("x_489_strides_0"), val = tensor([1])]; + tensor x_489_pad_0 = const()[name = string("x_489_pad_0"), val = tensor([0, 0])]; + tensor x_489_dilations_0 = const()[name = string("x_489_dilations_0"), val = tensor([1])]; + tensor encoder_layers_18_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(372934272))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(372943552))))[name = string("encoder_layers_18_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_489_cast_fp16 = conv(dilations = x_489_dilations_0, groups = x_489_groups_0, pad = x_489_pad_0, pad_type = x_489_pad_type_0, strides = x_489_strides_0, weight = encoder_layers_18_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_75_cast_fp16)[name = string("x_489_cast_fp16")]; + tensor input_991_perm_0 = const()[name = string("input_991_perm_0"), val = tensor([0, 2, 1])]; + tensor x_491_axes_0 = const()[name = string("x_491_axes_0"), val = tensor([-1])]; + tensor encoder_layers_18_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_18_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(372945664)))]; + tensor encoder_layers_18_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_18_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(372947776)))]; + tensor input_991_cast_fp16 = transpose(perm = input_991_perm_0, x = x_489_cast_fp16)[name = string("transpose_194")]; + tensor x_491_cast_fp16 = layer_norm(axes = x_491_axes_0, beta = encoder_layers_18_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_18_conv_batch_norm_weight_to_fp16, x = input_991_cast_fp16)[name = string("x_491_cast_fp16")]; + tensor input_993_perm_0 = const()[name = string("input_993_perm_0"), val = tensor([0, 2, 1])]; + tensor input_993_cast_fp16 = transpose(perm = input_993_perm_0, x = x_491_cast_fp16)[name = string("transpose_193")]; + tensor input_995_cast_fp16 = silu(x = input_993_cast_fp16)[name = string("input_995_cast_fp16")]; + string x_493_pad_type_0 = const()[name = string("x_493_pad_type_0"), val = string("valid")]; + tensor x_493_strides_0 = const()[name = string("x_493_strides_0"), val = tensor([1])]; + tensor x_493_pad_0 = const()[name = string("x_493_pad_0"), val = tensor([0, 0])]; + tensor x_493_dilations_0 = const()[name = string("x_493_dilations_0"), val = tensor([1])]; + int32 x_493_groups_0 = const()[name = string("x_493_groups_0"), val = int32(1)]; + tensor encoder_layers_18_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(372949888))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(373998528))))[name = string("encoder_layers_18_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_493_cast_fp16 = conv(dilations = x_493_dilations_0, groups = x_493_groups_0, pad = x_493_pad_0, pad_type = x_493_pad_type_0, strides = x_493_strides_0, weight = encoder_layers_18_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_995_cast_fp16)[name = string("x_493_cast_fp16")]; + tensor input_997_perm_0 = const()[name = string("input_997_perm_0"), val = tensor([0, 2, 1])]; + tensor input_997_cast_fp16 = transpose(perm = input_997_perm_0, x = x_493_cast_fp16)[name = string("transpose_192")]; + tensor input_999_cast_fp16 = add(x = input_983_cast_fp16, y = input_997_cast_fp16)[name = string("input_999_cast_fp16")]; + tensor input_1001_axes_0 = const()[name = string("input_1001_axes_0"), val = tensor([-1])]; + tensor encoder_layers_18_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_18_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(374000640)))]; + tensor encoder_layers_18_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_18_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(374002752)))]; + tensor input_1001_cast_fp16 = layer_norm(axes = input_1001_axes_0, beta = encoder_layers_18_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_18_norm_feed_forward2_weight_to_fp16, x = input_999_cast_fp16)[name = string("input_1001_cast_fp16")]; + tensor encoder_layers_18_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(374004864))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(378199232))))[name = string("encoder_layers_18_feed_forward2_linear1_weight_to_fp16_quantized")]; + tensor encoder_layers_18_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_18_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(378207488)))]; + tensor linear_170_cast_fp16 = linear(bias = encoder_layers_18_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_18_feed_forward2_linear1_weight_to_fp16_quantized, x = input_1001_cast_fp16)[name = string("linear_170_cast_fp16")]; + tensor input_1005_cast_fp16 = silu(x = linear_170_cast_fp16)[name = string("input_1005_cast_fp16")]; + tensor encoder_layers_18_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(378215744))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(382410112))))[name = string("encoder_layers_18_feed_forward2_linear2_weight_to_fp16_quantized")]; + tensor encoder_layers_18_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_18_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(382412224)))]; + tensor linear_171_cast_fp16 = linear(bias = encoder_layers_18_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_18_feed_forward2_linear2_weight_to_fp16_quantized, x = input_1005_cast_fp16)[name = string("linear_171_cast_fp16")]; + fp16 var_4465_to_fp16 = const()[name = string("op_4465_to_fp16"), val = fp16(0x1p-1)]; + tensor var_4466_cast_fp16 = mul(x = linear_171_cast_fp16, y = var_4465_to_fp16)[name = string("op_4466_cast_fp16")]; + tensor input_1011_cast_fp16 = add(x = input_999_cast_fp16, y = var_4466_cast_fp16)[name = string("input_1011_cast_fp16")]; + tensor input_1013_axes_0 = const()[name = string("input_1013_axes_0"), val = tensor([-1])]; + tensor encoder_layers_18_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_18_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(382414336)))]; + tensor encoder_layers_18_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_18_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(382416448)))]; + tensor input_1013_cast_fp16 = layer_norm(axes = input_1013_axes_0, beta = encoder_layers_18_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_18_norm_out_weight_to_fp16, x = input_1011_cast_fp16)[name = string("input_1013_cast_fp16")]; + tensor cache_77_begin_0 = const()[name = string("cache_77_begin_0"), val = tensor([19, 0, 0, 0])]; + tensor cache_77_end_0 = const()[name = string("cache_77_end_0"), val = tensor([20, 1, 42, 1024])]; + tensor cache_77_end_mask_0 = const()[name = string("cache_77_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_77_squeeze_mask_0 = const()[name = string("cache_77_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_77_cast_fp16 = slice_by_index(begin = cache_77_begin_0, end = cache_77_end_0, end_mask = cache_77_end_mask_0, squeeze_mask = cache_77_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_77_cast_fp16")]; + tensor cache_79_begin_0 = const()[name = string("cache_79_begin_0"), val = tensor([19, 0, 0, 0])]; + tensor cache_79_end_0 = const()[name = string("cache_79_end_0"), val = tensor([20, 1, 1024, 8])]; + tensor cache_79_end_mask_0 = const()[name = string("cache_79_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_79_squeeze_mask_0 = const()[name = string("cache_79_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_79_cast_fp16 = slice_by_index(begin = cache_79_begin_0, end = cache_79_end_0, end_mask = cache_79_end_mask_0, squeeze_mask = cache_79_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_79_cast_fp16")]; + tensor input_1015_axes_0 = const()[name = string("input_1015_axes_0"), val = tensor([-1])]; + tensor encoder_layers_19_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_19_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(382418560)))]; + tensor encoder_layers_19_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_19_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(382420672)))]; + tensor input_1015_cast_fp16 = layer_norm(axes = input_1015_axes_0, beta = encoder_layers_19_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_19_norm_feed_forward1_weight_to_fp16, x = input_1013_cast_fp16)[name = string("input_1015_cast_fp16")]; + tensor encoder_layers_19_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(382422784))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(386617152))))[name = string("encoder_layers_19_feed_forward1_linear1_weight_to_fp16_quantized")]; + tensor encoder_layers_19_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_19_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(386625408)))]; + tensor linear_172_cast_fp16 = linear(bias = encoder_layers_19_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_19_feed_forward1_linear1_weight_to_fp16_quantized, x = input_1015_cast_fp16)[name = string("linear_172_cast_fp16")]; + tensor input_1019_cast_fp16 = silu(x = linear_172_cast_fp16)[name = string("input_1019_cast_fp16")]; + tensor encoder_layers_19_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(386633664))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(390828032))))[name = string("encoder_layers_19_feed_forward1_linear2_weight_to_fp16_quantized")]; + tensor encoder_layers_19_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_19_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(390830144)))]; + tensor linear_173_cast_fp16 = linear(bias = encoder_layers_19_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_19_feed_forward1_linear2_weight_to_fp16_quantized, x = input_1019_cast_fp16)[name = string("linear_173_cast_fp16")]; + fp16 var_4502_to_fp16 = const()[name = string("op_4502_to_fp16"), val = fp16(0x1p-1)]; + tensor var_4503_cast_fp16 = mul(x = linear_173_cast_fp16, y = var_4502_to_fp16)[name = string("op_4503_cast_fp16")]; + tensor input_1025_cast_fp16 = add(x = input_1013_cast_fp16, y = var_4503_cast_fp16)[name = string("input_1025_cast_fp16")]; + tensor key_39_axes_0 = const()[name = string("key_39_axes_0"), val = tensor([-1])]; + tensor encoder_layers_19_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_19_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(390832256)))]; + tensor encoder_layers_19_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_19_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(390834368)))]; + tensor key_39_cast_fp16 = layer_norm(axes = key_39_axes_0, beta = encoder_layers_19_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_19_norm_self_att_weight_to_fp16, x = input_1025_cast_fp16)[name = string("key_39_cast_fp16")]; + bool input_1027_interleave_0 = const()[name = string("input_1027_interleave_0"), val = bool(false)]; + tensor input_1027_cast_fp16 = concat(axis = var_68, interleave = input_1027_interleave_0, values = (cache_77_cast_fp16, key_39_cast_fp16))[name = string("input_1027_cast_fp16")]; + tensor var_4525_begin_0 = const()[name = string("op_4525_begin_0"), val = tensor([0, 7, 0])]; + tensor var_4525_end_0 = const()[name = string("op_4525_end_0"), val = tensor([1, 42, 1024])]; + tensor var_4525_end_mask_0 = const()[name = string("op_4525_end_mask_0"), val = tensor([true, true, true])]; + tensor var_4525_cast_fp16 = slice_by_index(begin = var_4525_begin_0, end = var_4525_end_0, end_mask = var_4525_end_mask_0, x = cache_77_cast_fp16)[name = string("op_4525_cast_fp16")]; + bool var_4531_interleave_0 = const()[name = string("op_4531_interleave_0"), val = bool(false)]; + tensor var_4531_cast_fp16 = concat(axis = var_68, interleave = var_4531_interleave_0, values = (var_4525_cast_fp16, key_39_cast_fp16))[name = string("op_4531_cast_fp16")]; + tensor encoder_layers_19_self_attn_linear_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(390836480))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391885120))))[name = string("encoder_layers_19_self_attn_linear_q_weight_to_fp16_quantized")]; + tensor encoder_layers_19_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_19_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391887232)))]; + tensor linear_174_cast_fp16 = linear(bias = encoder_layers_19_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_19_self_attn_linear_q_weight_to_fp16_quantized, x = key_39_cast_fp16)[name = string("linear_174_cast_fp16")]; + tensor var_4536 = const()[name = string("op_4536"), val = tensor([1, -1, 8, 128])]; + tensor q_115_cast_fp16 = reshape(shape = var_4536, x = linear_174_cast_fp16)[name = string("q_115_cast_fp16")]; + tensor encoder_layers_19_self_attn_linear_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391889344))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(392937984))))[name = string("encoder_layers_19_self_attn_linear_k_weight_to_fp16_quantized")]; + tensor encoder_layers_19_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_19_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(392940096)))]; + tensor linear_175_cast_fp16 = linear(bias = encoder_layers_19_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_19_self_attn_linear_k_weight_to_fp16_quantized, x = input_1027_cast_fp16)[name = string("linear_175_cast_fp16")]; + tensor var_4541 = const()[name = string("op_4541"), val = tensor([1, -1, 8, 128])]; + tensor k_77_cast_fp16 = reshape(shape = var_4541, x = linear_175_cast_fp16)[name = string("k_77_cast_fp16")]; + tensor encoder_layers_19_self_attn_linear_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(392942208))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(393990848))))[name = string("encoder_layers_19_self_attn_linear_v_weight_to_fp16_quantized")]; + tensor encoder_layers_19_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_19_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(393992960)))]; + tensor linear_176_cast_fp16 = linear(bias = encoder_layers_19_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_19_self_attn_linear_v_weight_to_fp16_quantized, x = input_1027_cast_fp16)[name = string("linear_176_cast_fp16")]; + tensor var_4546 = const()[name = string("op_4546"), val = tensor([1, -1, 8, 128])]; + tensor v_39_cast_fp16 = reshape(shape = var_4546, x = linear_176_cast_fp16)[name = string("v_39_cast_fp16")]; + tensor value_47_perm_0 = const()[name = string("value_47_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_19_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_19_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(393995072)))]; + tensor var_4559_cast_fp16 = add(x = q_115_cast_fp16, y = encoder_layers_19_self_attn_pos_bias_u_to_fp16)[name = string("op_4559_cast_fp16")]; + tensor encoder_layers_19_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_19_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(393997184)))]; + tensor var_4561_cast_fp16 = add(x = q_115_cast_fp16, y = encoder_layers_19_self_attn_pos_bias_v_to_fp16)[name = string("op_4561_cast_fp16")]; + tensor q_with_bias_v_39_perm_0 = const()[name = string("q_with_bias_v_39_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_501_transpose_x_0 = const()[name = string("x_501_transpose_x_0"), val = bool(false)]; + bool x_501_transpose_y_0 = const()[name = string("x_501_transpose_y_0"), val = bool(false)]; + tensor op_4563_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(393999296))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(394098688))))[name = string("op_4563_to_fp16_quantized")]; + tensor q_with_bias_v_39_cast_fp16 = transpose(perm = q_with_bias_v_39_perm_0, x = var_4561_cast_fp16)[name = string("transpose_191")]; + tensor x_501_cast_fp16 = matmul(transpose_x = x_501_transpose_x_0, transpose_y = x_501_transpose_y_0, x = q_with_bias_v_39_cast_fp16, y = op_4563_to_fp16_quantized)[name = string("x_501_cast_fp16")]; + tensor x_503_pad_0 = const()[name = string("x_503_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_503_mode_0 = const()[name = string("x_503_mode_0"), val = string("constant")]; + fp16 const_326_to_fp16 = const()[name = string("const_326_to_fp16"), val = fp16(0x0p+0)]; + tensor x_503_cast_fp16 = pad(constant_val = const_326_to_fp16, mode = x_503_mode_0, pad = x_503_pad_0, x = x_501_cast_fp16)[name = string("x_503_cast_fp16")]; + tensor var_4571 = const()[name = string("op_4571"), val = tensor([1, 8, -1, 7])]; + tensor x_505_cast_fp16 = reshape(shape = var_4571, x = x_503_cast_fp16)[name = string("x_505_cast_fp16")]; + tensor var_4575_begin_0 = const()[name = string("op_4575_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_4575_end_0 = const()[name = string("op_4575_end_0"), val = tensor([1, 8, 98, 7])]; + tensor var_4575_end_mask_0 = const()[name = string("op_4575_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_4575_cast_fp16 = slice_by_index(begin = var_4575_begin_0, end = var_4575_end_0, end_mask = var_4575_end_mask_0, x = x_505_cast_fp16)[name = string("op_4575_cast_fp16")]; + tensor var_4576 = const()[name = string("op_4576"), val = tensor([1, 8, 7, 97])]; + tensor matrix_bd_77_cast_fp16 = reshape(shape = var_4576, x = var_4575_cast_fp16)[name = string("matrix_bd_77_cast_fp16")]; + bool matrix_ac_39_transpose_x_0 = const()[name = string("matrix_ac_39_transpose_x_0"), val = bool(false)]; + bool matrix_ac_39_transpose_y_0 = const()[name = string("matrix_ac_39_transpose_y_0"), val = bool(false)]; + tensor transpose_134_perm_0 = const()[name = string("transpose_134_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_135_perm_0 = const()[name = string("transpose_135_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_135 = transpose(perm = transpose_135_perm_0, x = k_77_cast_fp16)[name = string("transpose_189")]; + tensor transpose_134 = transpose(perm = transpose_134_perm_0, x = var_4559_cast_fp16)[name = string("transpose_190")]; + tensor matrix_ac_39_cast_fp16 = matmul(transpose_x = matrix_ac_39_transpose_x_0, transpose_y = matrix_ac_39_transpose_y_0, x = transpose_134, y = transpose_135)[name = string("matrix_ac_39_cast_fp16")]; + tensor matrix_bd_79_begin_0 = const()[name = string("matrix_bd_79_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_79_end_0 = const()[name = string("matrix_bd_79_end_0"), val = tensor([1, 8, 7, 49])]; + tensor matrix_bd_79_end_mask_0 = const()[name = string("matrix_bd_79_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_79_cast_fp16 = slice_by_index(begin = matrix_bd_79_begin_0, end = matrix_bd_79_end_0, end_mask = matrix_bd_79_end_mask_0, x = matrix_bd_77_cast_fp16)[name = string("matrix_bd_79_cast_fp16")]; + tensor var_4585_cast_fp16 = add(x = matrix_ac_39_cast_fp16, y = matrix_bd_79_cast_fp16)[name = string("op_4585_cast_fp16")]; + fp16 _inversed_scores_77_y_0_to_fp16 = const()[name = string("_inversed_scores_77_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_77_cast_fp16 = mul(x = var_4585_cast_fp16, y = _inversed_scores_77_y_0_to_fp16)[name = string("_inversed_scores_77_cast_fp16")]; + tensor scores_79_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_77_cast_fp16, cond = mask_11)[name = string("scores_79_cast_fp16")]; + tensor var_4591_cast_fp16 = softmax(axis = var_59, x = scores_79_cast_fp16)[name = string("op_4591_cast_fp16")]; + tensor input_1029_cast_fp16 = select(a = var_44_to_fp16, b = var_4591_cast_fp16, cond = mask_11)[name = string("input_1029_cast_fp16")]; + bool x_507_transpose_x_0 = const()[name = string("x_507_transpose_x_0"), val = bool(false)]; + bool x_507_transpose_y_0 = const()[name = string("x_507_transpose_y_0"), val = bool(false)]; + tensor value_47_cast_fp16 = transpose(perm = value_47_perm_0, x = v_39_cast_fp16)[name = string("transpose_188")]; + tensor x_507_cast_fp16 = matmul(transpose_x = x_507_transpose_x_0, transpose_y = x_507_transpose_y_0, x = input_1029_cast_fp16, y = value_47_cast_fp16)[name = string("x_507_cast_fp16")]; + tensor var_4595_perm_0 = const()[name = string("op_4595_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_4596 = const()[name = string("op_4596"), val = tensor([1, -1, 1024])]; + tensor var_4595_cast_fp16 = transpose(perm = var_4595_perm_0, x = x_507_cast_fp16)[name = string("transpose_187")]; + tensor input_1031_cast_fp16 = reshape(shape = var_4596, x = var_4595_cast_fp16)[name = string("input_1031_cast_fp16")]; + tensor encoder_layers_19_self_attn_linear_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(394099008))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(395147648))))[name = string("encoder_layers_19_self_attn_linear_out_weight_to_fp16_quantized")]; + tensor encoder_layers_19_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_19_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(395149760)))]; + tensor linear_178_cast_fp16 = linear(bias = encoder_layers_19_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_19_self_attn_linear_out_weight_to_fp16_quantized, x = input_1031_cast_fp16)[name = string("linear_178_cast_fp16")]; + tensor input_1035_cast_fp16 = add(x = input_1025_cast_fp16, y = linear_178_cast_fp16)[name = string("input_1035_cast_fp16")]; + tensor x_511_axes_0 = const()[name = string("x_511_axes_0"), val = tensor([-1])]; + tensor encoder_layers_19_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_19_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(395151872)))]; + tensor encoder_layers_19_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_19_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(395153984)))]; + tensor x_511_cast_fp16 = layer_norm(axes = x_511_axes_0, beta = encoder_layers_19_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_19_norm_conv_weight_to_fp16, x = input_1035_cast_fp16)[name = string("x_511_cast_fp16")]; + tensor input_1037_perm_0 = const()[name = string("input_1037_perm_0"), val = tensor([0, 2, 1])]; + string input_1039_pad_type_0 = const()[name = string("input_1039_pad_type_0"), val = string("valid")]; + tensor input_1039_strides_0 = const()[name = string("input_1039_strides_0"), val = tensor([1])]; + tensor input_1039_pad_0 = const()[name = string("input_1039_pad_0"), val = tensor([0, 0])]; + tensor input_1039_dilations_0 = const()[name = string("input_1039_dilations_0"), val = tensor([1])]; + int32 input_1039_groups_0 = const()[name = string("input_1039_groups_0"), val = int32(1)]; + tensor encoder_layers_19_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(395156096))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397253312))))[name = string("encoder_layers_19_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_1037_cast_fp16 = transpose(perm = input_1037_perm_0, x = x_511_cast_fp16)[name = string("transpose_186")]; + tensor input_1039_cast_fp16 = conv(dilations = input_1039_dilations_0, groups = input_1039_groups_0, pad = input_1039_pad_0, pad_type = input_1039_pad_type_0, strides = input_1039_strides_0, weight = encoder_layers_19_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_1037_cast_fp16)[name = string("input_1039_cast_fp16")]; + int32 x_513_split_num_splits_0 = const()[name = string("x_513_split_num_splits_0"), val = int32(2)]; + int32 x_513_split_axis_0 = const()[name = string("x_513_split_axis_0"), val = int32(1)]; + tensor x_513_split_cast_fp16_0, tensor x_513_split_cast_fp16_1 = split(axis = x_513_split_axis_0, num_splits = x_513_split_num_splits_0, x = input_1039_cast_fp16)[name = string("x_513_split_cast_fp16")]; + tensor x_513_split_1_sigmoid_cast_fp16 = sigmoid(x = x_513_split_cast_fp16_1)[name = string("x_513_split_1_sigmoid_cast_fp16")]; + tensor x_513_cast_fp16 = mul(x = x_513_split_cast_fp16_0, y = x_513_split_1_sigmoid_cast_fp16)[name = string("x_513_cast_fp16")]; + tensor input_1041_cast_fp16 = select(a = var_44_to_fp16, b = x_513_cast_fp16, cond = var_575)[name = string("input_1041_cast_fp16")]; + bool new_x_79_interleave_0 = const()[name = string("new_x_79_interleave_0"), val = bool(false)]; + tensor new_x_79_cast_fp16 = concat(axis = var_59, interleave = new_x_79_interleave_0, values = (cache_79_cast_fp16, input_1041_cast_fp16))[name = string("new_x_79_cast_fp16")]; + tensor var_4635_begin_0 = const()[name = string("op_4635_begin_0"), val = tensor([0, 0, 7])]; + tensor var_4635_end_0 = const()[name = string("op_4635_end_0"), val = tensor([1, 1024, 15])]; + tensor var_4635_end_mask_0 = const()[name = string("op_4635_end_mask_0"), val = tensor([true, true, true])]; + tensor var_4635_cast_fp16 = slice_by_index(begin = var_4635_begin_0, end = var_4635_end_0, end_mask = var_4635_end_mask_0, x = new_x_79_cast_fp16)[name = string("op_4635_cast_fp16")]; + string x_515_pad_type_0 = const()[name = string("x_515_pad_type_0"), val = string("valid")]; + int32 x_515_groups_0 = const()[name = string("x_515_groups_0"), val = int32(1024)]; + tensor x_515_strides_0 = const()[name = string("x_515_strides_0"), val = tensor([1])]; + tensor x_515_pad_0 = const()[name = string("x_515_pad_0"), val = tensor([0, 0])]; + tensor x_515_dilations_0 = const()[name = string("x_515_dilations_0"), val = tensor([1])]; + tensor encoder_layers_19_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397257472))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397266752))))[name = string("encoder_layers_19_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_515_cast_fp16 = conv(dilations = x_515_dilations_0, groups = x_515_groups_0, pad = x_515_pad_0, pad_type = x_515_pad_type_0, strides = x_515_strides_0, weight = encoder_layers_19_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_79_cast_fp16)[name = string("x_515_cast_fp16")]; + tensor input_1043_perm_0 = const()[name = string("input_1043_perm_0"), val = tensor([0, 2, 1])]; + tensor x_517_axes_0 = const()[name = string("x_517_axes_0"), val = tensor([-1])]; + tensor encoder_layers_19_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_19_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397268864)))]; + tensor encoder_layers_19_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_19_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397270976)))]; + tensor input_1043_cast_fp16 = transpose(perm = input_1043_perm_0, x = x_515_cast_fp16)[name = string("transpose_185")]; + tensor x_517_cast_fp16 = layer_norm(axes = x_517_axes_0, beta = encoder_layers_19_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_19_conv_batch_norm_weight_to_fp16, x = input_1043_cast_fp16)[name = string("x_517_cast_fp16")]; + tensor input_1045_perm_0 = const()[name = string("input_1045_perm_0"), val = tensor([0, 2, 1])]; + tensor input_1045_cast_fp16 = transpose(perm = input_1045_perm_0, x = x_517_cast_fp16)[name = string("transpose_184")]; + tensor input_1047_cast_fp16 = silu(x = input_1045_cast_fp16)[name = string("input_1047_cast_fp16")]; + string x_519_pad_type_0 = const()[name = string("x_519_pad_type_0"), val = string("valid")]; + tensor x_519_strides_0 = const()[name = string("x_519_strides_0"), val = tensor([1])]; + tensor x_519_pad_0 = const()[name = string("x_519_pad_0"), val = tensor([0, 0])]; + tensor x_519_dilations_0 = const()[name = string("x_519_dilations_0"), val = tensor([1])]; + int32 x_519_groups_0 = const()[name = string("x_519_groups_0"), val = int32(1)]; + tensor encoder_layers_19_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397273088))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(398321728))))[name = string("encoder_layers_19_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_519_cast_fp16 = conv(dilations = x_519_dilations_0, groups = x_519_groups_0, pad = x_519_pad_0, pad_type = x_519_pad_type_0, strides = x_519_strides_0, weight = encoder_layers_19_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_1047_cast_fp16)[name = string("x_519_cast_fp16")]; + tensor input_1049_perm_0 = const()[name = string("input_1049_perm_0"), val = tensor([0, 2, 1])]; + tensor input_1049_cast_fp16 = transpose(perm = input_1049_perm_0, x = x_519_cast_fp16)[name = string("transpose_183")]; + tensor input_1051_cast_fp16 = add(x = input_1035_cast_fp16, y = input_1049_cast_fp16)[name = string("input_1051_cast_fp16")]; + tensor input_1053_axes_0 = const()[name = string("input_1053_axes_0"), val = tensor([-1])]; + tensor encoder_layers_19_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_19_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(398323840)))]; + tensor encoder_layers_19_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_19_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(398325952)))]; + tensor input_1053_cast_fp16 = layer_norm(axes = input_1053_axes_0, beta = encoder_layers_19_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_19_norm_feed_forward2_weight_to_fp16, x = input_1051_cast_fp16)[name = string("input_1053_cast_fp16")]; + tensor encoder_layers_19_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(398328064))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(402522432))))[name = string("encoder_layers_19_feed_forward2_linear1_weight_to_fp16_quantized")]; + tensor encoder_layers_19_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_19_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(402530688)))]; + tensor linear_179_cast_fp16 = linear(bias = encoder_layers_19_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_19_feed_forward2_linear1_weight_to_fp16_quantized, x = input_1053_cast_fp16)[name = string("linear_179_cast_fp16")]; + tensor input_1057_cast_fp16 = silu(x = linear_179_cast_fp16)[name = string("input_1057_cast_fp16")]; + tensor encoder_layers_19_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(402538944))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406733312))))[name = string("encoder_layers_19_feed_forward2_linear2_weight_to_fp16_quantized")]; + tensor encoder_layers_19_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_19_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406735424)))]; + tensor linear_180_cast_fp16 = linear(bias = encoder_layers_19_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_19_feed_forward2_linear2_weight_to_fp16_quantized, x = input_1057_cast_fp16)[name = string("linear_180_cast_fp16")]; + fp16 var_4678_to_fp16 = const()[name = string("op_4678_to_fp16"), val = fp16(0x1p-1)]; + tensor var_4679_cast_fp16 = mul(x = linear_180_cast_fp16, y = var_4678_to_fp16)[name = string("op_4679_cast_fp16")]; + tensor input_1063_cast_fp16 = add(x = input_1051_cast_fp16, y = var_4679_cast_fp16)[name = string("input_1063_cast_fp16")]; + tensor input_1065_axes_0 = const()[name = string("input_1065_axes_0"), val = tensor([-1])]; + tensor encoder_layers_19_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_19_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406737536)))]; + tensor encoder_layers_19_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_19_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406739648)))]; + tensor input_1065_cast_fp16 = layer_norm(axes = input_1065_axes_0, beta = encoder_layers_19_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_19_norm_out_weight_to_fp16, x = input_1063_cast_fp16)[name = string("input_1065_cast_fp16")]; + tensor cache_81_begin_0 = const()[name = string("cache_81_begin_0"), val = tensor([20, 0, 0, 0])]; + tensor cache_81_end_0 = const()[name = string("cache_81_end_0"), val = tensor([21, 1, 42, 1024])]; + tensor cache_81_end_mask_0 = const()[name = string("cache_81_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_81_squeeze_mask_0 = const()[name = string("cache_81_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_81_cast_fp16 = slice_by_index(begin = cache_81_begin_0, end = cache_81_end_0, end_mask = cache_81_end_mask_0, squeeze_mask = cache_81_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_81_cast_fp16")]; + tensor cache_83_begin_0 = const()[name = string("cache_83_begin_0"), val = tensor([20, 0, 0, 0])]; + tensor cache_83_end_0 = const()[name = string("cache_83_end_0"), val = tensor([21, 1, 1024, 8])]; + tensor cache_83_end_mask_0 = const()[name = string("cache_83_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_83_squeeze_mask_0 = const()[name = string("cache_83_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_83_cast_fp16 = slice_by_index(begin = cache_83_begin_0, end = cache_83_end_0, end_mask = cache_83_end_mask_0, squeeze_mask = cache_83_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_83_cast_fp16")]; + tensor input_1067_axes_0 = const()[name = string("input_1067_axes_0"), val = tensor([-1])]; + tensor encoder_layers_20_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_20_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406741760)))]; + tensor encoder_layers_20_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_20_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406743872)))]; + tensor input_1067_cast_fp16 = layer_norm(axes = input_1067_axes_0, beta = encoder_layers_20_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_20_norm_feed_forward1_weight_to_fp16, x = input_1065_cast_fp16)[name = string("input_1067_cast_fp16")]; + tensor encoder_layers_20_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406745984))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(410940352))))[name = string("encoder_layers_20_feed_forward1_linear1_weight_to_fp16_quantized")]; + tensor encoder_layers_20_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_20_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(410948608)))]; + tensor linear_181_cast_fp16 = linear(bias = encoder_layers_20_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_20_feed_forward1_linear1_weight_to_fp16_quantized, x = input_1067_cast_fp16)[name = string("linear_181_cast_fp16")]; + tensor input_1071_cast_fp16 = silu(x = linear_181_cast_fp16)[name = string("input_1071_cast_fp16")]; + tensor encoder_layers_20_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(410956864))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(415151232))))[name = string("encoder_layers_20_feed_forward1_linear2_weight_to_fp16_quantized")]; + tensor encoder_layers_20_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_20_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(415153344)))]; + tensor linear_182_cast_fp16 = linear(bias = encoder_layers_20_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_20_feed_forward1_linear2_weight_to_fp16_quantized, x = input_1071_cast_fp16)[name = string("linear_182_cast_fp16")]; + fp16 var_4715_to_fp16 = const()[name = string("op_4715_to_fp16"), val = fp16(0x1p-1)]; + tensor var_4716_cast_fp16 = mul(x = linear_182_cast_fp16, y = var_4715_to_fp16)[name = string("op_4716_cast_fp16")]; + tensor input_1077_cast_fp16 = add(x = input_1065_cast_fp16, y = var_4716_cast_fp16)[name = string("input_1077_cast_fp16")]; + tensor key_41_axes_0 = const()[name = string("key_41_axes_0"), val = tensor([-1])]; + tensor encoder_layers_20_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_20_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(415155456)))]; + tensor encoder_layers_20_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_20_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(415157568)))]; + tensor key_41_cast_fp16 = layer_norm(axes = key_41_axes_0, beta = encoder_layers_20_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_20_norm_self_att_weight_to_fp16, x = input_1077_cast_fp16)[name = string("key_41_cast_fp16")]; + bool input_1079_interleave_0 = const()[name = string("input_1079_interleave_0"), val = bool(false)]; + tensor input_1079_cast_fp16 = concat(axis = var_68, interleave = input_1079_interleave_0, values = (cache_81_cast_fp16, key_41_cast_fp16))[name = string("input_1079_cast_fp16")]; + tensor var_4738_begin_0 = const()[name = string("op_4738_begin_0"), val = tensor([0, 7, 0])]; + tensor var_4738_end_0 = const()[name = string("op_4738_end_0"), val = tensor([1, 42, 1024])]; + tensor var_4738_end_mask_0 = const()[name = string("op_4738_end_mask_0"), val = tensor([true, true, true])]; + tensor var_4738_cast_fp16 = slice_by_index(begin = var_4738_begin_0, end = var_4738_end_0, end_mask = var_4738_end_mask_0, x = cache_81_cast_fp16)[name = string("op_4738_cast_fp16")]; + bool var_4744_interleave_0 = const()[name = string("op_4744_interleave_0"), val = bool(false)]; + tensor var_4744_cast_fp16 = concat(axis = var_68, interleave = var_4744_interleave_0, values = (var_4738_cast_fp16, key_41_cast_fp16))[name = string("op_4744_cast_fp16")]; + tensor encoder_layers_20_self_attn_linear_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(415159680))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(416208320))))[name = string("encoder_layers_20_self_attn_linear_q_weight_to_fp16_quantized")]; + tensor encoder_layers_20_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_20_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(416210432)))]; + tensor linear_183_cast_fp16 = linear(bias = encoder_layers_20_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_20_self_attn_linear_q_weight_to_fp16_quantized, x = key_41_cast_fp16)[name = string("linear_183_cast_fp16")]; + tensor var_4749 = const()[name = string("op_4749"), val = tensor([1, -1, 8, 128])]; + tensor q_121_cast_fp16 = reshape(shape = var_4749, x = linear_183_cast_fp16)[name = string("q_121_cast_fp16")]; + tensor encoder_layers_20_self_attn_linear_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(416212544))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(417261184))))[name = string("encoder_layers_20_self_attn_linear_k_weight_to_fp16_quantized")]; + tensor encoder_layers_20_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_20_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(417263296)))]; + tensor linear_184_cast_fp16 = linear(bias = encoder_layers_20_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_20_self_attn_linear_k_weight_to_fp16_quantized, x = input_1079_cast_fp16)[name = string("linear_184_cast_fp16")]; + tensor var_4754 = const()[name = string("op_4754"), val = tensor([1, -1, 8, 128])]; + tensor k_81_cast_fp16 = reshape(shape = var_4754, x = linear_184_cast_fp16)[name = string("k_81_cast_fp16")]; + tensor encoder_layers_20_self_attn_linear_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(417265408))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(418314048))))[name = string("encoder_layers_20_self_attn_linear_v_weight_to_fp16_quantized")]; + tensor encoder_layers_20_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_20_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(418316160)))]; + tensor linear_185_cast_fp16 = linear(bias = encoder_layers_20_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_20_self_attn_linear_v_weight_to_fp16_quantized, x = input_1079_cast_fp16)[name = string("linear_185_cast_fp16")]; + tensor var_4759 = const()[name = string("op_4759"), val = tensor([1, -1, 8, 128])]; + tensor v_41_cast_fp16 = reshape(shape = var_4759, x = linear_185_cast_fp16)[name = string("v_41_cast_fp16")]; + tensor value_49_perm_0 = const()[name = string("value_49_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_20_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_20_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(418318272)))]; + tensor var_4772_cast_fp16 = add(x = q_121_cast_fp16, y = encoder_layers_20_self_attn_pos_bias_u_to_fp16)[name = string("op_4772_cast_fp16")]; + tensor encoder_layers_20_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_20_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(418320384)))]; + tensor var_4774_cast_fp16 = add(x = q_121_cast_fp16, y = encoder_layers_20_self_attn_pos_bias_v_to_fp16)[name = string("op_4774_cast_fp16")]; + tensor q_with_bias_v_41_perm_0 = const()[name = string("q_with_bias_v_41_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_527_transpose_x_0 = const()[name = string("x_527_transpose_x_0"), val = bool(false)]; + bool x_527_transpose_y_0 = const()[name = string("x_527_transpose_y_0"), val = bool(false)]; + tensor op_4776_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(418322496))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(418421888))))[name = string("op_4776_to_fp16_quantized")]; + tensor q_with_bias_v_41_cast_fp16 = transpose(perm = q_with_bias_v_41_perm_0, x = var_4774_cast_fp16)[name = string("transpose_182")]; + tensor x_527_cast_fp16 = matmul(transpose_x = x_527_transpose_x_0, transpose_y = x_527_transpose_y_0, x = q_with_bias_v_41_cast_fp16, y = op_4776_to_fp16_quantized)[name = string("x_527_cast_fp16")]; + tensor x_529_pad_0 = const()[name = string("x_529_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_529_mode_0 = const()[name = string("x_529_mode_0"), val = string("constant")]; + fp16 const_339_to_fp16 = const()[name = string("const_339_to_fp16"), val = fp16(0x0p+0)]; + tensor x_529_cast_fp16 = pad(constant_val = const_339_to_fp16, mode = x_529_mode_0, pad = x_529_pad_0, x = x_527_cast_fp16)[name = string("x_529_cast_fp16")]; + tensor var_4784 = const()[name = string("op_4784"), val = tensor([1, 8, -1, 7])]; + tensor x_531_cast_fp16 = reshape(shape = var_4784, x = x_529_cast_fp16)[name = string("x_531_cast_fp16")]; + tensor var_4788_begin_0 = const()[name = string("op_4788_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_4788_end_0 = const()[name = string("op_4788_end_0"), val = tensor([1, 8, 98, 7])]; + tensor var_4788_end_mask_0 = const()[name = string("op_4788_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_4788_cast_fp16 = slice_by_index(begin = var_4788_begin_0, end = var_4788_end_0, end_mask = var_4788_end_mask_0, x = x_531_cast_fp16)[name = string("op_4788_cast_fp16")]; + tensor var_4789 = const()[name = string("op_4789"), val = tensor([1, 8, 7, 97])]; + tensor matrix_bd_81_cast_fp16 = reshape(shape = var_4789, x = var_4788_cast_fp16)[name = string("matrix_bd_81_cast_fp16")]; + bool matrix_ac_41_transpose_x_0 = const()[name = string("matrix_ac_41_transpose_x_0"), val = bool(false)]; + bool matrix_ac_41_transpose_y_0 = const()[name = string("matrix_ac_41_transpose_y_0"), val = bool(false)]; + tensor transpose_136_perm_0 = const()[name = string("transpose_136_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_137_perm_0 = const()[name = string("transpose_137_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_137 = transpose(perm = transpose_137_perm_0, x = k_81_cast_fp16)[name = string("transpose_180")]; + tensor transpose_136 = transpose(perm = transpose_136_perm_0, x = var_4772_cast_fp16)[name = string("transpose_181")]; + tensor matrix_ac_41_cast_fp16 = matmul(transpose_x = matrix_ac_41_transpose_x_0, transpose_y = matrix_ac_41_transpose_y_0, x = transpose_136, y = transpose_137)[name = string("matrix_ac_41_cast_fp16")]; + tensor matrix_bd_83_begin_0 = const()[name = string("matrix_bd_83_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_83_end_0 = const()[name = string("matrix_bd_83_end_0"), val = tensor([1, 8, 7, 49])]; + tensor matrix_bd_83_end_mask_0 = const()[name = string("matrix_bd_83_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_83_cast_fp16 = slice_by_index(begin = matrix_bd_83_begin_0, end = matrix_bd_83_end_0, end_mask = matrix_bd_83_end_mask_0, x = matrix_bd_81_cast_fp16)[name = string("matrix_bd_83_cast_fp16")]; + tensor var_4798_cast_fp16 = add(x = matrix_ac_41_cast_fp16, y = matrix_bd_83_cast_fp16)[name = string("op_4798_cast_fp16")]; + fp16 _inversed_scores_81_y_0_to_fp16 = const()[name = string("_inversed_scores_81_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_81_cast_fp16 = mul(x = var_4798_cast_fp16, y = _inversed_scores_81_y_0_to_fp16)[name = string("_inversed_scores_81_cast_fp16")]; + tensor scores_83_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_81_cast_fp16, cond = mask_11)[name = string("scores_83_cast_fp16")]; + tensor var_4804_cast_fp16 = softmax(axis = var_59, x = scores_83_cast_fp16)[name = string("op_4804_cast_fp16")]; + tensor input_1081_cast_fp16 = select(a = var_44_to_fp16, b = var_4804_cast_fp16, cond = mask_11)[name = string("input_1081_cast_fp16")]; + bool x_533_transpose_x_0 = const()[name = string("x_533_transpose_x_0"), val = bool(false)]; + bool x_533_transpose_y_0 = const()[name = string("x_533_transpose_y_0"), val = bool(false)]; + tensor value_49_cast_fp16 = transpose(perm = value_49_perm_0, x = v_41_cast_fp16)[name = string("transpose_179")]; + tensor x_533_cast_fp16 = matmul(transpose_x = x_533_transpose_x_0, transpose_y = x_533_transpose_y_0, x = input_1081_cast_fp16, y = value_49_cast_fp16)[name = string("x_533_cast_fp16")]; + tensor var_4808_perm_0 = const()[name = string("op_4808_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_4809 = const()[name = string("op_4809"), val = tensor([1, -1, 1024])]; + tensor var_4808_cast_fp16 = transpose(perm = var_4808_perm_0, x = x_533_cast_fp16)[name = string("transpose_178")]; + tensor input_1083_cast_fp16 = reshape(shape = var_4809, x = var_4808_cast_fp16)[name = string("input_1083_cast_fp16")]; + tensor encoder_layers_20_self_attn_linear_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(418422208))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(419470848))))[name = string("encoder_layers_20_self_attn_linear_out_weight_to_fp16_quantized")]; + tensor encoder_layers_20_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_20_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(419472960)))]; + tensor linear_187_cast_fp16 = linear(bias = encoder_layers_20_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_20_self_attn_linear_out_weight_to_fp16_quantized, x = input_1083_cast_fp16)[name = string("linear_187_cast_fp16")]; + tensor input_1087_cast_fp16 = add(x = input_1077_cast_fp16, y = linear_187_cast_fp16)[name = string("input_1087_cast_fp16")]; + tensor x_537_axes_0 = const()[name = string("x_537_axes_0"), val = tensor([-1])]; + tensor encoder_layers_20_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_20_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(419475072)))]; + tensor encoder_layers_20_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_20_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(419477184)))]; + tensor x_537_cast_fp16 = layer_norm(axes = x_537_axes_0, beta = encoder_layers_20_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_20_norm_conv_weight_to_fp16, x = input_1087_cast_fp16)[name = string("x_537_cast_fp16")]; + tensor input_1089_perm_0 = const()[name = string("input_1089_perm_0"), val = tensor([0, 2, 1])]; + string input_1091_pad_type_0 = const()[name = string("input_1091_pad_type_0"), val = string("valid")]; + tensor input_1091_strides_0 = const()[name = string("input_1091_strides_0"), val = tensor([1])]; + tensor input_1091_pad_0 = const()[name = string("input_1091_pad_0"), val = tensor([0, 0])]; + tensor input_1091_dilations_0 = const()[name = string("input_1091_dilations_0"), val = tensor([1])]; + int32 input_1091_groups_0 = const()[name = string("input_1091_groups_0"), val = int32(1)]; + tensor encoder_layers_20_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(419479296))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421576512))))[name = string("encoder_layers_20_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_1089_cast_fp16 = transpose(perm = input_1089_perm_0, x = x_537_cast_fp16)[name = string("transpose_177")]; + tensor input_1091_cast_fp16 = conv(dilations = input_1091_dilations_0, groups = input_1091_groups_0, pad = input_1091_pad_0, pad_type = input_1091_pad_type_0, strides = input_1091_strides_0, weight = encoder_layers_20_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_1089_cast_fp16)[name = string("input_1091_cast_fp16")]; + int32 x_539_split_num_splits_0 = const()[name = string("x_539_split_num_splits_0"), val = int32(2)]; + int32 x_539_split_axis_0 = const()[name = string("x_539_split_axis_0"), val = int32(1)]; + tensor x_539_split_cast_fp16_0, tensor x_539_split_cast_fp16_1 = split(axis = x_539_split_axis_0, num_splits = x_539_split_num_splits_0, x = input_1091_cast_fp16)[name = string("x_539_split_cast_fp16")]; + tensor x_539_split_1_sigmoid_cast_fp16 = sigmoid(x = x_539_split_cast_fp16_1)[name = string("x_539_split_1_sigmoid_cast_fp16")]; + tensor x_539_cast_fp16 = mul(x = x_539_split_cast_fp16_0, y = x_539_split_1_sigmoid_cast_fp16)[name = string("x_539_cast_fp16")]; + tensor input_1093_cast_fp16 = select(a = var_44_to_fp16, b = x_539_cast_fp16, cond = var_575)[name = string("input_1093_cast_fp16")]; + bool new_x_83_interleave_0 = const()[name = string("new_x_83_interleave_0"), val = bool(false)]; + tensor new_x_83_cast_fp16 = concat(axis = var_59, interleave = new_x_83_interleave_0, values = (cache_83_cast_fp16, input_1093_cast_fp16))[name = string("new_x_83_cast_fp16")]; + tensor var_4848_begin_0 = const()[name = string("op_4848_begin_0"), val = tensor([0, 0, 7])]; + tensor var_4848_end_0 = const()[name = string("op_4848_end_0"), val = tensor([1, 1024, 15])]; + tensor var_4848_end_mask_0 = const()[name = string("op_4848_end_mask_0"), val = tensor([true, true, true])]; + tensor var_4848_cast_fp16 = slice_by_index(begin = var_4848_begin_0, end = var_4848_end_0, end_mask = var_4848_end_mask_0, x = new_x_83_cast_fp16)[name = string("op_4848_cast_fp16")]; + string x_541_pad_type_0 = const()[name = string("x_541_pad_type_0"), val = string("valid")]; + int32 x_541_groups_0 = const()[name = string("x_541_groups_0"), val = int32(1024)]; + tensor x_541_strides_0 = const()[name = string("x_541_strides_0"), val = tensor([1])]; + tensor x_541_pad_0 = const()[name = string("x_541_pad_0"), val = tensor([0, 0])]; + tensor x_541_dilations_0 = const()[name = string("x_541_dilations_0"), val = tensor([1])]; + tensor encoder_layers_20_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421580672))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421589952))))[name = string("encoder_layers_20_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_541_cast_fp16 = conv(dilations = x_541_dilations_0, groups = x_541_groups_0, pad = x_541_pad_0, pad_type = x_541_pad_type_0, strides = x_541_strides_0, weight = encoder_layers_20_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_83_cast_fp16)[name = string("x_541_cast_fp16")]; + tensor input_1095_perm_0 = const()[name = string("input_1095_perm_0"), val = tensor([0, 2, 1])]; + tensor x_543_axes_0 = const()[name = string("x_543_axes_0"), val = tensor([-1])]; + tensor encoder_layers_20_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_20_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421592064)))]; + tensor encoder_layers_20_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_20_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421594176)))]; + tensor input_1095_cast_fp16 = transpose(perm = input_1095_perm_0, x = x_541_cast_fp16)[name = string("transpose_176")]; + tensor x_543_cast_fp16 = layer_norm(axes = x_543_axes_0, beta = encoder_layers_20_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_20_conv_batch_norm_weight_to_fp16, x = input_1095_cast_fp16)[name = string("x_543_cast_fp16")]; + tensor input_1097_perm_0 = const()[name = string("input_1097_perm_0"), val = tensor([0, 2, 1])]; + tensor input_1097_cast_fp16 = transpose(perm = input_1097_perm_0, x = x_543_cast_fp16)[name = string("transpose_175")]; + tensor input_1099_cast_fp16 = silu(x = input_1097_cast_fp16)[name = string("input_1099_cast_fp16")]; + string x_545_pad_type_0 = const()[name = string("x_545_pad_type_0"), val = string("valid")]; + tensor x_545_strides_0 = const()[name = string("x_545_strides_0"), val = tensor([1])]; + tensor x_545_pad_0 = const()[name = string("x_545_pad_0"), val = tensor([0, 0])]; + tensor x_545_dilations_0 = const()[name = string("x_545_dilations_0"), val = tensor([1])]; + int32 x_545_groups_0 = const()[name = string("x_545_groups_0"), val = int32(1)]; + tensor encoder_layers_20_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421596288))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(422644928))))[name = string("encoder_layers_20_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_545_cast_fp16 = conv(dilations = x_545_dilations_0, groups = x_545_groups_0, pad = x_545_pad_0, pad_type = x_545_pad_type_0, strides = x_545_strides_0, weight = encoder_layers_20_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_1099_cast_fp16)[name = string("x_545_cast_fp16")]; + tensor input_1101_perm_0 = const()[name = string("input_1101_perm_0"), val = tensor([0, 2, 1])]; + tensor input_1101_cast_fp16 = transpose(perm = input_1101_perm_0, x = x_545_cast_fp16)[name = string("transpose_174")]; + tensor input_1103_cast_fp16 = add(x = input_1087_cast_fp16, y = input_1101_cast_fp16)[name = string("input_1103_cast_fp16")]; + tensor input_1105_axes_0 = const()[name = string("input_1105_axes_0"), val = tensor([-1])]; + tensor encoder_layers_20_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_20_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(422647040)))]; + tensor encoder_layers_20_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_20_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(422649152)))]; + tensor input_1105_cast_fp16 = layer_norm(axes = input_1105_axes_0, beta = encoder_layers_20_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_20_norm_feed_forward2_weight_to_fp16, x = input_1103_cast_fp16)[name = string("input_1105_cast_fp16")]; + tensor encoder_layers_20_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(422651264))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(426845632))))[name = string("encoder_layers_20_feed_forward2_linear1_weight_to_fp16_quantized")]; + tensor encoder_layers_20_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_20_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(426853888)))]; + tensor linear_188_cast_fp16 = linear(bias = encoder_layers_20_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_20_feed_forward2_linear1_weight_to_fp16_quantized, x = input_1105_cast_fp16)[name = string("linear_188_cast_fp16")]; + tensor input_1109_cast_fp16 = silu(x = linear_188_cast_fp16)[name = string("input_1109_cast_fp16")]; + tensor encoder_layers_20_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(426862144))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(431056512))))[name = string("encoder_layers_20_feed_forward2_linear2_weight_to_fp16_quantized")]; + tensor encoder_layers_20_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_20_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(431058624)))]; + tensor linear_189_cast_fp16 = linear(bias = encoder_layers_20_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_20_feed_forward2_linear2_weight_to_fp16_quantized, x = input_1109_cast_fp16)[name = string("linear_189_cast_fp16")]; + fp16 var_4891_to_fp16 = const()[name = string("op_4891_to_fp16"), val = fp16(0x1p-1)]; + tensor var_4892_cast_fp16 = mul(x = linear_189_cast_fp16, y = var_4891_to_fp16)[name = string("op_4892_cast_fp16")]; + tensor input_1115_cast_fp16 = add(x = input_1103_cast_fp16, y = var_4892_cast_fp16)[name = string("input_1115_cast_fp16")]; + tensor input_1117_axes_0 = const()[name = string("input_1117_axes_0"), val = tensor([-1])]; + tensor encoder_layers_20_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_20_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(431060736)))]; + tensor encoder_layers_20_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_20_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(431062848)))]; + tensor input_1117_cast_fp16 = layer_norm(axes = input_1117_axes_0, beta = encoder_layers_20_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_20_norm_out_weight_to_fp16, x = input_1115_cast_fp16)[name = string("input_1117_cast_fp16")]; + tensor cache_85_begin_0 = const()[name = string("cache_85_begin_0"), val = tensor([21, 0, 0, 0])]; + tensor cache_85_end_0 = const()[name = string("cache_85_end_0"), val = tensor([22, 1, 42, 1024])]; + tensor cache_85_end_mask_0 = const()[name = string("cache_85_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_85_squeeze_mask_0 = const()[name = string("cache_85_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_85_cast_fp16 = slice_by_index(begin = cache_85_begin_0, end = cache_85_end_0, end_mask = cache_85_end_mask_0, squeeze_mask = cache_85_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_85_cast_fp16")]; + tensor cache_87_begin_0 = const()[name = string("cache_87_begin_0"), val = tensor([21, 0, 0, 0])]; + tensor cache_87_end_0 = const()[name = string("cache_87_end_0"), val = tensor([22, 1, 1024, 8])]; + tensor cache_87_end_mask_0 = const()[name = string("cache_87_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_87_squeeze_mask_0 = const()[name = string("cache_87_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_87_cast_fp16 = slice_by_index(begin = cache_87_begin_0, end = cache_87_end_0, end_mask = cache_87_end_mask_0, squeeze_mask = cache_87_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_87_cast_fp16")]; + tensor input_1119_axes_0 = const()[name = string("input_1119_axes_0"), val = tensor([-1])]; + tensor encoder_layers_21_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_21_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(431064960)))]; + tensor encoder_layers_21_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_21_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(431067072)))]; + tensor input_1119_cast_fp16 = layer_norm(axes = input_1119_axes_0, beta = encoder_layers_21_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_21_norm_feed_forward1_weight_to_fp16, x = input_1117_cast_fp16)[name = string("input_1119_cast_fp16")]; + tensor encoder_layers_21_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(431069184))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(435263552))))[name = string("encoder_layers_21_feed_forward1_linear1_weight_to_fp16_quantized")]; + tensor encoder_layers_21_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_21_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(435271808)))]; + tensor linear_190_cast_fp16 = linear(bias = encoder_layers_21_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_21_feed_forward1_linear1_weight_to_fp16_quantized, x = input_1119_cast_fp16)[name = string("linear_190_cast_fp16")]; + tensor input_1123_cast_fp16 = silu(x = linear_190_cast_fp16)[name = string("input_1123_cast_fp16")]; + tensor encoder_layers_21_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(435280064))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(439474432))))[name = string("encoder_layers_21_feed_forward1_linear2_weight_to_fp16_quantized")]; + tensor encoder_layers_21_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_21_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(439476544)))]; + tensor linear_191_cast_fp16 = linear(bias = encoder_layers_21_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_21_feed_forward1_linear2_weight_to_fp16_quantized, x = input_1123_cast_fp16)[name = string("linear_191_cast_fp16")]; + fp16 var_4928_to_fp16 = const()[name = string("op_4928_to_fp16"), val = fp16(0x1p-1)]; + tensor var_4929_cast_fp16 = mul(x = linear_191_cast_fp16, y = var_4928_to_fp16)[name = string("op_4929_cast_fp16")]; + tensor input_1129_cast_fp16 = add(x = input_1117_cast_fp16, y = var_4929_cast_fp16)[name = string("input_1129_cast_fp16")]; + tensor key_43_axes_0 = const()[name = string("key_43_axes_0"), val = tensor([-1])]; + tensor encoder_layers_21_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_21_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(439478656)))]; + tensor encoder_layers_21_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_21_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(439480768)))]; + tensor key_43_cast_fp16 = layer_norm(axes = key_43_axes_0, beta = encoder_layers_21_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_21_norm_self_att_weight_to_fp16, x = input_1129_cast_fp16)[name = string("key_43_cast_fp16")]; + bool input_1131_interleave_0 = const()[name = string("input_1131_interleave_0"), val = bool(false)]; + tensor input_1131_cast_fp16 = concat(axis = var_68, interleave = input_1131_interleave_0, values = (cache_85_cast_fp16, key_43_cast_fp16))[name = string("input_1131_cast_fp16")]; + tensor var_4951_begin_0 = const()[name = string("op_4951_begin_0"), val = tensor([0, 7, 0])]; + tensor var_4951_end_0 = const()[name = string("op_4951_end_0"), val = tensor([1, 42, 1024])]; + tensor var_4951_end_mask_0 = const()[name = string("op_4951_end_mask_0"), val = tensor([true, true, true])]; + tensor var_4951_cast_fp16 = slice_by_index(begin = var_4951_begin_0, end = var_4951_end_0, end_mask = var_4951_end_mask_0, x = cache_85_cast_fp16)[name = string("op_4951_cast_fp16")]; + bool var_4957_interleave_0 = const()[name = string("op_4957_interleave_0"), val = bool(false)]; + tensor var_4957_cast_fp16 = concat(axis = var_68, interleave = var_4957_interleave_0, values = (var_4951_cast_fp16, key_43_cast_fp16))[name = string("op_4957_cast_fp16")]; + tensor encoder_layers_21_self_attn_linear_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(439482880))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(440531520))))[name = string("encoder_layers_21_self_attn_linear_q_weight_to_fp16_quantized")]; + tensor encoder_layers_21_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_21_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(440533632)))]; + tensor linear_192_cast_fp16 = linear(bias = encoder_layers_21_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_21_self_attn_linear_q_weight_to_fp16_quantized, x = key_43_cast_fp16)[name = string("linear_192_cast_fp16")]; + tensor var_4962 = const()[name = string("op_4962"), val = tensor([1, -1, 8, 128])]; + tensor q_127_cast_fp16 = reshape(shape = var_4962, x = linear_192_cast_fp16)[name = string("q_127_cast_fp16")]; + tensor encoder_layers_21_self_attn_linear_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(440535744))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(441584384))))[name = string("encoder_layers_21_self_attn_linear_k_weight_to_fp16_quantized")]; + tensor encoder_layers_21_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_21_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(441586496)))]; + tensor linear_193_cast_fp16 = linear(bias = encoder_layers_21_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_21_self_attn_linear_k_weight_to_fp16_quantized, x = input_1131_cast_fp16)[name = string("linear_193_cast_fp16")]; + tensor var_4967 = const()[name = string("op_4967"), val = tensor([1, -1, 8, 128])]; + tensor k_85_cast_fp16 = reshape(shape = var_4967, x = linear_193_cast_fp16)[name = string("k_85_cast_fp16")]; + tensor encoder_layers_21_self_attn_linear_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(441588608))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(442637248))))[name = string("encoder_layers_21_self_attn_linear_v_weight_to_fp16_quantized")]; + tensor encoder_layers_21_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_21_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(442639360)))]; + tensor linear_194_cast_fp16 = linear(bias = encoder_layers_21_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_21_self_attn_linear_v_weight_to_fp16_quantized, x = input_1131_cast_fp16)[name = string("linear_194_cast_fp16")]; + tensor var_4972 = const()[name = string("op_4972"), val = tensor([1, -1, 8, 128])]; + tensor v_43_cast_fp16 = reshape(shape = var_4972, x = linear_194_cast_fp16)[name = string("v_43_cast_fp16")]; + tensor value_51_perm_0 = const()[name = string("value_51_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_21_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_21_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(442641472)))]; + tensor var_4985_cast_fp16 = add(x = q_127_cast_fp16, y = encoder_layers_21_self_attn_pos_bias_u_to_fp16)[name = string("op_4985_cast_fp16")]; + tensor encoder_layers_21_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_21_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(442643584)))]; + tensor var_4987_cast_fp16 = add(x = q_127_cast_fp16, y = encoder_layers_21_self_attn_pos_bias_v_to_fp16)[name = string("op_4987_cast_fp16")]; + tensor q_with_bias_v_43_perm_0 = const()[name = string("q_with_bias_v_43_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_553_transpose_x_0 = const()[name = string("x_553_transpose_x_0"), val = bool(false)]; + bool x_553_transpose_y_0 = const()[name = string("x_553_transpose_y_0"), val = bool(false)]; + tensor op_4989_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(442645696))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(442745088))))[name = string("op_4989_to_fp16_quantized")]; + tensor q_with_bias_v_43_cast_fp16 = transpose(perm = q_with_bias_v_43_perm_0, x = var_4987_cast_fp16)[name = string("transpose_173")]; + tensor x_553_cast_fp16 = matmul(transpose_x = x_553_transpose_x_0, transpose_y = x_553_transpose_y_0, x = q_with_bias_v_43_cast_fp16, y = op_4989_to_fp16_quantized)[name = string("x_553_cast_fp16")]; + tensor x_555_pad_0 = const()[name = string("x_555_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_555_mode_0 = const()[name = string("x_555_mode_0"), val = string("constant")]; + fp16 const_352_to_fp16 = const()[name = string("const_352_to_fp16"), val = fp16(0x0p+0)]; + tensor x_555_cast_fp16 = pad(constant_val = const_352_to_fp16, mode = x_555_mode_0, pad = x_555_pad_0, x = x_553_cast_fp16)[name = string("x_555_cast_fp16")]; + tensor var_4997 = const()[name = string("op_4997"), val = tensor([1, 8, -1, 7])]; + tensor x_557_cast_fp16 = reshape(shape = var_4997, x = x_555_cast_fp16)[name = string("x_557_cast_fp16")]; + tensor var_5001_begin_0 = const()[name = string("op_5001_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_5001_end_0 = const()[name = string("op_5001_end_0"), val = tensor([1, 8, 98, 7])]; + tensor var_5001_end_mask_0 = const()[name = string("op_5001_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_5001_cast_fp16 = slice_by_index(begin = var_5001_begin_0, end = var_5001_end_0, end_mask = var_5001_end_mask_0, x = x_557_cast_fp16)[name = string("op_5001_cast_fp16")]; + tensor var_5002 = const()[name = string("op_5002"), val = tensor([1, 8, 7, 97])]; + tensor matrix_bd_85_cast_fp16 = reshape(shape = var_5002, x = var_5001_cast_fp16)[name = string("matrix_bd_85_cast_fp16")]; + bool matrix_ac_43_transpose_x_0 = const()[name = string("matrix_ac_43_transpose_x_0"), val = bool(false)]; + bool matrix_ac_43_transpose_y_0 = const()[name = string("matrix_ac_43_transpose_y_0"), val = bool(false)]; + tensor transpose_138_perm_0 = const()[name = string("transpose_138_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_139_perm_0 = const()[name = string("transpose_139_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_139 = transpose(perm = transpose_139_perm_0, x = k_85_cast_fp16)[name = string("transpose_171")]; + tensor transpose_138 = transpose(perm = transpose_138_perm_0, x = var_4985_cast_fp16)[name = string("transpose_172")]; + tensor matrix_ac_43_cast_fp16 = matmul(transpose_x = matrix_ac_43_transpose_x_0, transpose_y = matrix_ac_43_transpose_y_0, x = transpose_138, y = transpose_139)[name = string("matrix_ac_43_cast_fp16")]; + tensor matrix_bd_87_begin_0 = const()[name = string("matrix_bd_87_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_87_end_0 = const()[name = string("matrix_bd_87_end_0"), val = tensor([1, 8, 7, 49])]; + tensor matrix_bd_87_end_mask_0 = const()[name = string("matrix_bd_87_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_87_cast_fp16 = slice_by_index(begin = matrix_bd_87_begin_0, end = matrix_bd_87_end_0, end_mask = matrix_bd_87_end_mask_0, x = matrix_bd_85_cast_fp16)[name = string("matrix_bd_87_cast_fp16")]; + tensor var_5011_cast_fp16 = add(x = matrix_ac_43_cast_fp16, y = matrix_bd_87_cast_fp16)[name = string("op_5011_cast_fp16")]; + fp16 _inversed_scores_85_y_0_to_fp16 = const()[name = string("_inversed_scores_85_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_85_cast_fp16 = mul(x = var_5011_cast_fp16, y = _inversed_scores_85_y_0_to_fp16)[name = string("_inversed_scores_85_cast_fp16")]; + tensor scores_87_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_85_cast_fp16, cond = mask_11)[name = string("scores_87_cast_fp16")]; + tensor var_5017_cast_fp16 = softmax(axis = var_59, x = scores_87_cast_fp16)[name = string("op_5017_cast_fp16")]; + tensor input_1133_cast_fp16 = select(a = var_44_to_fp16, b = var_5017_cast_fp16, cond = mask_11)[name = string("input_1133_cast_fp16")]; + bool x_559_transpose_x_0 = const()[name = string("x_559_transpose_x_0"), val = bool(false)]; + bool x_559_transpose_y_0 = const()[name = string("x_559_transpose_y_0"), val = bool(false)]; + tensor value_51_cast_fp16 = transpose(perm = value_51_perm_0, x = v_43_cast_fp16)[name = string("transpose_170")]; + tensor x_559_cast_fp16 = matmul(transpose_x = x_559_transpose_x_0, transpose_y = x_559_transpose_y_0, x = input_1133_cast_fp16, y = value_51_cast_fp16)[name = string("x_559_cast_fp16")]; + tensor var_5021_perm_0 = const()[name = string("op_5021_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_5022 = const()[name = string("op_5022"), val = tensor([1, -1, 1024])]; + tensor var_5021_cast_fp16 = transpose(perm = var_5021_perm_0, x = x_559_cast_fp16)[name = string("transpose_169")]; + tensor input_1135_cast_fp16 = reshape(shape = var_5022, x = var_5021_cast_fp16)[name = string("input_1135_cast_fp16")]; + tensor encoder_layers_21_self_attn_linear_out_weight_to_fp16 = const()[name = string("encoder_layers_21_self_attn_linear_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(442745408)))]; + tensor encoder_layers_21_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_21_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(444842624)))]; + tensor linear_196_cast_fp16 = linear(bias = encoder_layers_21_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_21_self_attn_linear_out_weight_to_fp16, x = input_1135_cast_fp16)[name = string("linear_196_cast_fp16")]; + tensor input_1139_cast_fp16 = add(x = input_1129_cast_fp16, y = linear_196_cast_fp16)[name = string("input_1139_cast_fp16")]; + tensor x_563_axes_0 = const()[name = string("x_563_axes_0"), val = tensor([-1])]; + tensor encoder_layers_21_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_21_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(444844736)))]; + tensor encoder_layers_21_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_21_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(444846848)))]; + tensor x_563_cast_fp16 = layer_norm(axes = x_563_axes_0, beta = encoder_layers_21_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_21_norm_conv_weight_to_fp16, x = input_1139_cast_fp16)[name = string("x_563_cast_fp16")]; + tensor input_1141_perm_0 = const()[name = string("input_1141_perm_0"), val = tensor([0, 2, 1])]; + string input_1143_pad_type_0 = const()[name = string("input_1143_pad_type_0"), val = string("valid")]; + tensor input_1143_strides_0 = const()[name = string("input_1143_strides_0"), val = tensor([1])]; + tensor input_1143_pad_0 = const()[name = string("input_1143_pad_0"), val = tensor([0, 0])]; + tensor input_1143_dilations_0 = const()[name = string("input_1143_dilations_0"), val = tensor([1])]; + int32 input_1143_groups_0 = const()[name = string("input_1143_groups_0"), val = int32(1)]; + tensor encoder_layers_21_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(444848960))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(446946176))))[name = string("encoder_layers_21_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_1141_cast_fp16 = transpose(perm = input_1141_perm_0, x = x_563_cast_fp16)[name = string("transpose_168")]; + tensor input_1143_cast_fp16 = conv(dilations = input_1143_dilations_0, groups = input_1143_groups_0, pad = input_1143_pad_0, pad_type = input_1143_pad_type_0, strides = input_1143_strides_0, weight = encoder_layers_21_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_1141_cast_fp16)[name = string("input_1143_cast_fp16")]; + int32 x_565_split_num_splits_0 = const()[name = string("x_565_split_num_splits_0"), val = int32(2)]; + int32 x_565_split_axis_0 = const()[name = string("x_565_split_axis_0"), val = int32(1)]; + tensor x_565_split_cast_fp16_0, tensor x_565_split_cast_fp16_1 = split(axis = x_565_split_axis_0, num_splits = x_565_split_num_splits_0, x = input_1143_cast_fp16)[name = string("x_565_split_cast_fp16")]; + tensor x_565_split_1_sigmoid_cast_fp16 = sigmoid(x = x_565_split_cast_fp16_1)[name = string("x_565_split_1_sigmoid_cast_fp16")]; + tensor x_565_cast_fp16 = mul(x = x_565_split_cast_fp16_0, y = x_565_split_1_sigmoid_cast_fp16)[name = string("x_565_cast_fp16")]; + tensor input_1145_cast_fp16 = select(a = var_44_to_fp16, b = x_565_cast_fp16, cond = var_575)[name = string("input_1145_cast_fp16")]; + bool new_x_87_interleave_0 = const()[name = string("new_x_87_interleave_0"), val = bool(false)]; + tensor new_x_87_cast_fp16 = concat(axis = var_59, interleave = new_x_87_interleave_0, values = (cache_87_cast_fp16, input_1145_cast_fp16))[name = string("new_x_87_cast_fp16")]; + tensor var_5061_begin_0 = const()[name = string("op_5061_begin_0"), val = tensor([0, 0, 7])]; + tensor var_5061_end_0 = const()[name = string("op_5061_end_0"), val = tensor([1, 1024, 15])]; + tensor var_5061_end_mask_0 = const()[name = string("op_5061_end_mask_0"), val = tensor([true, true, true])]; + tensor var_5061_cast_fp16 = slice_by_index(begin = var_5061_begin_0, end = var_5061_end_0, end_mask = var_5061_end_mask_0, x = new_x_87_cast_fp16)[name = string("op_5061_cast_fp16")]; + string x_567_pad_type_0 = const()[name = string("x_567_pad_type_0"), val = string("valid")]; + int32 x_567_groups_0 = const()[name = string("x_567_groups_0"), val = int32(1024)]; + tensor x_567_strides_0 = const()[name = string("x_567_strides_0"), val = tensor([1])]; + tensor x_567_pad_0 = const()[name = string("x_567_pad_0"), val = tensor([0, 0])]; + tensor x_567_dilations_0 = const()[name = string("x_567_dilations_0"), val = tensor([1])]; + tensor encoder_layers_21_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(446950336))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(446959616))))[name = string("encoder_layers_21_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_567_cast_fp16 = conv(dilations = x_567_dilations_0, groups = x_567_groups_0, pad = x_567_pad_0, pad_type = x_567_pad_type_0, strides = x_567_strides_0, weight = encoder_layers_21_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_87_cast_fp16)[name = string("x_567_cast_fp16")]; + tensor input_1147_perm_0 = const()[name = string("input_1147_perm_0"), val = tensor([0, 2, 1])]; + tensor x_569_axes_0 = const()[name = string("x_569_axes_0"), val = tensor([-1])]; + tensor encoder_layers_21_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_21_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(446961728)))]; + tensor encoder_layers_21_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_21_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(446963840)))]; + tensor input_1147_cast_fp16 = transpose(perm = input_1147_perm_0, x = x_567_cast_fp16)[name = string("transpose_167")]; + tensor x_569_cast_fp16 = layer_norm(axes = x_569_axes_0, beta = encoder_layers_21_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_21_conv_batch_norm_weight_to_fp16, x = input_1147_cast_fp16)[name = string("x_569_cast_fp16")]; + tensor input_1149_perm_0 = const()[name = string("input_1149_perm_0"), val = tensor([0, 2, 1])]; + tensor input_1149_cast_fp16 = transpose(perm = input_1149_perm_0, x = x_569_cast_fp16)[name = string("transpose_166")]; + tensor input_1151_cast_fp16 = silu(x = input_1149_cast_fp16)[name = string("input_1151_cast_fp16")]; + string x_571_pad_type_0 = const()[name = string("x_571_pad_type_0"), val = string("valid")]; + tensor x_571_strides_0 = const()[name = string("x_571_strides_0"), val = tensor([1])]; + tensor x_571_pad_0 = const()[name = string("x_571_pad_0"), val = tensor([0, 0])]; + tensor x_571_dilations_0 = const()[name = string("x_571_dilations_0"), val = tensor([1])]; + int32 x_571_groups_0 = const()[name = string("x_571_groups_0"), val = int32(1)]; + tensor encoder_layers_21_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(446965952))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(448014592))))[name = string("encoder_layers_21_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_571_cast_fp16 = conv(dilations = x_571_dilations_0, groups = x_571_groups_0, pad = x_571_pad_0, pad_type = x_571_pad_type_0, strides = x_571_strides_0, weight = encoder_layers_21_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_1151_cast_fp16)[name = string("x_571_cast_fp16")]; + tensor input_1153_perm_0 = const()[name = string("input_1153_perm_0"), val = tensor([0, 2, 1])]; + tensor input_1153_cast_fp16 = transpose(perm = input_1153_perm_0, x = x_571_cast_fp16)[name = string("transpose_165")]; + tensor input_1155_cast_fp16 = add(x = input_1139_cast_fp16, y = input_1153_cast_fp16)[name = string("input_1155_cast_fp16")]; + tensor input_1157_axes_0 = const()[name = string("input_1157_axes_0"), val = tensor([-1])]; + tensor encoder_layers_21_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_21_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(448016704)))]; + tensor encoder_layers_21_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_21_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(448018816)))]; + tensor input_1157_cast_fp16 = layer_norm(axes = input_1157_axes_0, beta = encoder_layers_21_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_21_norm_feed_forward2_weight_to_fp16, x = input_1155_cast_fp16)[name = string("input_1157_cast_fp16")]; + tensor encoder_layers_21_feed_forward2_linear1_weight_to_fp16 = const()[name = string("encoder_layers_21_feed_forward2_linear1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(448020928)))]; + tensor encoder_layers_21_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_21_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(456409600)))]; + tensor linear_197_cast_fp16 = linear(bias = encoder_layers_21_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_21_feed_forward2_linear1_weight_to_fp16, x = input_1157_cast_fp16)[name = string("linear_197_cast_fp16")]; + tensor input_1161_cast_fp16 = silu(x = linear_197_cast_fp16)[name = string("input_1161_cast_fp16")]; + tensor encoder_layers_21_feed_forward2_linear2_weight_to_fp16 = const()[name = string("encoder_layers_21_feed_forward2_linear2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(456417856)))]; + tensor encoder_layers_21_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_21_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(464806528)))]; + tensor linear_198_cast_fp16 = linear(bias = encoder_layers_21_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_21_feed_forward2_linear2_weight_to_fp16, x = input_1161_cast_fp16)[name = string("linear_198_cast_fp16")]; + fp16 var_5104_to_fp16 = const()[name = string("op_5104_to_fp16"), val = fp16(0x1p-1)]; + tensor var_5105_cast_fp16 = mul(x = linear_198_cast_fp16, y = var_5104_to_fp16)[name = string("op_5105_cast_fp16")]; + tensor input_1167_cast_fp16 = add(x = input_1155_cast_fp16, y = var_5105_cast_fp16)[name = string("input_1167_cast_fp16")]; + tensor input_1169_axes_0 = const()[name = string("input_1169_axes_0"), val = tensor([-1])]; + tensor encoder_layers_21_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_21_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(464808640)))]; + tensor encoder_layers_21_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_21_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(464810752)))]; + tensor input_1169_cast_fp16 = layer_norm(axes = input_1169_axes_0, beta = encoder_layers_21_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_21_norm_out_weight_to_fp16, x = input_1167_cast_fp16)[name = string("input_1169_cast_fp16")]; + tensor cache_89_begin_0 = const()[name = string("cache_89_begin_0"), val = tensor([22, 0, 0, 0])]; + tensor cache_89_end_0 = const()[name = string("cache_89_end_0"), val = tensor([23, 1, 42, 1024])]; + tensor cache_89_end_mask_0 = const()[name = string("cache_89_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_89_squeeze_mask_0 = const()[name = string("cache_89_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_89_cast_fp16 = slice_by_index(begin = cache_89_begin_0, end = cache_89_end_0, end_mask = cache_89_end_mask_0, squeeze_mask = cache_89_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_89_cast_fp16")]; + tensor cache_91_begin_0 = const()[name = string("cache_91_begin_0"), val = tensor([22, 0, 0, 0])]; + tensor cache_91_end_0 = const()[name = string("cache_91_end_0"), val = tensor([23, 1, 1024, 8])]; + tensor cache_91_end_mask_0 = const()[name = string("cache_91_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_91_squeeze_mask_0 = const()[name = string("cache_91_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_91_cast_fp16 = slice_by_index(begin = cache_91_begin_0, end = cache_91_end_0, end_mask = cache_91_end_mask_0, squeeze_mask = cache_91_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_91_cast_fp16")]; + tensor input_1171_axes_0 = const()[name = string("input_1171_axes_0"), val = tensor([-1])]; + tensor encoder_layers_22_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_22_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(464812864)))]; + tensor encoder_layers_22_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_22_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(464814976)))]; + tensor input_1171_cast_fp16 = layer_norm(axes = input_1171_axes_0, beta = encoder_layers_22_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_22_norm_feed_forward1_weight_to_fp16, x = input_1169_cast_fp16)[name = string("input_1171_cast_fp16")]; + tensor encoder_layers_22_feed_forward1_linear1_weight_to_fp16 = const()[name = string("encoder_layers_22_feed_forward1_linear1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(464817088)))]; + tensor encoder_layers_22_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_22_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(473205760)))]; + tensor linear_199_cast_fp16 = linear(bias = encoder_layers_22_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_22_feed_forward1_linear1_weight_to_fp16, x = input_1171_cast_fp16)[name = string("linear_199_cast_fp16")]; + tensor input_1175_cast_fp16 = silu(x = linear_199_cast_fp16)[name = string("input_1175_cast_fp16")]; + tensor encoder_layers_22_feed_forward1_linear2_weight_to_fp16 = const()[name = string("encoder_layers_22_feed_forward1_linear2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(473214016)))]; + tensor encoder_layers_22_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_22_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481602688)))]; + tensor linear_200_cast_fp16 = linear(bias = encoder_layers_22_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_22_feed_forward1_linear2_weight_to_fp16, x = input_1175_cast_fp16)[name = string("linear_200_cast_fp16")]; + fp16 var_5141_to_fp16 = const()[name = string("op_5141_to_fp16"), val = fp16(0x1p-1)]; + tensor var_5142_cast_fp16 = mul(x = linear_200_cast_fp16, y = var_5141_to_fp16)[name = string("op_5142_cast_fp16")]; + tensor input_1181_cast_fp16 = add(x = input_1169_cast_fp16, y = var_5142_cast_fp16)[name = string("input_1181_cast_fp16")]; + tensor key_45_axes_0 = const()[name = string("key_45_axes_0"), val = tensor([-1])]; + tensor encoder_layers_22_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_22_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481604800)))]; + tensor encoder_layers_22_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_22_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481606912)))]; + tensor key_45_cast_fp16 = layer_norm(axes = key_45_axes_0, beta = encoder_layers_22_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_22_norm_self_att_weight_to_fp16, x = input_1181_cast_fp16)[name = string("key_45_cast_fp16")]; + bool input_1183_interleave_0 = const()[name = string("input_1183_interleave_0"), val = bool(false)]; + tensor input_1183_cast_fp16 = concat(axis = var_68, interleave = input_1183_interleave_0, values = (cache_89_cast_fp16, key_45_cast_fp16))[name = string("input_1183_cast_fp16")]; + tensor var_5164_begin_0 = const()[name = string("op_5164_begin_0"), val = tensor([0, 7, 0])]; + tensor var_5164_end_0 = const()[name = string("op_5164_end_0"), val = tensor([1, 42, 1024])]; + tensor var_5164_end_mask_0 = const()[name = string("op_5164_end_mask_0"), val = tensor([true, true, true])]; + tensor var_5164_cast_fp16 = slice_by_index(begin = var_5164_begin_0, end = var_5164_end_0, end_mask = var_5164_end_mask_0, x = cache_89_cast_fp16)[name = string("op_5164_cast_fp16")]; + bool var_5170_interleave_0 = const()[name = string("op_5170_interleave_0"), val = bool(false)]; + tensor var_5170_cast_fp16 = concat(axis = var_68, interleave = var_5170_interleave_0, values = (var_5164_cast_fp16, key_45_cast_fp16))[name = string("op_5170_cast_fp16")]; + tensor encoder_layers_22_self_attn_linear_q_weight_to_fp16 = const()[name = string("encoder_layers_22_self_attn_linear_q_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481609024)))]; + tensor encoder_layers_22_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_22_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(483706240)))]; + tensor linear_201_cast_fp16 = linear(bias = encoder_layers_22_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_22_self_attn_linear_q_weight_to_fp16, x = key_45_cast_fp16)[name = string("linear_201_cast_fp16")]; + tensor var_5175 = const()[name = string("op_5175"), val = tensor([1, -1, 8, 128])]; + tensor q_133_cast_fp16 = reshape(shape = var_5175, x = linear_201_cast_fp16)[name = string("q_133_cast_fp16")]; + tensor encoder_layers_22_self_attn_linear_k_weight_to_fp16 = const()[name = string("encoder_layers_22_self_attn_linear_k_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(483708352)))]; + tensor encoder_layers_22_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_22_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(485805568)))]; + tensor linear_202_cast_fp16 = linear(bias = encoder_layers_22_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_22_self_attn_linear_k_weight_to_fp16, x = input_1183_cast_fp16)[name = string("linear_202_cast_fp16")]; + tensor var_5180 = const()[name = string("op_5180"), val = tensor([1, -1, 8, 128])]; + tensor k_89_cast_fp16 = reshape(shape = var_5180, x = linear_202_cast_fp16)[name = string("k_89_cast_fp16")]; + tensor encoder_layers_22_self_attn_linear_v_weight_to_fp16 = const()[name = string("encoder_layers_22_self_attn_linear_v_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(485807680)))]; + tensor encoder_layers_22_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_22_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(487904896)))]; + tensor linear_203_cast_fp16 = linear(bias = encoder_layers_22_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_22_self_attn_linear_v_weight_to_fp16, x = input_1183_cast_fp16)[name = string("linear_203_cast_fp16")]; + tensor var_5185 = const()[name = string("op_5185"), val = tensor([1, -1, 8, 128])]; + tensor v_45_cast_fp16 = reshape(shape = var_5185, x = linear_203_cast_fp16)[name = string("v_45_cast_fp16")]; + tensor value_53_perm_0 = const()[name = string("value_53_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_22_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_22_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(487907008)))]; + tensor var_5198_cast_fp16 = add(x = q_133_cast_fp16, y = encoder_layers_22_self_attn_pos_bias_u_to_fp16)[name = string("op_5198_cast_fp16")]; + tensor encoder_layers_22_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_22_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(487909120)))]; + tensor var_5200_cast_fp16 = add(x = q_133_cast_fp16, y = encoder_layers_22_self_attn_pos_bias_v_to_fp16)[name = string("op_5200_cast_fp16")]; + tensor q_with_bias_v_45_perm_0 = const()[name = string("q_with_bias_v_45_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_579_transpose_x_0 = const()[name = string("x_579_transpose_x_0"), val = bool(false)]; + bool x_579_transpose_y_0 = const()[name = string("x_579_transpose_y_0"), val = bool(false)]; + tensor op_5202_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(487911232))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(488010624))))[name = string("op_5202_to_fp16_quantized")]; + tensor q_with_bias_v_45_cast_fp16 = transpose(perm = q_with_bias_v_45_perm_0, x = var_5200_cast_fp16)[name = string("transpose_164")]; + tensor x_579_cast_fp16 = matmul(transpose_x = x_579_transpose_x_0, transpose_y = x_579_transpose_y_0, x = q_with_bias_v_45_cast_fp16, y = op_5202_to_fp16_quantized)[name = string("x_579_cast_fp16")]; + tensor x_581_pad_0 = const()[name = string("x_581_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_581_mode_0 = const()[name = string("x_581_mode_0"), val = string("constant")]; + fp16 const_365_to_fp16 = const()[name = string("const_365_to_fp16"), val = fp16(0x0p+0)]; + tensor x_581_cast_fp16 = pad(constant_val = const_365_to_fp16, mode = x_581_mode_0, pad = x_581_pad_0, x = x_579_cast_fp16)[name = string("x_581_cast_fp16")]; + tensor var_5210 = const()[name = string("op_5210"), val = tensor([1, 8, -1, 7])]; + tensor x_583_cast_fp16 = reshape(shape = var_5210, x = x_581_cast_fp16)[name = string("x_583_cast_fp16")]; + tensor var_5214_begin_0 = const()[name = string("op_5214_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_5214_end_0 = const()[name = string("op_5214_end_0"), val = tensor([1, 8, 98, 7])]; + tensor var_5214_end_mask_0 = const()[name = string("op_5214_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_5214_cast_fp16 = slice_by_index(begin = var_5214_begin_0, end = var_5214_end_0, end_mask = var_5214_end_mask_0, x = x_583_cast_fp16)[name = string("op_5214_cast_fp16")]; + tensor var_5215 = const()[name = string("op_5215"), val = tensor([1, 8, 7, 97])]; + tensor matrix_bd_89_cast_fp16 = reshape(shape = var_5215, x = var_5214_cast_fp16)[name = string("matrix_bd_89_cast_fp16")]; + bool matrix_ac_45_transpose_x_0 = const()[name = string("matrix_ac_45_transpose_x_0"), val = bool(false)]; + bool matrix_ac_45_transpose_y_0 = const()[name = string("matrix_ac_45_transpose_y_0"), val = bool(false)]; + tensor transpose_140_perm_0 = const()[name = string("transpose_140_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_141_perm_0 = const()[name = string("transpose_141_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_141 = transpose(perm = transpose_141_perm_0, x = k_89_cast_fp16)[name = string("transpose_162")]; + tensor transpose_140 = transpose(perm = transpose_140_perm_0, x = var_5198_cast_fp16)[name = string("transpose_163")]; + tensor matrix_ac_45_cast_fp16 = matmul(transpose_x = matrix_ac_45_transpose_x_0, transpose_y = matrix_ac_45_transpose_y_0, x = transpose_140, y = transpose_141)[name = string("matrix_ac_45_cast_fp16")]; + tensor matrix_bd_91_begin_0 = const()[name = string("matrix_bd_91_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_91_end_0 = const()[name = string("matrix_bd_91_end_0"), val = tensor([1, 8, 7, 49])]; + tensor matrix_bd_91_end_mask_0 = const()[name = string("matrix_bd_91_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_91_cast_fp16 = slice_by_index(begin = matrix_bd_91_begin_0, end = matrix_bd_91_end_0, end_mask = matrix_bd_91_end_mask_0, x = matrix_bd_89_cast_fp16)[name = string("matrix_bd_91_cast_fp16")]; + tensor var_5224_cast_fp16 = add(x = matrix_ac_45_cast_fp16, y = matrix_bd_91_cast_fp16)[name = string("op_5224_cast_fp16")]; + fp16 _inversed_scores_89_y_0_to_fp16 = const()[name = string("_inversed_scores_89_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_89_cast_fp16 = mul(x = var_5224_cast_fp16, y = _inversed_scores_89_y_0_to_fp16)[name = string("_inversed_scores_89_cast_fp16")]; + tensor scores_91_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_89_cast_fp16, cond = mask_11)[name = string("scores_91_cast_fp16")]; + tensor var_5230_cast_fp16 = softmax(axis = var_59, x = scores_91_cast_fp16)[name = string("op_5230_cast_fp16")]; + tensor input_1185_cast_fp16 = select(a = var_44_to_fp16, b = var_5230_cast_fp16, cond = mask_11)[name = string("input_1185_cast_fp16")]; + bool x_585_transpose_x_0 = const()[name = string("x_585_transpose_x_0"), val = bool(false)]; + bool x_585_transpose_y_0 = const()[name = string("x_585_transpose_y_0"), val = bool(false)]; + tensor value_53_cast_fp16 = transpose(perm = value_53_perm_0, x = v_45_cast_fp16)[name = string("transpose_161")]; + tensor x_585_cast_fp16 = matmul(transpose_x = x_585_transpose_x_0, transpose_y = x_585_transpose_y_0, x = input_1185_cast_fp16, y = value_53_cast_fp16)[name = string("x_585_cast_fp16")]; + tensor var_5234_perm_0 = const()[name = string("op_5234_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_5235 = const()[name = string("op_5235"), val = tensor([1, -1, 1024])]; + tensor var_5234_cast_fp16 = transpose(perm = var_5234_perm_0, x = x_585_cast_fp16)[name = string("transpose_160")]; + tensor input_1187_cast_fp16 = reshape(shape = var_5235, x = var_5234_cast_fp16)[name = string("input_1187_cast_fp16")]; + tensor encoder_layers_22_self_attn_linear_out_weight_to_fp16 = const()[name = string("encoder_layers_22_self_attn_linear_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(488010944)))]; + tensor encoder_layers_22_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_22_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(490108160)))]; + tensor linear_205_cast_fp16 = linear(bias = encoder_layers_22_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_22_self_attn_linear_out_weight_to_fp16, x = input_1187_cast_fp16)[name = string("linear_205_cast_fp16")]; + tensor input_1191_cast_fp16 = add(x = input_1181_cast_fp16, y = linear_205_cast_fp16)[name = string("input_1191_cast_fp16")]; + tensor x_589_axes_0 = const()[name = string("x_589_axes_0"), val = tensor([-1])]; + tensor encoder_layers_22_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_22_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(490110272)))]; + tensor encoder_layers_22_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_22_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(490112384)))]; + tensor x_589_cast_fp16 = layer_norm(axes = x_589_axes_0, beta = encoder_layers_22_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_22_norm_conv_weight_to_fp16, x = input_1191_cast_fp16)[name = string("x_589_cast_fp16")]; + tensor input_1193_perm_0 = const()[name = string("input_1193_perm_0"), val = tensor([0, 2, 1])]; + string input_1195_pad_type_0 = const()[name = string("input_1195_pad_type_0"), val = string("valid")]; + tensor input_1195_strides_0 = const()[name = string("input_1195_strides_0"), val = tensor([1])]; + tensor input_1195_pad_0 = const()[name = string("input_1195_pad_0"), val = tensor([0, 0])]; + tensor input_1195_dilations_0 = const()[name = string("input_1195_dilations_0"), val = tensor([1])]; + int32 input_1195_groups_0 = const()[name = string("input_1195_groups_0"), val = int32(1)]; + tensor encoder_layers_22_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(490114496))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(492211712))))[name = string("encoder_layers_22_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_1193_cast_fp16 = transpose(perm = input_1193_perm_0, x = x_589_cast_fp16)[name = string("transpose_159")]; + tensor input_1195_cast_fp16 = conv(dilations = input_1195_dilations_0, groups = input_1195_groups_0, pad = input_1195_pad_0, pad_type = input_1195_pad_type_0, strides = input_1195_strides_0, weight = encoder_layers_22_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_1193_cast_fp16)[name = string("input_1195_cast_fp16")]; + int32 x_591_split_num_splits_0 = const()[name = string("x_591_split_num_splits_0"), val = int32(2)]; + int32 x_591_split_axis_0 = const()[name = string("x_591_split_axis_0"), val = int32(1)]; + tensor x_591_split_cast_fp16_0, tensor x_591_split_cast_fp16_1 = split(axis = x_591_split_axis_0, num_splits = x_591_split_num_splits_0, x = input_1195_cast_fp16)[name = string("x_591_split_cast_fp16")]; + tensor x_591_split_1_sigmoid_cast_fp16 = sigmoid(x = x_591_split_cast_fp16_1)[name = string("x_591_split_1_sigmoid_cast_fp16")]; + tensor x_591_cast_fp16 = mul(x = x_591_split_cast_fp16_0, y = x_591_split_1_sigmoid_cast_fp16)[name = string("x_591_cast_fp16")]; + tensor input_1197_cast_fp16 = select(a = var_44_to_fp16, b = x_591_cast_fp16, cond = var_575)[name = string("input_1197_cast_fp16")]; + bool new_x_91_interleave_0 = const()[name = string("new_x_91_interleave_0"), val = bool(false)]; + tensor new_x_91_cast_fp16 = concat(axis = var_59, interleave = new_x_91_interleave_0, values = (cache_91_cast_fp16, input_1197_cast_fp16))[name = string("new_x_91_cast_fp16")]; + tensor var_5274_begin_0 = const()[name = string("op_5274_begin_0"), val = tensor([0, 0, 7])]; + tensor var_5274_end_0 = const()[name = string("op_5274_end_0"), val = tensor([1, 1024, 15])]; + tensor var_5274_end_mask_0 = const()[name = string("op_5274_end_mask_0"), val = tensor([true, true, true])]; + tensor var_5274_cast_fp16 = slice_by_index(begin = var_5274_begin_0, end = var_5274_end_0, end_mask = var_5274_end_mask_0, x = new_x_91_cast_fp16)[name = string("op_5274_cast_fp16")]; + string x_593_pad_type_0 = const()[name = string("x_593_pad_type_0"), val = string("valid")]; + int32 x_593_groups_0 = const()[name = string("x_593_groups_0"), val = int32(1024)]; + tensor x_593_strides_0 = const()[name = string("x_593_strides_0"), val = tensor([1])]; + tensor x_593_pad_0 = const()[name = string("x_593_pad_0"), val = tensor([0, 0])]; + tensor x_593_dilations_0 = const()[name = string("x_593_dilations_0"), val = tensor([1])]; + tensor encoder_layers_22_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(492215872))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(492225152))))[name = string("encoder_layers_22_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_593_cast_fp16 = conv(dilations = x_593_dilations_0, groups = x_593_groups_0, pad = x_593_pad_0, pad_type = x_593_pad_type_0, strides = x_593_strides_0, weight = encoder_layers_22_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_91_cast_fp16)[name = string("x_593_cast_fp16")]; + tensor input_1199_perm_0 = const()[name = string("input_1199_perm_0"), val = tensor([0, 2, 1])]; + tensor x_595_axes_0 = const()[name = string("x_595_axes_0"), val = tensor([-1])]; + tensor encoder_layers_22_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_22_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(492227264)))]; + tensor encoder_layers_22_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_22_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(492229376)))]; + tensor input_1199_cast_fp16 = transpose(perm = input_1199_perm_0, x = x_593_cast_fp16)[name = string("transpose_158")]; + tensor x_595_cast_fp16 = layer_norm(axes = x_595_axes_0, beta = encoder_layers_22_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_22_conv_batch_norm_weight_to_fp16, x = input_1199_cast_fp16)[name = string("x_595_cast_fp16")]; + tensor input_1201_perm_0 = const()[name = string("input_1201_perm_0"), val = tensor([0, 2, 1])]; + tensor input_1201_cast_fp16 = transpose(perm = input_1201_perm_0, x = x_595_cast_fp16)[name = string("transpose_157")]; + tensor input_1203_cast_fp16 = silu(x = input_1201_cast_fp16)[name = string("input_1203_cast_fp16")]; + string x_597_pad_type_0 = const()[name = string("x_597_pad_type_0"), val = string("valid")]; + tensor x_597_strides_0 = const()[name = string("x_597_strides_0"), val = tensor([1])]; + tensor x_597_pad_0 = const()[name = string("x_597_pad_0"), val = tensor([0, 0])]; + tensor x_597_dilations_0 = const()[name = string("x_597_dilations_0"), val = tensor([1])]; + int32 x_597_groups_0 = const()[name = string("x_597_groups_0"), val = int32(1)]; + tensor encoder_layers_22_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(492231488))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(493280128))))[name = string("encoder_layers_22_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_597_cast_fp16 = conv(dilations = x_597_dilations_0, groups = x_597_groups_0, pad = x_597_pad_0, pad_type = x_597_pad_type_0, strides = x_597_strides_0, weight = encoder_layers_22_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_1203_cast_fp16)[name = string("x_597_cast_fp16")]; + tensor input_1205_perm_0 = const()[name = string("input_1205_perm_0"), val = tensor([0, 2, 1])]; + tensor input_1205_cast_fp16 = transpose(perm = input_1205_perm_0, x = x_597_cast_fp16)[name = string("transpose_156")]; + tensor input_1207_cast_fp16 = add(x = input_1191_cast_fp16, y = input_1205_cast_fp16)[name = string("input_1207_cast_fp16")]; + tensor input_1209_axes_0 = const()[name = string("input_1209_axes_0"), val = tensor([-1])]; + tensor encoder_layers_22_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_22_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(493282240)))]; + tensor encoder_layers_22_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_22_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(493284352)))]; + tensor input_1209_cast_fp16 = layer_norm(axes = input_1209_axes_0, beta = encoder_layers_22_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_22_norm_feed_forward2_weight_to_fp16, x = input_1207_cast_fp16)[name = string("input_1209_cast_fp16")]; + tensor encoder_layers_22_feed_forward2_linear1_weight_to_fp16 = const()[name = string("encoder_layers_22_feed_forward2_linear1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(493286464)))]; + tensor encoder_layers_22_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_22_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(501675136)))]; + tensor linear_206_cast_fp16 = linear(bias = encoder_layers_22_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_22_feed_forward2_linear1_weight_to_fp16, x = input_1209_cast_fp16)[name = string("linear_206_cast_fp16")]; + tensor input_1213_cast_fp16 = silu(x = linear_206_cast_fp16)[name = string("input_1213_cast_fp16")]; + tensor encoder_layers_22_feed_forward2_linear2_weight_to_fp16 = const()[name = string("encoder_layers_22_feed_forward2_linear2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(501683392)))]; + tensor encoder_layers_22_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_22_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(510072064)))]; + tensor linear_207_cast_fp16 = linear(bias = encoder_layers_22_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_22_feed_forward2_linear2_weight_to_fp16, x = input_1213_cast_fp16)[name = string("linear_207_cast_fp16")]; + fp16 var_5317_to_fp16 = const()[name = string("op_5317_to_fp16"), val = fp16(0x1p-1)]; + tensor var_5318_cast_fp16 = mul(x = linear_207_cast_fp16, y = var_5317_to_fp16)[name = string("op_5318_cast_fp16")]; + tensor input_1219_cast_fp16 = add(x = input_1207_cast_fp16, y = var_5318_cast_fp16)[name = string("input_1219_cast_fp16")]; + tensor input_1221_axes_0 = const()[name = string("input_1221_axes_0"), val = tensor([-1])]; + tensor encoder_layers_22_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_22_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(510074176)))]; + tensor encoder_layers_22_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_22_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(510076288)))]; + tensor input_1221_cast_fp16 = layer_norm(axes = input_1221_axes_0, beta = encoder_layers_22_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_22_norm_out_weight_to_fp16, x = input_1219_cast_fp16)[name = string("input_1221_cast_fp16")]; + tensor cache_93_begin_0 = const()[name = string("cache_93_begin_0"), val = tensor([23, 0, 0, 0])]; + tensor cache_93_end_0 = const()[name = string("cache_93_end_0"), val = tensor([24, 1, 42, 1024])]; + tensor cache_93_end_mask_0 = const()[name = string("cache_93_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_93_squeeze_mask_0 = const()[name = string("cache_93_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_93_cast_fp16 = slice_by_index(begin = cache_93_begin_0, end = cache_93_end_0, end_mask = cache_93_end_mask_0, squeeze_mask = cache_93_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_93_cast_fp16")]; + tensor cache_begin_0 = const()[name = string("cache_begin_0"), val = tensor([23, 0, 0, 0])]; + tensor cache_end_0 = const()[name = string("cache_end_0"), val = tensor([24, 1, 1024, 8])]; + tensor cache_end_mask_0 = const()[name = string("cache_end_mask_0"), val = tensor([false, true, true, true])]; + tensor cache_squeeze_mask_0 = const()[name = string("cache_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor cache_cast_fp16 = slice_by_index(begin = cache_begin_0, end = cache_end_0, end_mask = cache_end_mask_0, squeeze_mask = cache_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_cast_fp16")]; + tensor input_1223_axes_0 = const()[name = string("input_1223_axes_0"), val = tensor([-1])]; + tensor encoder_layers_23_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_23_norm_feed_forward1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(510078400)))]; + tensor encoder_layers_23_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_23_norm_feed_forward1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(510080512)))]; + tensor input_1223_cast_fp16 = layer_norm(axes = input_1223_axes_0, beta = encoder_layers_23_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_23_norm_feed_forward1_weight_to_fp16, x = input_1221_cast_fp16)[name = string("input_1223_cast_fp16")]; + tensor encoder_layers_23_feed_forward1_linear1_weight_to_fp16 = const()[name = string("encoder_layers_23_feed_forward1_linear1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(510082624)))]; + tensor encoder_layers_23_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_23_feed_forward1_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(518471296)))]; + tensor linear_208_cast_fp16 = linear(bias = encoder_layers_23_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_23_feed_forward1_linear1_weight_to_fp16, x = input_1223_cast_fp16)[name = string("linear_208_cast_fp16")]; + tensor input_1227_cast_fp16 = silu(x = linear_208_cast_fp16)[name = string("input_1227_cast_fp16")]; + tensor encoder_layers_23_feed_forward1_linear2_weight_to_fp16 = const()[name = string("encoder_layers_23_feed_forward1_linear2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(518479552)))]; + tensor encoder_layers_23_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_23_feed_forward1_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(526868224)))]; + tensor linear_209_cast_fp16 = linear(bias = encoder_layers_23_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_23_feed_forward1_linear2_weight_to_fp16, x = input_1227_cast_fp16)[name = string("linear_209_cast_fp16")]; + fp16 var_5354_to_fp16 = const()[name = string("op_5354_to_fp16"), val = fp16(0x1p-1)]; + tensor var_5355_cast_fp16 = mul(x = linear_209_cast_fp16, y = var_5354_to_fp16)[name = string("op_5355_cast_fp16")]; + tensor input_1233_cast_fp16 = add(x = input_1221_cast_fp16, y = var_5355_cast_fp16)[name = string("input_1233_cast_fp16")]; + tensor key_axes_0 = const()[name = string("key_axes_0"), val = tensor([-1])]; + tensor encoder_layers_23_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_23_norm_self_att_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(526870336)))]; + tensor encoder_layers_23_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_23_norm_self_att_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(526872448)))]; + tensor key_cast_fp16 = layer_norm(axes = key_axes_0, beta = encoder_layers_23_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_23_norm_self_att_weight_to_fp16, x = input_1233_cast_fp16)[name = string("key_cast_fp16")]; + bool input_1235_interleave_0 = const()[name = string("input_1235_interleave_0"), val = bool(false)]; + tensor input_1235_cast_fp16 = concat(axis = var_68, interleave = input_1235_interleave_0, values = (cache_93_cast_fp16, key_cast_fp16))[name = string("input_1235_cast_fp16")]; + tensor var_5377_begin_0 = const()[name = string("op_5377_begin_0"), val = tensor([0, 7, 0])]; + tensor var_5377_end_0 = const()[name = string("op_5377_end_0"), val = tensor([1, 42, 1024])]; + tensor var_5377_end_mask_0 = const()[name = string("op_5377_end_mask_0"), val = tensor([true, true, true])]; + tensor var_5377_cast_fp16 = slice_by_index(begin = var_5377_begin_0, end = var_5377_end_0, end_mask = var_5377_end_mask_0, x = cache_93_cast_fp16)[name = string("op_5377_cast_fp16")]; + bool cache_last_channel_cur_interleave_0 = const()[name = string("cache_last_channel_cur_interleave_0"), val = bool(false)]; + tensor cache_last_channel_cur_cast_fp16 = concat(axis = var_68, interleave = cache_last_channel_cur_interleave_0, values = (var_5377_cast_fp16, key_cast_fp16))[name = string("cache_last_channel_cur_cast_fp16")]; + tensor encoder_layers_23_self_attn_linear_q_weight_to_fp16 = const()[name = string("encoder_layers_23_self_attn_linear_q_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(526874560)))]; + tensor encoder_layers_23_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_23_self_attn_linear_q_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(528971776)))]; + tensor linear_210_cast_fp16 = linear(bias = encoder_layers_23_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_23_self_attn_linear_q_weight_to_fp16, x = key_cast_fp16)[name = string("linear_210_cast_fp16")]; + tensor var_5388 = const()[name = string("op_5388"), val = tensor([1, -1, 8, 128])]; + tensor q_139_cast_fp16 = reshape(shape = var_5388, x = linear_210_cast_fp16)[name = string("q_139_cast_fp16")]; + tensor encoder_layers_23_self_attn_linear_k_weight_to_fp16 = const()[name = string("encoder_layers_23_self_attn_linear_k_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(528973888)))]; + tensor encoder_layers_23_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_23_self_attn_linear_k_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(531071104)))]; + tensor linear_211_cast_fp16 = linear(bias = encoder_layers_23_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_23_self_attn_linear_k_weight_to_fp16, x = input_1235_cast_fp16)[name = string("linear_211_cast_fp16")]; + tensor var_5393 = const()[name = string("op_5393"), val = tensor([1, -1, 8, 128])]; + tensor k_93_cast_fp16 = reshape(shape = var_5393, x = linear_211_cast_fp16)[name = string("k_93_cast_fp16")]; + tensor encoder_layers_23_self_attn_linear_v_weight_to_fp16 = const()[name = string("encoder_layers_23_self_attn_linear_v_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(531073216)))]; + tensor encoder_layers_23_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_23_self_attn_linear_v_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(533170432)))]; + tensor linear_212_cast_fp16 = linear(bias = encoder_layers_23_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_23_self_attn_linear_v_weight_to_fp16, x = input_1235_cast_fp16)[name = string("linear_212_cast_fp16")]; + tensor var_5398 = const()[name = string("op_5398"), val = tensor([1, -1, 8, 128])]; + tensor v_cast_fp16 = reshape(shape = var_5398, x = linear_212_cast_fp16)[name = string("v_cast_fp16")]; + tensor value_perm_0 = const()[name = string("value_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor encoder_layers_23_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_23_self_attn_pos_bias_u_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(533172544)))]; + tensor var_5411_cast_fp16 = add(x = q_139_cast_fp16, y = encoder_layers_23_self_attn_pos_bias_u_to_fp16)[name = string("op_5411_cast_fp16")]; + tensor encoder_layers_23_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_23_self_attn_pos_bias_v_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(533174656)))]; + tensor var_5413_cast_fp16 = add(x = q_139_cast_fp16, y = encoder_layers_23_self_attn_pos_bias_v_to_fp16)[name = string("op_5413_cast_fp16")]; + tensor q_with_bias_v_perm_0 = const()[name = string("q_with_bias_v_perm_0"), val = tensor([0, 2, -3, -1])]; + bool x_605_transpose_x_0 = const()[name = string("x_605_transpose_x_0"), val = bool(false)]; + bool x_605_transpose_y_0 = const()[name = string("x_605_transpose_y_0"), val = bool(false)]; + tensor op_5415_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(533176768))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(533276160))))[name = string("op_5415_to_fp16_quantized")]; + tensor q_with_bias_v_cast_fp16 = transpose(perm = q_with_bias_v_perm_0, x = var_5413_cast_fp16)[name = string("transpose_155")]; + tensor x_605_cast_fp16 = matmul(transpose_x = x_605_transpose_x_0, transpose_y = x_605_transpose_y_0, x = q_with_bias_v_cast_fp16, y = op_5415_to_fp16_quantized)[name = string("x_605_cast_fp16")]; + tensor x_607_pad_0 = const()[name = string("x_607_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + string x_607_mode_0 = const()[name = string("x_607_mode_0"), val = string("constant")]; + fp16 const_378_to_fp16 = const()[name = string("const_378_to_fp16"), val = fp16(0x0p+0)]; + tensor x_607_cast_fp16 = pad(constant_val = const_378_to_fp16, mode = x_607_mode_0, pad = x_607_pad_0, x = x_605_cast_fp16)[name = string("x_607_cast_fp16")]; + tensor var_5423 = const()[name = string("op_5423"), val = tensor([1, 8, -1, 7])]; + tensor x_609_cast_fp16 = reshape(shape = var_5423, x = x_607_cast_fp16)[name = string("x_609_cast_fp16")]; + tensor var_5427_begin_0 = const()[name = string("op_5427_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_5427_end_0 = const()[name = string("op_5427_end_0"), val = tensor([1, 8, 98, 7])]; + tensor var_5427_end_mask_0 = const()[name = string("op_5427_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_5427_cast_fp16 = slice_by_index(begin = var_5427_begin_0, end = var_5427_end_0, end_mask = var_5427_end_mask_0, x = x_609_cast_fp16)[name = string("op_5427_cast_fp16")]; + tensor var_5428 = const()[name = string("op_5428"), val = tensor([1, 8, 7, 97])]; + tensor matrix_bd_93_cast_fp16 = reshape(shape = var_5428, x = var_5427_cast_fp16)[name = string("matrix_bd_93_cast_fp16")]; + bool matrix_ac_transpose_x_0 = const()[name = string("matrix_ac_transpose_x_0"), val = bool(false)]; + bool matrix_ac_transpose_y_0 = const()[name = string("matrix_ac_transpose_y_0"), val = bool(false)]; + tensor transpose_142_perm_0 = const()[name = string("transpose_142_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_143_perm_0 = const()[name = string("transpose_143_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_143 = transpose(perm = transpose_143_perm_0, x = k_93_cast_fp16)[name = string("transpose_153")]; + tensor transpose_142 = transpose(perm = transpose_142_perm_0, x = var_5411_cast_fp16)[name = string("transpose_154")]; + tensor matrix_ac_cast_fp16 = matmul(transpose_x = matrix_ac_transpose_x_0, transpose_y = matrix_ac_transpose_y_0, x = transpose_142, y = transpose_143)[name = string("matrix_ac_cast_fp16")]; + tensor matrix_bd_begin_0 = const()[name = string("matrix_bd_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor matrix_bd_end_0 = const()[name = string("matrix_bd_end_0"), val = tensor([1, 8, 7, 49])]; + tensor matrix_bd_end_mask_0 = const()[name = string("matrix_bd_end_mask_0"), val = tensor([true, true, true, false])]; + tensor matrix_bd_cast_fp16 = slice_by_index(begin = matrix_bd_begin_0, end = matrix_bd_end_0, end_mask = matrix_bd_end_mask_0, x = matrix_bd_93_cast_fp16)[name = string("matrix_bd_cast_fp16")]; + tensor var_5437_cast_fp16 = add(x = matrix_ac_cast_fp16, y = matrix_bd_cast_fp16)[name = string("op_5437_cast_fp16")]; + fp16 _inversed_scores_93_y_0_to_fp16 = const()[name = string("_inversed_scores_93_y_0_to_fp16"), val = fp16(0x1.6ap-4)]; + tensor _inversed_scores_93_cast_fp16 = mul(x = var_5437_cast_fp16, y = _inversed_scores_93_y_0_to_fp16)[name = string("_inversed_scores_93_cast_fp16")]; + tensor scores_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_93_cast_fp16, cond = mask_11)[name = string("scores_cast_fp16")]; + tensor var_5443_cast_fp16 = softmax(axis = var_59, x = scores_cast_fp16)[name = string("op_5443_cast_fp16")]; + tensor input_1237_cast_fp16 = select(a = var_44_to_fp16, b = var_5443_cast_fp16, cond = mask_11)[name = string("input_1237_cast_fp16")]; + bool x_611_transpose_x_0 = const()[name = string("x_611_transpose_x_0"), val = bool(false)]; + bool x_611_transpose_y_0 = const()[name = string("x_611_transpose_y_0"), val = bool(false)]; + tensor value_cast_fp16 = transpose(perm = value_perm_0, x = v_cast_fp16)[name = string("transpose_152")]; + tensor x_611_cast_fp16 = matmul(transpose_x = x_611_transpose_x_0, transpose_y = x_611_transpose_y_0, x = input_1237_cast_fp16, y = value_cast_fp16)[name = string("x_611_cast_fp16")]; + tensor var_5447_perm_0 = const()[name = string("op_5447_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_5448 = const()[name = string("op_5448"), val = tensor([1, -1, 1024])]; + tensor var_5447_cast_fp16 = transpose(perm = var_5447_perm_0, x = x_611_cast_fp16)[name = string("transpose_151")]; + tensor input_1239_cast_fp16 = reshape(shape = var_5448, x = var_5447_cast_fp16)[name = string("input_1239_cast_fp16")]; + tensor encoder_layers_23_self_attn_linear_out_weight_to_fp16 = const()[name = string("encoder_layers_23_self_attn_linear_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(533276480)))]; + tensor encoder_layers_23_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_23_self_attn_linear_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(535373696)))]; + tensor linear_214_cast_fp16 = linear(bias = encoder_layers_23_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_23_self_attn_linear_out_weight_to_fp16, x = input_1239_cast_fp16)[name = string("linear_214_cast_fp16")]; + tensor input_1243_cast_fp16 = add(x = input_1233_cast_fp16, y = linear_214_cast_fp16)[name = string("input_1243_cast_fp16")]; + tensor x_615_axes_0 = const()[name = string("x_615_axes_0"), val = tensor([-1])]; + tensor encoder_layers_23_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_23_norm_conv_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(535375808)))]; + tensor encoder_layers_23_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_23_norm_conv_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(535377920)))]; + tensor x_615_cast_fp16 = layer_norm(axes = x_615_axes_0, beta = encoder_layers_23_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_23_norm_conv_weight_to_fp16, x = input_1243_cast_fp16)[name = string("x_615_cast_fp16")]; + tensor input_1245_perm_0 = const()[name = string("input_1245_perm_0"), val = tensor([0, 2, 1])]; + string input_1247_pad_type_0 = const()[name = string("input_1247_pad_type_0"), val = string("valid")]; + tensor input_1247_strides_0 = const()[name = string("input_1247_strides_0"), val = tensor([1])]; + tensor input_1247_pad_0 = const()[name = string("input_1247_pad_0"), val = tensor([0, 0])]; + tensor input_1247_dilations_0 = const()[name = string("input_1247_dilations_0"), val = tensor([1])]; + int32 input_1247_groups_0 = const()[name = string("input_1247_groups_0"), val = int32(1)]; + tensor encoder_layers_23_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(535380032))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(537477248))))[name = string("encoder_layers_23_conv_pointwise_conv1_weight_to_fp16_quantized")]; + tensor input_1245_cast_fp16 = transpose(perm = input_1245_perm_0, x = x_615_cast_fp16)[name = string("transpose_150")]; + tensor input_1247_cast_fp16 = conv(dilations = input_1247_dilations_0, groups = input_1247_groups_0, pad = input_1247_pad_0, pad_type = input_1247_pad_type_0, strides = input_1247_strides_0, weight = encoder_layers_23_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_1245_cast_fp16)[name = string("input_1247_cast_fp16")]; + int32 x_617_split_num_splits_0 = const()[name = string("x_617_split_num_splits_0"), val = int32(2)]; + int32 x_617_split_axis_0 = const()[name = string("x_617_split_axis_0"), val = int32(1)]; + tensor x_617_split_cast_fp16_0, tensor x_617_split_cast_fp16_1 = split(axis = x_617_split_axis_0, num_splits = x_617_split_num_splits_0, x = input_1247_cast_fp16)[name = string("x_617_split_cast_fp16")]; + tensor x_617_split_1_sigmoid_cast_fp16 = sigmoid(x = x_617_split_cast_fp16_1)[name = string("x_617_split_1_sigmoid_cast_fp16")]; + tensor x_617_cast_fp16 = mul(x = x_617_split_cast_fp16_0, y = x_617_split_1_sigmoid_cast_fp16)[name = string("x_617_cast_fp16")]; + tensor input_1249_cast_fp16 = select(a = var_44_to_fp16, b = x_617_cast_fp16, cond = var_575)[name = string("input_1249_cast_fp16")]; + bool new_x_interleave_0 = const()[name = string("new_x_interleave_0"), val = bool(false)]; + tensor new_x_cast_fp16 = concat(axis = var_59, interleave = new_x_interleave_0, values = (cache_cast_fp16, input_1249_cast_fp16))[name = string("new_x_cast_fp16")]; + tensor cache_last_time_cur_begin_0 = const()[name = string("cache_last_time_cur_begin_0"), val = tensor([0, 0, 7])]; + tensor cache_last_time_cur_end_0 = const()[name = string("cache_last_time_cur_end_0"), val = tensor([1, 1024, 15])]; + tensor cache_last_time_cur_end_mask_0 = const()[name = string("cache_last_time_cur_end_mask_0"), val = tensor([true, true, true])]; + tensor cache_last_time_cur_cast_fp16 = slice_by_index(begin = cache_last_time_cur_begin_0, end = cache_last_time_cur_end_0, end_mask = cache_last_time_cur_end_mask_0, x = new_x_cast_fp16)[name = string("cache_last_time_cur_cast_fp16")]; + string x_619_pad_type_0 = const()[name = string("x_619_pad_type_0"), val = string("valid")]; + int32 x_619_groups_0 = const()[name = string("x_619_groups_0"), val = int32(1024)]; + tensor x_619_strides_0 = const()[name = string("x_619_strides_0"), val = tensor([1])]; + tensor x_619_pad_0 = const()[name = string("x_619_pad_0"), val = tensor([0, 0])]; + tensor x_619_dilations_0 = const()[name = string("x_619_dilations_0"), val = tensor([1])]; + tensor encoder_layers_23_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(537481408))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(537490688))))[name = string("encoder_layers_23_conv_depthwise_conv_weight_to_fp16_quantized")]; + tensor x_619_cast_fp16 = conv(dilations = x_619_dilations_0, groups = x_619_groups_0, pad = x_619_pad_0, pad_type = x_619_pad_type_0, strides = x_619_strides_0, weight = encoder_layers_23_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_cast_fp16)[name = string("x_619_cast_fp16")]; + tensor input_1251_perm_0 = const()[name = string("input_1251_perm_0"), val = tensor([0, 2, 1])]; + tensor x_621_axes_0 = const()[name = string("x_621_axes_0"), val = tensor([-1])]; + tensor encoder_layers_23_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_23_conv_batch_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(537492800)))]; + tensor encoder_layers_23_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_23_conv_batch_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(537494912)))]; + tensor input_1251_cast_fp16 = transpose(perm = input_1251_perm_0, x = x_619_cast_fp16)[name = string("transpose_149")]; + tensor x_621_cast_fp16 = layer_norm(axes = x_621_axes_0, beta = encoder_layers_23_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_23_conv_batch_norm_weight_to_fp16, x = input_1251_cast_fp16)[name = string("x_621_cast_fp16")]; + tensor input_1253_perm_0 = const()[name = string("input_1253_perm_0"), val = tensor([0, 2, 1])]; + tensor input_1253_cast_fp16 = transpose(perm = input_1253_perm_0, x = x_621_cast_fp16)[name = string("transpose_148")]; + tensor input_1255_cast_fp16 = silu(x = input_1253_cast_fp16)[name = string("input_1255_cast_fp16")]; + string x_623_pad_type_0 = const()[name = string("x_623_pad_type_0"), val = string("valid")]; + tensor x_623_strides_0 = const()[name = string("x_623_strides_0"), val = tensor([1])]; + tensor x_623_pad_0 = const()[name = string("x_623_pad_0"), val = tensor([0, 0])]; + tensor x_623_dilations_0 = const()[name = string("x_623_dilations_0"), val = tensor([1])]; + int32 x_623_groups_0 = const()[name = string("x_623_groups_0"), val = int32(1)]; + tensor encoder_layers_23_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(537497024))), scale = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(538545664))))[name = string("encoder_layers_23_conv_pointwise_conv2_weight_to_fp16_quantized")]; + tensor x_623_cast_fp16 = conv(dilations = x_623_dilations_0, groups = x_623_groups_0, pad = x_623_pad_0, pad_type = x_623_pad_type_0, strides = x_623_strides_0, weight = encoder_layers_23_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_1255_cast_fp16)[name = string("x_623_cast_fp16")]; + tensor input_1257_perm_0 = const()[name = string("input_1257_perm_0"), val = tensor([0, 2, 1])]; + tensor input_1257_cast_fp16 = transpose(perm = input_1257_perm_0, x = x_623_cast_fp16)[name = string("transpose_147")]; + tensor input_1259_cast_fp16 = add(x = input_1243_cast_fp16, y = input_1257_cast_fp16)[name = string("input_1259_cast_fp16")]; + tensor input_1261_axes_0 = const()[name = string("input_1261_axes_0"), val = tensor([-1])]; + tensor encoder_layers_23_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_23_norm_feed_forward2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(538547776)))]; + tensor encoder_layers_23_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_23_norm_feed_forward2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(538549888)))]; + tensor input_1261_cast_fp16 = layer_norm(axes = input_1261_axes_0, beta = encoder_layers_23_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_23_norm_feed_forward2_weight_to_fp16, x = input_1259_cast_fp16)[name = string("input_1261_cast_fp16")]; + tensor encoder_layers_23_feed_forward2_linear1_weight_to_fp16 = const()[name = string("encoder_layers_23_feed_forward2_linear1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(538552000)))]; + tensor encoder_layers_23_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_23_feed_forward2_linear1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(546940672)))]; + tensor linear_215_cast_fp16 = linear(bias = encoder_layers_23_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_23_feed_forward2_linear1_weight_to_fp16, x = input_1261_cast_fp16)[name = string("linear_215_cast_fp16")]; + tensor input_1265_cast_fp16 = silu(x = linear_215_cast_fp16)[name = string("input_1265_cast_fp16")]; + tensor encoder_layers_23_feed_forward2_linear2_weight_to_fp16 = const()[name = string("encoder_layers_23_feed_forward2_linear2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(546948928)))]; + tensor encoder_layers_23_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_23_feed_forward2_linear2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(555337600)))]; + tensor linear_216_cast_fp16 = linear(bias = encoder_layers_23_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_23_feed_forward2_linear2_weight_to_fp16, x = input_1265_cast_fp16)[name = string("linear_216_cast_fp16")]; + fp16 var_5530_to_fp16 = const()[name = string("op_5530_to_fp16"), val = fp16(0x1p-1)]; + tensor var_5531_cast_fp16 = mul(x = linear_216_cast_fp16, y = var_5530_to_fp16)[name = string("op_5531_cast_fp16")]; + tensor input_1271_cast_fp16 = add(x = input_1259_cast_fp16, y = var_5531_cast_fp16)[name = string("input_1271_cast_fp16")]; + tensor audio_signal_axes_0 = const()[name = string("audio_signal_axes_0"), val = tensor([-1])]; + tensor encoder_layers_23_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_23_norm_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(555339712)))]; + tensor encoder_layers_23_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_23_norm_out_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(555341824)))]; + tensor audio_signal_cast_fp16 = layer_norm(axes = audio_signal_axes_0, beta = encoder_layers_23_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_23_norm_out_weight_to_fp16, x = input_1271_cast_fp16)[name = string("audio_signal_cast_fp16")]; + int32 obj_5_axis_0 = const()[name = string("obj_5_axis_0"), val = int32(0)]; + tensor obj_5_cast_fp16 = stack(axis = obj_5_axis_0, values = (var_484_cast_fp16, var_697_cast_fp16, var_910_cast_fp16, var_1123_cast_fp16, var_1336_cast_fp16, var_1549_cast_fp16, var_1762_cast_fp16, var_1975_cast_fp16, var_2188_cast_fp16, var_2401_cast_fp16, var_2614_cast_fp16, var_2827_cast_fp16, var_3040_cast_fp16, var_3253_cast_fp16, var_3466_cast_fp16, var_3679_cast_fp16, var_3892_cast_fp16, var_4105_cast_fp16, var_4318_cast_fp16, var_4531_cast_fp16, var_4744_cast_fp16, var_4957_cast_fp16, var_5170_cast_fp16, cache_last_channel_cur_cast_fp16))[name = string("obj_5_cast_fp16")]; + int32 obj_7_axis_0 = const()[name = string("obj_7_axis_0"), val = int32(0)]; + tensor obj_7_cast_fp16 = stack(axis = obj_7_axis_0, values = (var_588_cast_fp16, var_801_cast_fp16, var_1014_cast_fp16, var_1227_cast_fp16, var_1440_cast_fp16, var_1653_cast_fp16, var_1866_cast_fp16, var_2079_cast_fp16, var_2292_cast_fp16, var_2505_cast_fp16, var_2718_cast_fp16, var_2931_cast_fp16, var_3144_cast_fp16, var_3357_cast_fp16, var_3570_cast_fp16, var_3783_cast_fp16, var_3996_cast_fp16, var_4209_cast_fp16, var_4422_cast_fp16, var_4635_cast_fp16, var_4848_cast_fp16, var_5061_cast_fp16, var_5274_cast_fp16, cache_last_time_cur_cast_fp16))[name = string("obj_7_cast_fp16")]; + tensor var_5547 = add(x = cache_len, y = max_audio_length_1)[name = string("op_5547")]; + string var_5547_promoted_to_fp16_dtype_0 = const()[name = string("op_5547_promoted_to_fp16_dtype_0"), val = string("fp16")]; + fp16 const_384_to_fp16 = const()[name = string("const_384_to_fp16"), val = fp16(-inf)]; + fp16 var_49_promoted_to_fp16 = const()[name = string("op_49_promoted_to_fp16"), val = fp16(0x1.5p+5)]; + tensor var_5547_to_fp16 = cast(dtype = var_5547_promoted_to_fp16_dtype_0, x = var_5547)[name = string("cast_9")]; + tensor clip_1_cast_fp16 = clip(alpha = const_384_to_fp16, beta = var_49_promoted_to_fp16, x = var_5547_to_fp16)[name = string("clip_1_cast_fp16")]; + int32 one_hot_1_batch_dims_0 = const()[name = string("one_hot_1_batch_dims_0"), val = int32(0)]; + bool one_hot_1_validate_indices_0 = const()[name = string("one_hot_1_validate_indices_0"), val = bool(false)]; + tensor to_onehot_identity_table_to_fp16 = const()[name = string("to_onehot_identity_table_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(555343936)))]; + string prompt_id_to_int16_dtype_0 = const()[name = string("prompt_id_to_int16_dtype_0"), val = string("int16")]; + string cast_230_dtype_0 = const()[name = string("cast_230_dtype_0"), val = string("int32")]; + int32 greater_equal_0_y_0 = const()[name = string("greater_equal_0_y_0"), val = int32(0)]; + tensor prompt_id_to_int16 = cast(dtype = prompt_id_to_int16_dtype_0, x = prompt_id)[name = string("cast_8")]; + tensor cast_230 = cast(dtype = cast_230_dtype_0, x = prompt_id_to_int16)[name = string("cast_7")]; + tensor greater_equal_0 = greater_equal(x = cast_230, y = greater_equal_0_y_0)[name = string("greater_equal_0")]; + int32 slice_by_index_2 = const()[name = string("slice_by_index_2"), val = int32(128)]; + tensor add_0 = add(x = cast_230, y = slice_by_index_2)[name = string("add_0")]; + tensor select_0 = select(a = cast_230, b = add_0, cond = greater_equal_0)[name = string("select_0")]; + string select_0_to_int16_dtype_0 = const()[name = string("select_0_to_int16_dtype_0"), val = string("int16")]; + string cast_0_dtype_0 = const()[name = string("cast_0_dtype_0"), val = string("int32")]; + int32 greater_equal_0_y_0_1 = const()[name = string("greater_equal_0_y_0_1"), val = int32(0)]; + tensor select_0_to_int16 = cast(dtype = select_0_to_int16_dtype_0, x = select_0)[name = string("cast_6")]; + tensor cast_0 = cast(dtype = cast_0_dtype_0, x = select_0_to_int16)[name = string("cast_5")]; + tensor greater_equal_0_1 = greater_equal(x = cast_0, y = greater_equal_0_y_0_1)[name = string("greater_equal_0_1")]; + int32 slice_by_index_0 = const()[name = string("slice_by_index_0"), val = int32(128)]; + tensor add_0_1 = add(x = cast_0, y = slice_by_index_0)[name = string("add_0_1")]; + tensor select_0_1 = select(a = cast_0, b = add_0_1, cond = greater_equal_0_1)[name = string("select_0_1")]; + int32 greater_equal_0_y_0_2 = const()[name = string("greater_equal_0_y_0_2"), val = int32(0)]; + tensor greater_equal_0_2 = greater_equal(x = select_0_1, y = greater_equal_0_y_0_2)[name = string("greater_equal_0_2")]; + int32 slice_by_index_0_1 = const()[name = string("slice_by_index_0_1"), val = int32(128)]; + tensor add_0_2 = add(x = select_0_1, y = slice_by_index_0_1)[name = string("add_0_2")]; + tensor select_0_2 = select(a = select_0_1, b = add_0_2, cond = greater_equal_0_2)[name = string("select_0_2")]; + int32 one_hot_1_cast_fp16_cast_uint16_cast_uint16_axis_0 = const()[name = string("one_hot_1_cast_fp16_cast_uint16_cast_uint16_axis_0"), val = int32(0)]; + tensor one_hot_1_cast_fp16_cast_uint16_cast_uint16 = gather(axis = one_hot_1_cast_fp16_cast_uint16_cast_uint16_axis_0, batch_dims = one_hot_1_batch_dims_0, indices = select_0_2, validate_indices = one_hot_1_validate_indices_0, x = to_onehot_identity_table_to_fp16)[name = string("one_hot_1_cast_fp16_cast_uint16_cast_uint16")]; + tensor var_5593_axes_0 = const()[name = string("op_5593_axes_0"), val = tensor([1])]; + tensor var_5593_cast_fp16 = expand_dims(axes = var_5593_axes_0, x = one_hot_1_cast_fp16_cast_uint16_cast_uint16)[name = string("op_5593_cast_fp16")]; + tensor one_hot_reps_0 = const()[name = string("one_hot_reps_0"), val = tensor([1, 7, 1])]; + tensor one_hot_cast_fp16 = tile(reps = one_hot_reps_0, x = var_5593_cast_fp16)[name = string("one_hot_cast_fp16")]; + int32 var_5602 = const()[name = string("op_5602"), val = int32(-1)]; + bool input_1273_interleave_0 = const()[name = string("input_1273_interleave_0"), val = bool(false)]; + tensor input_1273_cast_fp16 = concat(axis = var_5602, interleave = input_1273_interleave_0, values = (audio_signal_cast_fp16, one_hot_cast_fp16))[name = string("input_1273_cast_fp16")]; + tensor prompt_kernel_0_weight_to_fp16 = const()[name = string("prompt_kernel_0_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(555376768)))]; + tensor prompt_kernel_0_bias_to_fp16 = const()[name = string("prompt_kernel_0_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(560095424)))]; + tensor linear_217_cast_fp16 = linear(bias = prompt_kernel_0_bias_to_fp16, weight = prompt_kernel_0_weight_to_fp16, x = input_1273_cast_fp16)[name = string("linear_217_cast_fp16")]; + tensor input_cast_fp16 = relu(x = linear_217_cast_fp16)[name = string("input_cast_fp16")]; + tensor prompt_kernel_2_weight_to_fp16 = const()[name = string("prompt_kernel_2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(560099584)))]; + tensor prompt_kernel_2_bias_to_fp16 = const()[name = string("prompt_kernel_2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(564293952)))]; + tensor linear_218_cast_fp16 = linear(bias = prompt_kernel_2_bias_to_fp16, weight = prompt_kernel_2_weight_to_fp16, x = input_cast_fp16)[name = string("linear_218_cast_fp16")]; + tensor var_5615_perm_0 = const()[name = string("op_5615_perm_0"), val = tensor([0, 2, 1])]; + string var_5615_cast_fp16_to_fp32_dtype_0 = const()[name = string("op_5615_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + string var_5620_dtype_0 = const()[name = string("op_5620_dtype_0"), val = string("int32")]; + tensor var_5623_perm_0 = const()[name = string("op_5623_perm_0"), val = tensor([1, 0, 2, 3])]; + string var_5623_cast_fp16_to_fp32_dtype_0 = const()[name = string("op_5623_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + tensor var_5626_perm_0 = const()[name = string("op_5626_perm_0"), val = tensor([1, 0, 2, 3])]; + string var_5626_cast_fp16_to_fp32_dtype_0 = const()[name = string("op_5626_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + string var_5631_dtype_0 = const()[name = string("op_5631_dtype_0"), val = string("int32")]; + tensor cache_len_out = cast(dtype = var_5631_dtype_0, x = clip_1_cast_fp16)[name = string("cast_0")]; + tensor var_5626_cast_fp16 = transpose(perm = var_5626_perm_0, x = obj_7_cast_fp16)[name = string("transpose_144")]; + tensor cache_time_out = cast(dtype = var_5626_cast_fp16_to_fp32_dtype_0, x = var_5626_cast_fp16)[name = string("cast_1")]; + tensor var_5623_cast_fp16 = transpose(perm = var_5623_perm_0, x = obj_5_cast_fp16)[name = string("transpose_145")]; + tensor cache_channel_out = cast(dtype = var_5623_cast_fp16_to_fp32_dtype_0, x = var_5623_cast_fp16)[name = string("cast_2")]; + tensor encoded_length = cast(dtype = var_5620_dtype_0, x = clip_0_cast_fp16)[name = string("cast_3")]; + tensor var_5615_cast_fp16 = transpose(perm = var_5615_perm_0, x = linear_218_cast_fp16)[name = string("transpose_146")]; + tensor encoded = cast(dtype = var_5615_cast_fp16_to_fp32_dtype_0, x = var_5615_cast_fp16)[name = string("cast_4")]; + } -> (encoded, encoded_length, cache_channel_out, cache_time_out, cache_len_out); +} \ No newline at end of file diff --git a/ja/560ms/encoder.mlmodelc/weights/weight.bin b/ja/560ms/encoder.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..7b1ca90fb42a5ee3e6703488fa2d418b42814a01 --- /dev/null +++ b/ja/560ms/encoder.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:beaf958ec4cca237d94bf72f5afded9f1b3f6b93439b5d8e3b309e4ee1560e83 +size 564296064 diff --git a/ja/560ms/encoder.mlpackage/Data/com.apple.CoreML/model.mlmodel b/ja/560ms/encoder.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..21020cc3a69303cc019d3c163864648766ab26b9 --- /dev/null +++ b/ja/560ms/encoder.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1e6d4ec3401a73a4b4c0409170717e2caf7bc829f85aa844dbb5f70309dc61ea +size 802813 diff --git a/ja/560ms/encoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/ja/560ms/encoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..7b1ca90fb42a5ee3e6703488fa2d418b42814a01 --- /dev/null +++ b/ja/560ms/encoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:beaf958ec4cca237d94bf72f5afded9f1b3f6b93439b5d8e3b309e4ee1560e83 +size 564296064 diff --git a/ja/560ms/encoder.mlpackage/Manifest.json b/ja/560ms/encoder.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..44bfa973b64e3a625c9a52d15b370d5a8baa0bbd --- /dev/null +++ b/ja/560ms/encoder.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "5BB04C05-9A8D-40F1-9561-33AB8C6304FC": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + }, + "7CFD7E55-2E07-4224-889C-214DD954BA80": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + } + }, + "rootModelIdentifier": "7CFD7E55-2E07-4224-889C-214DD954BA80" +} diff --git a/ja/560ms/joint.mlmodelc/analytics/coremldata.bin b/ja/560ms/joint.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..a7ba5a90671da5c40e03362f44f23df528bc6d93 --- /dev/null +++ b/ja/560ms/joint.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e342ce20383866520d2c6c860c2bf14d887b9e7fef53606661b41a23ad09472e +size 243 diff --git a/ja/560ms/joint.mlmodelc/coremldata.bin b/ja/560ms/joint.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..219c9bac9ed82b5626e34705f215643647b64d90 --- /dev/null +++ b/ja/560ms/joint.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1df3d770386f24c28a1a187cd341e4f2527d0b1c7f2959e5f606383e2ba9ddc6 +size 401 diff --git a/ja/560ms/joint.mlmodelc/model.mil b/ja/560ms/joint.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..827f6cd71b5910ea07d4f6ba43462967d8b86410 --- /dev/null +++ b/ja/560ms/joint.mlmodelc/model.mil @@ -0,0 +1,31 @@ +program(1.3) +[buildInfo = dict({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.10.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor decoder, tensor encoder) { + tensor input_1_perm_0 = const()[name = string("input_1_perm_0"), val = tensor([0, 2, 1])]; + string encoder_to_fp16_dtype_0 = const()[name = string("encoder_to_fp16_dtype_0"), val = string("fp16")]; + tensor input_3_perm_0 = const()[name = string("input_3_perm_0"), val = tensor([0, 2, 1])]; + string decoder_to_fp16_dtype_0 = const()[name = string("decoder_to_fp16_dtype_0"), val = string("fp16")]; + tensor module_enc_weight_to_fp16 = const()[name = string("module_enc_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))]; + tensor module_enc_bias_to_fp16 = const()[name = string("module_enc_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1310848)))]; + tensor encoder_to_fp16 = cast(dtype = encoder_to_fp16_dtype_0, x = encoder)[name = string("cast_2")]; + tensor input_1_cast_fp16 = transpose(perm = input_1_perm_0, x = encoder_to_fp16)[name = string("transpose_1")]; + tensor linear_0_cast_fp16 = linear(bias = module_enc_bias_to_fp16, weight = module_enc_weight_to_fp16, x = input_1_cast_fp16)[name = string("linear_0_cast_fp16")]; + tensor module_pred_weight_to_fp16 = const()[name = string("module_pred_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1312192)))]; + tensor module_pred_bias_to_fp16 = const()[name = string("module_pred_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2131456)))]; + tensor decoder_to_fp16 = cast(dtype = decoder_to_fp16_dtype_0, x = decoder)[name = string("cast_1")]; + tensor input_3_cast_fp16 = transpose(perm = input_3_perm_0, x = decoder_to_fp16)[name = string("transpose_0")]; + tensor linear_1_cast_fp16 = linear(bias = module_pred_bias_to_fp16, weight = module_pred_weight_to_fp16, x = input_3_cast_fp16)[name = string("linear_1_cast_fp16")]; + tensor var_23_axes_0 = const()[name = string("op_23_axes_0"), val = tensor([2])]; + tensor var_23_cast_fp16 = expand_dims(axes = var_23_axes_0, x = linear_0_cast_fp16)[name = string("op_23_cast_fp16")]; + tensor var_25_axes_0 = const()[name = string("op_25_axes_0"), val = tensor([1])]; + tensor var_25_cast_fp16 = expand_dims(axes = var_25_axes_0, x = linear_1_cast_fp16)[name = string("op_25_cast_fp16")]; + tensor input_5_cast_fp16 = add(x = var_23_cast_fp16, y = var_25_cast_fp16)[name = string("input_5_cast_fp16")]; + tensor input_7_cast_fp16 = relu(x = input_5_cast_fp16)[name = string("input_7_cast_fp16")]; + tensor module_joint_net_2_weight_to_fp16 = const()[name = string("module_joint_net_2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2132800)))]; + tensor module_joint_net_2_bias_to_fp16 = const()[name = string("module_joint_net_2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3929984)))]; + tensor linear_2_cast_fp16 = linear(bias = module_joint_net_2_bias_to_fp16, weight = module_joint_net_2_weight_to_fp16, x = input_7_cast_fp16)[name = string("linear_2_cast_fp16")]; + string linear_2_cast_fp16_to_fp32_dtype_0 = const()[name = string("linear_2_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + tensor logits = cast(dtype = linear_2_cast_fp16_to_fp32_dtype_0, x = linear_2_cast_fp16)[name = string("cast_0")]; + } -> (logits); +} \ No newline at end of file diff --git a/ja/560ms/joint.mlmodelc/weights/weight.bin b/ja/560ms/joint.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..31da4412e1214e0b52e77c023e0490150a12e242 --- /dev/null +++ b/ja/560ms/joint.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:44c15c8be3c89e5d531258b840e560e467084d55ed400453a9d21325757fb111 +size 3932856 diff --git a/ja/560ms/joint.mlpackage/Data/com.apple.CoreML/model.mlmodel b/ja/560ms/joint.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..1bc98711fb995d49c835ce43242e33bc518a943d --- /dev/null +++ b/ja/560ms/joint.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1df49e8dd61d86e19d95b935020748db2eda7ae1273f39988001a30535c5be45 +size 4545 diff --git a/ja/560ms/joint.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/ja/560ms/joint.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..31da4412e1214e0b52e77c023e0490150a12e242 --- /dev/null +++ b/ja/560ms/joint.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:44c15c8be3c89e5d531258b840e560e467084d55ed400453a9d21325757fb111 +size 3932856 diff --git a/ja/560ms/joint.mlpackage/Manifest.json b/ja/560ms/joint.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..ef5aca0b538b9de8e541f27a29fff913203df8c7 --- /dev/null +++ b/ja/560ms/joint.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "8348189F-7CB9-4961-A35F-4049C53D63B6": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + }, + "AA6A8B4F-747E-4EC1-87E1-2B387F1149D8": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + } + }, + "rootModelIdentifier": "8348189F-7CB9-4961-A35F-4049C53D63B6" +} diff --git a/ja/560ms/metadata.json b/ja/560ms/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..7498278692154e2ceb928e04b3b87d8fe6317763 --- /dev/null +++ b/ja/560ms/metadata.json @@ -0,0 +1,199 @@ +{ + "model": "nvidia/nemotron-asr-streaming-multilingual-0.6b", + "model_class": "nemo.collections.asr.models.rnnt_bpe_models_prompt.EncDecRNNTBPEModelWithPrompt", + "sample_rate": 16000, + "mel_features": 128, + "chunk_mel_frames": 56, + "pre_encode_cache": 9, + "total_mel_frames": 65, + "att_context_size": [ + 42, + 13 + ], + "vocab_size": 1403, + "blank_idx": 1403, + "vocab_pruned": true, + "vocab_pruned_original_size": 13087, + "cache_channel_shape": [ + 1, + 24, + 42, + 1024 + ], + "cache_time_shape": [ + 1, + 24, + 1024, + 8 + ], + "decoder_hidden": 640, + "decoder_layers": 2, + "encoder_dim": 1024, + "num_prompts": 128, + "prompt_dictionary": { + "af-ZA": 54, + "am-ET": 49, + "ar": 7, + "ar-AR": 7, + "auto": 101, + "ay-BO": 81, + "az-AZ": 66, + "bg": 30, + "bg-BG": 30, + "bn-IN": 36, + "cs": 22, + "cs-CZ": 22, + "da": 25, + "da-DK": 25, + "de": 9, + "de-DE": 9, + "el": 21, + "el-GR": 21, + "en": 0, + "en-GB": 1, + "en-US": 0, + "enGB": 1, + "es": 3, + "es-ES": 2, + "es-US": 3, + "esES": 2, + "et": 60, + "et-EE": 60, + "fa-IR": 38, + "fi": 26, + "fi-FI": 26, + "fr": 8, + "fr-CA": 100, + "fr-FR": 8, + "gn-PY": 82, + "gu-IN": 42, + "ha-NG": 50, + "haw-US": 97, + "he-IL": 64, + "hi": 6, + "hi-HI": 6, + "hi-IN": 6, + "hr": 29, + "hr-HR": 29, + "hu": 23, + "hu-HU": 23, + "hy-AM": 68, + "id-ID": 34, + "ig-NG": 53, + "it": 15, + "it-IT": 15, + "ja-JA": 10, + "ja-JP": 10, + "ka-GE": 67, + "km-KH": 47, + "kn-IN": 43, + "ko": 14, + "ko-KO": 14, + "ko-KR": 14, + "ku-TR": 65, + "ky-KG": 71, + "ln-CD": 58, + "lt": 31, + "lt-LT": 31, + "lv": 61, + "lv-LV": 61, + "mi-NZ": 96, + "ml-IN": 44, + "mr-IN": 41, + "ms-MY": 35, + "mt-MT": 102, + "nah-MX": 83, + "nb": 103, + "nb-NO": 103, + "ne-NP": 46, + "nl": 16, + "nl-NL": 16, + "nn": 104, + "nn-NO": 104, + "no": 27, + "no-NO": 27, + "ny-MW": 57, + "or-KE": 59, + "pl": 17, + "pl-PL": 17, + "pt": 13, + "pt-BR": 12, + "pt-PT": 13, + "qu-PE": 80, + "ro": 20, + "ro-RO": 20, + "ru": 11, + "ru-RU": 11, + "rw-RW": 55, + "si-LK": 45, + "sk": 28, + "sk-SK": 28, + "sl": 62, + "sl-SI": 62, + "sm-WS": 98, + "so-SO": 56, + "sv": 24, + "sv-SE": 24, + "sw-KE": 48, + "ta-IN": 39, + "te-IN": 40, + "tg-TJ": 70, + "th-TH": 32, + "to-TO": 99, + "tr": 18, + "tr-TR": 18, + "uk": 19, + "uk-UA": 19, + "ur-PK": 37, + "uz-UZ": 69, + "vi-VN": 33, + "yo-NG": 52, + "zh-CN": 4, + "zh-TW": 5, + "zh-ZH": 4, + "zu-ZA": 51 + }, + "default_prompt_id": 101, + "lang_tag_token_ids": [ + 1, + 52, + 62, + 66, + 69, + 70, + 75, + 76, + 77, + 79, + 81, + 83, + 86, + 88, + 89, + 90, + 92, + 94, + 95, + 96, + 97, + 99, + 100, + 103, + 107, + 109, + 111, + 112, + 114, + 115, + 117, + 1389, + 1390, + 1391, + 1392, + 1393, + 1394, + 1395, + 1402 + ], + "chunk_ms": 560 +} \ No newline at end of file diff --git a/ja/560ms/preprocessor.mlmodelc/analytics/coremldata.bin b/ja/560ms/preprocessor.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..d4b1561ab413a9d87db506bc842f077779dcbded --- /dev/null +++ b/ja/560ms/preprocessor.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3b8033bec5ee01649f325b8f4c5aeef1b31c99b469ce56d46039c1b73f09585d +size 243 diff --git a/ja/560ms/preprocessor.mlmodelc/coremldata.bin b/ja/560ms/preprocessor.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..bac73ca18a242859bea870ffafad4dab2fb941b1 --- /dev/null +++ b/ja/560ms/preprocessor.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:eb1828619de97e8d9cc2ac31f370c94dcc9b30a828c497b32918f4ad00096a7d +size 431 diff --git a/ja/560ms/preprocessor.mlmodelc/model.mil b/ja/560ms/preprocessor.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..b1a0b2b9193c992de42e51245fc1ef433d345afc --- /dev/null +++ b/ja/560ms/preprocessor.mlmodelc/model.mil @@ -0,0 +1,122 @@ +program(1.3) +[buildInfo = dict({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.10.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor audio, tensor audio_length) [FlexibleShapeInformation = tuple>>, tuple, ?>>>>((("DefaultShapes", {{"audio", [1, 1]}}), ("RangeDims", {{"audio", [[1, 1], [1, 1280000]]}})))] { + int32 var_9 = const()[name = string("op_9"), val = int32(1)]; + int32 var_10 = const()[name = string("op_10"), val = int32(160)]; + int32 var_12 = const()[name = string("op_12"), val = int32(0)]; + int32 var_33 = const()[name = string("op_33"), val = int32(512)]; + tensor var_34 = add(x = audio_length, y = var_33)[name = string("op_34")]; + int32 var_35 = const()[name = string("op_35"), val = int32(512)]; + tensor var_36 = sub(x = var_34, y = var_35)[name = string("op_36")]; + tensor floor_div_0 = floor_div(x = var_36, y = var_10)[name = string("floor_div_0")]; + tensor var_39 = equal(x = audio_length, y = var_12)[name = string("op_39")]; + tensor var_40 = const()[name = string("op_40"), val = tensor([0])]; + tensor mel_length = select(a = var_40, b = floor_div_0, cond = var_39)[name = string("seq_len")]; + string audio_to_fp16_dtype_0 = const()[name = string("audio_to_fp16_dtype_0"), val = string("fp16")]; + tensor audio_to_fp16 = cast(dtype = audio_to_fp16_dtype_0, x = audio)[name = string("cast_10")]; + tensor var_42_shape_cast_fp16 = shape(x = audio_to_fp16)[name = string("op_42_shape_cast_fp16")]; + int32 gather_0_axis_0 = const()[name = string("gather_0_axis_0"), val = int32(0)]; + int32 gather_0_batch_dims_0 = const()[name = string("gather_0_batch_dims_0"), val = int32(0)]; + bool gather_0_validate_indices_0 = const()[name = string("gather_0_validate_indices_0"), val = bool(false)]; + string var_42_shape_cast_fp16_to_int16_dtype_0 = const()[name = string("op_42_shape_cast_fp16_to_int16_dtype_0"), val = string("int16")]; + uint16 gather_0_indices_0_to_uint16 = const()[name = string("gather_0_indices_0_to_uint16"), val = uint16(1)]; + tensor var_42_shape_cast_fp16_to_int16 = cast(dtype = var_42_shape_cast_fp16_to_int16_dtype_0, x = var_42_shape_cast_fp16)[name = string("cast_9")]; + int16 gather_0_cast_uint16 = gather(axis = gather_0_axis_0, batch_dims = gather_0_batch_dims_0, indices = gather_0_indices_0_to_uint16, validate_indices = gather_0_validate_indices_0, x = var_42_shape_cast_fp16_to_int16)[name = string("gather_0_cast_uint16")]; + string gather_0_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_0_cast_uint16_to_int32_dtype_0"), val = string("int32")]; + int32 const_0 = const()[name = string("const_0"), val = int32(0)]; + int32 const_1 = const()[name = string("const_1"), val = int32(1)]; + int32 gather_0_cast_uint16_to_int32 = cast(dtype = gather_0_cast_uint16_to_int32_dtype_0, x = gather_0_cast_uint16)[name = string("cast_8")]; + tensor var_43 = range_1d(end = gather_0_cast_uint16_to_int32, start = const_0, step = const_1)[name = string("op_43")]; + tensor var_44_axes_0 = const()[name = string("op_44_axes_0"), val = tensor([0])]; + tensor var_44 = expand_dims(axes = var_44_axes_0, x = var_43)[name = string("op_44")]; + tensor var_45_axes_0 = const()[name = string("op_45_axes_0"), val = tensor([1])]; + tensor var_45 = expand_dims(axes = var_45_axes_0, x = audio_length)[name = string("op_45")]; + tensor timemask = less(x = var_44, y = var_45)[name = string("timemask")]; + tensor var_48_begin_0 = const()[name = string("op_48_begin_0"), val = tensor([0, 0])]; + tensor var_48_end_0 = const()[name = string("op_48_end_0"), val = tensor([1, 1])]; + tensor var_48_end_mask_0 = const()[name = string("op_48_end_mask_0"), val = tensor([true, false])]; + tensor var_48_squeeze_mask_0 = const()[name = string("op_48_squeeze_mask_0"), val = tensor([false, true])]; + tensor var_48_cast_fp16 = slice_by_index(begin = var_48_begin_0, end = var_48_end_0, end_mask = var_48_end_mask_0, squeeze_mask = var_48_squeeze_mask_0, x = audio_to_fp16)[name = string("op_48_cast_fp16")]; + tensor var_49_axes_0 = const()[name = string("op_49_axes_0"), val = tensor([1])]; + tensor var_49_cast_fp16 = expand_dims(axes = var_49_axes_0, x = var_48_cast_fp16)[name = string("op_49_cast_fp16")]; + tensor var_51_begin_0 = const()[name = string("op_51_begin_0"), val = tensor([0, 1])]; + tensor var_51_end_0 = const()[name = string("op_51_end_0"), val = tensor([1, 0])]; + tensor var_51_end_mask_0 = const()[name = string("op_51_end_mask_0"), val = tensor([true, true])]; + tensor var_51_cast_fp16 = slice_by_index(begin = var_51_begin_0, end = var_51_end_0, end_mask = var_51_end_mask_0, x = audio_to_fp16)[name = string("op_51_cast_fp16")]; + tensor var_53_begin_0 = const()[name = string("op_53_begin_0"), val = tensor([0, 0])]; + tensor var_53_end_0 = const()[name = string("op_53_end_0"), val = tensor([1, -1])]; + tensor var_53_end_mask_0 = const()[name = string("op_53_end_mask_0"), val = tensor([true, false])]; + tensor var_53_cast_fp16 = slice_by_index(begin = var_53_begin_0, end = var_53_end_0, end_mask = var_53_end_mask_0, x = audio_to_fp16)[name = string("op_53_cast_fp16")]; + fp16 var_54_to_fp16 = const()[name = string("op_54_to_fp16"), val = fp16(0x1.f0cp-1)]; + tensor var_55_cast_fp16 = mul(x = var_53_cast_fp16, y = var_54_to_fp16)[name = string("op_55_cast_fp16")]; + tensor var_56_cast_fp16 = sub(x = var_51_cast_fp16, y = var_55_cast_fp16)[name = string("op_56_cast_fp16")]; + bool x_3_interleave_0 = const()[name = string("x_3_interleave_0"), val = bool(false)]; + tensor x_3_cast_fp16 = concat(axis = var_9, interleave = x_3_interleave_0, values = (var_49_cast_fp16, var_56_cast_fp16))[name = string("x_3_cast_fp16")]; + tensor var_59 = logical_not(x = timemask)[name = string("op_59")]; + fp16 var_16_to_fp16 = const()[name = string("op_16_to_fp16"), val = fp16(0x0p+0)]; + tensor input_1_cast_fp16 = select(a = var_16_to_fp16, b = x_3_cast_fp16, cond = var_59)[name = string("input_1_cast_fp16")]; + tensor concat_1x = const()[name = string("concat_1x"), val = tensor([1, 1, -1])]; + tensor input_3_cast_fp16 = reshape(shape = concat_1x, x = input_1_cast_fp16)[name = string("input_3_cast_fp16")]; + tensor input_5_pad_0 = const()[name = string("input_5_pad_0"), val = tensor([0, 0, 0, 0, 256, 256])]; + string input_5_mode_0 = const()[name = string("input_5_mode_0"), val = string("constant")]; + fp16 const_3_to_fp16 = const()[name = string("const_3_to_fp16"), val = fp16(0x0p+0)]; + tensor input_5_cast_fp16 = pad(constant_val = const_3_to_fp16, mode = input_5_mode_0, pad = input_5_pad_0, x = input_3_cast_fp16)[name = string("input_5_cast_fp16")]; + tensor concat_2x = const()[name = string("concat_2x"), val = tensor([1, -1])]; + tensor input_cast_fp16 = reshape(shape = concat_2x, x = input_5_cast_fp16)[name = string("input_cast_fp16")]; + tensor expand_dims_3 = const()[name = string("expand_dims_3"), val = tensor([160])]; + tensor expand_dims_4_axes_0 = const()[name = string("expand_dims_4_axes_0"), val = tensor([1])]; + tensor expand_dims_4_cast_fp16 = expand_dims(axes = expand_dims_4_axes_0, x = input_cast_fp16)[name = string("expand_dims_4_cast_fp16")]; + string conv_0_pad_type_0 = const()[name = string("conv_0_pad_type_0"), val = string("valid")]; + tensor conv_0_pad_0 = const()[name = string("conv_0_pad_0"), val = tensor([0, 0])]; + tensor conv_0_dilations_0 = const()[name = string("conv_0_dilations_0"), val = tensor([1])]; + int32 conv_0_groups_0 = const()[name = string("conv_0_groups_0"), val = int32(1)]; + tensor expand_dims_1_to_fp16 = const()[name = string("expand_dims_1_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))]; + tensor conv_0_cast_fp16 = conv(dilations = conv_0_dilations_0, groups = conv_0_groups_0, pad = conv_0_pad_0, pad_type = conv_0_pad_type_0, strides = expand_dims_3, weight = expand_dims_1_to_fp16, x = expand_dims_4_cast_fp16)[name = string("conv_0_cast_fp16")]; + string conv_1_pad_type_0 = const()[name = string("conv_1_pad_type_0"), val = string("valid")]; + tensor conv_1_pad_0 = const()[name = string("conv_1_pad_0"), val = tensor([0, 0])]; + tensor conv_1_dilations_0 = const()[name = string("conv_1_dilations_0"), val = tensor([1])]; + int32 conv_1_groups_0 = const()[name = string("conv_1_groups_0"), val = int32(1)]; + tensor expand_dims_2_to_fp16 = const()[name = string("expand_dims_2_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(263296)))]; + tensor conv_1_cast_fp16 = conv(dilations = conv_1_dilations_0, groups = conv_1_groups_0, pad = conv_1_pad_0, pad_type = conv_1_pad_type_0, strides = expand_dims_3, weight = expand_dims_2_to_fp16, x = expand_dims_4_cast_fp16)[name = string("conv_1_cast_fp16")]; + int32 stack_0_axis_0 = const()[name = string("stack_0_axis_0"), val = int32(-1)]; + tensor stack_0_cast_fp16 = stack(axis = stack_0_axis_0, values = (conv_0_cast_fp16, conv_1_cast_fp16))[name = string("stack_0_cast_fp16")]; + fp16 var_19_promoted_to_fp16 = const()[name = string("op_19_promoted_to_fp16"), val = fp16(0x1p+1)]; + tensor var_74_cast_fp16 = pow(x = stack_0_cast_fp16, y = var_19_promoted_to_fp16)[name = string("op_74_cast_fp16")]; + tensor var_76_axes_0 = const()[name = string("op_76_axes_0"), val = tensor([-1])]; + bool var_76_keep_dims_0 = const()[name = string("op_76_keep_dims_0"), val = bool(false)]; + tensor var_76_cast_fp16 = reduce_sum(axes = var_76_axes_0, keep_dims = var_76_keep_dims_0, x = var_74_cast_fp16)[name = string("op_76_cast_fp16")]; + tensor x_11_cast_fp16 = identity(x = var_76_cast_fp16)[name = string("x_11_cast_fp16")]; + bool x_13_transpose_x_0 = const()[name = string("x_13_transpose_x_0"), val = bool(false)]; + bool x_13_transpose_y_0 = const()[name = string("x_13_transpose_y_0"), val = bool(false)]; + tensor const_4_to_fp16 = const()[name = string("const_4_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(526528)))]; + tensor x_13_cast_fp16 = matmul(transpose_x = x_13_transpose_x_0, transpose_y = x_13_transpose_y_0, x = const_4_to_fp16, y = x_11_cast_fp16)[name = string("x_13_cast_fp16")]; + fp16 var_83_to_fp16 = const()[name = string("op_83_to_fp16"), val = fp16(0x1p-24)]; + tensor var_84_cast_fp16 = add(x = x_13_cast_fp16, y = var_83_to_fp16)[name = string("op_84_cast_fp16")]; + fp32 x_epsilon_0 = const()[name = string("x_epsilon_0"), val = fp32(0x1p-149)]; + tensor x_cast_fp16 = log(epsilon = x_epsilon_0, x = var_84_cast_fp16)[name = string("x_cast_fp16")]; + tensor var_86_shape_cast_fp16 = shape(x = x_cast_fp16)[name = string("op_86_shape_cast_fp16")]; + int32 gather_5_batch_dims_0 = const()[name = string("gather_5_batch_dims_0"), val = int32(0)]; + bool gather_5_validate_indices_0 = const()[name = string("gather_5_validate_indices_0"), val = bool(false)]; + string var_86_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_86_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")]; + int32 gather_5_cast_uint16_axis_0 = const()[name = string("gather_5_cast_uint16_axis_0"), val = int32(0)]; + uint16 select_0_to_uint16 = const()[name = string("select_0_to_uint16"), val = uint16(2)]; + tensor var_86_shape_cast_fp16_to_uint16 = cast(dtype = var_86_shape_cast_fp16_to_uint16_dtype_0, x = var_86_shape_cast_fp16)[name = string("cast_7")]; + uint16 gather_5_cast_uint16_cast_uint16 = gather(axis = gather_5_cast_uint16_axis_0, batch_dims = gather_5_batch_dims_0, indices = select_0_to_uint16, validate_indices = gather_5_validate_indices_0, x = var_86_shape_cast_fp16_to_uint16)[name = string("gather_5_cast_uint16_cast_uint16")]; + string gather_5_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_5_cast_uint16_to_int32_dtype_0"), val = string("int32")]; + int32 const_5 = const()[name = string("const_5"), val = int32(0)]; + int32 const_6 = const()[name = string("const_6"), val = int32(1)]; + int32 gather_5_cast_uint16_to_int32 = cast(dtype = gather_5_cast_uint16_to_int32_dtype_0, x = gather_5_cast_uint16_cast_uint16)[name = string("cast_6")]; + tensor mask_1 = range_1d(end = gather_5_cast_uint16_to_int32, start = const_5, step = const_6)[name = string("mask_1")]; + tensor expand_dims_0_axes_0 = const()[name = string("expand_dims_0_axes_0"), val = tensor([0])]; + tensor expand_dims_0 = expand_dims(axes = expand_dims_0_axes_0, x = mask_1)[name = string("expand_dims_0")]; + tensor var_91_axes_0 = const()[name = string("op_91_axes_0"), val = tensor([1])]; + tensor var_91 = expand_dims(axes = var_91_axes_0, x = mel_length)[name = string("op_91")]; + tensor mask = greater_equal(x = expand_dims_0, y = var_91)[name = string("mask")]; + tensor var_93_axes_0 = const()[name = string("op_93_axes_0"), val = tensor([1])]; + tensor var_93 = expand_dims(axes = var_93_axes_0, x = mask)[name = string("op_93")]; + tensor processed_signal_cast_fp16 = select(a = var_16_to_fp16, b = x_cast_fp16, cond = var_93)[name = string("processed_signal_cast_fp16")]; + string processed_signal_cast_fp16_to_fp32_dtype_0 = const()[name = string("processed_signal_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; + tensor mel = cast(dtype = processed_signal_cast_fp16_to_fp32_dtype_0, x = processed_signal_cast_fp16)[name = string("cast_5")]; + } -> (mel, mel_length); +} \ No newline at end of file diff --git a/ja/560ms/preprocessor.mlmodelc/weights/weight.bin b/ja/560ms/preprocessor.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..86dd375f6649d262d58c9cf8b89006ceab216411 --- /dev/null +++ b/ja/560ms/preprocessor.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:297514e2b211d14b0e53cb97193d679bb89ead98d28e578f3f1d049ddbcc36b3 +size 592384 diff --git a/ja/560ms/preprocessor.mlpackage/Data/com.apple.CoreML/model.mlmodel b/ja/560ms/preprocessor.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..050fa97ca7a2aa4b7c4fa318f4fa2a51914287c4 --- /dev/null +++ b/ja/560ms/preprocessor.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0feb90f2cf9a5bc749fa763b6eb78a23755b6184b41a262fe08faebe4e709b3e +size 16035 diff --git a/ja/560ms/preprocessor.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/ja/560ms/preprocessor.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..86dd375f6649d262d58c9cf8b89006ceab216411 --- /dev/null +++ b/ja/560ms/preprocessor.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:297514e2b211d14b0e53cb97193d679bb89ead98d28e578f3f1d049ddbcc36b3 +size 592384 diff --git a/ja/560ms/preprocessor.mlpackage/Manifest.json b/ja/560ms/preprocessor.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..11f4407dd2c2a1fafa63dd69a93da5d4e2866fba --- /dev/null +++ b/ja/560ms/preprocessor.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "571B03EE-25FB-448D-B106-8C0685101326": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + }, + "B1E3045B-7172-4628-BD63-F2960CEFBA16": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + } + }, + "rootModelIdentifier": "571B03EE-25FB-448D-B106-8C0685101326" +} diff --git a/ja/560ms/tokenizer.json b/ja/560ms/tokenizer.json new file mode 100644 index 0000000000000000000000000000000000000000..b05c35042625729d82c27359c8b8cf9ae4e2e4c9 --- /dev/null +++ b/ja/560ms/tokenizer.json @@ -0,0 +1,1406 @@ +{ + "0": "", + "1": "", + "2": "▁", + "3": ".", + "4": ",", + "5": "e", + "6": "t", + "7": "a", + "8": "s", + "9": "o", + "10": "i", + "11": "r", + "12": "h", + "13": "m", + "14": "p", + "15": "n", + "16": "g", + "17": "en", + "18": "on", + "19": "y", + "20": "0", + "21": "b", + "22": "2", + "23": "1", + "24": "k", + "25": "3", + "26": "5", + "27": "I", + "28": "4", + "29": "z", + "30": "A", + "31": "6", + "32": "E", + "33": "7", + "34": "T", + "35": "8", + "36": "9", + "37": "S", + "38": "C", + "39": "M", + "40": "P", + "41": "D", + "42": "B", + "43": "U", + "44": "N", + "45": "O", + "46": "G", + "47": "F", + "48": "L", + "49": "R", + "50": "H", + "51": "K", + "52": "", + "53": "le", + "54": "li", + "55": "to", + "56": "na", + "57": "ci", + "58": "▁A", + "59": "vi", + "60": "Z", + "61": "X", + "62": "", + "63": "et", + "64": "ar", + "65": "el", + "66": "", + "67": "ie", + "68": "▁G", + "69": "", + "70": "", + "71": "ta", + "72": "ks", + "73": "gi", + "74": "di", + "75": "", + "76": "", + "77": "", + "78": "cs", + "79": "", + "80": "sti", + "81": "", + "82": "pe", + "83": "", + "84": "as", + "85": "▁Tai", + "86": "", + "87": "▁20", + "88": "", + "89": "", + "90": "", + "91": "-", + "92": "", + "93": "▁M", + "94": "", + "95": "", + "96": "", + "97": "", + "98": "pp", + "99": "", + "100": "", + "101": "▁N", + "102": "mm", + "103": "", + "104": "▁and", + "105": "th", + "106": "ex", + "107": "", + "108": "ng", + "109": "", + "110": "▁La", + "111": "", + "112": "", + "113": "avec", + "114": "", + "115": "", + "116": "!", + "117": "", + "118": "▁あ", + "119": "▁す", + "120": "▁し", + "121": "▁そ", + "122": "▁ま", + "123": "▁こ", + "124": "以", + "125": "仕", + "126": "▁文", + "127": "▁強", + "128": "▁月", + "129": "▁連", + "130": "▁ペ", + "131": "▁海", + "132": "▁科", + "133": "▁ゴ", + "134": "▁原", + "135": "▁ギ", + "136": "▁ヤ", + "137": "▁特", + "138": "▁参", + "139": "▁男", + "140": "普", + "141": "受", + "142": "▁都", + "143": "▁水", + "144": "▁終", + "145": "▁南", + "146": "▁例", + "147": "▁利", + "148": "▁現", + "149": "▁真", + "150": "▁古", + "151": "▁彼", + "152": "▁交", + "153": "専", + "154": "▁悪", + "155": "▁旅", + "156": "▁技", + "157": "▁再", + "158": "企", + "159": "登", + "160": "総", + "161": "改", + "162": "探", + "163": "異", + "164": "完", + "165": "頑", + "166": "深", + "167": "推", + "168": "昨", + "169": "危", + "170": "詳", + "171": "余", + "172": "描", + "173": "被", + "174": "固", + "175": "副", + "176": "混", + "177": "救", + "178": "互", + "179": "閉", + "180": "亜", + "181": "植", + "182": "責", + "183": "仮", + "184": "脱", + "185": "恐", + "186": "契", + "187": "隠", + "188": "麻", + "189": "哲", + "190": "宣", + "191": "浮", + "192": "諸", + "193": "巨", + "194": "午", + "195": "占", + "196": "距", + "197": "汚", + "198": "延", + "199": "訪", + "200": "沿", + "201": "唯", + "202": "批", + "203": "傾", + "204": "邦", + "205": "購", + "206": "維", + "207": "燃", + "208": "既", + "209": "緊", + "210": "獲", + "211": "炭", + "212": "搭", + "213": "努", + "214": "顧", + "215": "狭", + "216": "雰", + "217": "塗", + "218": "排", + "219": "奪", + "220": "埋", + "221": "履", + "222": "侵", + "223": "該", + "224": "頻", + "225": "勧", + "226": "魅", + "227": "抑", + "228": "驚", + "229": "衝", + "230": "耐", + "231": "披", + "232": "透", + "233": "促", + "234": "撤", + "235": "潜", + "236": "吐", + "237": "睡", + "238": "孤", + "239": "昆", + "240": "遭", + "241": "溶", + "242": "貿", + "243": "疾", + "244": "拒", + "245": "脅", + "246": "挟", + "247": "漏", + "248": "覆", + "249": "畜", + "250": "紛", + "251": "絞", + "252": "擬", + "253": "疎", + "254": "跳", + "255": "廷", + "256": "郊", + "257": "克", + "258": "措", + "259": "虐", + "260": "陥", + "261": "徐", + "262": "漠", + "263": "愚", + "264": "噴", + "265": "卿", + "266": "胡", + "267": "溢", + "268": "妨", + "269": "沸", + "270": "阻", + "271": "妃", + "272": "舶", + "273": "濡", + "274": "剥", + "275": "洪", + "276": "鋭", + "277": "娯", + "278": "腎", + "279": "訂", + "280": "肖", + "281": "枢", + "282": "髄", + "283": "傑", + "284": "砕", + "285": "幣", + "286": "勾", + "287": "臆", + "288": "歪", + "289": "喝", + "290": "彗", + "291": "礁", + "292": "蚊", + "293": "焉", + "294": "塵", + "295": "脊", + "296": "鱗", + "297": "瞑", + "298": "峻", + "299": "徊", + "300": "徘", + "301": "猥", + "302": "竣", + "303": "穣", + "304": "隻", + "305": "睨", + "306": "瀕", + "307": "褻", + "308": "踵", + "309": "奢", + "310": "董", + "311": "毅", + "312": "醒", + "313": "墜", + "314": "窟", + "315": "綿", + "316": "遮", + "317": "騰", + "318": "顎", + "319": "遍", + "320": "蓋", + "321": "澄", + "322": "棲", + "323": "虎", + "324": "粒", + "325": "胎", + "326": "爪", + "327": "殻", + "328": "鉛", + "329": "枕", + "330": "堤", + "331": "膝", + "332": "峠", + "333": "搬", + "334": "晋", + "335": "餌", + "336": "狼", + "337": "衰", + "338": "閲", + "339": "筒", + "340": "軸", + "341": "緯", + "342": "暦", + "343": "刃", + "344": "潔", + "345": "裂", + "346": "叫", + "347": "殖", + "348": "媒", + "349": "癌", + "350": "鎖", + "351": "磁", + "352": "洞", + "353": "遇", + "354": "枝", + "355": "紫", + "356": "縦", + "357": "胞", + "358": "釈", + "359": "威", + "360": "晶", + "361": "砲", + "362": "焦", + "363": "尿", + "364": "魂", + "365": "潮", + "366": "旬", + "367": "慎", + "368": "噂", + "369": "隔", + "370": "穴", + "371": "慮", + "372": "即", + "373": "滑", + "374": "雷", + "375": "摘", + "376": "鏡", + "377": "棒", + "378": "悟", + "379": "葬", + "380": "序", + "381": "貫", + "382": "氷", + "383": "針", + "384": "煮", + "385": "棄", + "386": "銃", + "387": "汁", + "388": "封", + "389": "湿", + "390": "靴", + "391": "豚", + "392": "締", + "393": "豪", + "394": "票", + "395": "皮", + "396": "縮", + "397": "徹", + "398": "較", + "399": "忍", + "400": "核", + "401": "儀", + "402": "到", + "403": "削", + "404": "駆", + "405": "繁", + "406": "陰", + "407": "浄", + "408": "脈", + "409": "滞", + "410": "至", + "411": "枚", + "412": "偉", + "413": "致", + "414": "貨", + "415": "漢", + "416": "己", + "417": "握", + "418": "欧", + "419": "薄", + "420": "献", + "421": "預", + "422": "龍", + "423": "快", + "424": "句", + "425": "縁", + "426": "微", + "427": "妙", + "428": "晩", + "429": "粉", + "430": "卓", + "431": "圏", + "432": "兼", + "433": "脳", + "434": "竜", + "435": "鳴", + "436": "騒", + "437": "請", + "438": "卵", + "439": "唱", + "440": "嵐", + "441": "臓", + "442": "箱", + "443": "祖", + "444": "浴", + "445": "壁", + "446": "析", + "447": "厚", + "448": "筆", + "449": "承", + "450": "均", + "451": "律", + "452": "否", + "453": "脚", + "454": "湖", + "455": "乳", + "456": "揮", + "457": "滅", + "458": "乾", + "459": "羽", + "460": "候", + "461": "拡", + "462": "貸", + "463": "砂", + "464": "敬", + "465": "庁", + "466": "煙", + "467": "底", + "468": "露", + "469": "骨", + "470": "倍", + "471": "殿", + "472": "易", + "473": "層", + "474": "幕", + "475": "毛", + "476": "爆", + "477": "暇", + "478": "械", + "479": "隣", + "480": "輸", + "481": "柄", + "482": "範", + "483": "掲", + "484": "嘘", + "485": "剤", + "486": "墓", + "487": "衣", + "488": "射", + "489": "菓", + "490": "募", + "491": "乱", + "492": "迎", + "493": "抱", + "494": "酸", + "495": "雄", + "496": "虫", + "497": "複", + "498": "為", + "499": "泳", + "500": "宝", + "501": "激", + "502": "暑", + "503": "疑", + "504": "誘", + "505": "暴", + "506": "聖", + "507": "捨", + "508": "破", + "509": "革", + "510": "希", + "511": "折", + "512": "惑", + "513": "測", + "514": "紀", + "515": "舎", + "516": "署", + "517": "患", + "518": "岸", + "519": "秀", + "520": "免", + "521": "禁", + "522": "躍", + "523": "聴", + "524": "抗", + "525": "税", + "526": "奏", + "527": "弾", + "528": "礼", + "529": "童", + "530": "裏", + "531": "吹", + "532": "眠", + "533": "歯", + "534": "拠", + "535": "慣", + "536": "触", + "537": "飼", + "538": "群", + "539": "宗", + "540": "傷", + "541": "額", + "542": "塩", + "543": "静", + "544": "留", + "545": "罪", + "546": "純", + "547": "壊", + "548": "闘", + "549": "弱", + "550": "刻", + "551": "航", + "552": "栄", + "553": "姿", + "554": "亡", + "555": "織", + "556": "敗", + "557": "章", + "558": "吸", + "559": "令", + "560": "捜", + "561": "模", + "562": "絵", + "563": "申", + "564": "盤", + "565": "積", + "566": "標", + "567": "階", + "568": "省", + "569": "項", + "570": "猫", + "571": "従", + "572": "非", + "573": "季", + "574": "捕", + "575": "党", + "576": "圧", + "577": "香", + "578": "操", + "579": "暗", + "580": "症", + "581": "散", + "582": "突", + "583": "適", + "584": "雑", + "585": "跡", + "586": "厳", + "587": "鳥", + "588": "逃", + "589": "講", + "590": "晴", + "591": "徴", + "592": "困", + "593": "短", + "594": "婦", + "595": "略", + "596": "齢", + "597": "震", + "598": "敵", + "599": "博", + "600": "血", + "601": "満", + "602": "舗", + "603": "宙", + "604": "寿", + "605": "遺", + "606": "極", + "607": "里", + "608": "因", + "609": "典", + "610": "染", + "611": "徒", + "612": "巻", + "613": "頂", + "614": "超", + "615": "河", + "616": "盛", + "617": "犬", + "618": "豊", + "619": "端", + "620": "紹", + "621": "首", + "622": "陽", + "623": "歳", + "624": "印", + "625": "紙", + "626": "払", + "627": "求", + "628": "障", + "629": "簡", + "630": "途", + "631": "創", + "632": "船", + "633": "菜", + "634": "ゥ", + "635": "勤", + "636": "痛", + "637": "並", + "638": "景", + "639": "雪", + "640": "節", + "641": "浜", + "642": "清", + "643": "抜", + "644": "勢", + "645": "暮", + "646": "銀", + "647": "盟", + "648": "魚", + "649": "率", + "650": "洋", + "651": "渡", + "652": "順", + "653": "況", + "654": "談", + "655": "舞", + "656": "案", + "657": "岩", + "658": "負", + "659": "旧", + "660": "財", + "661": "故", + "662": "冬", + "663": "横", + "664": "奥", + "665": "比", + "666": "囲", + "667": "停", + "668": "築", + "669": "波", + "670": "林", + "671": "暖", + "672": "索", + "673": "赤", + "674": "給", + "675": "末", + "676": "催", + "677": "遅", + "678": "述", + "679": "黒", + "680": "細", + "681": "与", + "682": "減", + "683": "級", + "684": "費", + "685": "越", + "686": "差", + "687": "領", + "688": "衛", + "689": "隊", + "690": "薬", + "691": "氏", + "692": "望", + "693": "似", + "694": "就", + "695": "条", + "696": "処", + "697": "谷", + "698": "策", + "699": "効", + "700": "熱", + "701": "復", + "702": "ヌ", + "703": "振", + "704": "規", + "705": "港", + "706": "注", + "707": "森", + "708": "防", + "709": "継", + "710": "退", + "711": "火", + "712": "陸", + "713": "去", + "714": "視", + "715": "整", + "716": "準", + "717": "庭", + "718": "ゾ", + "719": "独", + "720": "撃", + "721": "児", + "722": "橋", + "723": "換", + "724": "念", + "725": "識", + "726": "打", + "727": "津", + "728": "雨", + "729": "幸", + "730": "含", + "731": "響", + "732": "労", + "733": "官", + "734": "追", + "735": "遠", + "736": "未", + "737": "販", + "738": "街", + "739": "曜", + "740": "程", + "741": "提", + "742": "玉", + "743": "判", + "744": "移", + "745": "攻", + "746": "低", + "747": "装", + "748": "断", + "749": "及", + "750": "証", + "751": "象", + "752": "守", + "753": "戻", + "754": "詞", + "755": "投", + "756": "載", + "757": "具", + "758": "除", + "759": "環", + "760": "展", + "761": "争", + "762": "失", + "763": "春", + "764": "挙", + "765": "返", + "766": "馬", + "767": "欲", + "768": "材", + "769": "図", + "770": "養", + "771": "焼", + "772": "導", + "773": "夢", + "774": "米", + "775": "冷", + "776": "息", + "777": "兵", + "778": "済", + "779": "劇", + "780": "央", + "781": "険", + "782": "服", + "783": "態", + "784": "走", + "785": "評", + "786": "権", + "787": "論", + "788": "ゅ", + "789": "境", + "790": "察", + "791": "授", + "792": "頼", + "793": "派", + "794": "撮", + "795": "素", + "796": "修", + "797": "第", + "798": "質", + "799": "告", + "800": "興", + "801": "秒", + "802": "宇", + "803": "肉", + "804": "像", + "805": "称", + "806": "値", + "807": "頭", + "808": "週", + "809": "督", + "810": "消", + "811": "芸", + "812": "顔", + "813": "読", + "814": "仲", + "815": "遊", + "816": "試", + "817": "酒", + "818": "離", + "819": "増", + "820": "殺", + "821": "鉄", + "822": "害", + "823": "割", + "824": "石", + "825": "夏", + "826": "助", + "827": "英", + "828": "想", + "829": "管", + "830": "急", + "831": "頃", + "832": "づ", + "833": "造", + "834": "史", + "835": "量", + "836": "製", + "837": "府", + "838": "足", + "839": "王", + "840": "委", + "841": "両", + "842": "辺", + "843": "残", + "844": "逆", + "845": "備", + "846": "軍", + "847": "警", + "848": "査", + "849": "列", + "850": "編", + "851": "段", + "852": "反", + "853": "ゼ", + "854": "携", + "855": "歩", + "856": "座", + "857": "飛", + "858": "丈", + "859": "価", + "860": "監", + "861": "ヘ", + "862": "周", + "863": "毎", + "864": "統", + "865": "収", + "866": "落", + "867": "星", + "868": "降", + "869": "側", + "870": "療", + "871": "師", + "872": "写", + "873": "類", + "874": "命", + "875": "介", + "876": "護", + "877": "死", + "878": "果", + "879": "任", + "880": "更", + "881": "常", + "882": "検", + "883": "過", + "884": "資", + "885": "働", + "886": "認", + "887": "般", + "888": "示", + "889": "客", + "890": "習", + "891": "究", + "892": "半", + "893": "録", + "894": "字", + "895": "昔", + "896": "影", + "897": "覚", + "898": "型", + "899": "声", + "900": "件", + "901": "義", + "902": "施", + "903": "容", + "904": "路", + "905": "呼", + "906": "役", + "907": "単", + "908": "状", + "909": "建", + "910": "由", + "911": "属", + "912": "土", + "913": "葉", + "914": "起", + "915": "覧", + "916": "配", + "917": "張", + "918": "接", + "919": "込", + "920": "待", + "921": "室", + "922": "病", + "923": "帯", + "924": "婚", + "925": "光", + "926": "個", + "927": "職", + "928": "営", + "929": "ぼ", + "930": "研", + "931": "計", + "932": "直", + "933": "難", + "934": "絶", + "935": "ヨ", + "936": "照", + "937": "西", + "938": "約", + "939": "存", + "940": "験", + "941": "治", + "942": "解", + "943": "転", + "944": "商", + "945": "進", + "946": "係", + "947": "説", + "948": "観", + "949": "球", + "950": "乗", + "951": "支", + "952": "得", + "953": "議", + "954": "門", + "955": "止", + "956": "重", + "957": "温", + "958": "着", + "959": "飲", + "960": "母", + "961": "士", + "962": "ざ", + "963": "集", + "964": "万", + "965": "太", + "966": "続", + "967": "線", + "968": "種", + "969": "格", + "970": "位", + "971": "ユ", + "972": "歌", + "973": "夜", + "974": "共", + "975": "正", + "976": "必", + "977": "ヒ", + "978": "色", + "979": "問", + "980": "再", + "981": "域", + "982": "ゆ", + "983": "勝", + "984": "台", + "985": "技", + "986": "旅", + "987": "引", + "988": "系", + "989": "院", + "990": "悪", + "991": "基", + "992": "神", + "993": "産", + "994": "決", + "995": "民", + "996": "交", + "997": "政", + "998": "賞", + "999": "空", + "1000": "医", + "1001": "彼", + "1002": "夫", + "1003": "可", + "1004": "誰", + "1005": "古", + "1006": "帰", + "1007": "術", + "1008": "相", + "1009": "団", + "1010": "伝", + "1011": "住", + "1012": "題", + "1013": "平", + "1014": "予", + "1015": "音", + "1016": "朝", + "1017": "指", + "1018": "真", + "1019": "ヴ", + "1020": "務", + "1021": "点", + "1022": "各", + "1023": "館", + "1024": "応", + "1025": "現", + "1026": "利", + "1027": "天", + "1028": "等", + "1029": "木", + "1030": "白", + "1031": "形", + "1032": "供", + "1033": "経", + "1034": "族", + "1035": "早", + "1036": "例", + "1037": "不", + "1038": "切", + "1039": "南", + "1040": "加", + "1041": "際", + "1042": "終", + "1043": "様", + "1044": "放", + "1045": "和", + "1046": "州", + "1047": "水", + "1048": "協", + "1049": "在", + "1050": "組", + "1051": "向", + "1052": "広", + "1053": "身", + "1054": "界", + "1055": "工", + "1056": "選", + "1057": "始", + "1058": "元", + "1059": "々", + "1060": "親", + "1061": "美", + "1062": "信", + "1063": "都", + "1064": "置", + "1065": "局", + "1066": "運", + "1067": "送", + "1068": "風", + "1069": "口", + "1070": "演", + "1071": "調", + "1072": "ぎ", + "1073": "優", + "1074": "次", + "1075": "ォ", + "1076": "他", + "1077": "園", + "1078": "保", + "1079": "男", + "1080": "参", + "1081": "少", + "1082": "百", + "1083": "特", + "1084": "考", + "1085": "無", + "1086": "七", + "1087": "ヤ", + "1088": "ギ", + "1089": "良", + "1090": "ザ", + "1091": "制", + "1092": "売", + "1093": "能", + "1094": "原", + "1095": "ゲ", + "1096": "有", + "1097": "安", + "1098": "ゴ", + "1099": "育", + "1100": "科", + "1101": "要", + "1102": "料", + "1103": "書", + "1104": "語", + "1105": "設", + "1106": "海", + "1107": "期", + "1108": "流", + "1109": "確", + "1110": "ペ", + "1111": "区", + "1112": "む", + "1113": "連", + "1114": "買", + "1115": "ひ", + "1116": "ふ", + "1117": "付", + "1118": "町", + "1119": "活", + "1120": "情", + "1121": "月", + "1122": "表", + "1123": "曲", + "1124": "強", + "1125": "世", + "1126": "明", + "1127": "成", + "1128": "ノ", + "1129": "ァ", + "1130": "文", + "1131": "違", + "1132": "東", + "1133": "友", + "1134": "意", + "1135": "力", + "1136": "式", + "1137": "法", + "1138": "報", + "1139": "員", + "1140": "心", + "1141": "屋", + "1142": "品", + "1143": "北", + "1144": "先", + "1145": "島", + "1146": "味", + "1147": "川", + "1148": "開", + "1149": "千", + "1150": "関", + "1151": "電", + "1152": "然", + "1153": "度", + "1154": "達", + "1155": "面", + "1156": "九", + "1157": "数", + "1158": "取", + "1159": "楽", + "1160": "金", + "1161": "性", + "1162": "野", + "1163": "別", + "1164": "戦", + "1165": "公", + "1166": "機", + "1167": "道", + "1168": "目", + "1169": "記", + "1170": "び", + "1171": "発", + "1172": "対", + "1173": "立", + "1174": "初", + "1175": "化", + "1176": "ソ", + "1177": "ワ", + "1178": "田", + "1179": "持", + "1180": "ガ", + "1181": "車", + "1182": "番", + "1183": "ピ", + "1184": "聞", + "1185": "回", + "1186": "ぶ", + "1187": "ベ", + "1188": "げ", + "1189": "実", + "1190": "ボ", + "1191": "店", + "1192": "小", + "1193": "定", + "1194": "モ", + "1195": "長", + "1196": "新", + "1197": "ハ", + "1198": "ケ", + "1199": "外", + "1200": "ポ", + "1201": "近", + "1202": "所", + "1203": "へ", + "1204": "同", + "1205": "ネ", + "1206": "内", + "1207": "女", + "1208": "ホ", + "1209": "体", + "1210": "好", + "1211": "ツ", + "1212": "セ", + "1213": "知", + "1214": "山", + "1215": "来", + "1216": "ェ", + "1217": "使", + "1218": "ョ", + "1219": "ズ", + "1220": "主", + "1221": "動", + "1222": "理", + "1223": "物", + "1224": "映", + "1225": "者", + "1226": "ぐ", + "1227": "的", + "1228": "代", + "1229": "変", + "1230": "教", + "1231": "社", + "1232": "用", + "1233": "話", + "1234": "名", + "1235": "構", + "1236": "高", + "1237": "最", + "1238": "ず", + "1239": "ミ", + "1240": "校", + "1241": "ダ", + "1242": "食", + "1243": "後", + "1244": "手", + "1245": "三", + "1246": "通", + "1247": "感", + "1248": "合", + "1249": "多", + "1250": "業", + "1251": "入", + "1252": "エ", + "1253": "場", + "1254": "べ", + "1255": "上", + "1256": "家", + "1257": "私", + "1258": "年", + "1259": "間", + "1260": "画", + "1261": "前", + "1262": "下", + "1263": "ャ", + "1264": "地", + "1265": "二", + "1266": "ウ", + "1267": "ナ", + "1268": "ビ", + "1269": "自", + "1270": "全", + "1271": "パ", + "1272": "結", + "1273": "ブ", + "1274": "ュ", + "1275": "市", + "1276": "サ", + "1277": "気", + "1278": "方", + "1279": "デ", + "1280": "十", + "1281": "キ", + "1282": "当", + "1283": "国", + "1284": "作", + "1285": "ィ", + "1286": "部", + "1287": "オ", + "1288": "ニ", + "1289": "チ", + "1290": "ム", + "1291": "グ", + "1292": "メ", + "1293": "ご", + "1294": "子", + "1295": "ば", + "1296": "生", + "1297": "ほ", + "1298": "せ", + "1299": "何", + "1300": "出", + "1301": "言", + "1302": "今", + "1303": "バ", + "1304": "事", + "1305": "中", + "1306": "プ", + "1307": "時", + "1308": "コ", + "1309": "見", + "1310": "テ", + "1311": "会", + "1312": "マ", + "1313": "カ", + "1314": "思", + "1315": "ロ", + "1316": "ジ", + "1317": "フ", + "1318": "シ", + "1319": "め", + "1320": "レ", + "1321": "ド", + "1322": "分", + "1323": "ょ", + "1324": "ろ", + "1325": "学", + "1326": "行", + "1327": "タ", + "1328": "大", + "1329": "つ", + "1330": "本", + "1331": "日", + "1332": "わ", + "1333": "一", + "1334": "ク", + "1335": "み", + "1336": "リ", + "1337": "ア", + "1338": "ッ", + "1339": "人", + "1340": "ラ", + "1341": "お", + "1342": "じ", + "1343": "イ", + "1344": "ル", + "1345": "ト", + "1346": "ゃ", + "1347": "き", + "1348": "さ", + "1349": "ち", + "1350": "や", + "1351": "ス", + "1352": "ど", + "1353": "け", + "1354": "く", + "1355": "え", + "1356": "を", + "1357": "り", + "1358": "よ", + "1359": "こ", + "1360": "ン", + "1361": "だ", + "1362": "れ", + "1363": "ら", + "1364": "ね", + "1365": "が", + "1366": "ま", + "1367": "ー", + "1368": "も", + "1369": "そ", + "1370": "し", + "1371": "に", + "1372": "は", + "1373": "る", + "1374": "す", + "1375": "と", + "1376": "た", + "1377": "あ", + "1378": "て", + "1379": "っ", + "1380": "で", + "1381": "か", + "1382": "な", + "1383": "ん", + "1384": "う", + "1385": "の", + "1386": "、", + "1387": "。", + "1388": "い", + "1389": "", + "1390": "", + "1391": "", + "1392": "", + "1393": "", + "1394": "", + "1395": "", + "1396": "▁香", + "1397": "▁群", + "1398": "▁米", + "1399": "咆", + "1400": "哮", + "1401": "翅", + "1402": "", + "1403": "" +} \ No newline at end of file